AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
70
Papers today
8h
Update frequency
7
Days of history
Benchmarking World Models for Continual Learning on Compositional Tasks
Reinforcement Learning
Robotics
- Introduction of a compositional continual learning benchmark for world models in robot manipulation.
- Separation of world models into task-agnostic backbones and task-specific heads to enhance continual learning.
- Evaluation of both monolithic and modular world models to assess their performance in compositional tasks.
- Modular world models demonstrate improved knowledge reuse but do not completely solve the forgetting problem.
Read more
Benchmarking World Models for Continual Learning on Compositional Tasks
Summary
This paper introduces a novel benchmark for evaluating world models in the context of continual learning, specifically targeting robot manipulation tasks. The authors argue that a key aspect of effective world models is their ability to learn continually across tasks while retaining knowledge from previous experiences. To address the challenges of knowledge reuse and adaptation, they propose a compositional continual learning benchmark that isolates these aspects by designing task sequences that culminate in composition tasks. These tasks are structured to allow for the examination of knowledge reuse across different input modalities—action and perception. The study evaluates state-of-the-art world models alongside a modular variant that incorporates reusable components. Results indicate that while modularity improves the balance between knowledge reuse and forgetting, existing methods still fall short, highlighting the need for further advancements in continual world models that can efficiently reuse knowledge without forgetting.
Methodology
The authors designed a benchmark that includes a series of tasks structured to test the ability of world models to reuse knowledge without forgetting. They factorized the tasks along action and perception axes and evaluated various world models, including a modular architecture that explicitly incorporates reusable components. The evaluation involved applying conventional continual learning methods to assess the models' performance in adapting to new tasks while retaining previously learned knowledge.
Results
The evaluation revealed that modular world models performed better in balancing knowledge reuse and forgetting compared to conventional monolithic models. However, neither approach fully addressed the challenge of continual learning, indicating a significant opportunity for future research in developing world models that can effectively reuse knowledge without forgetting.
Implications
The findings suggest that enhancing modularity in world models could lead to more effective continual learning systems in robotics, potentially improving the efficiency of robots in adapting to new tasks and environments. This work lays the groundwork for future research aimed at developing world models that can better leverage prior knowledge in dynamic settings.
Common Cause, Not Cross-Attention: Blocking Visual Shortcuts in Audio-Video Generation
Generative Models
Multimodal
Theory
- Models that use cross-attention or shared latents often rely on visual shortcuts, leading to incorrect audio predictions when appearance-event correlations are broken.
- Simply sharing a latent representation does not eliminate the visual shortcut problem.
- Counterfactual invariance is necessary and sufficient to identify the causal predictor in AV generation.
- Interventions on nuisance variables are required to block visual shortcuts effectively.
Read more
Common Cause, Not Cross-Attention: Blocking Visual Shortcuts in Audio-Video Generation
Summary
This paper investigates the issue of spurious correlations between visual appearance and audio in joint audio-video (AV) generation models. The authors present a controlled causal study demonstrating that models relying on direct cross-attention or shared latent representations often learn to predict sound based on visual appearance rather than the actual causal event. This leads to catastrophic failures when the appearance-event correlation is disrupted. The study introduces a structural causal model (SCM) to formalize this issue and shows that simply sharing a latent representation does not resolve the problem. Instead, the authors propose that interventions on nuisance variables are necessary to block visual shortcuts. They establish that counterfactual invariance is both necessary and sufficient for identifying the causal predictor, and validate their findings across various experimental settings, including real images and audio. The paper concludes by highlighting the need for further research to apply these insights to real-world AV systems.
Methodology
The authors develop a structural causal model (SCM) to analyze the relationship between audio and video in AV generation. They conduct controlled experiments to demonstrate the failure modes of existing models and validate the necessity of counterfactual invariance through various interventions on nuisance variables.
Results
The study finds that models relying on visual appearance for audio generation perform poorly when the appearance-event correlation is disrupted, with errors increasing significantly. The proposed interventions based on counterfactual invariance restore robustness without harming performance on unconfounded data.
Implications
The findings suggest that future AV generation models should incorporate mechanisms to address visual shortcuts, particularly through interventions on nuisance variables. This could lead to more robust models capable of generalizing better to unseen data and real-world applications.
Overlay_dx - Automating forecasting evaluation
Time Series
Optimization
Interpretability
- Introduction of overlay_dx as a visual evaluation metric for time series predictions.
- Combines visual interpretability with quantitative scoring through area under the overlay curve.
- Demonstrates effectiveness across various time series prediction scenarios.
- Enhances communication of model performance to non-technical stakeholders.
Read more
Overlay_dx - Automating forecasting evaluation
Summary
The paper introduces overlay_dx, a novel evaluation metric designed to enhance the assessment of time series prediction models. Traditional evaluation metrics often provide numerical values that lack comprehensibility, making it difficult for practitioners to differentiate model performances effectively. Overlay_dx addresses this issue by offering a visual representation of prediction accuracy through confidence intervals around actual values. It also computes the area under the overlay curve, providing a quantitative measure of alignment between predicted and actual values across various thresholds. The authors conducted extensive experiments demonstrating that overlay_dx offers a unified evaluation framework that combines visual and numerical assessments, facilitating improved model comparison and offering valuable insights for optimization in time series prediction. The implementation of overlay_dx as an open-source project ensures reproducibility and community adoption, with utilities for visualizing overlay curves and computing Area Under Curve (AUC) scores.
Methodology
The authors developed overlay_dx by creating a visual metric that represents prediction accuracy through confidence intervals. They also introduced a quantitative scoring mechanism based on the area under the overlay curve. Extensive experiments were conducted to validate the effectiveness of the metric across different time series prediction scenarios.
Results
The results indicated that overlay_dx provides a more intuitive understanding of model performance compared to traditional metrics. It was shown to be robust against outliers and capable of capturing performance characteristics across varying prediction thresholds, thus facilitating better model comparisons.
Implications
The overlay_dx metric has the potential to improve the evaluation of time series prediction models in various domains, such as finance and industrial monitoring, by providing clearer insights into model performance. Its visual nature makes it particularly useful for communicating results to stakeholders with varying levels of technical expertise.
SupportCal: Label-Free Calibration of Post-Trained LLMs via Reference Support and Corroboration
NLP
Large Language Models
- Introduces SUPPORTCAL, a label-free calibration method for PoLMs using PLM references.
- Demonstrates that moderate inclusion of disagreement examples can improve calibration.
- Retains unit weight for agreement examples while assigning continuous weights to disagreements.
- Empirical results show lower Expected Calibration Error (ECE) compared to baseline methods.
Read more
SupportCal: Label-Free Calibration of Post-Trained LLMs via Reference Support and Corroboration
Summary
The paper addresses the issue of confidence calibration in post-trained language models (PoLMs), which often become overconfident compared to their pretrained counterparts (PLMs). The authors propose a novel method called SUPPORTCAL, which utilizes a label-free calibration approach by leveraging the better-calibrated PLM as a reference. Unlike previous methods that discard disagreement examples entirely, SUPPORTCAL introduces a continuous weighting scheme where agreement examples retain unit weight and disagreement examples are assigned weights based on the PLM's relative support and corroboration from a candidate pool of pretrained references. This method allows for a more nuanced calibration process. The authors conduct a controlled diagnostic experiment that reveals a non-monotonic effect of including disagreement examples, showing that a moderate inclusion can enhance calibration. The empirical results demonstrate that SUPPORTCAL consistently outperforms the agreement-only baseline across various configurations in tasks such as MedMCQA and MathQA, indicating its effectiveness in improving calibration without the need for labeled data.
Methodology
SUPPORTCAL employs a continuous weighting mechanism for calibration, where agreement examples are given unit weight and disagreement examples receive weights based on the own-base PLM's support for the PoLM's predictions. The method fits a single temperature using weighted KL alignment to the PLM reference, allowing for a more flexible calibration process that incorporates both agreement and disagreement examples.
Results
The empirical evaluation across datasets like MedMCQA and MathQA shows that SUPPORTCAL achieves lower mean ECE than the agreement-only baseline for nearly all target-model configurations. Additional results from the TweetEval Sentiment task confirm the method's effectiveness in a fixed-label classification context.
Implications
SUPPORTCAL's approach to calibration could be beneficial in scenarios where labeled calibration data is scarce or expensive, allowing for improved model reliability in various applications of language models. This method could enhance the deployment of PoLMs in real-world tasks, ensuring that their confidence levels are more aligned with actual performance.
K-TRAIL: Simulator-Guided Generative Design of EM/RF Circuits
Generative Models
Optimization
- K-TRAIL combines diffusion models with Kalman correction for RF circuit design.
- The framework allows for automated synthesis from both S-parameter targets and RF constraints.
- Simulator-guided generation improves layout response accuracy and diversity.
- K-TRAIL addresses the non-uniqueness of inverse EM design effectively.
Read more
K-TRAIL: Simulator-Guided Generative Design of EM/RF Circuits
Summary
The paper introduces K-TRAIL, a novel simulator-guided generative framework aimed at automating the design of electromagnetic (EM) and radio frequency (RF) circuits. The challenge of inverse design in this domain stems from the non-unique relationship between circuit layouts and their electrical responses, compounded by the computationally intensive nature of full-wave simulations. K-TRAIL addresses these challenges by integrating diffusion-based layout generation with derivative-free ensemble Kalman guidance, allowing for real-time feedback from a black-box EM simulator to refine candidate layouts during the generation process. This framework can synthesize layouts based on either prescribed S-parameter responses or RF performance constraints. Experimental results demonstrate that K-TRAIL significantly enhances the alignment of generated layouts with target responses and can discover structurally distinct layouts that meet circuit-level design requirements. The approach represents a significant advancement towards generative, verification-aware RF circuit design while maintaining the flexibility to explore a wide range of layout topologies.
Methodology
K-TRAIL employs a conditional diffusion model to generate a distribution of plausible circuit layouts based on desired RF responses. It incorporates a Kalman correction mechanism that utilizes feedback from an EM simulator to refine these layouts iteratively, without requiring gradient information or differentiable solver models. This approach allows for the generation of layouts that can satisfy specific design constraints and targets.
Results
Experiments conducted on multi-layer RFIC structures showed that K-TRAIL significantly improves the agreement between generated layouts and target S-parameter responses. The framework successfully identifies diverse and structurally distinct layouts that fulfill the required circuit-level specifications, demonstrating its effectiveness in practical design scenarios.
Implications
K-TRAIL has the potential to revolutionize the design process of RF circuits by providing a more efficient and automated approach to layout generation. Its ability to incorporate real-time feedback from EM simulations could lead to faster design cycles and more innovative circuit designs, ultimately benefiting industries reliant on advanced RF technologies.
Knowledge-Graph-Augmented Chronos-2 for HEC-RAS Surrogate Forecasting
Time Series
Graph Learning
Optimization
- KG-Chronos-2 integrates a frozen time-series model with hydraulic project knowledge for improved forecasting.
- The method outperformed existing models, achieving the lowest RMSE among the evaluated systems.
- The study demonstrates the effectiveness of knowledge graphs in enhancing surrogate modeling for hydraulic forecasting.
- Results indicate significant reductions in forecasting errors compared to traditional models.
Read more
Knowledge-Graph-Augmented Chronos-2 for HEC-RAS Surrogate Forecasting
Summary
This paper explores the integration of a time-series foundation model with hydraulic project knowledge to enhance surrogate forecasting of water-surface elevation (WSE) in the Hydrologic Engineering Center’s River Analysis System (HEC-RAS). The proposed method, KG-Chronos-2, combines a frozen Chronos-2 predictor with techniques such as exact-state residual decoding, graph-conditioned historical retrieval, and input-aligned correction. The authors evaluate KG-Chronos-2 against several benchmarks, including persistence, residual LSTM, project-conditioned recurrent GeoFNO, hydraulic DCRNN-style models, and the frozen Chronos-2. The evaluation utilizes data from simulations covering 4,675 cross sections across 71 reaches. Results indicate that KG-Chronos-2 significantly outperforms the other models, achieving a root-mean-square error (RMSE) of 0.246970 in native WSE units, which is a 14.13% improvement over frozen Chronos-2 and a 39.54% improvement over recurrent GeoFNO. The findings suggest that coupling a frozen temporal predictor with project knowledge can effectively enhance forecasting accuracy for HEC-RAS applications.
Methodology
The authors developed KG-Chronos-2 by coupling a frozen Chronos-2 predictor with hydraulic project knowledge. This involved using exact-state residual decoding, graph-conditioned historical retrieval, and input-aligned correction to enhance the forecasting process. The model was evaluated against various benchmarks using data from multiple simulations, focusing on the accuracy of WSE predictions across numerous cross sections.
Results
KG-Chronos-2 achieved an event-balanced RMSE of 0.246970, representing a 14.13% reduction compared to frozen Chronos-2, a 29.38% reduction compared to the hydraulic DCRNN-style model, and a 39.54% reduction compared to recurrent GeoFNO. The model also demonstrated the lowest active-window and final-lead RMSE among the evaluated systems, with a significant bootstrap interval favoring its performance over frozen Chronos-2.
Implications
The findings suggest that integrating knowledge graphs with time-series models can significantly enhance the accuracy of hydraulic forecasting. This approach could be applied to other domains requiring surrogate modeling, potentially leading to more efficient and accurate predictions in complex systems.
Particle Competition and Cooperation for Robust Graph Convolutional Network Learning Under Label Noise
Graph Learning
- PCC refines noisy labels before GCN training without modifying the GCN architecture.
- PCC+GCN shows improved robustness across multiple conventional label-noise models.
- Achieved the best overall average rank in the NoisyGL benchmark.
- Remains competitive under instance-dependent label noise.
Read more
Particle Competition and Cooperation for Robust Graph Convolutional Network Learning Under Label Noise
Summary
This paper presents a novel framework called PCC+GCN, which integrates Particle Competition and Cooperation (PCC) with Graph Convolutional Networks (GCNs) to enhance robustness against label noise in graph learning tasks. The PCC component acts as a preprocessing stage that refines noisy labels by identifying suspicious labeled nodes through particle dynamics, allowing for the preservation, removal, or reassignment of labels before GCN training. This hybrid approach leverages the strengths of both PCC for label correction and GCN for effective representation learning. The authors conducted extensive experiments on ten datasets from the NoisyGL benchmark, evaluating the method under various conventional label noise models, including Uniform, Pair, and Random noise, as well as instance-dependent noise. The results demonstrate that PCC+GCN significantly improves classification accuracy and ranks highest among competing methods, while also being the fastest robust method on the majority of datasets tested. The findings suggest that PCC-based label refinement is a viable and efficient strategy for enhancing GCN performance in the presence of noisy supervision.
Methodology
The proposed PCC+GCN framework combines a label-refinement stage using Particle Competition and Cooperation with the training of a Graph Convolutional Network. PCC identifies and refines labels of nodes in the graph, which is then used by GCN for final classification. The framework also allows for the augmentation of the graph with feature-based k-nearest-neighbor edges, while the GCN operates on the original graph structure.
Results
PCC+GCN achieved the highest overall average accuracy and the best average rank among evaluated methods on the NoisyGL benchmark, with an average gain of 1.67 percentage points over baseline GCNs. Under instance-dependent noise, it remained competitive with leading robust methods while being the fastest on eight out of ten datasets.
Implications
The findings indicate that integrating label refinement strategies like PCC with GCNs can significantly enhance performance in real-world applications where label noise is prevalent. This approach can be particularly useful in domains such as social network analysis, biological networks, and any graph-structured data scenarios where accurate labeling is challenging.
Efficient Mixture-of-Experts with Speculative Decoding via Expert Coactivation
NLP
Large Language Models
Efficient ML
- Combining MoEs with SD can lead to significant inference speed improvements.
- High expert coactivation during training enhances runtime efficiency.
- Four specific training modifications can improve expert coactivation.
- The proposed method achieves a 21% increase in throughput over standard MoEs.
Read more
Efficient Mixture-of-Experts with Speculative Decoding via Expert Coactivation
Summary
This paper addresses the challenges of combining Mixture-of-Experts (MoE) models with Speculative Decoding (SD) to enhance inference speed. While SD can accelerate dense models by verifying multiple tokens in parallel, its effectiveness with MoEs is limited by the number of tokens verified, which increases memory transfer costs from DRAM to the Neural Processing Unit (NPU). The authors investigate how the design of the MoE router during training affects the runtime of MoEs when used with SD. They discover that routers promoting high expert coactivation lead to significantly improved runtimes. The study proposes four training modifications: a global load-balancing loss, shared experts, a consistency loss, and an autoregressive expert selection mechanism. These changes enhance expert coactivation, resulting in a 21% increase in throughput compared to baseline MoEs while maintaining comparable accuracy. The findings suggest that optimizing the training process can effectively mitigate the memory transfer bottleneck associated with MoEs and SD.
Methodology
The authors analyze the interaction between MoE routing and SD, focusing on training-time modifications to improve expert coactivation. They implement a global load-balancing loss, shared experts, a consistency loss, and an autoregressive expert selection mechanism to encourage neighboring tokens to select similar experts, thereby reducing memory transfer costs during inference.
Results
The proposed modifications resulted in a model that improved throughput by 21% over MoE baselines while maintaining accuracy. The study demonstrated that enhancing expert coactivation effectively mitigates the memory transfer bottleneck associated with using SD in MoE architectures.
Implications
The findings suggest that optimizing the training process of MoEs can lead to more efficient inference in large-scale language models, making them more practical for real-time applications. This work could influence future designs of MoE architectures and their integration with decoding strategies.
Kinks vs. Smoothness: Identifiability of Real Analytic nICA for Laplace-like Sources
Theory
Generative Models
Interpretability
- Proves identifiability of nICA for real analytic functions with Laplace-like sources.
- Introduces Real Analytic Decoders (RAD) that can be easily integrated into existing frameworks.
- Demonstrates the method's effectiveness through experiments on synthetic and real datasets.
- Highlights the importance of smoothness in functions for achieving identifiability in nICA.
Read more
Kinks vs. Smoothness: Identifiability of Real Analytic nICA for Laplace-like Sources
Summary
This paper addresses the challenge of identifying independent factors in complex data through nonlinear Independent Component Analysis (nICA). The authors prove that nICA can achieve identifiability for real analytic generating functions when the source probability density functions exhibit a finite number of discontinuities in their first derivative, with the Laplace distribution being a key example. The proof contrasts the presence of kinks in the source distribution with the smoothness of real analytic functions. The authors introduce a method called Real Analytic Decoders (RAD), which can be implemented with minimal modifications to existing training pipelines, such as Normalizing Flows and Variational Autoencoders. The paper includes experiments on both synthetic and real datasets, including the CelebA dataset, demonstrating the method's ability to recover interpretable latent factors. The findings suggest that the proposed approach enhances the stability and reproducibility of training in nICA, addressing issues of non-identifiability that can hinder deep learning performance.
Methodology
The authors utilize theoretical proofs to establish identifiability conditions for nICA under specific assumptions about the source distributions. They also conduct empirical experiments using Normalizing Flows and Variational Autoencoders to validate their theoretical findings, focusing on the recovery of latent factors from complex datasets.
Results
The paper successfully proves that nICA can recover true latent factors from Laplace-like sources with a finite number of discontinuities in their first derivative. Experimental results on the CelebA dataset show that RAD can recover interpretable latent variables, confirming the theoretical claims of identifiability and demonstrating good training stability.
Implications
The findings have significant implications for unsupervised learning and representation learning, particularly in enhancing the interpretability and reliability of models that rely on nICA. The ability to recover independent factors accurately can improve the performance of machine learning systems in various applications, including image analysis and financial modeling.
Blind Thermodynamic Ontology Discovery from Anonymous Experiments
Theory
- Establishes an operational foundation for thermodynamic ontology discovery from anonymous experiments.
- Develops a polynomial-time algorithm that extracts extensive and intensive scaling sectors.
- Validates the framework through blind evaluations on simulated and real fluid datasets.
- Demonstrates that omitting any of the four experimental operations leads to unresolved physical ambiguities.
Read more
Blind Thermodynamic Ontology Discovery from Anonymous Experiments
Summary
This paper addresses the challenge of discovering the underlying thermodynamic ontology from anonymous experimental data, where measurements are obscured by an unknown linear mixture of extensive states and intensive responses. The authors propose an operational identifiability theory and a polynomial-time algorithm that systematically extracts extensive and intensive scaling sectors from experimental contrasts. The methodology involves four key physical interventions: replication, thermal contact, integrability and concavity, and reservoir coupling, which progressively reduce the observational symmetry group to a precise thermodynamic gauge. The authors validate their framework through blind evaluations on van der Waals fluids and Curie–Weiss magnets, demonstrating the ability to resolve phase transitions without prior knowledge of coordinate systems. Additionally, external validation across six real fluids from the NIST WebBook confirms the robustness of the operational ontology discovery across different physical substances, effectively addressing the non-identifiability barrier in thermodynamic systems.
Methodology
The authors formulate an operational identifiability hierarchy and develop a direct contrast estimator. They utilize four physical interventions to reduce the observational symmetry group, enabling the recovery of extensive and intensive quantities from anonymous measurements. The methodology is grounded in thermodynamic principles and employs mathematical constructs from differential geometry.
Results
The proposed framework successfully recovers thermodynamic quantities from anonymous data, resolving phase transitions in both simulated and real fluids. The blind evaluations confirm the algorithm's robustness under ill-conditioned mixing, while external validation demonstrates the transferability of the operational ontology discovery across various physical substances.
Implications
This work has significant implications for scientific machine learning, particularly in fields where understanding thermodynamic properties is crucial. The ability to autonomously discover thermodynamic ontologies from uncalibrated experiments could enhance the modeling of complex systems in physics, chemistry, and engineering.
Triggers and Diagnostics for LLM-Based Interpretability Failures in Active Inference Agents
NLP
Large Language Models
Interpretability
- LLM explainers can produce fluent but incorrect narratives, undermining operator oversight.
- Significant interpretability failures were observed, including belief-drift blindness and sycophantic rationalization.
- The study highlights the need for auditing LLM explainers in agentic deployments.
- Proposed mitigations for identified failures were not evaluated, indicating a gap for future research.
Read more
Triggers and Diagnostics for LLM-Based Interpretability Failures in Active Inference Agents
Summary
This paper investigates the reliability of Large Language Model (LLM) explainers used in conjunction with Active Inference (AIF) agents, particularly in safety-critical applications. The authors audit the performance of LLM explainers attached to an AIF agent managing German energy grid demand, examining how these explainers respond to various triggers that could lead to interpretability failures. The study employs three LLM backends (GPT-4o, Claude-3-Opus, Gemini) and tests them against three black-box triggers. The findings reveal significant shortcomings: the explainers fail to flag substantial deviations in the agent's beliefs caused by corrupted observation streams, and they often produce rationalizations for objectively incorrect actions. The paper highlights the lack of mechanisms in the explainer architecture to verify the truthfulness of the explanations provided, raising concerns about the effectiveness of LLM explainers in ensuring operator oversight. While the authors propose potential mitigations for the identified failures, they do not evaluate these solutions, suggesting that future work should focus on verifying the effectiveness of proposed fixes.
Methodology
The authors constructed an LLM explainer integrated with an Active Inference agent managing energy demand data. They conducted a series of tests using black-box triggers to assess the explainers' responses across three different LLM backends, measuring the accuracy and reliability of the explanations generated during these tests.
Results
The study found that none of the 30 explanations flagged significant belief drifts caused by observation corruption, and 80-95% of the explanations rationalized objectively wrong actions. Additionally, data exfiltration attacks were successful across all three LLM providers, demonstrating vulnerabilities in the metadata channel.
Implications
The findings underscore the critical need for robust auditing mechanisms for LLM explainers in autonomous systems, particularly in safety-critical domains. The results suggest that reliance on LLM-generated explanations without verification can lead to misguided actions by operators, potentially resulting in harmful consequences.
A Distributional Optimisation Perspective on Combining Models in Deep Learning
Optimization
Theory
Large Language Models
- Distributional optimisation provides a unified framework for combining models in deep learning.
- The paper reformulates ensemble methods and LoRA averaging as entropy-regularised optimisation problems.
- The ensemble method is convex, allowing for stronger convergence guarantees compared to LoRA averaging.
- Empirical studies demonstrate the effectiveness of the proposed methods on various tasks.
Read more
A Distributional Optimisation Perspective on Combining Models in Deep Learning
Summary
This paper explores the integration of distributional optimisation into the model combination process in deep learning, aiming to improve performance by jointly training models rather than relying on ad hoc methods. The authors present two common strategies for model combination: ensembles and low-rank adapter (LoRA) averaging, and reformulate these approaches as entropy-regularised distributional optimisation problems. They demonstrate that while the ensemble method leads to a convex optimisation problem, the LoRA averaging does not, which affects convergence guarantees. The paper assesses various algorithms, including a functional variant of variational gradient descent (VGD), and conducts empirical studies on synthetic classification tasks and fine-tuning large language models on commonsense reasoning benchmarks. The findings suggest that a principled approach to model combination can enhance predictive performance and robustness, addressing the limitations of traditional methods that often lack systematic frameworks.
Methodology
The authors cast model combination strategies as optimisation tasks over probability distributions, formulating loss functions for ensembles and LoRA averaging. They explore various algorithms, including a functional variant of variational gradient descent (VGD), and conduct empirical evaluations on synthetic tasks and language model fine-tuning.
Results
The study shows that the proposed distributional optimisation framework can lead to improved performance in model combination tasks. The empirical results indicate that joint training of models enhances predictive accuracy and robustness, particularly in the context of large language models.
Implications
This work suggests that adopting a distributional optimisation perspective can lead to more effective model combination strategies, potentially influencing future research in ensemble methods and model fine-tuning. It opens avenues for more systematic approaches in machine learning that leverage model diversity for better generalisation.
What Must Survive? Exact Task-Information--State Frontiers for Resource-Sufficient Learning
Theory
Efficient ML
Optimization
- Introduces exact task-information-state frontiers for resource-efficient learning.
- Demonstrates how advance task information can reduce the required state dimension.
- Establishes NP-hardness of finding optimal advice partitions.
- Presents theoretical results supported by practical examples.
Read more
What Must Survive? Exact Task-Information--State Frontiers for Resource-Sufficient Learning
Summary
This paper investigates the relationship between task information and state retention in the context of resource-efficient learning. It addresses how much state can be saved when limited advance task information is available before state formation. The author introduces the concept of exact task-information-state frontiers, which are defined for a finite family of linear tasks. The main result is the formulation of the exact frontier as a minimization problem over partitions of tasks, revealing that advance task information can significantly reduce the required state dimension through low-rank joint task operators. The paper also discusses the NP-hardness of finding optimal advice partitions and presents several examples illustrating the theoretical findings. Notably, the paper provides insights into task-aware compression and predictive-state representations, emphasizing the importance of understanding the geometry of task relationships in determining state sufficiency.
Methodology
The methodology involves defining a framework for task-aware compression where an oracle reveals task information before state formation. The author formulates the problem as a minimization over partitions of tasks, utilizing concepts from linear algebra such as rank and singular values to derive exact and approximate frontiers. The paper also includes proofs of NP-hardness for optimal partitioning problems and explores the implications of these findings through theoretical and practical examples.
Results
The main results include the exact formulation of the task-information-state frontier, which shows that the minimum retained dimension for exact recovery of tasks can be computed through task partitions. The paper also establishes a common-core law and demonstrates the NP-hardness of optimal advice partitioning. Additionally, it provides robust approximate frontiers for scenarios with nonzero recovery error, further extending the theoretical framework.
Implications
The findings have significant implications for designing efficient learning systems that can adapt to varying task information availability. The insights into task-aware compression and state retention can inform the development of more efficient algorithms in machine learning, particularly in multi-task learning and predictive modeling. The NP-hardness results also highlight the computational challenges in optimizing task representations, suggesting the need for heuristic or approximate methods in practical applications.
Discovering Physical Representation Languages
Theory
- Introduces the concept of physical representation-language discovery from controlled experiments.
- Develops a polynomial-time procedure for recovering measurement types and their relationships.
- Establishes an end-to-end measurement bound and matching minimax rates for finite experiments.
- Demonstrates robustness of the framework across various physical conditions and complexities.
Read more
Discovering Physical Representation Languages
Summary
This paper addresses the challenge of discovering physical representation languages, which involves identifying the underlying ontology of measurements from anonymous controlled experiments. The authors propose a theoretical framework and a polynomial-time algorithm to recover various elements such as carrier and differential sequences, measurement types, and dual relations. They introduce an identifiability theory that transforms material nuisances into manageable components and separates quantities from coordinates. The framework is validated through blind Maxwell experiments, demonstrating its robustness across different physical regimes, including non-Markovian memory and nonlinearities. The authors aim to shift scientific machine learning from merely learning laws in a predefined language to discovering the language in which these laws can be expressed, establishing theoretical limits on observational identifiability.
Methodology
The authors utilize a combination of theoretical analysis and constructive algorithms to recover the hidden ontology of measurements. They define a representation language as a tuple that includes various components like primal carriers and differential sequences. The methodology incorporates identifiability theory and uses blind Maxwell experiments to validate the framework under different conditions.
Results
The proposed framework successfully recovers complete primal and relative-dual ontologies from anonymous observations. It scales effectively to large systems and demonstrates robustness against severe physical challenges. The authors provide theoretical guarantees on the identifiability of the measurement framework and establish a two-stage measurement bound with a matching minimax rate.
Implications
This work has significant implications for scientific machine learning, particularly in the fields of physics and engineering, where understanding the underlying representation of physical laws can lead to more accurate modeling and predictions. It opens avenues for further research into automated discovery of physical laws and their representations.
Multiple latent orderings better predict language model preferences
NLP
Large Language Models
Theory
- Intransitivity in LLM preferences indicates multiple latent orderings rather than sampling noise.
- The proposed MBT model outperforms traditional single-utility models in predicting LLM preferences.
- Aggregate evaluations can obscure the underlying preference heterogeneity among models.
- Plural preferences in LLMs necessitate a reevaluation of alignment and evaluation methodologies.
Read more
Multiple latent orderings better predict language model preferences
Summary
This paper addresses the inconsistencies observed in language model (LLM) preferences, particularly the phenomenon of intransitivity where a model may prefer item A over B, B over C, and yet C over A. Traditional approaches treat these inconsistencies as noise around a single latent ordering. The authors propose that such intransitivity actually reflects the aggregation of multiple latent, internally consistent orderings. They introduce a noise-augmented mixture Bradley–Terry (MBT) model that infers latent preference components from repeated pairwise comparisons across various tasks. The study demonstrates that a mixture of orderings can better explain structural inconsistencies than single-utility models. A case study on Moral Machine dilemmas illustrates that models with differing aggregate preferences can still share latent components. The findings suggest that LLMs embody plural preferences, indicating that evaluation frameworks treating LLM preferences as singular may overlook significant underlying diversity.
Methodology
The authors conducted experiments using seven open-weight LLMs across four pairwise comparison tasks, including both factual and subjective comparisons. They employed a noise-augmented mixture Bradley–Terry model to analyze the pairwise preferences and identify latent preference components.
Results
The study found that no single preference utility could account for the observed intransitivities, confirming the presence of structured noise. The MBT model demonstrated superior predictive capabilities compared to a single-component Bradley–Terry baseline. Additionally, the analysis revealed that distinct components of models aligned in aggregate could disagree when disaggregated, highlighting the complexity of LLM preferences.
Implications
The findings suggest that LLMs may not represent a singular preference structure, which has significant implications for how these models are evaluated and aligned. Understanding the plurality of preferences can enhance the design of LLM applications in sensitive domains such as ethics and social decision-making.
Computationally efficient safe exploration in reinforcement learning
Reinforcement Learning
Robotics
Efficient ML
- Introduction of COLSAFE-MDP for safe exploration in constrained MDPs.
- Utilization of the Nadaraya-Watson estimator for constant-time updates.
- Theoretical guarantees of safety and near-optimality with mild assumptions.
- Significant performance improvements over GP-based methods in terms of safety and computational efficiency.
Read more
Computationally efficient safe exploration in reinforcement learning
Summary
This paper addresses the challenge of safe exploration in reinforcement learning (RL), particularly in constrained environments where safety is paramount, such as robotics and space exploration. The authors introduce COLSAFE-MDP, a novel algorithm that utilizes the Nadaraya-Watson estimator to achieve computational efficiency in safe exploration of constrained Markov decision processes (CMDPs). Unlike traditional Gaussian process-based methods that scale cubically with the number of observations, COLSAFE-MDP operates in constant time per observation, significantly enhancing computational efficiency. The algorithm constructs confidence intervals to define safe sets and guarantees that safety constraints are not violated with high probability while converging to a near-optimal policy. The performance of COLSAFE-MDP is evaluated in both synthetic grid environments and real Martian terrain data, demonstrating its ability to identify nearly three times as many safe states compared to GP-based methods, all while maintaining safety throughout the exploration process.
Methodology
The authors propose COLSAFE-MDP, which employs a recursive formulation of the Nadaraya-Watson estimator tailored for CMDPs. This estimator allows for constant-time updates, enabling efficient computation of confidence intervals that define safe exploration sets. The algorithm ensures that safety constraints are adhered to with high probability while optimizing for rewards.
Results
COLSAFE-MDP was tested in grid-based environments and on observational data from Martian terrain. The results indicated that the algorithm discovered nearly three times as many safe states compared to Gaussian process-based baselines, all while executing in a fraction of the time and without any safety violations.
Implications
The findings suggest that COLSAFE-MDP can be effectively applied in real-world scenarios where safety is critical, such as autonomous navigation in unknown environments, including planetary exploration and robotics. The computational efficiency of the algorithm makes it suitable for real-time decision-making in resource-constrained settings.
On Emergent Capabilities and Model Merging
Theory
Interpretability
- Merging two models with shared emergent capabilities preserves those capabilities.
- Emergent capabilities that are superadditive cannot be recreated through merging.
- When only one model possesses an emergent capability, merging dilutes it faster than trained capabilities.
- The behavior of emergent capabilities under merging differs significantly from that of trained capabilities.
Read more
On Emergent Capabilities and Model Merging
Summary
This paper investigates the effects of model merging on emergent capabilities—behaviors that arise from fine-tuning but were not explicit training targets. The authors analyze two independent testbeds across three model families to understand how emergent capabilities are preserved, created, or diluted through the merging process. They identify three main outcomes: (1) merging retains emergent capabilities when both parent models possess them, (2) emergent capabilities that are superadditive cannot be recreated through merging, and (3) when only one parent carries the emergent capability, merging leads to a faster dilution of that capability compared to trained skills. The findings challenge the assumption that emergent behaviors will survive merging in a manner similar to trained capabilities, highlighting the distinct nature of emergent behaviors in the context of model combination.
Methodology
The authors conducted experiments using two independent testbeds (activation oracles and emergent-misaligned models) across three different model families. They analyzed the effects of merging on emergent capabilities by comparing the performance of merged models against their parent models in terms of capability retention and dilution.
Results
The results indicate that merging retains 83-96% of emergent behavior when both parents share the capability. However, no weighted merge can recreate superadditive emergent capabilities, and when only one parent has the capability, it decays faster than the trained capability. These findings reveal that emergent behaviors do not compose in the same way as trained capabilities.
Implications
The implications of this research are significant for model deployment, particularly in contexts where emergent capabilities are critical for safety and interpretability. Understanding how merging affects these capabilities can guide practitioners in selecting and combining models more effectively.
TTSE: A Two-Track Online Self-Evolution Framework
Large Language Models
Reinforcement Learning
Theory
- TTSE introduces a dual-track evolution mechanism for LLM agents, separating environmental knowledge and execution strategies.
- The framework allows continuous verification and adaptation of environmental facts and task execution procedures.
- A decision-theoretic analysis provides insights into the conditions under which environment-conditioned policies outperform others.
- TTSE shows improved adaptability and performance on standard benchmarks compared to existing methods.
Read more
TTSE: A Two-Track Online Self-Evolution Framework
Summary
The paper introduces TTSE (Two-Track Self-Evolution), a novel framework designed for Large Language Model (LLM) agents operating in dynamic environments. Traditional reinforcement learning approaches often treat environmental knowledge as static, limiting the agent's ability to adapt over time. TTSE addresses this by separating knowledge into two tracks: FACT (environmental facts) and TIP (task-conditioned implementation procedures). This dual-track system allows agents to continuously verify and update their understanding of the environment while simultaneously refining their execution strategies. The authors provide a decision-theoretic framework that decomposes the agent's risk into two components: environment-representation regret and conditional-execution regret. Through extensive experiments, TTSE demonstrates superior adaptability on benchmarks like ALFWorld and ScienceWorld, showing improved performance over existing methods. The framework's compatibility with other self-evolution techniques is also validated, leading to significant performance gains across multiple domains.
Methodology
The authors propose the TTSE framework, which consists of two independent tracks: FACT for environmental knowledge and TIP for execution strategies. They employ a decision-theoretic approach to analyze the risks associated with each track and conduct ablation experiments to validate the framework's effectiveness. The framework is tested on various benchmarks, including ALFWorld and ScienceWorld, and integrated with the Bayesian-Agent algorithm to assess its compatibility and performance improvements.
Results
TTSE outperformed baseline methods on the ALFWorld and ScienceWorld benchmarks, demonstrating superior task adaptation. The integration with the Bayesian-Agent algorithm resulted in a 15.4 percentage point improvement across five major domains of SOPBench. Additionally, TTSE consistently outperformed baseline models in real-world tasks on the PinchBench benchmark across three independent runs.
Implications
The TTSE framework has significant implications for the development of autonomous agents capable of long-term operation in dynamic environments. By enabling continuous self-evolution, it enhances the adaptability and reliability of LLM agents in various applications, including software development, scientific exploration, and interactive systems.
Prioritized Rollouts for Efficient World Model-based Vision-Language-Action Policy Optimization
Reinforcement Learning
Robotics
Optimization
- Introduces U-GROW, a method that prioritizes high-uncertainty states for policy optimization in VLA models.
- Demonstrates that policy uncertainty can effectively identify states with greater potential for improvement.
- Shows that U-GROW can be integrated into existing MBRL pipelines without changing optimization objectives.
- Achieves improved sample efficiency and success rates in both simulated and real-world tasks.
Read more
Prioritized Rollouts for Efficient World Model-based Vision-Language-Action Policy Optimization
Summary
This paper addresses the challenges of fine-tuning Vision-Language-Action (VLA) models using reinforcement learning (RL) in real-world robotic applications, where the cost of interaction is high. The authors propose a novel method called U-GROW (Uncertainty-Guided Rollouts for policy Optimization with World models), which leverages policy uncertainty to prioritize states that offer greater potential for policy improvement. The key insight is that high uncertainty in the policy often occurs at critical decision-making stages, where small changes can significantly affect outcomes. U-GROW introduces a lightweight sampling layer that modifies the rollout-start distribution to focus on these high-uncertainty states, thereby enhancing the efficiency of model-based reinforcement learning (MBRL). The method can be easily integrated into existing MBRL frameworks without altering their optimization objectives. The authors validate U-GROW through experiments on both simulated and real-world manipulation tasks, demonstrating improved sample efficiency and success rates compared to traditional uniform sampling methods.
Methodology
The authors developed U-GROW, which utilizes a short-horizon branched rollout approach combined with an uncertainty-guided sampling layer. This layer filters out low-value states and prioritizes high-uncertainty states for rollout initiation. The method modifies only the rollout-start distribution, allowing for seamless integration into existing MBRL frameworks.
Results
Experiments conducted on RoboTwin 2.0 and LIBERO benchmarks, as well as four real-world manipulation tasks, showed that U-GROW significantly enhances sample efficiency and final success rates compared to uniform rollout-start sampling, all while using the same world model and rollout budget.
Implications
The findings suggest that incorporating policy uncertainty into the rollout process can lead to more efficient training of VLA models in robotics, potentially reducing the costs and time associated with real-world interactions. This approach could be applied to various robotic tasks requiring fine-tuning and optimization.
A Lightweight Plug-in Gate for Transformer-Based Time-Series Forecasters
Time Series
- Introduction of a lightweight pre-encoder gate for regulating covariate admission in Transformer-based forecasting models.
- Implementation of a usage-regularized variant to control average admission without redesigning the forecasting backbone.
- Comprehensive evaluation across multiple datasets and Transformer architectures, demonstrating the effectiveness of the proposed method.
Read more
A Lightweight Plug-in Gate for Transformer-Based Time-Series Forecasters
Summary
This paper addresses the challenge of incorporating external covariates into Transformer-based time-series forecasting models. Traditional approaches often lack an explicit mechanism for regulating how these covariates influence the forecasting process. The authors propose a lightweight plug-in interface, termed a pre-encoder gate, which allows for the admission of covariate representations before they enter the encoder. This gate assigns sigmoid scores to each representation unit, effectively controlling their contribution to the forecasting model. Additionally, a usage-regularized variant is introduced to penalize excessive admission of covariates, aiming to maintain forecasting accuracy while reducing the average admission scores. The proposed method is evaluated on several Transformer architectures, including TimeXer, iTransformer, and Patch Time Series Transformer (PatchTST), using various datasets such as ETTm1, ETTm2, Traffic, Energy, and influenza-like illness (ILI). The experiments involve paired forecasting comparisons, ablation studies, and diagnostic analyses to assess the effectiveness of the gate. The results indicate that the pre-encoder gate maintains competitive forecasting performance compared to baseline models while providing a mechanism to control covariate admission.
Methodology
The authors implemented a pre-encoder gate that computes scores for covariate representation units before they enter the encoder. This gate uses a two-layer multilayer perceptron (MLP) scoring rule and is evaluated under a zero-extra-tuning protocol. The study includes paired forecasting comparisons, gate-placement and initialization ablation, controlled covariate-admission analysis, and a variance inflation factor (VIF)-informed permutation feature importance (PFI) diagnostic case study.
Results
The pre-encoder gate was found to be competitive with baseline models in terms of forecasting accuracy. The usage penalty effectively reduced average admission scores while keeping forecasting errors comparable to the unpenalized settings. The analysis revealed insights into the relationship between learned gate weights and the redundancy of covariates, as assessed by VIF and PFI diagnostics.
Implications
The proposed pre-encoder gate offers a novel approach to enhance the performance of Transformer-based time-series forecasting models by allowing for more controlled integration of external covariates. This could have significant applications in fields such as energy management, transportation planning, and public health analysis, where accurate long-term forecasting is crucial.
$λ$-Controlled GRPO: Turning Flow-Matching Ratio Instability into a Budgeted Resource
Reinforcement Learning
Generative Models
Optimization
- Introduces path variance as a key factor in the instability of reinforcement learning for flow-matching models.
- Proposes λ-Controlled GRPO, which uses analytical calibration instead of empirical stabilizers.
- Demonstrates significant improvements in text accuracy and preference rewards in image generation tasks.
- Establishes a method for budgeting gradient effort based on predicted path variance costs.
Read more
$λ$-Controlled GRPO: Turning Flow-Matching Ratio Instability into a Budgeted Resource
Summary
This paper introduces λ-Controlled GRPO, a novel approach to stabilize reinforcement learning in flow-matching models, particularly in the context of image generation. The authors identify a common source of instability in the training process, termed 'path variance', which arises from the Gaussian transition kernel of the sampler. Unlike previous methods that treat instability as separate issues requiring hand-tuned stabilizers, this work reframes instability as a measurable resource that can be budgeted. The proposed method calibrates the importance ratio behavior analytically rather than relying on noisy empirical statistics, allowing for a more principled allocation of gradient effort across denoising steps. The authors demonstrate that λ-Controlled GRPO significantly improves text accuracy and preference rewards in a text-to-image model compared to existing empirical stabilizers, while effectively managing late-step path variance within intended limits. This approach not only enhances the quality of generated outputs but also provides a systematic framework for understanding and controlling the dynamics of policy updates in reinforcement learning.
Methodology
The authors derive a finite-grid importance-ratio law to predict log-ratio statistics from path variance. They develop a cheap online estimator for path variance and implement λ-Controlled GRPO, which allocates gradient effort based on predicted costs rather than empirical observations. The method is validated through experiments on a text-to-image model under two reward settings.
Results
λ-Controlled GRPO outperforms the strongest empirical stabilizer in terms of text accuracy and preference rewards. It effectively maintains late-step path variance within the intended budget, addressing the systematic overshooting observed in baseline methods.
Implications
The findings suggest that understanding and controlling path variance can lead to more stable and effective reinforcement learning algorithms for generative models. This could have broad applications in image generation and other areas where reinforcement learning is applied to align models with complex reward signals.
On the Limits of Maximal Coding Rate Reduction for Out-of-Distribution Generalisation
Theory
Interpretability
- MCR2 can lead to complete prediction failures under distribution shifts.
- Representations based on unstable environmental features may achieve optimal coding rates but fail in prediction accuracy.
- Incorporating invariance principles from IRM does not eliminate OOD prediction failures.
- The study highlights the need for new assumptions or learning principles for reliable OOD generalization.
Read more
On the Limits of Maximal Coding Rate Reduction for Out-of-Distribution Generalisation
Summary
This paper investigates the limitations of the Maximal Coding Rate Reduction (MCR2) framework in the context of out-of-distribution (OOD) generalization. While MCR2 aims to learn structured, discriminative representations of class-wise submanifolds, the authors demonstrate that it can lead to complete prediction failures when faced with distribution shifts. They identify two primary limitations: first, the MCR2 objective can yield representations based solely on unstable environmental features, which may achieve optimal coding rates but fail to predict accurately when the correlation between features and labels reverses. Second, incorporating principles from invariant risk minimization (IRM) does not resolve these failures, as the same coding operator can remain optimal across different training environments while still leading to poor predictions. The authors provide a systematic analysis of these issues, highlighting the need for additional assumptions or learning principles to ensure stable predictive relationships across environments. This work raises important questions about the reliability of MCR2 representations in practical applications, particularly in scenarios involving distribution shifts.
Methodology
The authors analyze the OOD generalization capabilities of the MCR2 framework using a simple two-class model. They isolate failure mechanisms by demonstrating scenarios where optimal coding geometry does not guarantee reliable predictions. The study also examines the effects of incorporating invariance principles from existing OOD learning approaches.
Results
The analysis reveals that MCR2 can lead to high prediction errors even when coding quality is near optimal. Specifically, representations can rely on unstable features that, while achieving optimal coding rates, result in incorrect predictions when the relationship between features and labels changes. The incorporation of invariance principles does not mitigate these failures.
Implications
The findings suggest that while MCR2 offers a promising framework for representation learning, its limitations in OOD scenarios necessitate further research into additional principles or assumptions that can enhance predictive reliability. This has implications for the deployment of deep learning systems in real-world applications where distribution shifts are common.
Robust Market Making with Hawkes Order Flow and Price Impact via Adversarial Reinforcement Learning
Reinforcement Learning
Time Series
Optimization
- Extension of adversarial reinforcement learning to include Hawkes order arrivals and price impact.
- Introduction of LSTM to handle increased non-stationarity in market conditions.
- Characterization of equilibrium properties from both theoretical and empirical perspectives.
- Development of a robustness evaluation protocol focusing on left-tail return metrics.
Read more
Robust Market Making with Hawkes Order Flow and Price Impact via Adversarial Reinforcement Learning
Summary
This paper addresses the challenges faced by market-making strategies in limit-order-book markets, particularly model uncertainty and regime-shift risk. Traditional models, such as the Avellaneda–Stoikov framework, often rely on simplified assumptions that do not capture the complexities of high-frequency trading environments, such as clustered order flow and price impact. The authors propose an extension of adversarial reinforcement learning (ARL) to incorporate Hawkes self-exciting order arrivals and trade-induced price impact, thereby creating a more realistic market-making environment. They introduce an LSTM module to manage the increased non-stationarity resulting from these enhancements. The paper characterizes the equilibrium properties of the proposed framework and establishes a robustness evaluation protocol focused on left-tail performance metrics. Experimental results demonstrate that the new method significantly improves left-tail return performance in complex microstructure environments without increasing terminal directional inventory bias, indicating a more robust trading strategy.
Methodology
The authors reformulate the market-making problem as a zero-sum game between a market maker and an environmental adversary. They extend the ARLMM framework to include Hawkes processes for order arrivals and trade-induced price impacts. An LSTM model is employed to enhance regime identification amidst increased non-stationarity. The robustness of the proposed method is evaluated using left-tail metrics, specifically CVaR10% and CVaR30%, to assess performance under adverse conditions.
Results
The proposed method outperforms existing market-making strategies in terms of left-tail return performance across various complex microstructure environments. The results indicate that the improvements are not due to increased terminal directional inventory bias, confirming the robustness of the strategy.
Implications
The findings suggest that incorporating more realistic market dynamics into reinforcement learning frameworks can lead to more robust trading strategies. This has potential applications in algorithmic trading and financial market modeling, where understanding and adapting to market microstructure is crucial for performance.
Not All Ranks Are Equal: Budget-Aware LoRA Merging Across Tasks
Efficient ML
Computer Vision
NLP
- Identifies uniform rank allocation as a major source of performance gaps in LoRA merging.
- Introduces Net Utility, a data-free metric for optimal rank allocation across tasks.
- Demonstrates significant performance improvements in multi-task learning with non-uniform rank allocation.
- Tests the method across diverse vision and language tasks, showing consistent enhancements.
Read more
Not All Ranks Are Equal: Budget-Aware LoRA Merging Across Tasks
Summary
This paper addresses the inefficiencies in merging low-rank adapters (LoRAs) for multi-task learning, particularly focusing on the uniform rank allocation assumption that leads to performance gaps between merged and per-task LoRAs. The authors introduce a novel metric called Net Utility, which evaluates the utility of singular directions derived from the Singular Value Decomposition (SVD) of task-specific LoRAs. By scoring these directions based on their contribution to their respective tasks and their interference with others, the method allows for a more effective allocation of rank budgets across tasks and layers. The proposed approach is tested across vision and language tasks, demonstrating that non-uniform rank allocation significantly enhances performance compared to traditional methods that apply uniform budgets. The results indicate an average improvement of +2.1% for vision tasks and +2.2% for language tasks, with some configurations achieving up to +3.8% improvement.
Methodology
The authors propose a greedy approach to rank allocation using the Net Utility metric, which scores singular directions from the SVD of LoRAs based on their utility for their own tasks and their interference with other tasks. This method is applied on top of five different merging techniques across three merging spaces, allowing for a globally budget-constrained selection of singular directions.
Results
The Net Utility based rank allocation outperforms traditional uniform allocation methods, achieving an average performance improvement of +2.1% on vision tasks and +2.2% on language tasks. Some specific method-space combinations even reach improvements of +3.8%.
Implications
The findings suggest that more tailored rank allocation strategies can lead to better performance in multi-task learning scenarios, particularly in applications requiring efficient model deployment, such as on-device or edge computing. This approach could enhance the usability of LoRA in real-world applications where computational resources are limited.
Learning to Move Cities: Deep Meta-Models and Reinforcement Policies for Calibration and Control in Urban Networks
Reinforcement Learning
Optimization
- Introduces a shared latent-space framework for urban transportation calibration and control.
- Develops a combinatorial MLP-autoencoder for efficient simulator calibration.
- Implements a deep Q-learning agent for dynamic traffic optimization.
- Achieves up to 51% reduction in system-wide travel times in empirical evaluations.
Read more
Learning to Move Cities: Deep Meta-Models and Reinforcement Policies for Calibration and Control in Urban Networks
Summary
This paper addresses the complex optimization challenges in urban transportation networks, specifically focusing on the calibration of high-fidelity simulators and real-time operational control. The authors propose a shared latent-space framework that integrates simulator calibration with reinforcement learning control through a common learned representation of urban traffic dynamics. The framework utilizes a combinatorial MLP-autoencoder architecture to create low-dimensional manifolds that link simulator inputs (like origin-destination demand and network parameters) to outputs (such as travel times and congestion patterns). This approach enhances sample efficiency compared to traditional methods, enabling better fitting to observational data within fixed computational budgets. Additionally, a deep Q-learning agent is implemented to optimize dynamic traffic assignments through scheduling and routing adjustments. Empirical evaluations on benchmark networks demonstrate that this approach can reduce system-wide travel times by up to 51% compared to baseline operations. The shared latent representation not only aids in Bayesian calibration but also enhances the reinforcement learning state representation, allowing for more effective control policies. Overall, the paper highlights the transformative potential of deep learning methods in urban mobility planning and management, particularly for large-scale networks where traditional optimization methods struggle.
Methodology
The methodology involves a two-component framework: a deep learning architecture (combinatorial MLP-autoencoder) for simulator calibration and a reinforcement learning system (deep Q-learning) for dynamic network control. The latent representation learned during calibration is reused in the reinforcement learning state space, allowing both modules to operate on a compressed representation of transportation dynamics.
Results
The proposed framework significantly improves sample efficiency in simulator calibration and achieves a reduction in system-wide travel times by up to 51% in benchmark evaluations, demonstrating its effectiveness over traditional methods.
Implications
The findings suggest that integrating deep learning techniques into urban transportation planning can enhance operational efficiency and adaptability, particularly in large-scale networks. This approach could lead to more effective traffic management strategies and improved urban mobility.
When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation
Large Language Models
NLP
Theory
- Identification of 'scientific-judgment collapse' in AI reviewer training.
- Controlled experiments demonstrate the impact of synthetic reviews on judgment diversity.
- Introduction of TrustReviewer, an LLM-based system to mitigate collapse effects.
- Findings suggest that increased synthetic review exposure compresses rating distributions.
Read more
When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation
Summary
This paper investigates the recursive training of AI reviewers using large language models (LLMs) and identifies a phenomenon termed 'scientific-judgment collapse.' The authors explore how AI-generated reviews can influence subsequent AI reviewers, leading to a compression of rating distributions and a decrease in semantic diversity in judgments. The study employs a controlled experimental framework, starting with the Llama 3.1 8B model, which is fine-tuned on official ICLR reviews from 2018-2023. Successor models are then trained on ICLR 2024 data with varying proportions of official and synthetic reviews. The results indicate that increased exposure to synthetic reviews leads to a significant reduction in both same-paper and corpus-level semantic diversity. To address this issue, the authors propose TrustReviewer, an open-source LLM-based system that intervenes at both training and test times to mitigate the effects of scientific-judgment collapse. The paper highlights the risks associated with recursive reviewer training and offers practical solutions to enhance judgment diversity in AI-assisted scientific evaluations.
Methodology
The authors conducted a controlled experiment using the Llama 3.1 8B model, fine-tuning it on official ICLR reviews and generating synthetic reviews for ICLR 2024 papers. Successor models were trained with varying mixtures of official and synthetic reviews to isolate the effects of synthetic exposure on judgment diversity.
Results
The study found that introducing synthetic reviews led to a compression of rating distributions and a decrease in semantic diversity, with reductions of approximately 11% and 5% for same-paper and corpus-level diversity, respectively, as synthetic exposure increased from 0% to 100%. This indicates a homogenization of judgments rather than a shift towards leniency or harshness.
Implications
The findings raise concerns about the recursive training of AI reviewers and its potential to limit the diversity of scientific judgments. The proposed TrustReviewer system offers a framework for improving the quality of AI-generated peer reviews, which could enhance the reliability of AI in scientific evaluation processes.
From Latent Biomarkers to Clinical Rules: Embedding-Guided Rule Mining and Attribution-Based Translation for Interpretable Tabular Learning
Interpretability
- Introduces the concept of latent biomarkers for tabular clinical data.
- Develops a combined attribution method for translating embedding-space rules into raw clinical features.
- Demonstrates improved performance of translated rules over raw-feature rules in five out of six clinical datasets.
- Highlights the importance of interpretable machine learning in clinical decision-making.
Read more
From Latent Biomarkers to Clinical Rules: Embedding-Guided Rule Mining and Attribution-Based Translation for Interpretable Tabular Learning
Summary
This paper addresses the need for interpretable clinical decision support tools that not only provide accurate predictions but also explain the reasoning behind them. Traditional rule-based models often fail to capture complex interactions among clinical variables, leading to suboptimal performance. The authors propose a four-step pipeline that begins with a Feature Tokenizer and Transformer (FT-Transformer) to learn a compact representation of patients from raw tabular features. This representation allows for the identification of latent biomarkers, which serve as data-derived proxies for disease states. Decision rules are then mined in this latent space using small decision trees, and these rules are translated back into measurable clinical features through a combination of gradient-input saliency and CLS attention attribution. The framework was evaluated on six clinical datasets, demonstrating that the translated rules significantly outperformed those derived from raw features in five out of six cases, with notable AUROC improvements. The study highlights the potential of using latent biomarkers for enhanced interpretability in clinical settings, bridging the gap between complex model outputs and actionable clinical insights.
Methodology
The proposed methodology consists of a four-step pipeline: (1) using an FT-Transformer to create a compact patient representation from raw clinical features, (2) identifying embedding dimensions as latent biomarkers, (3) mining decision rules in the latent space with small decision trees, and (4) translating these rules back to raw clinical features using gradient saliency and attention attribution techniques.
Results
The framework was tested on six datasets, showing that translated rules outperformed raw-feature rules in five datasets, with AUROC gains ranging from 0.04 to 0.23. The heart disease dataset showed a drop in AUROC from 0.98 to 0.72 after translation, indicating challenges in compressing high-fidelity latent rules into raw features.
Implications
This work has significant implications for the development of clinical decision support systems that are both predictive and interpretable, enabling clinicians to understand and act on model outputs effectively. It suggests a pathway for integrating advanced machine learning techniques into clinical practice while maintaining transparency and trust.
Explainable Neuro-Fuzzy Prediction for Trustworthy Decision-Making in Maritime
Interpretability
- Introduces a novel explainable decision-making framework for predictive maintenance in maritime systems.
- Integrates a neuro-fuzzy prediction model with a two-stage explainable component.
- Achieves an average AUC-ROC value of up to 99% on benchmark datasets.
- Provides both feature-level and local rule-based explanations for predictions.
Read more
Explainable Neuro-Fuzzy Prediction for Trustworthy Decision-Making in Maritime
Summary
This paper presents a novel explainable decision-making framework that integrates a neuro-fuzzy prediction model with a two-stage explainable component aimed at enhancing predictive maintenance in maritime systems. The framework addresses the critical need for early fault detection in naval propulsion systems by providing both feature-level and local rule-based explanations of predictions made by black box models. The first stage of the explainable component utilizes gradient-based saliency maps to identify influential input features, while the second stage employs a fuzzy decision tree to extract human-readable rules. This approach not only improves the interpretability of predictions but also fosters trust in machine learning systems by making their inferences understandable to users. The framework is evaluated using a deep residual-based neural backbone on various benchmark datasets, demonstrating superior predictive performance with an average AUC-ROC value reaching up to 99%, while also maintaining explainability. This work represents a significant advancement in the field of explainable AI, particularly in the context of maritime applications, where decision-making can have critical safety implications.
Methodology
The proposed framework consists of a neural network-based fuzzy classifier as the primary predictive model, complemented by a two-stage explainable component. The first stage employs gradient-based saliency maps to highlight important features influencing predictions, while the second stage utilizes a fuzzy decision tree to extract interpretable rules from the model's predictions. This architecture allows for automatic rule extraction without prior domain knowledge, enhancing the model's usability and interpretability.
Results
The framework demonstrated exceptional predictive performance, achieving an average AUC-ROC value of up to 99% across various benchmark datasets. The explainability features provided insights into the decision-making process, allowing users to understand the rationale behind predictions, which is crucial for fostering trust in automated systems.
Implications
The proposed framework has significant implications for predictive maintenance in maritime operations, where understanding the reasoning behind machine predictions can lead to better decision-making and enhanced safety. It can be applied to various maritime systems to improve operational efficiency and reduce the risk of failures.
Federated Deep Clustering Networks for High-Dimensional and Heterogeneous Data
Federated Learning
- FedDCN is the first to generalize DCN to the federated setting.
- It operates without a centralized dataset, enhancing privacy.
- The framework employs synthetic data augmentation and geometric regularization for improved robustness.
- Experimental results show superior performance on benchmark datasets under IID and non-IID conditions.
Read more
Federated Deep Clustering Networks for High-Dimensional and Heterogeneous Data
Summary
This paper addresses the challenge of clustering high-dimensional data in a federated learning (FL) context, where data is distributed across clients and privacy is a concern. The authors introduce FedDCN, a novel federated deep clustering framework that generalizes Deep Clustering Networks (DCN) for decentralized data. Unlike previous methods, FedDCN does not require access to a centralized dataset for pretraining, making it suitable for privacy-sensitive applications. The framework optimizes both reconstruction and clustering losses while incorporating synthetic data augmentations and a geometric regularization term to enhance robustness against non-identically independently distributed (non-IID) data. The experimental results demonstrate that FedDCN achieves state-of-the-art performance on benchmark datasets, effectively handling data heterogeneity and improving clustering quality compared to existing federated deep clustering methods.
Methodology
The authors propose a federated protocol for learning Deep Clustering Networks (DCN) that optimizes a combination of reconstruction loss and clustering loss. To tackle the challenges posed by non-IID data, they introduce synthetic data augmentations and a geometric regularization term to align latent spaces across clients. The method is evaluated through experiments that assess its performance under both IID and non-IID scenarios.
Results
FedDCN demonstrated state-of-the-art performance on benchmark datasets, showing significant improvements in clustering quality and robustness against data heterogeneity compared to existing federated deep clustering methods. The experimental evaluation confirmed the effectiveness of the proposed synthetic data augmentations and geometric regularization.
Implications
The findings suggest that FedDCN can be effectively applied in various privacy-sensitive domains where data is decentralized, such as healthcare, finance, and logistics. The framework's ability to handle high-dimensional and heterogeneous data opens avenues for further research and practical applications in federated learning.
MOSAIC-SR: Transformer-Guided Symbolic Regression for Scientific Equation Recovery
Interpretability
Optimization
Theory
- MOSAIC-SR combines neural generation with symbolic search for improved equation recovery.
- The method uses a pretrained Transformer to propose initial equation sketches, avoiding random initialization.
- It achieves the highest symbolic solution rates across multiple datasets while maintaining high predictive accuracy.
- The framework emphasizes the importance of structural repair and constant fitting for accurate symbolic recovery.
Read more
MOSAIC-SR: Transformer-Guided Symbolic Regression for Scientific Equation Recovery
Summary
MOSAIC-SR introduces a novel approach to symbolic regression, which aims to recover closed-form equations from observational data. Traditional methods often struggle to balance flexible structural search with efficient inference, leading to high computational costs and symbolic errors. MOSAIC-SR leverages a pretrained Transformer model to generate multiple initial sketches of equations, which serve as starting points for a more focused search in the expression space. This method combines symbolic search with scale-aware constant optimization and local symbolic repair, allowing for the recovery of both structure and constants in the equations. The framework is evaluated on the SRSD-Feynman dataset and six additional benchmarks, demonstrating superior performance in symbolic solution rates and predictive accuracy compared to existing methods. The results indicate that the use of learned priors can significantly enhance the search process, making it more efficient and effective in recovering scientific equations.
Methodology
MOSAIC-SR employs a hybrid framework that utilizes a pretrained Transformer to generate symbolic sketches based on a subsample of observations. These sketches are then refined through Monte Carlo tree search (MCTS) and local search methods, which alternately optimize the structure and constants of the equations. The approach includes a recovery-oriented pretraining scheme that focuses on learning normalized structural skeletons, deferring specific variable assignments and constant estimations to the inference stage, where observed data can guide these decisions.
Results
MOSAIC-SR outperformed existing symbolic regression methods on the SRSD-Feynman dataset and six additional benchmarks, achieving the highest symbolic solution rates and ranking among the top two methods in predictive accuracy. The framework's ability to focus on promising equation structures and its effective numerical optimization and symbolic repair processes were key to its success.
Implications
The findings suggest that MOSAIC-SR can significantly enhance the process of scientific discovery by providing interpretable models that accurately recover underlying equations from data. This has potential applications in various scientific fields where understanding the relationships between variables is crucial.
GEM-MPC: Balancing Exploration and Exploitation through Expert-Guided Planning
Reinforcement Learning
Robotics
Optimization
- GEM-MPC improves the alignment between planning and learning in reinforcement learning.
- The method combines a greedy planner-cloning policy with an exploratory KL-regularized policy.
- Gated Prior Distillation filters stale planning targets, reducing computational costs.
- GEM-MPC consistently outperforms existing baselines in continuous-control tasks.
Read more
GEM-MPC: Balancing Exploration and Exploitation through Expert-Guided Planning
Summary
The paper addresses the challenge of effective exploration in high-dimensional continuous control within reinforcement learning (RL). The authors propose GEM-MPC, a method that enhances the interaction between planning and learning by using Model Predictive Path Integral control (MPPI). GEM-MPC combines a policy that clones the planner with a KL-regularized policy that explores around it, thus balancing exploitation and guided exploration. The method also introduces Gated Prior Distillation, which selectively learns from stored planning distributions only when they provide better targets than the current prior, mitigating the issues of stale planning data without the computational cost of full reanalysis. The authors demonstrate that GEM-MPC consistently outperforms existing planning-based RL methods across various continuous-control benchmarks while operating under lower computational budgets.
Methodology
GEM-MPC utilizes MPPI to integrate two distinct policies: a greedy policy trained via Behavior Cloning to replicate the planner's actions and a KL-regularized policy that explores around the planner's distribution. The Gated Prior Distillation mechanism selectively updates the learning process based on the quality of stored planning distributions, avoiding the need for full reanalysis.
Results
The experiments conducted show that GEM-MPC outperforms existing planning-based reinforcement learning methods under lower computational budgets, demonstrating improved efficiency and effectiveness in exploration and exploitation.
Implications
The findings suggest that GEM-MPC can be applied in various high-dimensional continuous control tasks, potentially enhancing performance in robotics and other applications requiring efficient exploration strategies in reinforcement learning.
ARM: Attention with Routed-Memory for Learnable Sparse Control
NLP
Large Language Models
Efficient ML
- ARM introduces a learnable soft eviction policy for KV-cache management, enhancing memory preservation.
- The method employs a dynamic retrieval policy that adapts to input complexity, improving attention sparsity.
- Experimental results show ARM's superior performance and efficiency compared to traditional KV-caching methods.
- ARM's design allows for scalable long-context inference, addressing the growing memory demands of LLMs.
Read more
ARM: Attention with Routed-Memory for Learnable Sparse Control
Summary
This paper introduces Attention with Routed Memory (ARM), a novel key-value (KV) caching structure designed to enhance the efficiency of long-context inference in large language models (LLMs). Traditional KV caching mechanisms face limitations due to their reliance on fixed eviction policies, which can lead to information loss and increased latency. ARM addresses these issues by implementing a fully differentiable memory system organized as a hierarchical router. Utilizing a Gumbel-Softmax approach, ARM learns to select memory slots and perform sigmoid-gated updates, allowing for soft integration of new and stored information without hard evictions. This method not only preserves critical information but also adapts the number of memory slots accessed based on the complexity of the input context. The experimental results demonstrate that ARM outperforms existing fixed KV-caching approaches in terms of performance and efficiency, providing a scalable solution for both short and long-context reasoning tasks. The findings suggest that ARM's innovative design can significantly improve memory usage and reduce generation latency, paving the way for more efficient LLM deployment.
Methodology
The authors propose a hierarchical routing mechanism for managing KV caches, utilizing Gumbel-Softmax for memory slot selection and sigmoid gating for soft updates. This allows for a learnable eviction policy and adaptive retrieval based on input context complexity.
Results
ARM demonstrated significant improvements in performance on commonsense and long-context reasoning benchmarks, achieving better efficiency in memory usage and reduced generation latency compared to fixed KV-caching approaches.
Implications
The ARM framework could lead to more efficient deployment of large language models, enabling them to handle longer contexts without the prohibitive memory and latency costs associated with traditional caching methods. This has potential applications in various NLP tasks requiring extensive context management.
Multi-Armed Bernoulli Bandits via Minimax Single-Arm Stopping
Theory
Optimization
- Introduces an index policy for MAB derived from SAB minimax solutions.
- Establishes a semi-infinite linear programming formulation for optimizing stopping policies.
- Achieves a distribution-free regret bound matching the minimax-optimal order.
- Demonstrates superior performance of the proposed policy in numerical experiments.
Read more
Multi-Armed Bernoulli Bandits via Minimax Single-Arm Stopping
Summary
This paper presents a novel index policy for finite-horizon Bernoulli multi-armed bandits (MAB) derived from minimax solutions to single-arm bandit (SAB) problems. The authors focus on minimizing worst-case regret in SAB scenarios, where the decision-maker compares an unknown Bernoulli arm to a known reward. They formulate the minimax problem as a semi-infinite linear programming problem, allowing for the derivation of stopping policies that effectively compare arms based on their cumulative continuation probabilities. The proposed indices are adjusted to ensure monotonicity and incorporate a reward-shortfall cap. The authors establish a distribution-free regret bound of 4.45√KT + 10.75K for K arms over a horizon T, which aligns with the minimax-optimal regret order found in existing literature. The paper also details a finite-grid implementation that quantifies approximation loss and demonstrates through numerical experiments that the SAB-based index policy consistently outperforms benchmark policies in terms of worst-case regret, particularly in two-arm settings where it closely matches the grid-based MAB minimax policy.
Methodology
The authors develop a semi-infinite linear programming formulation to optimize stopping policies in the SAB context. They derive indices based on cumulative continuation probabilities and adjust them for monotonicity and reward-shortfall caps. The methodology includes a finite-grid approximation for practical implementation.
Results
The proposed SAB-based index policy achieves a distribution-free regret bound of 4.45√KT + 10.75K, matching the minimax-optimal regret order. In numerical experiments, it outperforms all benchmark policies across various arm counts and horizons, closely aligning with the minimax policy in two-arm scenarios.
Implications
The findings suggest that the proposed index policy can be effectively applied in various sequential decision-making contexts, such as clinical trials, online experimentation, and resource allocation, where minimizing regret is crucial. The approach provides a robust framework for decision-makers facing uncertainty in arm rewards.
Reinforcement Learning Inspired Black-box Adversarial Attacks for Computer Vision
Computer Vision
Reinforcement Learning
Optimization
- RIBA formulates adversarial perturbation generation as a Markov decision process, enabling the use of reinforcement learning techniques.
- The proposed method significantly reduces the number of queries needed to generate adversarial examples compared to existing black-box attacks.
- RIBA achieves performance comparable to white-box attacks on adversarially trained models, demonstrating its effectiveness.
- The approach is particularly relevant for real-world applications where query costs are a concern, such as in web APIs.
Read more
Reinforcement Learning Inspired Black-box Adversarial Attacks for Computer Vision
Summary
This paper addresses the vulnerability of neural networks in computer vision to adversarial attacks, particularly under the black-box threat model where the attacker has limited access to the target model. The authors propose a novel approach called Reinforcement Learning Inspired Black-box Adversarial attack (RIBA), which formulates the problem of generating adversarial perturbations as a Markov decision process (MDP). By leveraging concepts from reinforcement learning, specifically a soft actor-critic algorithm, RIBA optimizes perturbations efficiently with minimal queries to the target model. The effectiveness of RIBA is demonstrated through experiments on standard datasets, including CIFAR-10 and ImageNet, showing significant reductions in the median number of queries required to generate successful adversarial examples compared to existing state-of-the-art attacks. The results indicate that RIBA can match the performance of white-box attacks even against adversarially trained models, highlighting its potential for practical applications in security-critical AI systems.
Methodology
The authors developed RIBA by modeling the adversarial perturbation generation as an MDP, utilizing a soft actor-critic algorithm to optimize perturbations. The actor generates perturbations while the critic evaluates their effectiveness based on the model's predictions, allowing for efficient exploration of the perturbation space with minimal queries.
Results
RIBA demonstrated a 25.4% reduction in median queries needed to attack a ResNet-18 model on CIFAR-10 and a 22.5% reduction for a ViT-B/16 model on ImageNet. Additionally, RIBA matched the performance of white-box attacks against adversarially trained models, showcasing its robustness and efficiency.
Implications
The findings suggest that RIBA could enhance the security of AI systems in safety-critical applications by providing a more efficient method for generating adversarial examples. This could lead to better defenses against adversarial attacks in real-world deployments, particularly in scenarios where access to the model is limited.
Lifted Bellman Linear Programming for Offline Reinforcement Learning
Reinforcement Learning
Optimization
Theory
- Introduction of the Lifted Bellman Linear Program (LBLP) for offline RL, focusing on in-sample Bellman optimality.
- LBLP avoids regression losses and target networks, stabilizing the training process.
- The unique minimizer of LBLP provides bounds between the best dataset return and the optimal value.
- Approximate Lifted Bellman Unconstrained Minimization (ALBUM) implements LBLP with neural networks, detaching K-step rollout targets.
Read more
Lifted Bellman Linear Programming for Offline Reinforcement Learning
Summary
This paper introduces the Lifted Bellman Linear Program (LBLP) as a novel approach to offline reinforcement learning (RL). Traditional methods often rely on regression losses against bootstrapped value targets, which can lead to instability and value overestimation. The authors propose LBLP, which imposes in-sample Bellman optimality through inequality constraints on the critic, avoiding the need for target networks and exponential moving averages. By formulating LBLP in the joint (Q, V) space, the method ensures that all constraints are based solely on state-action pairs from the dataset. This unique minimizer provides a stable solution that lies between the best dataset return and the optimal value. The paper further discusses the Approximate Lifted Bellman Unconstrained Minimization (ALBUM) method, which implements the LBLP using neural networks and allows for stable training without off-policy corrections. The results demonstrate that ALBUM matches the performance of existing methods while using fewer parameters and less GPU memory, indicating its efficiency and effectiveness in offline RL scenarios.
Methodology
The authors formulate the LBLP as a primal linear program over the joint (Q, V) space, utilizing inequality constraints based on state-action pairs from the dataset. They also introduce hinge penalties to relax the constraints for neural network implementation in ALBUM, allowing for stable training without the need for target networks or EMA updates.
Results
The LBLP formulation provides a unique minimizer that is bounded between the best return in the dataset and the optimal value. ALBUM, when tested on OGBench, matches the average performance of FQL while using fewer parameters and less peak GPU memory, demonstrating its efficiency.
Implications
The proposed LBLP and ALBUM methods could significantly enhance the stability and efficiency of offline reinforcement learning, making it more applicable in real-world scenarios where data is limited and value overestimation is a concern.
DPTM-DT: Dual-Pretrained Transformer Multitask Representation Learning for Drug-Target Prediction
Multimodal
- DPTM-DT integrates multiple representation modalities for improved drug-target prediction.
- The model employs a dual-pretrained approach with GROVER, ESM, and CTD features.
- It utilizes a shared representation for multitask learning across regression and classification tasks.
- Experimental results show superior performance compared to existing methods on benchmark datasets.
Read more
DPTM-DT: Dual-Pretrained Transformer Multitask Representation Learning for Drug-Target Prediction
Summary
The paper introduces DPTM-DT, a dual-pretrained Transformer framework designed for multitask drug-target prediction, addressing limitations in existing models that often use incomplete representations and separate training for related tasks. DPTM-DT integrates GROVER molecular graph embeddings, ESM protein language-model embeddings, and CTD physicochemical descriptors, facilitating a bidirectional cross-modal attention mechanism to exchange information between drug and target modalities. The architecture employs a shared pair representation to simultaneously predict continuous affinity regression, high-affinity binary classification, and six-level affinity classification. Experimental evaluations on the Davis and KIBA datasets demonstrate that DPTM-DT outperforms existing methods across various metrics, achieving significant improvements in mean squared error (MSE), correlation index (CI), area under the precision-recall curve (AUPR), Matthews correlation coefficient (MCC), and Macro-F1 scores. Ablation studies confirm the contributions of dual target representation, gated fusion, and cross-modal attention to the model's performance. This work highlights the effectiveness of multitask learning in drug-target interactions and sets a new benchmark for future research in computational drug discovery.
Methodology
DPTM-DT employs a multitask architecture that combines dual pretrained representations from GROVER for molecular graphs and ESM for protein sequences, along with CTD descriptors. The model uses a shared 256-dimensional space for feature fusion and cross-modal interaction, enabling simultaneous predictions for continuous affinity regression, binary interaction classification, and six-class affinity-level classification.
Results
On the random 80/20 split of the Davis dataset, DPTM-DT achieved MSE/CI values of 0.193/0.917, and on KIBA, 0.120/0.918. For binary classification, it reported AUPR/MCC values of 0.727/0.654 on Davis and 0.798/0.689 on KIBA. In multiclass classification, it achieved Macro-F1/Top-2 scores of 0.800/0.932 and 0.815/0.962, respectively, demonstrating the best overall performance among compared methods.
Implications
The findings suggest that DPTM-DT can significantly enhance drug-target interaction predictions, which is crucial for drug discovery processes such as candidate screening and drug repositioning. The model's multitask learning approach may serve as a foundation for future research in computational biology and pharmacology.
Riemannian Neural Hamiltonian Flows: Geodesic Symplectic Transport and Interpretability
Generative Models
Theory
Interpretability
- Introduction of Riemannian Neural Hamiltonian Flows (RNHF) for generative modeling on Riemannian manifolds.
- RNHF combines fixed kinetic energy from Riemannian metrics with a learned scalar potential and a geodesic integrator.
- The model is symplectic, reversible, and volume-preserving, addressing the challenges of applying generative models to curved spaces.
- A framework for interpretability is established, linking learned potentials to implicit profiles and target distributions.
Read more
Riemannian Neural Hamiltonian Flows: Geodesic Symplectic Transport and Interpretability
Summary
This paper introduces Riemannian Neural Hamiltonian Flows (RNHF), an innovative extension of Neural Hamiltonian Flows (NHF) that operates on Riemannian manifolds rather than Euclidean spaces. The RNHF model integrates a fixed kinetic energy derived from the Riemannian metric with a learned scalar potential and employs a geodesic leapfrog integrator. This approach ensures that the resulting transformation is symplectic, reversible, and volume-preserving, addressing the limitations of existing generative models that typically operate in flat spaces. The paper also provides a framework for interpretability, explaining how the learned Hamiltonian can be understood through the competition between potential forces and momentum distribution pressures. The authors demonstrate that every normalizable potential defines an implicit profile, with the matched potential serving as an interpretable specialization. The paper includes numerical experiments across various geometries (Euclidean, hyperbolic, and spherical) that showcase the competitive sample quality and computational efficiency of RNHF, confirming the interpretability of the learned potential. Overall, this work expands the applicability of Hamiltonian dynamics in generative modeling to curved spaces, offering a robust theoretical foundation and practical tools for future research.
Methodology
The paper develops RNHF by extending NHF to Riemannian manifolds, utilizing a learned scalar potential and a geodesic leapfrog integrator. The model is trained using an evidence lower bound without requiring Jacobian determinants or divergence estimates, allowing for efficient likelihood-based training.
Results
Numerical experiments demonstrate that RNHF achieves competitive sample quality and computational efficiency compared to existing Riemannian continuous normalizing flows. The interpretability of the learned potential is confirmed through the analysis of implicit profiles and matched potentials.
Implications
The introduction of RNHF opens new avenues for generative modeling in complex geometrical spaces, enhancing the interpretability of learned models and potentially improving applications in areas such as physics, computer graphics, and data analysis where data naturally resides on curved manifolds.
ValueDiff: Value-Geometric KV Cache Eviction for Sink-Suppressed LLMs
NLP
Large Language Models
Efficient ML
- ValueDiff is a value-geometric eviction method that improves token retention in sink-suppressed LLMs.
- The method scores tokens based on the dispersion of their value vectors, rather than relying on key-side information.
- Empirical results show that ValueDiff achieves superior retention rates compared to prior eviction methods across multiple benchmarks.
- The study highlights the importance of value geometry as a reliable eviction signal in modern LLM architectures.
Read more
ValueDiff: Value-Geometric KV Cache Eviction for Sink-Suppressed LLMs
Summary
This paper introduces ValueDiff, a novel method for key-value (KV) cache eviction in large language models (LLMs) that exhibit weaker persistent attention sinks. Traditional eviction methods rely heavily on key-side information, which is becoming less effective in modern architectures that utilize techniques like QK-normalization and gated attention. The authors observe that as attention sinks weaken, the dispersion of value vectors increases relative to key vectors. This leads to the development of ValueDiff, which ranks tokens based on the L2 deviation of their value vectors from the cache mean. The method is evaluated under fixed cache budgets during both prefill and generation phases. Results demonstrate that ValueDiff retains a high percentage of dense information across various models, outperforming existing methods significantly. The findings suggest that value geometry serves as a more reliable eviction signal in sink-suppressed models, providing a new perspective on cache management in LLMs.
Methodology
ValueDiff scores tokens by calculating the L2 deviation of their value vectors from the mean value vector in the cache. This approach is based on the observation that models with weaker attention sinks exhibit greater value vector dispersion. The method is evaluated through experiments on various benchmarks, including RULER, LongBench, and MATH-500, under fixed cache budgets during both prefill and generation phases.
Results
ValueDiff retains 88-99% of dense information on the RULER benchmark at a 2k token budget, achieving the best performance on 6 out of 7 models. On LongBench, it averages 92% retention at a 4k budget, outperforming the strongest prior baseline by 9%. Additionally, on MATH-500, ValueDiff is the top-performing non-dense method across all tested models at a 25% cache budget, exceeding previous methods by up to 20 points on gated-attention models.
Implications
The findings from this research could lead to more efficient cache management strategies in LLMs, particularly in resource-constrained environments. By leveraging value geometry for eviction decisions, future models may achieve better performance and memory efficiency, enhancing their deployment in real-world applications.
G-NAC: Graph Neural Automata Clustering via Emergent Domain Formation
Graph Learning
- Introduction of G-NAC, an unsupervised clustering framework leveraging recurrent graph-neural cellular dynamics.
- Formulation of a domain-formation objective that encourages coherent representations without cluster-label supervision.
- Development of a relational inference procedure based on graph-edge distance-rank stability.
- Evaluation against multiple clustering baselines, demonstrating competitive performance and robustness.
Read more
G-NAC: Graph Neural Automata Clustering via Emergent Domain Formation
Summary
The paper presents Graph Neural Automata Clustering (G-NAC), an innovative unsupervised clustering method that partitions numerical feature vectors without relying on cluster-label supervision. G-NAC represents each observation as a cell on a fixed neighborhood graph, where a shared recurrent graph-neural cellular rule evolves hidden and domain states through local interactions. The method emphasizes graph smoothness while incorporating variance, covariance, dispersion, and temporal regularization to prevent representational collapse. The final clustering output is derived from a rank-based sparse spectral affinity generated from the evolved relational state. The authors evaluated G-NAC across 73 dataset/K tasks from 57 datasets, achieving a mean Adjusted Rand Index (ARI) of 0.7951, which is comparable to the baseline Genie (0.7941) and superior to other evaluated methods. The robustness of G-NAC was tested through controlled corruption experiments, revealing minimal degradation under node neighbor deletion but progressive decline under edge rewiring and measurement noise. Additionally, the model demonstrated efficient scaling with empirical fit time and GPU memory usage, and the learned transition rules showed effective transferability across different graph scales. Overall, G-NAC represents a significant advancement in unsupervised clustering by utilizing emergent dynamics to organize observations into coherent domains.
Methodology
G-NAC employs a fixed neighborhood graph to represent observations, utilizing a shared recurrent graph-neural cellular rule to evolve hidden and domain states through local interactions. The method incorporates regularization techniques to maintain coherence in representations and uses a rank-based approach to derive cluster assignments from the evolved relational state.
Results
G-NAC achieved a mean ARI of 0.7951 across 73 dataset/K tasks, outperforming several classical and modern clustering methods. The model showed resilience to node neighbor deletion but was sensitive to edge rewiring and measurement noise. Additionally, it demonstrated efficient scaling in terms of fit time and GPU memory usage, with learned rules effectively transferring across different graph sizes.
Implications
G-NAC has potential applications in various fields requiring unsupervised clustering of complex data structures, such as social network analysis, biological data clustering, and any domain where emergent patterns from local interactions are of interest. Its ability to operate without labeled data makes it particularly valuable in scenarios where obtaining labels is challenging.
$t_0$: A Time-Series Foundation Model for Forecasting with Context
Time Series
- Introduction of $t_0$, a family of time-series foundation models for multivariate forecasting.
- Models $t_0$-alpha and $t_0$-beta utilize transformer architecture with alternating attention mechanisms.
- Significant performance improvements observed when using known-future covariates.
- Models demonstrate strong performance on public benchmarks, ranking third in multiple evaluations.
Read more
$t_0$: A Time-Series Foundation Model for Forecasting with Context
Summary
The paper introduces $t_0$, a family of open-weights foundation models designed for time-series forecasting that incorporates multivariate context. The authors present two initial models, $t_0$-alpha and $t_0$-beta, which feature 102M and 256M parameters, respectively. These models generate forecasts based on historical target data, past covariates, and known-future covariates without requiring task-specific retraining. The architecture employs transformer layers that alternate attention across time and variates, enabling the models to produce probabilistic forecasts through quantile predictions. The pretraining process utilizes a combination of curated public datasets and synthetic generator families to capture covariate-to-target dependencies. Evaluation results show that $t_0$-alpha achieves a Continuous Ranked Probability Score (CRPS) of 0.4941, while $t_0$-beta scores 0.4738 CRPS and 0.6865 Mean Absolute Scaled Error (MASE), ranking third in both metrics on the GIFT-Eval benchmark. The models demonstrate significant improvements in forecasting accuracy when incorporating known-future covariates, with $t_0$-alpha showing a 6.3 percentage point increase in skill across 30 tasks. The paper also discusses the models' robustness to missing data and their performance on real-world benchmarks, including electricity demand forecasting.
Methodology
The authors developed $t_0$ models using a transformer architecture that alternates attention across time and variates. The models were pretrained on a diverse dataset that included both real and synthetic data to capture dependencies between covariates and targets. They evaluated the models on various benchmarks, focusing on their performance in zero-shot forecasting scenarios and their robustness to missing data.
Results
On the GIFT-Eval benchmark, $t_0$-alpha achieved a CRPS of 0.4941, while $t_0$-beta reached a CRPS of 0.4738 and a MASE of 0.6865, both ranking third among evaluated models. The inclusion of known-future covariates improved $t_0$-alpha's skill by 6.3 percentage points across 30 tasks. In independent evaluations, both models reduced the Mean Absolute Error (MAE) of a lagged-price baseline by 38% in forecasting hourly ERCOT prices.
Implications
The $t_0$ models have significant implications for various forecasting applications, including energy management, inventory control, and demand forecasting in retail. Their ability to leverage multivariate context and perform well in zero-shot scenarios makes them suitable for deployment in real-world settings where data may be limited or variable.
NAVIR: Neuromorphic Audio-Visual Speech Recognition for Robust Human-Robot Interaction on Edge Hardware
Multimodal
Audio & Speech
Robotics
- Development of a neuromorphic AVSR system for noisy industrial environments.
- Separation of spatial and temporal encoding into distinct modules for efficiency.
- Significant reduction in word error rates compared to audio-only systems.
- Energy-efficient operation on neuromorphic hardware with substantial power savings.
Read more
NAVIR: Neuromorphic Audio-Visual Speech Recognition for Robust Human-Robot Interaction on Edge Hardware
Summary
The NAVIR project addresses the challenges of voice-controlled interaction in noisy industrial environments by developing a neuromorphic audio-visual speech recognition (AVSR) system. Traditional audio-only speech recognition struggles with acoustic noise, leading to the need for manual interfaces that can slow workflows and compromise safety. NAVIR leverages audio-visual cues, specifically lip motion, to enhance recognition robustness. The proposed system is designed for the BrainChip Akida neuromorphic processor, which imposes strict architectural constraints, necessitating a novel approach that separates spatial and temporal encoding into distinct modules. The architecture includes a per-frame visual encoder, a temporal video encoder, and a spectrogram audio encoder, all integrated with a lightweight predictor head and a constrained beam search decoder. Training involves connectionist temporal classification on noise-augmented audio data, followed by quantization-aware fine-tuning. The results demonstrate significant improvements in word error rates (WER) on the GRID benchmark, achieving 14.0% WER under noise and 3.3% WER on overlapped speakers, compared to audio-only baselines. The system also shows remarkable energy efficiency, with a 13-fold advantage over traditional artificial neural networks and significantly lower energy consumption than conventional CPUs and GPUs. This work represents the first complete multimodal AVSR pipeline operational on neuromorphic hardware, showcasing its potential for real-world human-robot interaction applications.
Methodology
The NAVIR system utilizes an end-to-end architecture that separates spatial and temporal processing into distinct modules compatible with the BrainChip Akida processor. It employs a per-frame visual encoder, a temporal video encoder, and a spectrogram audio encoder, followed by a lightweight predictor head and a constrained beam search decoder. Training is conducted using connectionist temporal classification on noise-augmented datasets, with subsequent quantization-aware fine-tuning.
Results
The quantized audio-visual model achieved a 14.0% word error rate (WER) under noise conditions and 3.3% WER on overlapped speakers on the GRID benchmark, outperforming audio-only baselines. It also attained 98.6% command accuracy at 1.5% WER on a task-specific industrial-command corpus. Energy analysis indicated a 13-fold energy efficiency advantage over traditional ANN models, with on-board measurements showing 5-fold lower energy consumption than a Raspberry Pi CPU and over 100-fold lower than a laptop GPU.
Implications
The NAVIR system has significant implications for enhancing human-robot interaction in industrial settings, providing a robust solution for voice-controlled interfaces in noisy environments. Its energy-efficient design makes it suitable for deployment on edge hardware, potentially transforming how robots interact with humans in various applications.
GraphToolbox: A Configurable Python Framework for Graph Neural Network Forecasting
Graph Learning
Time Series
Interpretability
- GraphToolbox unifies various stages of GNN forecasting into a single framework.
- It supports data-driven graph construction and empirical selection methods.
- The framework includes an adapter for 51 convolution operators and recurrent cells.
- Online expert aggregation improves forecasting accuracy significantly.
Read more
GraphToolbox: A Configurable Python Framework for Graph Neural Network Forecasting
Summary
The paper introduces GraphToolbox, an open-source Python framework designed to streamline the process of forecasting electricity demand using Graph Neural Networks (GNNs). Traditional forecasting methods often struggle with the complexities of spatially related signals, which are inherent in electricity demand across various regions. GraphToolbox addresses this by integrating multiple stages of the forecasting workflow—graph construction, model selection, training, aggregation, and interpretability—into a single, configurable pipeline built on PyTorch Geometric. The framework allows users to easily modify configurations for graph construction and convolution methods, facilitating systematic architectural evaluations. The authors demonstrate the effectiveness of GraphToolbox through two case studies on French regional load and net-load forecasting, showcasing its ability to outperform classical forecasting methods and improve accuracy through online expert aggregation.
Methodology
GraphToolbox employs a configuration-driven approach to streamline the forecasting workflow. It includes methods for data-driven graph construction, an adapter for various convolution operators, online expert aggregation for model forecasts, and tools for interpretability and statistical significance testing. The framework is built on PyTorch Geometric, allowing for flexible integration of GNN components.
Results
In the evaluation of the framework, the forecasting error for French regional load using 48 convolutions ranged from 1.14% to 1.60%, which improved to 0.98% with online aggregation. For net-load forecasting, while direct graph models were less accurate than classical additive models, forecasting physical components separately showed improvements, although the gap remained.
Implications
GraphToolbox has the potential to enhance the accuracy and efficiency of electricity demand forecasting, making it easier for practitioners to implement GNNs in real-world applications. Its systematic approach can facilitate further research and development in spatio-temporal modeling and forecasting in various domains beyond electricity.
Prescriptive SVD-Inspired Attention via Spectral Energy Retention
Computer Vision
Interpretability
Efficient ML
- Introduces a diagnosis–intervention–verification framework for SVD-Inspired Attention (SVDA).
- Demonstrates that spectral energy retention can effectively prune low-energy attention directions.
- Empirical results show a reduction in parameters and computational costs without significant accuracy loss.
- Establishes a structured approach to modify attention mechanisms based on spectral diagnostics.
Read more
Prescriptive SVD-Inspired Attention via Spectral Energy Retention
Summary
This paper introduces a novel framework for enhancing the interpretability and operational utility of self-attention mechanisms in Transformer architectures through a prescriptive approach to SVD-Inspired Attention (SVDA). Traditional self-attention mechanisms are often opaque, making it challenging to discern which attention directions are crucial. SVDA addresses this by incorporating a learned diagonal spectrum into the query-key interaction, allowing for the inspection of attention directions using metrics such as spectral entropy and effective rank. The authors propose a diagnosis–intervention–verification framework that transitions from merely diagnosing attention spectra to implementing actionable interventions. The primary intervention explored is spectral energy retention, which aims to prune low-energy directions in the attention-score pathway. Empirical evaluations on datasets such as FashionMNIST, CIFAR-10, CIFAR-100, and Food-101 demonstrate that this approach can significantly reduce the number of score directions and parameters while maintaining model accuracy. The findings suggest that SVDA can serve as an interpretable mechanism that not only reveals the internal structure of attention but also facilitates targeted modifications to improve model efficiency.
Methodology
The authors developed a framework that integrates diagnostic metrics from SVDA to inform targeted interventions in the attention mechanism. They specifically evaluated spectral energy retention as a method for pruning low-energy attention directions. The methodology involved training SVDA models, diagnosing their attention spectra, proposing interventions based on spectral insights, and verifying the effects of these interventions on model performance and structure.
Results
The empirical evaluation across multiple datasets revealed that applying a spectral energy retention prescription led to a reduction of 24.5–53.7% in score directions, a decrease in parameters by 2.6–4.3%, and a reduction in estimated MACs by 2.8–5.4%. The mean accuracy change of the dimension-reduced models ranged from -0.03 to +0.05 percentage points, indicating that the interventions did not significantly compromise model performance.
Implications
The findings suggest that SVDA can enhance the interpretability of attention mechanisms in Transformers while enabling efficient model modifications. This approach could be beneficial in various applications where model efficiency and interpretability are critical, such as in computer vision tasks and resource-constrained environments.
CTRL: Control-Based Time Series Forecasting with LLM-Guided Residual Learning
Time Series
Large Language Models
- CTRL separates semantic reasoning from numerical prediction, enhancing stability in time series forecasting.
- Specialized LLM agents analyze prediction errors through decomposed components, generating effective correction policies.
- The framework allows for test-time adaptation without relying on ground truth, making it efficient in dynamic environments.
- CTRL shows significant performance gains in non-stationary settings while remaining competitive in stationary scenarios.
Read more
CTRL: Control-Based Time Series Forecasting with LLM-Guided Residual Learning
Summary
The paper introduces CTRL, a novel framework for time series forecasting that leverages large language models (LLMs) in a unique way. Unlike traditional methods that either reduce LLMs to mere numerical predictors or allow them to generate forecasts directly, CTRL separates semantic reasoning from quantitative prediction. It employs a frozen backbone to generate base forecasts while specialized LLM agents act as controllers that analyze prediction errors by decomposing them into trend, seasonal, and irregular components. This approach allows for the generation of compact control signals that are translated into forecast corrections by a lightweight residual decoder. Notably, CTRL incorporates a label-free test-time adaptation mechanism that detects distribution shifts based solely on input statistics, enabling it to adjust control signals efficiently with minimal LLM calls. The framework is designed to enhance robustness in non-stationary environments while maintaining competitiveness in stationary settings. The authors demonstrate that CTRL achieves significant improvements in forecasting accuracy, particularly under non-stationary dynamics, while requiring only a few LLM calls for adaptation.
Methodology
CTRL utilizes a frozen backbone for initial forecasting and employs specialized LLM agents as controllers to analyze prediction errors. These agents generate compact control signals based on decomposed time series components, which are then refined by a lightweight residual decoder. The framework also implements a test-time adaptation mechanism that detects distribution shifts using input statistics alone, facilitating efficient adjustments without ground truth or parameter updates.
Results
The CTRL framework demonstrates substantial improvements in forecasting accuracy, particularly in non-stationary environments, outperforming traditional LLM-based methods. The results indicate that CTRL effectively adapts to distribution shifts with minimal LLM calls, achieving robust performance across various time series datasets.
Implications
CTRL's approach to decoupling reasoning from prediction could lead to more stable and interpretable time series forecasting models. Its ability to adapt to non-stationary dynamics without requiring extensive retraining may have significant applications in fields such as finance, public health, and energy management, where timely and accurate forecasting is critical.
COMPLEX: A Closed-Form Certified Embedding of Multiparameter Persistence Modules
Theory
Optimization
Efficient ML
- Introduces COMPLEX, a closed-form embedding for multiparameter persistence modules.
- Establishes the first two-sided distortion bound for multiparameter feature maps.
- Achieves state-of-the-art performance on Orbit benchmarks without training.
- Demonstrates that separated modules remain distinct in the embedding space.
Read more
COMPLEX: A Closed-Form Certified Embedding of Multiparameter Persistence Modules
Summary
The paper introduces COMPLEX, a novel closed-form embedding for multiparameter persistence modules that addresses the lack of a lower gauge in existing vectorizations. By slicing the persistence module along a near-diagonal net and embedding each slice using the PLACE/PALACE landmark map, COMPLEX achieves a two-sided distortion bound, marking the first instance of a certified feature map for multiparameter persistence. The method guarantees that separated modules remain distinct in the embedding space, thus providing a measure of faithfulness. The authors demonstrate that COMPLEX outperforms existing methods on benchmark datasets, achieving state-of-the-art results without requiring training or calibration. The paper also discusses the implications of the embedding in terms of local per-prediction certification and the effectiveness of various kernel choices in maintaining the certification. Overall, COMPLEX represents a significant advancement in the field of multiparameter persistence and its applications in machine learning.
Methodology
The methodology involves slicing the multiparameter persistence module along a fixed near-diagonal net, embedding each slice using the PLACE/PALACE landmark map, and concatenating the results. The paper establishes a checkable witnessing-slice coherence condition to ensure the closed-form lower gauge, allowing for a two-sided distortion bound.
Results
COMPLEX achieves 91.95% accuracy on the Orbit5k benchmark and 92.98% on Orbit100k, outperforming existing methods including Euler-characteristic surfaces and transformers. It also exceeds the performance of Gril on molecular benchmarks, demonstrating its effectiveness across various applications.
Implications
The findings suggest that COMPLEX can be utilized in applications requiring reliable embeddings of multiparameter persistence modules, such as in topological data analysis and machine learning tasks where faithfulness and certification are critical. The method's training-free nature also makes it suitable for scenarios with limited data availability.
SWE-Proof: Can Language Models Resolve Real-World Issues with Machine-Checked Proofs?
Large Language Models
Theory
Generative Models
- Introduces BENCHPROOFER, a pipeline for formal verification of LLM-generated code.
- SWE-PROOF benchmark includes 500 real-world coding tasks with formal correctness proofs.
- Demonstrates that many test-passing patches are flawed, emphasizing the limitations of traditional testing.
- Finds that the quality of specifications significantly impacts verification success rates.
Read more
SWE-Proof: Can Language Models Resolve Real-World Issues with Machine-Checked Proofs?
Summary
The paper addresses the challenge of ensuring correctness in code generated by large language models (LLMs) in software engineering. Traditional benchmarks rely on test suites that are often incomplete and susceptible to memorization, leading to incorrect solutions being accepted. The authors propose BENCHPROOFER, a pipeline that transforms coding tasks with known correct patches into formally verified tasks. This involves creating specifications for new code, summarizing existing functions with axioms, and requiring agreement from both mechanical and adversarial verification gates before admitting a solution. The resulting SWE-PROOF benchmark includes 500 real-world coding issues with formal correctness proofs, significantly improving upon existing methods. The evaluation of two advanced models, Claude Opus 4.8 and GPT-5.5, reveals that a substantial portion of test-passing patches are flawed, and that the quality of specifications is critical for successful verification. The study highlights the importance of faithful specification synthesis as a key challenge in verified code generation.
Methodology
The authors developed BENCHPROOFER, which constructs formal specifications and proofs for coding tasks by summarizing existing functions and using a combination of mechanical verification and adversarial audits. The SWE-PROOF benchmark was created to evaluate the performance of LLMs on real-world coding tasks, providing ground truth specifications and correctness proofs.
Results
The evaluation showed that between 25% to 50% of patches that passed hidden tests were flawed. Supplying correct formal specifications improved resolution rates from 85% to 95% for Opus 4.8 and from 81.2% to 94.7% for GPT-5.5. However, only 62% of synthesized specifications passed the audit, with faithfulness being the primary failure mode.
Implications
The findings suggest that formal verification can significantly enhance the reliability of code generated by LLMs, but the challenge of creating high-quality specifications remains a critical barrier. This work could lead to improved practices in software engineering and the development of more robust AI systems for code generation.
OpenMAS-GCom. A Diagnostic Benchmark for Graph-enhanced Multi-Agent Systems
NLP
Large Language Models
Graph Learning
- OpenMAS-GCom provides a controlled environment for diagnosing performance in G-MAS by isolating specific components.
- The benchmark evaluates 17 configurations across 29 datasets, introducing 400 complex tasks for comprehensive assessment.
- Experiments reveal significant performance degradation when specialist agents are removed compared to critic agents.
- Different configurations achieve varying levels of accuracy, highlighting the importance of communication structures and role assignments.
Read more
OpenMAS-GCom. A Diagnostic Benchmark for Graph-enhanced Multi-Agent Systems
Summary
The paper introduces OpenMAS-GCom, a benchmark designed to diagnose the performance of graph-enhanced multi-agent systems (G-MAS) by isolating the effects of communication structures, role assignments, and information flows. G-MAS utilizes communication graphs and role assignments to coordinate large language model agents, but existing evaluations struggle to attribute performance differences to specific components due to intertwined factors like model strength and communication patterns. OpenMAS-GCom addresses this by allowing controlled interventions on G-MAS configurations while keeping other variables constant. The benchmark evaluates 17 configurations across 29 datasets and introduces 400 G-MAS-Complex tasks that require agents to synthesize information from multiple sources. The authors conduct experiments demonstrating that the removal of specialist agents significantly impacts performance more than critic removal, and that different configurations yield varying accuracy levels on complex tasks. The benchmark is open-sourced, providing a library for reproducible evaluations and insights into the organizational dynamics of G-MAS.
Methodology
The authors developed OpenMAS-GCom as a diagnostic benchmark that employs controlled interventions to assess the impact of various components in G-MAS. This includes modifying communication links, removing specific agent roles, and altering the content of messages while keeping other factors constant. The benchmark evaluates the performance of different configurations using a common interface that records outputs and resource usage.
Results
The experiments conducted using OpenMAS-GCom showed that the removal of specialist agents led to a mean score reduction of 3.58 percentage points, while critic removal resulted in a smaller reduction of 0.58 percentage points. Additionally, different configurations displayed distinct responses to incorrect messages and worker failures, with varying accuracy levels achieved on the G-MAS-Complex tasks.
Implications
The findings from OpenMAS-GCom can inform the design and evaluation of future G-MAS configurations, helping researchers understand the critical factors that influence performance. The open-sourced nature of the benchmark encourages further experimentation and development in the field of multi-agent systems, particularly in enhancing communication strategies and role assignments.
OneBid: A Unified Auto-Bidding Foundation Model for Diverse oCPX Advertising Scenarios
Reinforcement Learning
Generative Models
Optimization
- OneBid unifies multiple oCPX advertising scenarios into a single auto-bidding model.
- The model employs a Mixture-of-Experts architecture to balance shared and scenario-specific knowledge.
- Introduces a novel optimization method (CROP) for safe offline policy improvement.
- Achieves significant performance gains in conversion scenarios, with an overall +2.2% improvement in ADVV.
Read more
OneBid: A Unified Auto-Bidding Foundation Model for Diverse oCPX Advertising Scenarios
Summary
The paper introduces OneBid, a unified auto-bidding foundation model designed to optimize advertising strategies across diverse cost-per-X (oCPX) scenarios, such as registration and purchase. Traditional auto-bidding methods have evolved from rule-based systems to reinforcement learning and generative models, yet they often operate in a scenario-specific manner, leading to inefficiencies and underutilization of cross-scenario insights. OneBid addresses these challenges by integrating multiple oCPX scenarios into a single model, overcoming issues related to multi-objective control, scalability under latency constraints, and safe offline policy improvement. The model employs a sequence-level Mixture-of-Experts architecture to capture both shared knowledge and scenario-specific patterns while maintaining low latency. It enhances the Decision Transformer framework by introducing two atomic signals—Return-to-Go (RTG) for conversion value and Cost-to-Go (CTG) for cost efficiency. The post-training phase aligns the model with specific deployment preferences using a novel Critic-guided Relative Offline Policy optimization method (CROP), which mitigates risks associated with online exploration. The effectiveness of OneBid is validated through online A/B tests, demonstrating significant performance improvements across various conversion scenarios.
Methodology
OneBid utilizes a sequence-level Mixture-of-Experts architecture to learn from heterogeneous oCPX logs. It extends the Decision Transformer framework by incorporating two control signals (RTG and CTG) and employs a critic-guided optimization method (CROP) for post-training alignment with scenario-specific preferences. The model is pre-trained on diverse data and fine-tuned for specific deployment contexts, ensuring low latency and effective performance.
Results
OneBid has shown a consistent improvement in performance during online A/B testing, achieving an overall +2.2% increase in Average Daily Value per Visitor (ADVV) on oCPX ads, with peak improvements of +13.1% in Return on Ad Spend (ROAS) scenarios. These results validate the model's effectiveness in real-world advertising applications.
Implications
The development of OneBid suggests a shift towards unified models in computational advertising, potentially reducing engineering overhead and enhancing the ability to leverage cross-scenario insights. This approach could lead to more efficient bidding strategies and improved advertiser outcomes across various conversion goals.
PROSE: A Theory of Optimal Stopping with Perishable Evidence for Peer Selection in Intermittently Connected Decentralised Learning
Federated Learning
Theory
Optimization
- Develops a theory of optimal stopping for peer selection in decentralized federated learning.
- Introduces a reservation-value threshold rule for optimal decision-making under perishable evidence.
- Establishes confidence bounds and a maximin certification rule for peer selection.
- Demonstrates the impact of mobility on the value of information and stopping decisions.
Read more
PROSE: A Theory of Optimal Stopping with Perishable Evidence for Peer Selection in Intermittently Connected Decentralised Learning
Summary
This paper presents a theoretical framework for optimal stopping in decentralized federated learning (FL) systems, where peer availability is transient and evidence about peers is perishable. The author formulates the peer-selection problem as a finite-horizon Markov optimal-stopping problem, incorporating costly information acquisition and future-arrival options. The study establishes an optimal policy characterized by a reservation value and provides several key contributions: (i) it derives confidence bounds and a maximin certification rule for peer selection; (ii) it introduces a mobility-aware value-of-information stopping rule that accounts for link hazards; (iii) it presents a closed-form value of waiting under marked-Poisson contact arrivals; and (iv) it demonstrates a myopic-optimality theorem indicating that a one-step confidence-safe rule is effective in volatile mobility regimes. The framework is instantiated as PROSE (Perishable-evidence Reservation-value Optimal Stopping for Exchange), a lightweight policy that operates locally and recovers classical decision-making limits under specific conditions. The analytical nature of the development emphasizes the importance of understanding when and with whom model exchanges should occur in decentralized learning environments.
Methodology
The methodology involves formulating the peer-selection problem as a finite-horizon Markov optimal-stopping problem. The author employs analytical techniques to derive optimal policies, confidence bounds, and stopping rules, considering the dynamics of peer availability and the perishable nature of evidence.
Results
The paper proves the existence of an optimal policy characterized by a reservation value, derives confidence bounds for peer selection, and establishes a mobility-aware stopping rule. It also provides a closed-form expression for the value of waiting and demonstrates that a myopic-optimality theorem holds in volatile mobility regimes.
Implications
The findings have significant implications for decentralized federated learning systems, particularly in mobile and intermittently connected environments. The developed theory and the PROSE policy can enhance decision-making processes for peer selection, improving collaboration efficiency and model exchange in practical applications.
Rethinking Class Imbalance for Single-Cell Foundation Models: A Systematic Benchmark Across Architectures and Long-Tail Loss Functions
Theory
Optimization
- High aggregate accuracy in single-cell models often masks poor performance on rare cell types.
- Systematic benchmarking reveals that loss function optimization alone is insufficient for certain rare classes.
- Absolute training-set size predicts the effectiveness of reweighting strategies, not relative frequency.
- Class-balanced loss and LDAM are the most effective long-tail loss functions across various settings.
Read more
Rethinking Class Imbalance for Single-Cell Foundation Models: A Systematic Benchmark Across Architectures and Long-Tail Loss Functions
Summary
This paper addresses the challenge of class imbalance in single-cell foundation models, which can achieve high aggregate accuracy but often fail to classify rare, disease-relevant cell populations. The authors conduct a systematic benchmark of six long-tail loss functions (cross-entropy, weighted CE, class-balanced loss, focal loss, LDAM, logit-adjusted softmax) across three model architectures and three datasets, totaling 162 controlled training runs. They find that the gap between overall accuracy and rare-class recall is consistent across different settings, driven more by dataset structure than by pre-training. The study reveals that rare-class failures can be categorized into two regimes based on embedding geometry, with some classes being recoverable through appropriate loss functions while others remain poorly classified regardless of the loss used. The effectiveness of reweighting strategies is shown to depend on the absolute training-set size of classes rather than their relative frequency. Among the evaluated losses, class-balanced loss and LDAM consistently perform well, while logit adjustment tends to favor recall over precision. The findings provide a benchmark and practical guidelines for effectively combining foundation models with imbalanced biological data.
Methodology
The authors conducted a systematic benchmark involving 162 controlled training runs across three architectures and three datasets, evaluating six long-tail loss functions. They analyzed the performance of these loss functions in terms of overall accuracy, Macro-F1 score, and rare-class recall, while also examining the embedding geometry of the classes.
Results
The study found consistent gaps between overall accuracy and rare-class recall across all settings, indicating that dataset structure plays a significant role. Certain classes were recoverable with the right loss function, while others remained poorly classified. Class-balanced loss and LDAM emerged as the most reliable choices, while logit adjustment was found to compromise precision for recall.
Implications
The findings highlight the need for tailored approaches in single-cell RNA sequencing tasks, particularly for rare cell populations. The benchmark and guidelines provided can help researchers select appropriate loss functions to improve classification performance in imbalanced biological datasets.
ExpBoN: Exponential-Noise Best-of-$n$ for Efficient Test-Time LLM Alignment
Large Language Models
Efficient ML
Theory
- Introduction of ExpBoN, a soft Best-of-n method using exponential noise for improved LLM alignment.
- Theoretical guarantees of exponentially fast convergence and regret behavior for ExpBoN.
- Integration of ExpBoN into the GSI framework, leading to significant computational savings.
- Empirical results demonstrating substantial reductions in computational costs while preserving accuracy.
Read more
ExpBoN: Exponential-Noise Best-of-$n$ for Efficient Test-Time LLM Alignment
Summary
The paper introduces ExpBoN, a novel soft Best-of-n (BoN) sampling method designed for efficient test-time alignment of large language models (LLMs). Traditional BoN sampling, while effective, suffers from limitations in controlling the trade-off between reward maximization and distribution shift due to its hard maximization approach. ExpBoN addresses these issues by employing an exponential-noise report-noisy-max mechanism, which allows for smoother control and guarantees exponentially fast convergence in total variation, expected reward, and KL divergence. The authors provide a thorough theoretical analysis of ExpBoN's convergence and regret behavior, demonstrating its advantages over existing methods. Additionally, ExpBoN is integrated into the guided speculative inference (GSI) framework, resulting in ExpGSI, which significantly reduces computational costs while maintaining accuracy. Experimental results on datasets such as MATH500, MMLU-STEM, and Minerva Math show that ExpGSI can reduce computational requirements by 14%-39% for the Qwen2.5-Math model and up to 45% for the Qwen3 model at n=16. Overall, the paper establishes a strong theoretical and algorithmic foundation for exponential-noise BoN and its application in efficient LLM alignment.
Methodology
The authors developed ExpBoN by utilizing an exponential-noise report-noisy-max mechanism, which allows for a more sample-efficient approach to BoN sampling. They conducted theoretical analyses to establish convergence properties and regret guarantees. The integration of ExpBoN into the GSI framework was also performed to enhance efficiency in LLM alignment.
Results
Experiments showed that ExpGSI reduced computational costs by 14%-39% for the Qwen2.5-Math model and up to 45% for the Qwen3 model at n=16, while maintaining comparable accuracy to existing methods.
Implications
The findings suggest that ExpBoN can be a powerful tool for improving the efficiency of LLM alignment during inference, making it easier to deploy LLMs in real-world applications without the need for extensive retraining.
Dissecting Hierarchical Reasoning Models: A Mechanistic Study
Theory
Interpretability
- HRM outperforms one-pass Transformer baselines in reasoning tasks.
- The contributions of high- and low-level states in HRM vary by task and inference stage.
- Linearly decodable features do not necessarily indicate causal relevance in HRM's reasoning process.
- SAE ablations produce larger behavioral changes than linear probe direction ablations.
Read more
Dissecting Hierarchical Reasoning Models: A Mechanistic Study
Summary
This paper investigates the Hierarchical Reasoning Model (HRM), a hierarchical Transformer-based reasoning model, to understand its internal mechanisms and the information it encodes during reasoning tasks. The authors focus on three tasks: Sudoku, Maze navigation, and ARC-AGI-2, employing various methodologies to analyze HRM's performance compared to Transformer baselines. They conduct causal interventions on recurrent states and utilize linear probes and sparse autoencoders (SAEs) for feature analysis. The findings reveal that recurrent models outperform one-pass baselines, and the contributions of high- and low-level states vary across tasks and inference stages. Notably, while task variables can be linearly decoded from HRM's recurrent states, the causal relevance of these features is not supported by directed ablation experiments. The study concludes that HRM implements constraint-aware iterative refinement without relying on a compact set of causally important features, emphasizing the need for improved mechanistic interpretability techniques for latent-space reasoning models.
Methodology
The authors compare HRM against Transformer baselines with and without recurrent modules, apply causal interventions on latent representations, and conduct representational analysis using linear probes and sparse autoencoders to discover features and validate their causal impact.
Results
The study finds that HRM iteratively refines a solution state with varying contributions from high- and low-level modules. While task-relevant information can be decoded from HRM's states, directed ablation does not show significant causal relevance. Additionally, the top features identified by SAEs do not consistently outperform random feature sets, indicating a distributed computation model.
Implications
The findings suggest that understanding the mechanisms of latent-space reasoning models like HRM is crucial for developing more effective interpretability techniques, which could enhance the application of such models in complex reasoning tasks across various domains.
Token Utility Is Selection-Conditioned: Coupled Selection of Prompt Context and Response Supervision for Efficient Instruction Tuning
NLP
Large Language Models
Efficient ML
- Introduces the concept of selection-conditioned token utility, emphasizing the interdependence of prompt and response selections.
- Develops BRIDGE, a method that coordinates prompt and response selections through a shared validation-directed interaction surrogate.
- Demonstrates that BRIDGE outperforms existing independent selection methods across multiple model families and tasks.
- Finds that the performance advantage of BRIDGE increases with stronger compression in mathematical reasoning tasks.
Read more
Token Utility Is Selection-Conditioned: Coupled Selection of Prompt Context and Response Supervision for Efficient Instruction Tuning
Summary
This paper addresses the challenge of efficient instruction tuning for large language models (LLMs) by proposing a novel method called BRIDGE (Budgeted Response-Prompt Interaction via Directional Gradient-guided Efficient Token Selection). The authors highlight that the utility of tokens in prompts and responses is not intrinsic but is conditioned on the selection state of the opposite side. Existing methods often evaluate prompts and responses independently, leading to a mismatch in selection states. BRIDGE captures the interdependence of token utility by employing a shared validation-directed interaction surrogate that values prompt context based on the selected supervision it supports and vice versa. The method alternates between selecting prompt and response tokens, updating conditional scores through precomputed interactions. This coordinated selection process is shown to improve the overall performance of instruction tuning across three model families in tasks such as mathematical reasoning, code generation, and instruction following. The findings reveal that the advantage of BRIDGE increases with stronger compression, demonstrating its effectiveness in optimizing instruction tuning processes.
Methodology
The authors propose BRIDGE, which utilizes a shared validation-directed interaction surrogate to evaluate prompt and response tokens conditionally based on the retained state of the opposite side. The method alternates between selecting prompt and response tokens while updating conditional scores by aggregating precomputed interaction coefficients. A structure-aware projection is employed to convert conditional response scores into coherent supervision spans, ensuring effective supervision during training.
Results
BRIDGE achieves the best target-task accuracy and general capability average among compared selection methods across three model families in tasks such as mathematical reasoning, code generation, and instruction following. The method shows increasing performance gains over independent selection as compression strengthens, particularly in mathematical reasoning tasks.
Implications
The findings suggest that optimizing the selection process in instruction tuning can lead to more efficient training of large language models, potentially reducing computational costs and improving model performance. This approach could be applied to various tasks requiring prompt-response interactions, enhancing the effectiveness of LLMs in real-world applications.
Multi-Domain Clustering via Measure Quantization
Optimization
Theory
Multimodal
- Introduction of a general framework for multi-domain clustering via measure quantization.
- Utilization of probability metrics like Sinkhorn divergence and Maximum Mean Discrepancy for clustering.
- Scalable mini-batch optimization strategy for efficient clustering.
- Demonstrated superior performance on multiple benchmarks compared to classical methods.
Read more
Multi-Domain Clustering via Measure Quantization
Summary
This paper presents a novel framework for multi-domain clustering that extends traditional clustering methods, such as K-means, to handle heterogeneous data from multiple domains. The authors propose a measure quantization approach that minimizes a probability metric, specifically the Sinkhorn divergence or Maximum Mean Discrepancy, between the probability measures of different domains and a shared set of cluster prototypes. The method allows for data points to be assigned to clusters using either nearest centroid or optimal transport strategies, facilitating a collaborative approach that considers all samples within a domain. A mini-batch optimization technique is introduced to enhance scalability, reducing memory and computational costs while maintaining clustering performance. The experimental results demonstrate that the proposed Sinkhorn-based method consistently outperforms classical and existing multi-domain clustering baselines across various datasets, including image, audio, and sensor data, even when scaling to large sample sizes.
Methodology
The authors develop a framework that generalizes the measure quantization problem to multiple heterogeneous domains. They minimize a dissimilarity metric between the empirical probability measures of the domains and the cluster prototypes. The optimization is performed using gradient descent, and data points are assigned to clusters through either nearest centroid or optimal transport methods. A mini-batch approach is employed to enhance computational efficiency.
Results
The experimental evaluation on five multi-domain benchmarks shows that the proposed method consistently outperforms classical clustering algorithms and existing multi-domain clustering techniques. The performance advantage is maintained even when the number of samples scales to hundreds of thousands, indicating the robustness and effectiveness of the approach.
Implications
This work has significant implications for clustering tasks in scenarios where data is drawn from multiple domains with different statistical properties. The proposed framework can be applied in various fields such as image processing, audio analysis, and sensor data interpretation, enhancing the ability to analyze and interpret complex datasets.
From Regional to Global: Transfer Learning for Atmospheric Transport Emulators
Graph Learning
Efficient ML
Theory
- Introduction of GATES, a machine learning emulator for atmospheric transport, which operates 1,000 times faster than traditional LPDMs.
- Evaluation of model performance across four distinct global regions to assess spatial transferability and generalization capabilities.
- Use of leave-one-region-out experiments to investigate the effectiveness of transfer learning in atmospheric transport modeling.
- Characterization of regional differences in atmospheric transport to inform model training and improve global applicability.
Read more
From Regional to Global: Transfer Learning for Atmospheric Transport Emulators
Summary
This paper addresses the challenge of estimating greenhouse gas emissions using atmospheric concentration observations and chemical transport models, specifically focusing on the limitations of traditional Lagrangian Particle Dispersion Models (LPDMs) due to their computational expense and regional specificity. The authors propose a machine learning-based emulator, GATES (Graph Neural Network Atmospheric Transport Emulation System), which approximates LPDM outputs at a significantly reduced computational cost. The study evaluates the spatial transferability of atmospheric transport emulators across four diverse regions: South America, East Asia, South Asia, and North Africa. By conducting leave-one-region-out experiments, the authors analyze the generalizability of models trained on specific regions to new regions, thereby enhancing the understanding of regional differences in atmospheric transport. The findings contribute to the development of efficient, scalable, and globally applicable atmospheric transport emulators, facilitating improved top-down estimates of greenhouse gas emissions.
Methodology
The authors trained the GATES model using meteorological inputs derived from the Met Office Unified Model and LPDM-generated footprints across four regions. They implemented both region-specific and multi-region models, conducting leave-one-region-out experiments to assess generalization capabilities. The model training involved temporal data splitting and subsampling to manage dataset sizes, with performance evaluated quantitatively across regions.
Results
The results indicate that while region-specific models perform well within their trained regions, multi-region models show promise in generalizing across different regions. The leave-one-region-out experiments revealed varying degrees of transferability, highlighting the importance of understanding regional heterogeneity in atmospheric transport.
Implications
The findings suggest that machine learning emulators can significantly enhance the efficiency of atmospheric transport modeling, enabling more accurate and timely estimates of greenhouse gas emissions. This has implications for climate monitoring and policy-making, particularly in the context of international climate agreements.
RS-Claw-Evolution: Environment-Feedback-Driven Evolution for Lightweight Remote Sensing Agents in Long-Horizon Tasks
Reinforcement Learning
Large Language Models
NLP
- Introduction of RS-Claw-Evolution framework for lightweight RS agents.
- Three-stage evolution process: interaction, experience, and decision evolution.
- Implementation of programming-based interaction to manage states and reduce redundancy.
- Feedback-driven fine-tuning strategy to enhance learning from errors.
Read more
RS-Claw-Evolution: Environment-Feedback-Driven Evolution for Lightweight Remote Sensing Agents in Long-Horizon Tasks
Summary
The paper introduces RS-Claw-Evolution, a novel framework designed to enhance the capabilities of lightweight remote sensing (RS) agents in executing long-horizon tasks. Traditional lightweight RS agents face challenges such as loss of long-horizon states, ineffective use of environmental feedback, and sparse optimization signals. RS-Claw-Evolution addresses these issues through a three-stage evolution process: interaction evolution, experience evolution, and decision evolution. In the interaction evolution stage, a programming-based interaction paradigm is implemented to manage observations and maintain intermediate states, reducing context redundancy. The experience evolution stage employs a feedback-driven supervised fine-tuning strategy that generates informative error patterns while preserving recovery behaviors. Finally, the decision evolution stage utilizes reinforcement learning to optimize tool-use behaviors by incorporating multi-dimensional environment rewards. Experiments conducted on the Earth-Bench benchmark demonstrate the effectiveness of the proposed framework, with the optimized Qwen3-4B-based agent achieving a 65.9% accuracy in Autonomous Planning mode, significantly outperforming both the untrained Qwen3-32B baseline and other large-scale models. The results indicate that effective integration of environmental feedback can significantly enhance the performance of lightweight RS agents, narrowing the gap with larger models in complex remote sensing tasks.
Methodology
The methodology involves a three-stage evolution framework: 1) Interaction evolution with a programming-based paradigm to manage observations; 2) Experience evolution using a feedback-driven supervised fine-tuning strategy to generate informative error patterns; 3) Decision evolution employing reinforcement learning with multi-dimensional rewards to optimize tool-use behaviors.
Results
The RS-Claw-Evolution framework was validated on the Earth-Bench benchmark, resulting in the Qwen3-4B-based agent achieving 65.9% accuracy in Autonomous Planning mode, outperforming the untrained Qwen3-32B baseline (43.8%) and other models like DeepSeek-V3.1 (60.8%), while approaching GPT-5 (71.6%).
Implications
The findings suggest that integrating environmental feedback effectively can enhance the capabilities of lightweight RS agents, making them more competitive with larger models in long-horizon tasks. This has potential applications in automating complex geospatial analyses and improving the reliability of remote sensing operations.
Task-Aware Hybrid QUBO Optimization for Structured Neural Network Pruning
Optimization
Efficient ML
Computer Vision
- Introduces a hybrid global-local framework for task-aware mixed-precision quantization.
- Utilizes a Task-Aware QUBO formulation that incorporates quantization error and layer sensitivity.
- Implements graph-aware structural constraints to enhance compatibility in activation precisions.
- Demonstrates improved performance on a compact image-denoising model compared to existing methods.
Read more
Task-Aware Hybrid QUBO Optimization for Structured Neural Network Pruning
Summary
This paper addresses the challenge of mixed-precision quantization in deep neural networks, particularly for compact image restoration models deployed on resource-constrained hardware. The authors propose a hybrid global-local combinatorial framework that formulates a Task-Aware Quadratic Unconstrained Binary Optimization (QUBO) problem to optimize layer-wise bit-width assignments. The framework incorporates measured quantization errors and layer sensitivities, along with graph-aware structural constraints to ensure compatible activation precisions. The proposed method, PROTES, initializes from a high-quality QUBO solution and refines the precision assignment through direct network evaluation, applying an asymmetric Return-on-Investment (ROI) penalty to discourage unnecessary computational expansion. Experimental results on a NAFBlock-based image-denoising model demonstrate significant improvements in accuracy-compute and accuracy-storage trade-offs compared to uniform and unconstrained routing baselines.
Methodology
The methodology involves formulating a Task-Aware QUBO that integrates quantization error and layer sensitivity metrics, alongside graph-aware constraints. The optimization process is initialized with a high-quality QUBO solution and refined using PROTES, which evaluates the network's performance directly while applying penalties to control computational costs.
Results
The experiments conducted on the HalfUNet model, evaluated on the Smartphone Image Denoising Dataset (SIDD), show that the proposed method achieves better accuracy-compute and accuracy-storage trade-offs compared to uniform and unconstrained routing baselines, indicating its effectiveness in optimizing mixed-precision configurations.
Implications
The findings suggest that the proposed hybrid QUBO framework can significantly enhance the deployment of deep learning models on edge devices by optimizing their computational efficiency without compromising accuracy. This has potential applications in various domains requiring efficient neural network implementations, such as mobile computing and IoT devices.
Trading Depth for Time in Recurrent Transformers
NLP
Large Language Models
Efficient ML
- Latent Recurrent Transformers (LRTs) introduce a thought token for hidden state refinement between vocabulary tokens.
- The thought token model achieves comparable performance to deeper models while using approximately 48% fewer parameters.
- Temporal computation through thought tokens can recover a significant portion of the benefits associated with increased physical depth.
- The study provides insights into the depth-time trade-off in recurrent Transformer architectures.
Read more
Trading Depth for Time in Recurrent Transformers
Summary
This paper investigates the trade-off between increasing computational depth and extending temporal steps in Recurrent Transformers, specifically through the introduction of Latent Recurrent Transformers (LRTs). LRTs utilize a latent thought token that refines the hidden state between vocabulary tokens, allowing for a comparison of parameter efficiency between temporal and physical depth. The authors propose a model that retains one backbone forward pass per vocabulary token while inserting a thought token that shares parameters with the vocabulary tokens. The study compares the performance of an L-layer LRT with thought tokens against a 2L-layer LRT without them, both executing 2L Transformer blocks per vocabulary token. The results indicate that the thought token model achieves significant performance improvements with fewer parameters, suggesting that temporal computation can effectively substitute for increased physical depth in recurrent architectures.
Methodology
The authors developed Latent Recurrent Transformers (LRTs) that incorporate a thought token between vocabulary tokens to refine hidden states. They compared the performance of L-layer LRTs with thought tokens against 2L-layer LRTs without thought tokens, maintaining the same number of block executions during decoding. The methodology involved analyzing the impact of temporal versus physical depth on model performance and parameter efficiency.
Results
The introduction of a single thought token allowed the shallower LRT model to recover 67% and 81% of the performance improvement achieved by doubling the depth in 16-layer and 20-layer NanoChat backbones, respectively. This was accomplished with approximately 48% fewer parameters. Further improvements were noted with additional thought tokens, although deeper models still outperformed them at the same decoding block count.
Implications
The findings suggest that models can achieve high performance with fewer parameters by leveraging temporal computation, which could lead to more efficient architectures in natural language processing tasks. This approach may also inform future designs of recurrent models in various applications, emphasizing the importance of feedback mechanisms.
GLR-MM: Graph-Based Global-Local Reconstruction for Robust Multimodal Chest X-ray and EHR Representation Learning under Missing Modalities
Multimodal
Graph Learning
Computer Vision
- GLR-MM effectively reconstructs missing modalities using both local patient data and global cohort information.
- The framework employs a graph-based approach to enhance the prediction of ICU mortality under varying levels of missing data.
- GLR-MM shows superior performance compared to existing models, particularly in scenarios with high missingness (50%).
- The model integrates adaptive fusion techniques to optimize the use of available data for improved predictions.
Read more
GLR-MM: Graph-Based Global-Local Reconstruction for Robust Multimodal Chest X-ray and EHR Representation Learning under Missing Modalities
Summary
The paper introduces GLR-MM, a novel framework designed to enhance early ICU mortality prediction by effectively handling missing modalities in multimodal clinical data, specifically chest X-rays (CXR) and electronic health records (EHR). Traditional models often struggle with incomplete data, as they either align observed representations or reconstruct missing modalities without leveraging both intra-patient and inter-patient evidence. GLR-MM addresses this gap by employing a graph-based approach that integrates local and global reconstruction strategies. The framework consists of two branches: a local branch that predicts missing embeddings using the patient's remaining modalities, and a global branch that retrieves information from similar patients in a constructed graph. The model adaptively fuses these estimates and optimizes for class-balanced prediction, reconstruction, and contrastive objectives. Evaluations on a dataset of 9,620 ICU stays demonstrate that while existing models like MUSE perform better under mild and moderate missingness, GLR-MM significantly outperforms at 50% missingness, indicating its robustness in scenarios with severe data incompleteness.
Methodology
GLR-MM utilizes a Graph-Based Global-Local Reconstruction framework that consists of two main branches: a local branch for predicting missing embeddings from the patient's remaining modalities and a global branch for retrieving information from similar patients in a graph structure. The model constructs an observed-only patient graph and employs attention mechanisms for adaptive fusion of reconstructed embeddings, optimizing for various objectives including prediction accuracy and reconstruction quality.
Results
In experiments with 9,620 ICU stays, GLR-MM demonstrated a higher Area Under the Receiver Operating Characteristic (AUROC) and Area Under the Precision-Recall Curve (AUPRC) at 50% missingness, outperforming the MUSE model by 0.0088 and 0.0249, respectively. These results indicate that GLR-MM is particularly effective in handling severe missing data scenarios.
Implications
The findings suggest that GLR-MM can significantly improve clinical decision-making processes in critical care settings by providing robust predictions even when faced with incomplete data. This approach could be extended to other multimodal clinical applications, enhancing the reliability of predictions in various healthcare contexts.
iSDFT: Information-Proximal Self-Distillation for Continual Learning in LLMs
Large Language Models
NLP
Theory
- iSDFT allows controlled information transfer from teacher to student models, improving continual learning.
- The method introduces a new target-selection principle based on information constraints, enhancing local adaptation.
- iSDFT consistently outperforms vanilla SDFT across multiple tasks and models, demonstrating its robustness.
- The approach maintains a high retention rate of original capabilities while improving task-specific performance.
Read more
iSDFT: Information-Proximal Self-Distillation for Continual Learning in LLMs
Summary
The paper introduces Information-Proximal Self-Distillation Fine-Tuning (iSDFT), a novel approach for continual learning in large language models (LLMs). Traditional self-distillation fine-tuning (SDFT) learns from demonstrations but lacks control over the amount of information transferred from the teacher model to the student model, which can lead to forgetting previously acquired skills. iSDFT addresses this by treating the teacher as a budgeted source of information, allowing for a more controlled transfer of knowledge. At each token generation, iSDFT selects a target distribution that is closest to the current student while adhering to a specified information constraint. This results in a locally determined target that mitigates cumulative drift from the base policy. The method was evaluated across four different LLM architectures and two specialization tasks, demonstrating improvements in task acquisition and retention of original capabilities. iSDFT outperformed vanilla SDFT in seven out of eight settings and maintained tighter retention on the original benchmark suite, showcasing its effectiveness in balancing specialization and broader capability retention.
Methodology
iSDFT employs an information-proximal approach to self-distillation, where the target distribution for the student model is selected based on its proximity to the current student while satisfying a prescribed information constraint. This is achieved through KL I-projection onto a teacher-information constraint set. Additionally, a KL penalty is used to anchor the student model to a frozen base policy, ensuring long-term retention of previously learned behaviors.
Results
iSDFT improved task acquisition in 7 out of 8 model-task settings compared to vanilla SDFT and matched performance in the remaining setting. It maintained 73% of evaluations within 0.5 points of the base model on the original SDFT retention suite, compared to 52% for the strongest baseline. Furthermore, iSDFT achieved the largest mean improvement across ten additional benchmarks in mathematics, coding, and competition-related tasks.
Implications
The findings suggest that controlled information transfer can enhance the specialization of LLMs while preserving their broader capabilities. This has significant implications for applications requiring continual learning and adaptation, such as personalized AI systems, educational tools, and dynamic task environments.
PAGE: Partition-Aware Gated KV-Cache Eviction
NLP
Large Language Models
Efficient ML
- PAGE reframes KV-cache eviction as an input-dependent admission decision.
- It identifies two classes of inputs: capacity-bound (harmful eviction) and dilution-prone (beneficial eviction).
- The method utilizes a label-free scalar from prefill attention to predict eviction safety.
- PAGE significantly reduces the harm rate of eviction from 0.75 to 0.026, a 29× reduction.
Read more
PAGE: Partition-Aware Gated KV-Cache Eviction
Summary
The paper introduces PAGE (Partition-Aware Gated KV-Cache Eviction), a novel approach to managing key-value (KV) cache eviction in large language models (LLMs) during inference. Traditional KV-cache eviction methods focus on which tokens to retain under a fixed memory budget but do not consider whether eviction is appropriate for each input. This can lead to significant accuracy degradation for certain inputs. The authors propose reframing eviction as an input-dependent admission decision, identifying two classes of inputs: capacity-bound, where eviction is detrimental, and dilution-prone, where it is beneficial. They utilize a label-free scalar derived from prefill attention, specifically the early-to-late drop in pairwise top-k head agreement, to predict the class of each input. PAGE applies an existing eviction method when this drop exceeds a threshold and retains the full cache otherwise, effectively safeguarding against harmful eviction. The method is evaluated across multiple models and benchmarks, demonstrating a substantial reduction in the harm rate associated with eviction, thus maintaining high accuracy without retraining the eviction mechanism.
Methodology
The authors developed PAGE as a training-free gating mechanism that leverages a scalar derived from prefill attention to assess the safety of eviction for each input. They empirically partition inputs into two classes based on the early-to-late drop in pairwise top-k head agreement. PAGE applies an existing eviction method when this drop exceeds a predefined threshold and retains the full cache otherwise, ensuring that harmful evictions are avoided.
Results
PAGE was shown to cut the rate of catastrophic eviction from 0.75 to 0.026 across various models and benchmarks, translating to a 29× reduction in harmful evictions. The method effectively maintains accuracy levels, achieving a consistent performance of approximately 89% accuracy, even under aggressive compression scenarios.
Implications
The findings suggest that PAGE can enhance the reliability of KV-cache eviction in LLMs, particularly for applications requiring long-context processing. By preventing harmful evictions, PAGE can improve the performance of LLMs in real-world tasks, ensuring that critical information is retained during inference.
Talk to Me, Jarvis: An Open-Source Edge-Deployable Voice Assistant Framework for Autonomous Racecars
NLP
Audio & Speech
Robotics
- Jarvis is an offline voice assistant specifically designed for autonomous racing applications.
- The framework integrates speech recognition and command classification, achieving low latency and high accuracy.
- Experimental results demonstrate superior performance compared to larger online-hosted models.
- The authors provide an open-source implementation to support further research.
Read more
Talk to Me, Jarvis: An Open-Source Edge-Deployable Voice Assistant Framework for Autonomous Racecars
Summary
This paper presents 'Jarvis', an offline voice assistant framework designed for autonomous racecars, addressing the challenges of latency and network dependency associated with online-hosted models. The framework integrates speech recognition, synthesis, and natural language command classification into a lightweight local architecture. The core component is a text-to-command classifier, fine-tuned from the Mistral 7B model, which achieves high intent recognition accuracy and low processing latency. Experimental results show that Jarvis outperforms larger online models, achieving 97.63% intent recognition accuracy with an average latency of 1.39 seconds. The authors provide an open-source implementation to facilitate further research and fine-tuning, emphasizing the importance of rapid command processing in time-critical autonomous driving scenarios.
Methodology
The authors developed a local voice assistant framework that combines wake-word detection, speech transcription, and command classification. They fine-tuned the Mistral 7B model for intent recognition and utilized the Whisper architecture for speech-to-text processing, ensuring low latency and operational independence from network connectivity.
Results
Jarvis achieved an intent recognition accuracy of 97.63% with an average processing latency of 1.39 seconds, outperforming larger online models. This performance is critical for the rapid decision-making required in autonomous racing.
Implications
The development of Jarvis has significant implications for enhancing human-machine interaction in autonomous vehicles, particularly in racing contexts where quick responses are essential. The open-source nature of the framework encourages further advancements in voice command systems for various autonomous applications.
Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention
NLP
Large Language Models
Efficient ML
- CKDA enhances KDA's expressivity by enabling 2D rotations through a combination of delta-rule transformations and channel-wise reflections.
- The model maintains non-expansiveness and computational efficiency while achieving the expressivity of DeltaProduct2.
- CKDA can track finite groups isomorphic to subgroups of SO(3) with fewer layers than existing models.
- Empirical results show CKDA outperforms Transformers and other linear RNNs in language modeling tasks.
Read more
Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention
Summary
This paper introduces Complex Kimi Delta Attention (CKDA), an enhancement of Kimi Delta Attention (KDA) that improves expressivity in linear recurrent neural networks (RNNs). The authors demonstrate that CKDA can perform 2D rotations by combining a delta-rule transformation with a reflection from a channel-wise gate. By extending the parameter ranges of KDA, allowing gate values in [-1, 1] and the delta-rule coefficient β in [0, 2], CKDA maintains stability and efficiency while achieving the expressivity of DeltaProduct2. The paper characterizes CKDA's transition matrices, proving that they can represent every orthogonal diagonal-plus-rank-one matrix. CKDA can track finite groups isomorphic to subgroups of SO(3) with fewer layers than other models, showcasing superior performance in tasks such as length extrapolation and language modeling compared to Transformers and other linear RNNs. The implementation retains high throughput, making CKDA a promising candidate for efficient sequence modeling tasks.
Methodology
The authors extend the Kimi Delta Attention framework by allowing gate values in [-1, 1] and adjusting the delta-rule coefficient β to [0, 2]. They analyze the spectral properties of CKDA's transition matrices and demonstrate its ability to perform complex transformations such as 2D rotations. The expressivity of CKDA is characterized through theoretical proofs and empirical evaluations on various sequence modeling tasks.
Results
CKDA achieves state-tracking expressivity comparable to DeltaProduct2 while requiring fewer layers. It demonstrates the strongest length extrapolation performance among tested KDA settings and matches or exceeds the performance of Transformers in language modeling tasks with a significant parameter count.
Implications
CKDA's advancements in expressivity and efficiency make it a valuable contribution to the field of sequence modeling, particularly in applications involving language processing and time series analysis. Its ability to perform complex transformations efficiently could lead to improved performance in various machine learning tasks.
AirGC-CD: Gaussian-Circulant Precoding for Exactly Debiasable PAPR Reduction in Over-the-Air Federated Learning
Federated Learning
Optimization
Efficient ML
- AirGC-CD effectively reduces PAPR in over-the-air federated learning without introducing bias.
- The scheme utilizes Gaussian-circulant precoding to ensure unbiased aggregation of model updates.
- It compresses data transmission efficiently, reducing the number of channel uses required.
- The convergence analysis shows a favorable rate without a bias floor, enhancing learning performance.
Read more
AirGC-CD: Gaussian-Circulant Precoding for Exactly Debiasable PAPR Reduction in Over-the-Air Federated Learning
Summary
The paper introduces AirGC-CD, a novel scheme designed to enhance over-the-air federated learning (FL) by addressing the challenges posed by high peak-to-average power ratio (PAPR) during the transmission of local model updates from edge devices. Traditional methods either clip the peaks, introducing bias, or back off on transmit power, which degrades the signal-to-noise ratio (SNR). AirGC-CD employs a Gaussian-circulant precoding technique that ensures the output is Gaussian-distributed, allowing for effective clipping that can be inverted on average, thus maintaining unbiased aggregation. The method compresses the transmission from model dimension d to sketch dimension m efficiently, using O(d log d) complexity through fast Fourier transforms. The authors provide a convergence analysis under peak power constraints, demonstrating a convergence rate of O(1/√T) without a bias floor. Experimental results across five image datasets indicate that AirGC-CD significantly outperforms existing over-the-air FL schemes, particularly in low SNR scenarios, while utilizing fewer channel resources per round.
Methodology
The authors propose AirGC-CD, which involves precoding local updates with a Gaussian-circulant matrix before clipping. This approach ensures that the clipped output is Gaussian-distributed, allowing for a known clipping function that can be inverted on average. The method also incorporates fast Fourier transforms for efficient compression of model updates.
Results
AirGC-CD was shown to outperform baseline over-the-air FL schemes in various experimental settings, particularly at low SNR, while requiring fewer channel uses per communication round. The convergence analysis confirmed an O(1/√T) convergence rate without a bias floor.
Implications
The findings suggest that AirGC-CD can facilitate more efficient federated learning in environments with limited communication resources, making it suitable for applications in edge computing and 6G networks where data privacy and transmission efficiency are critical.
Scaling Discovery through Test-Time Communication
Optimization
Large Language Models
Theory
- Test-time communication among agents can lead to significant performance improvements over independent attempts.
- A team of k communicating agents can match the success rate of 4k independent agents, with benefits compounding as the team size increases.
- Communicating agents achieved new state-of-the-art results in polyomino packing and MNIST classifier compression tasks.
- The effectiveness of communication is contingent on having sufficient compute resources and clear feedback mechanisms.
Read more
Scaling Discovery through Test-Time Communication
Summary
This paper investigates the benefits of test-time communication among multiple agents in solving complex tasks. The authors demonstrate that a team of communicating agents can significantly outperform independent agents, particularly in challenging environments where sharing discoveries can lead to collective progress. They focus on the ARC-AGI-3 benchmark, where a team of k agents (team@k) achieves success rates comparable to 4k independent agents, with performance improvements scaling with the number of agents. The study also explores the application of this communication strategy in research-oriented tasks, such as polyomino packing and MNIST classifier compression, where communicating agents set new state-of-the-art results. The findings suggest that communication is most effective when agents have sufficient computational resources and clear measures of progress, enabling them to build on each other's discoveries. However, the benefits of communication diminish when feedback is limited or when agents cannot verify progress, indicating that the context of the task plays a crucial role in the effectiveness of multi-agent communication.
Methodology
The authors conducted experiments comparing the performance of teams of communicating agents (team@k) against the best results from independent agents (best@k) across various tasks. They utilized a shared workspace for agents to exchange intermediate results and artifacts, allowing for collaborative problem-solving without predefined roles. The study focused on benchmarks like ARC-AGI-3, polyomino packing, and MNIST classifier compression.
Results
The results showed that team@k agents matched or exceeded the performance of significantly larger groups of independent agents. Specifically, team@3 matched the success of best@13, and team@5 matched best@33 on ARC-AGI-3. In polyomino packing, team@4 achieved a score of 0.922, surpassing previous best-known results. For MNIST classifier compression, a team of four agents produced a 1,957-byte classifier with 99.4% accuracy, outperforming both the best-known human solution and single-agent results.
Implications
The findings suggest that fostering communication among agents can enhance collaborative problem-solving in AI systems, particularly in complex and open-ended tasks. This approach could be applied in various domains, including optimization problems, research tasks, and scenarios where continuous learning and adaptation are crucial.
Neural Cellular Automata Learn General Features in their Hidden Channels
Efficient ML
Theory
Computer Vision
- NCAs provide a parameter-efficient alternative to traditional deep learning models, reducing the risk of overfitting.
- The paper introduces a novel transfer-learning mechanism using hidden states instead of weights, enhancing few-shot learning capabilities.
- NCAs outperform recurrent and feed-forward architectures on MNIST benchmarks with fewer than 10,000 parameters.
- Hidden channels in NCAs capture general topological features, allowing for effective transfer of knowledge across classes.
Read more
Neural Cellular Automata Learn General Features in their Hidden Channels
Summary
This paper explores the internal dynamics of Neural Cellular Automata (NCAs) and their potential for few-shot learning through a novel transfer-learning mechanism. While traditional deep learning models often rely on over-parameterization, leading to overfitting, NCAs provide a parameter-efficient alternative. The authors investigate how the hidden channels of NCAs can be utilized for transfer learning by injecting a pretrained teacher's hidden states into a student model. This approach allows the student to leverage well-structured topological features without requiring identical architectures or parameters. The study evaluates the performance of NCAs against recurrent and feed-forward models on few-shot and scale-variant MNIST benchmarks, demonstrating that NCAs achieve superior generalization with significantly fewer parameters. Mechanistic analysis reveals that the hidden channels effectively decouple feature extraction from classification, capturing general, scale-invariant topological primitives rather than class-specific templates. This enables strong few-shot performance on unseen classes, highlighting the potential of hidden-state dynamics as a robust computational substrate for efficient transfer learning.
Methodology
The authors conducted experiments comparing NCAs with recurrent and feed-forward models on the MNIST dataset, focusing on few-shot learning scenarios. They implemented a transfer-learning mechanism where a teacher NCA's hidden states were injected into a student NCA during training. The models were evaluated based on their ability to generalize from limited examples, using metrics such as mean squared error and cross-entropy loss.
Results
The results indicated that NCAs significantly outperformed the other architectures in few-shot learning tasks, achieving robust generalization with a minimal parameter budget. The hidden channels were found to effectively absorb morphological complexity and converge to orthogonal states, facilitating the capture of scale-invariant topological primitives.
Implications
The findings suggest that leveraging hidden-state dynamics in NCAs can lead to more efficient and interpretable models for few-shot learning. This approach could be applied in various domains where data is scarce, enabling the development of robust machine learning systems that generalize well from limited training data.
Physics-residual machine learning predicts oxygen-evolution catalyst activity beyond the training range from sparse polarization measurements
Optimization
Theory
Efficient ML
- Introduces a physics-residual machine learning approach for predicting catalyst activity.
- Achieves significant error reduction compared to traditional data-driven models.
- Demonstrates that a small number of labeled catalysts can effectively train the model.
- Enables predictions beyond the training range, facilitating better candidate selection.
Read more
Physics-residual machine learning predicts oxygen-evolution catalyst activity beyond the training range from sparse polarization measurements
Summary
This paper presents a novel approach to predicting the activity of oxygen-evolution catalysts using a physics-residual machine learning model. The authors address the challenge of selecting candidates for further measurement in combinatorial libraries, where the activity often lies beyond the range of previously measured data. The proposed method combines the Tafel equation for extrapolating current density with a learned residual correction based on feature-space distance, along with an applicability-domain score to identify predictions that exceed the training range. The model was validated on a library of 322 candidates, achieving a mean absolute error of 0.203 mA cm−2, significantly lower than the 1.330 mA cm−2 error from a traditional data-driven model. The results indicate that only 35 labeled catalysts are sufficient for effective model training, and the approach can reduce overpotential errors by 29 to 52% in independent datasets. This method allows for more efficient screening of catalysts, potentially shortening measurement times while still accurately ranking the most active compositions.
Methodology
The methodology involves using the Tafel equation to extrapolate current density from sparse polarization measurements, applying a learned residual correction based on feature-space distance, and employing an applicability-domain score to identify predictions that exceed the training range. This hybrid approach combines physics-based modeling with machine learning techniques.
Results
The physics-residual machine learning model achieved a mean absolute error of 0.203 mA cm−2 in predicting catalyst activity, compared to 1.330 mA cm−2 from a conventional data-driven model. The model was validated on a library of 322 candidates, with 282 candidates exceeding the training maximum. Additionally, the method reduced overpotential errors by 29 to 52% in independent datasets.
Implications
The findings suggest that this approach can significantly enhance the efficiency of catalyst screening processes, allowing researchers to identify and rank the most promising candidates with fewer measurements. This could accelerate the development of more effective oxygen-evolution catalysts, contributing to advancements in water electrolysis and hydrogen production technologies.
The Limits of Speculation: Bounding Speculative Decoding in Mixture-of-Experts
NLP
Large Language Models
Optimization
- Introduces the Sequence-Conditioned Offline Oracle methodology for estimating MoE verification costs.
- Demonstrates diminishing returns when increasing the speculation budget in MoE models.
- Establishes a linear relationship in Delta Space Analysis for decision-making in speculative decoding.
- Provides a theoretical framework for understanding the limits of speculative decoding in MoE architectures.
Read more
The Limits of Speculation: Bounding Speculative Decoding in Mixture-of-Experts
Summary
This paper addresses the challenges of speculative decoding in Mixture-of-Experts (MoE) models, particularly the instability of verification costs due to input-dependent expert loading. The authors reformulate the problem of selecting a speculation budget as an offline Stochastic Shortest Path (SSP) problem over reference sequences. They develop a diagnostic Oracle that utilizes counterfactual simulation to accurately estimate the verification costs associated with erroneous tokens in MoE models. Through detailed analysis, the authors demonstrate that the decisions made by the Oracle reveal a linear boundary for rejected candidates, indicating that the optimization process is governed by a balance between marginal costs and expected progress. The findings suggest that while increasing the speculation budget can lead to diminishing returns in MoE models, a simple linear balance can be maintained to inform lightweight online heuristics for adaptive decoding strategies.
Methodology
The authors formulated the speculation budget selection as an offline Stochastic Shortest Path (SSP) problem and developed a hybrid offline Oracle to simulate counterfactual scenarios. This approach allowed them to isolate algorithmic losses from system-level ones and analyze the decision-making process in MoE models.
Results
The analysis revealed that increasing the speculation budget in MoE models leads to diminishing returns due to the increased routing penalties for low-probability tokens. The Oracle's decisions indicated a linear boundary for rejected candidates, simplifying the complexity of the global search to a manageable linear balance.
Implications
The findings provide a theoretical foundation for designing more efficient speculative decoding strategies in MoE models, potentially leading to improved performance in applications that rely on autoregressive generation, such as natural language processing and large language models.
SolarFlowRefiner: Refinement-Aware Flow Matching for Surface Solar Radiation Downscaling
Generative Models
Optimization
Time Series
- Introduces SolarFlowRefiner, a refinement-aware framework for SSR downscaling.
- Addresses the challenge of reconstructing high-resolution SSR fields from coarse ERA5 data.
- Utilizes a conditional FlowMatch generator and a refiner trained on structured errors.
- Demonstrates significant improvements over traditional standalone and post-hoc refinement methods.
Read more
SolarFlowRefiner: Refinement-Aware Flow Matching for Surface Solar Radiation Downscaling
Summary
The paper presents SolarFlowRefiner, a novel framework designed for downscaling surface solar radiation (SSR) from coarse ERA5 reanalysis data to high-resolution SolarCube fields. The challenge lies in the inherent ambiguity of coarse data, where a single grid cell may represent both sunlit and cloud-shadowed areas, leading to oversmoothing in traditional models. SolarFlowRefiner addresses this by employing a refinement-aware flow-matching approach, which integrates a conditional FlowMatch generator that predicts a normalized correction to an upsampled ERA5 baseline. The refiner is trained on the prediction-conditioned states, allowing it to learn from the structured errors produced by the generator. This joint optimization enhances the accuracy of the final SSR reconstruction. The authors demonstrate the effectiveness of their approach through experiments on a day-blocked ERA5–SolarCube benchmark, showing consistent improvements over standalone generation and post-hoc refinement methods. The proposed framework not only advances SSR downscaling but also offers a general strategy for coupling generative predictors with iterative correctors.
Methodology
The methodology involves a two-stage process where a conditional FlowMatch generator first predicts a normalized correction to the coarse ERA5 baseline. This is followed by a refinement stage where the refiner iteratively corrects the generated residual. The training of both components is conducted end-to-end, allowing the generator to produce outputs that are optimized for refinement, thus addressing the mismatch seen in traditional approaches.
Results
Experiments conducted on the ERA5–SolarCube benchmark indicate that SolarFlowRefiner consistently outperforms both standalone generation models and traditional post-hoc refinement techniques, leading to more accurate high-resolution SSR reconstructions.
Implications
The findings suggest that SolarFlowRefiner can significantly enhance solar forecasting and grid management by providing more accurate and spatially detailed SSR fields. This has broader implications for renewable energy integration into electricity grids and climate applications, where precise solar radiation data is crucial.
A Comparative Framework for Evaluating Foundation Models on Tabular Data: A Case Study in Healthcare
Theory
Interpretability
Efficient ML
- Introduction of OpTFM, a framework for evaluating foundation models on tabular data in healthcare.
- Evaluation across six dimensions relevant to clinical applications, allowing for nuanced model comparisons.
- Demonstration of varying model rankings based on specific healthcare use cases.
- Provision of a taxonomy of 45 foundation models categorized by architecture.
Read more
A Comparative Framework for Evaluating Foundation Models on Tabular Data: A Case Study in Healthcare
Summary
This paper introduces OpTFM, a comparative evaluation framework designed to assess foundation models for tabular data, particularly in healthcare applications. The framework evaluates models across six clinically relevant dimensions: generalization to new datasets, patient privacy protection, data requirements, scalability, interpretability, and fairness across patient subgroups. Each dimension is further broken down into measurable sub-components, allowing for the creation of compound scores, or super-metrics, that provide a comprehensive view of model performance. The authors demonstrate the framework's utility through two healthcare use cases: screening for iron deficiency and predicting heart failure, revealing that model rankings can vary significantly based on the specific clinical context and priorities. Additionally, the paper presents a taxonomy of 45 foundation models categorized by their underlying architectures, aiding researchers and practitioners in navigating the expanding landscape of tabular foundation models. The study highlights the need for a structured approach to model selection in healthcare, addressing gaps in existing surveys that fail to provide actionable insights for clinicians and data scientists.
Methodology
The authors developed the OpTFM framework, which scores and ranks foundation models based on six clinically meaningful dimensions. They applied this framework to two healthcare use cases, analyzing the performance of various models and demonstrating the impact of context on model selection. A taxonomy of models was also created to categorize them by architecture.
Results
The application of the OpTFM framework revealed that model rankings differ significantly based on the healthcare task at hand. The framework's structured approach allowed for a comprehensive evaluation of models, highlighting strengths and weaknesses across multiple dimensions. The taxonomy provided a clearer understanding of the landscape of foundation models for tabular data.
Implications
The findings suggest that clinicians and data scientists can make more informed decisions when selecting foundation models for specific healthcare applications. The OpTFM framework can facilitate better model deployment in clinical settings by addressing privacy, interpretability, and generalization concerns. The taxonomy serves as a valuable resource for navigating the rapidly evolving field of foundation models.