AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
48
Papers today
8h
Update frequency
7
Days of history
CalibForge: Adversarial Solver Calibration for Scaling Learnable Terminal Tasks
Large Language Models
Reinforcement Learning
Optimization
- CalibForge synthesizes terminal tasks using adversarial solver calibration to ensure tasks are appropriately challenging.
- Two calibration strategies (multi-solver and contrastive) enhance the task validation process by focusing on solver behavior.
- The system generated 5,431 calibrated tasks, significantly improving model performance on multiple benchmarks.
- Models trained on calibrated tasks outperformed baseline models by substantial margins, demonstrating the effectiveness of the approach.
Read more
CalibForge: Adversarial Solver Calibration for Scaling Learnable Terminal Tasks
Summary
The paper introduces CalibForge, an autonomous terminal-task synthesis system designed to enhance the training of terminal agents by ensuring tasks are not only executable but also appropriately challenging. The authors argue that traditional validation methods do not adequately assess task difficulty relative to solver capabilities. CalibForge employs adversarial solver calibration, utilizing verified solver behavior to refine candidate tasks through two strategies: multi-solver calibration, which targets disagreement among a diverse pool of solvers, and contrastive solver calibration, which focuses on a strong-pass/weak-fail relationship. This approach operationalizes a 'solver-relative learnable zone' where tasks are solvable yet not uniformly solved by all solvers. The system successfully constructs 5,431 calibrated terminal tasks, demonstrating that both calibration strategies provide more effective supervision than conventional methods. Experimental results show significant improvements in task performance, with models trained on the calibrated tasks achieving notable accuracy on various benchmarks, thereby supporting the notion of solver-relative learnability as a viable target for developing effective training data.
Methodology
CalibForge operates through an adversarial author-solver loop where solver behavior informs the task construction process. It employs two calibration strategies: multi-solver calibration, which retains tasks based on solver disagreement, and contrastive solver calibration, which focuses on a strong-pass/weak-fail relationship. The system iteratively revises candidate tasks based on solver outcomes and trajectories, ensuring they fall within a learnable zone.
Results
Models trained on the full collection of 5,431 calibrated tasks achieved accuracies of 32.58% and 47.57% on Terminal-Bench 2.0, with the largest improvements over baseline models reaching 24.71 percentage points. Additionally, the 35B model scored 44.32% on SWE-bench Pro and 48.77% on Doc2Repo, showcasing the effectiveness of the calibrated tasks across various benchmarks.
Implications
The findings suggest that adversarial calibration can significantly enhance the quality of training data for terminal agents, leading to improved performance on complex tasks. This approach could be applied in various domains requiring robust task generation and validation, such as software engineering and AI training.
Capacity Confounds and Coverage Guarantees in Adaptive Sub-model Federated Learning
Federated Learning
- Client heterogeneity estimation from updates is confounded by device capacity.
- Adaptive capacity allocation can lead to model corruption if clients are under-resourced.
- A coverage guarantee can prevent failure modes in adaptive allocation strategies.
- Uniform allocation may perform comparably to heterogeneity-aware policies under certain conditions.
Read more
Capacity Confounds and Coverage Guarantees in Adaptive Sub-model Federated Learning
Summary
This paper investigates the challenges of adaptive capacity allocation in sub-model federated learning (FL), where clients with limited resources train reduced-width versions of a global model. The authors introduce the Heterogeneity-Aware Sub-model Federated Learning (HAS-FL) framework to analyze whether client data heterogeneity can be estimated from observed updates. The findings reveal that existing methods for estimating client heterogeneity are confounded by device capacity, leading to a misinterpretation of data signals. The study identifies a hidden failure mode in adaptive allocation when clients are restricted below full model capacity, resulting in corrupted global model performance. A coverage guarantee is proposed to mitigate this issue, demonstrating that uniform allocation can collapse under certain conditions. The research concludes that while sub-model training is cost-effective for resource-constrained clients, the accuracy trade-off is significant, and the benefits of adaptive allocation stem more from capacity budgeting than from effective heterogeneity estimation.
Methodology
The authors employed a systematic analysis of the HAS-FL framework, validating their findings against ground-truth label-distribution divergence across multiple datasets. They explored various estimators and conducted empirical evaluations to assess the impact of capacity on heterogeneity estimation and model performance.
Results
The study found that estimates of client heterogeneity were strongly negatively correlated with device capacity, indicating that capacity confounds the data signal. Additionally, when clients were limited in model capacity, uncovered parameters corrupted the global model. The proposed coverage guarantee successfully addressed this failure, and matched-budget controls showed that random allocation performed similarly to heterogeneity-aware strategies in certain scenarios.
Implications
The findings suggest a need for future federated learning designs to incorporate heterogeneity signals that can be disentangled from capacity effects. This could lead to more effective adaptive allocation strategies that maintain model accuracy while accommodating resource-constrained clients.
ArchEGraph: A Large-Scale Graph Dataset for Geometry-Topology-Physics Aligned Building Energy Modeling
Graph Learning
- ArchEGraph is a large-scale dataset that aligns building geometry, topology, and physics for energy modeling.
- The dataset includes 5,481 buildings and over 49,000 simulation cases, capturing complex geometric and topological features.
- Two benchmark tasks are defined: graph reconstruction from meshes and topology-informed load prediction.
- Standardized evaluation protocols are introduced to assess model performance across various conditions.
Read more
ArchEGraph: A Large-Scale Graph Dataset for Geometry-Topology-Physics Aligned Building Energy Modeling
Summary
The paper introduces ArchEGraph, a large-scale benchmark dataset designed to enhance building energy modeling (BEM) by explicitly mapping building geometry to performance metrics. The dataset comprises 5,481 buildings and 49,326 validated building-weather simulation cases, featuring over 1.33 Γ 10^5 space nodes and 1.44 Γ 10^6 face nodes, which reflect significant geometric and topological complexity. ArchEGraph supports two primary benchmark tasks: (i) graph reconstruction from polygonal meshes, focusing on recovering topological structures from geometric representations, and (ii) topology-informed load prediction, which utilizes graph structures and temporal weather data to forecast zone-level energy demand. The authors establish standardized evaluation protocols for these tasks and conduct experiments to assess model robustness across different buildings and climates. By providing a unified testbed for studying the coupling of geometry, topology, and physics in building energy modeling, ArchEGraph aims to facilitate the development of scalable and generalizable surrogate models for energy prediction.
Methodology
The authors generated the ArchEGraph dataset through physics-based simulations to ensure consistency across geometry, topology, and physics. They defined two benchmark tasks: M2G (Mesh to Graph) for reconstructing building topology from geometric inputs, and G2E (Graph to Energy) for predicting zone-level load dynamics based on temporal weather conditions. Standardized evaluation protocols were established, and experiments were conducted to evaluate model performance across different buildings and climates.
Results
The introduction of ArchEGraph allows for improved modeling of building energy use, addressing limitations of existing datasets that lack comprehensive geometry and physics alignment. The benchmark tasks demonstrated the potential for enhanced prediction accuracy and robustness of machine learning models in building energy applications.
Implications
ArchEGraph has significant implications for sustainable building design, enabling architects and engineers to receive rapid feedback on design decisions. It also provides a foundation for future research in structured physical learning and the development of more effective surrogate models for building energy prediction.
Faster Query-Key Learning Sharpens Attention in Self-Attention Models
NLP
Theory
Interpretability
- The parameterization of query-key and output-value circuits significantly affects attention patterns in self-attention models.
- Faster learning rates for query-key parameters lead to sharper attention on task-relevant tokens.
- The study provides closed-form dynamics that explain how relative learning speeds shape attention during training.
- Experiments confirm that attention structures can be altered without compromising predictive performance.
Read more
Faster Query-Key Learning Sharpens Attention in Self-Attention Models
Summary
This paper investigates the dynamics of self-attention mechanisms in transformer models, focusing on the interaction between the query-key and output-value circuits during training. The authors demonstrate that the parameterization of these circuits significantly influences attention patterns, particularly how attention is allocated to task-relevant tokens. By analyzing the training dynamics of a single-layer transformer model trained for next-token prediction, the authors show that varying the learning rates of the query-key and output-value parameters leads to qualitatively different attention structures, even when predictive performance remains similar. The study derives closed-form dynamics that reveal how the relative learning speeds of these circuits affect attention sharpening, with faster query-key learning resulting in sharper attention on relevant tokens. Experiments conducted on both synthetic and real-world datasets validate the theoretical findings, showing that changes in optimization speed and parameterization can alter attention structures while maintaining comparable predictive performance.
Methodology
The authors analyze the training dynamics of a single-layer transformer model using gradient-flow analysis. They derive closed-form dynamics under controlled synthetic settings, comparing factorized and collapsed parameterizations of the query-key and output-value circuits. The study employs both theoretical derivations and empirical experiments on synthetic and real-world datasets to validate the proposed hypotheses.
Results
The results indicate that the relative learning rates of the query-key and output-value circuits govern attention concentration. Faster query-key learning leads to sharper attention on relevant tokens, while maintaining similar predictive performance across different parameterizations. The derived dynamics show that the effective parameters evolve at different rates, providing a mechanistic explanation for the observed attention sharpening.
Implications
The findings suggest that careful tuning of learning rates and parameterization in self-attention models can enhance attention interpretability and effectiveness in various applications, such as natural language processing and other domains utilizing transformer architectures. This work could lead to improved model designs that leverage attention mechanisms more effectively.
CohortHijack: Robustness of Single Cell Annotation to Companion Cell Removal
Theory
- CohortHijack identifies companion-cell removal as a threat to single-cell annotation integrity.
- Structured removal methods consistently outperform random removal strategies.
- Multi-start search techniques can significantly alter target annotations with minimal collateral impact.
- Cohort composition is a critical factor in the reliability of single-cell annotations.
Read more
CohortHijack: Robustness of Single Cell Annotation to Companion Cell Removal
Summary
The paper introduces CohortHijack, a robustness audit designed to evaluate the vulnerability of single-cell annotation methods to the removal of non-target companion cells. The authors investigate how the removal of selected cells can manipulate the refined annotations of a target cell, despite the target's gene expression profile remaining unchanged. The study employs various removal strategies, including random and structured methods, and evaluates their impact on two single-cell datasets (PBMC3K and Paul15) using logistic regression and calibrated linear SVM classifiers. The findings reveal that structured removal methods outperform random ones, with multi-start search techniques significantly altering target annotations while maintaining low collateral changes. The research highlights the importance of cohort composition in single-cell annotation and identifies it as a potential attack surface for adversarial manipulation. The study also validates the findings using the CellTypist majority voting framework, demonstrating that independent predictions remain stable while refined labels can change with minor cohort adjustments. Overall, the paper emphasizes the need for robust annotation methods that account for cohort dependencies.
Methodology
The authors developed a framework for evaluating single-cell annotation robustness by removing selected non-target cells while preserving the target cell's expression profile. They employed various removal strategies, including random, nearest-neighbor, same-class, greedy search, multi-start greedy search, and beam search, and conducted controlled experiments on two single-cell datasets using logistic regression and calibrated linear SVM classifiers.
Results
The results indicated that structured removal methods were more effective than random removals, particularly in the Paul15 dataset. The multi-start search method altered 24.33% of linear-SVM targets and 19.67% of logistic-regression targets while keeping collateral changes below 0.4%. The findings also showed that when neighborhood refinement was disabled, the effects of companion-cell removal disappeared, underscoring the dependency of annotations on cohort composition.
Implications
The findings suggest that single-cell annotation methods need to be robust against changes in cohort composition to ensure reliable predictions. This research could inform the development of more resilient annotation tools and highlight the importance of considering cohort dependencies in single-cell analyses.
Reasoning Errors Have a Region and a Direction in the Residual-Stream Trajectory of LLMs
NLP
Large Language Models
Interpretability
- Introduces a three-stream detector for improved reasoning error detection in LLMs.
- Combines motion with coarse and fine location readings to enhance context interpretation.
- Achieves up to 12% accuracy improvement over existing displacement-only methods.
- Demonstrates effectiveness across various reasoning benchmarks, including factual tasks.
Read more
Reasoning Errors Have a Region and a Direction in the Residual-Stream Trajectory of LLMs
Summary
This paper addresses the challenge of distinguishing sound reasoning from flawed reasoning in large language models (LLMs), which is increasingly important as these models are used for tasks requiring verifiable reasoning. The authors critique existing trajectory-based methods that utilize layerwise residual-stream displacements to capture representation changes, noting that while these methods suppress stable, label-correlated signals, they may also omit crucial reasoning-related information. To overcome this limitation, the authors propose a novel three-stream detector that integrates motion with two restricted views of location: a coarse region reader using vector quantization and a fine direction reader analyzing normalized multi-layer states. This approach allows for sufficient context restoration to interpret motion without reverting to full-state probing. The proposed method demonstrates significant improvements in reasoning accuracy on unseen benchmarks, outperforming existing state-of-the-art methods and single-layer probing baselines. Additionally, it shows superior performance in tasks related to factual completion and fact verification, indicating its robustness across different reasoning tasks. The findings suggest that reasoning validity is more effectively assessed through state-conditioned motion rather than static states or decontextualized trajectories alone.
Methodology
The authors developed a three-stream validity detector that incorporates three components: a motion reader that captures layerwise displacements, a coarse region reader that uses vector quantization to identify the location of states, and a fine direction reader that analyzes normalized multi-layer states. This design allows for the selective restoration of reasoning-related state information while maintaining the advantages of trajectory-based methods.
Results
The proposed method improved selection accuracy by up to 12% compared to the displacement-only state of the art and 21% over single-layer probing baselines on reasoning benchmarks. It also outperformed other detectors in tasks related to factual completion and fact verification, demonstrating its effectiveness in identifying reasoning validity.
Implications
The findings suggest that enhancing the interpretability of LLMs through improved reasoning error detection could lead to more reliable applications in critical areas such as healthcare, legal reasoning, and automated decision-making, where sound reasoning is essential.
Sub-Quadratic Bisimulation Metrics via Approximate Nearest Neighbors: Coverage-Augmented Guarantees and Computable Two-Sided Certificates
Reinforcement Learning
Theory
Efficient ML
- Introduces a sub-quadratic algorithm for computing bisimulation metrics in MDPs.
- Utilizes an approximate nearest neighbor index for efficient state pair selection.
- Establishes a coverage-augmented anytime error bound for the proposed method.
- Demonstrates significant computational efficiency improvements in experiments.
Read more
Sub-Quadratic Bisimulation Metrics via Approximate Nearest Neighbors: Coverage-Augmented Guarantees and Computable Two-Sided Certificates
Summary
This paper addresses the computational challenges associated with bisimulation metrics in Markov Decision Processes (MDPs), which traditionally require quadratic time complexity due to the need to update every state pair. The authors propose a sub-quadratic algorithm that utilizes an approximate nearest neighbor (ANN) index to selectively update state pairs based on a low-dimensional embedding. This method not only reduces the computational burden but also provides a certificate-carrying framework that ensures the accuracy of the bisimulation metric. The main analytical contribution is a coverage-augmented anytime error bound that accounts for the limitations of local index quality in controlling global error. The authors demonstrate that while uncovered pairs can retain their initialization gap, the algorithm can still provide a computable sandwich width that bounds the error. Experimental results validate the proposed method, showing significant improvements in efficiency and accuracy compared to existing baselines, particularly in larger state spaces.
Methodology
The authors develop a certificate-carrying sub-quadratic algorithm that employs an approximate nearest neighbor index over a low-dimensional embedding of the state space. The algorithm selectively updates state pairs using the exact Kantorovich operator, while bounding the global error through a coverage-augmented approach. Theoretical results include an anytime error bound and a computable sandwich width that certifies the accuracy of the metric without requiring the exact metric computation.
Results
The proposed method achieves a significant reduction in computational complexity, demonstrating sub-quadratic scaling in experiments compared to traditional quadratic methods. In benchmark tests, the algorithm reaches the exact-metric skyline with fewer than half of the state pairs covered, outperforming existing methods by a factor of 22-33x in terms of efficiency. The results also show that the method can effectively handle large state spaces, such as a 2500-state gridworld, improving over reward-only metrics by 28.6% using only a fraction of the computational resources.
Implications
The findings of this paper have potential implications for reinforcement learning applications where efficient state representation and similarity measurement are crucial. The proposed method can enhance the scalability of algorithms that rely on bisimulation metrics, making them applicable to larger and more complex environments. Additionally, the framework for certifying metric accuracy could be beneficial in various domains requiring reliable state abstraction and representation learning.
How Far Do Simple Transformations Translate Across Text Embedding Models?
NLP
Theory
Interpretability
- Simple transformations can recover shared structures in some text embedding models but not universally.
- Compatibility of transformations is influenced by model architecture, training objectives, and pooling strategies.
- The study employs a multi-diagnostic approach, evaluating geometric similarity, retrieval, and downstream transfer.
- Findings challenge the literature's assumption of universal latent compatibility across different models.
Read more
How Far Do Simple Transformations Translate Across Text Embedding Models?
Summary
This paper investigates the effectiveness of simple transformations, specifically linear mappings, in translating representations across various text embedding models. The authors aim to understand the compatibility of independently trained models in organizing semantic information, which is crucial for enabling AI-to-AI communication without the need for human-readable text. The study evaluates nine different embedding models that vary in architecture, pooling strategy, and training objectives. Using multiple diagnostic methods including Centered Kernel Alignment (CKA), downstream transfer, fidelity, and retrieval, the authors find that while simple transformations can recover meaningful shared structures for some model pairs, they fail for others. The compatibility of transformations is influenced by the models' architecture, training objectives, pooling methods, and data distributions. Overall, the findings challenge the notion of universal compatibility among heterogeneous embedding spaces, suggesting that simple mappings are not as universally applicable as previously thought.
Methodology
The authors conducted a systematic empirical study using nine text embedding models with varying characteristics. They applied linear mappings as simple translators and evaluated their effectiveness through multiple metrics, including CKA for geometric similarity, downstream transfer performance, fidelity, and retrieval capabilities. The study included both constrained linear maps from small anchor sets and more flexible linear maps trained on larger paired corpora.
Results
The results indicate that while simple transformations can facilitate meaningful translations between some compatible model pairs, they fail significantly for others. The degree of compatibility is shown to depend on the specific architectural and training characteristics of the models involved. The study provides evidence that the notion of universal compatibility among heterogeneous embedding spaces is overly simplistic.
Implications
The findings suggest that while simple transformations can enhance the modularity and reusability of pretrained components across different models, they are not universally applicable. This has important implications for the development of AI systems that rely on latent communication, as understanding the limits of these transformations is crucial for future advancements in AI-to-AI interactions.
Mathematical Principles and Experimental Discoveries of the Emergence of Symbolic Patterns in Artificial Neural Networks
Theory
Interpretability
- Symbolic patterns can emerge in ANNs, providing a means to explain their inference logic.
- Two mathematical criteriaβmonotonicity and smoothnessβlead to the emergence of sparse symbolic interactions.
- The emergent symbolic patterns demonstrate strong transferability across different input samples and models.
- The study suggests a shift towards communicative learning, allowing for direct inspection and tuning of ANN logic.
Read more
Mathematical Principles and Experimental Discoveries of the Emergence of Symbolic Patterns in Artificial Neural Networks
Summary
This paper addresses the challenge of explainability in artificial neural networks (ANNs) by investigating the emergence of symbolic patterns within these models. The authors propose that the inference logic of ANNs can be reformulated as sparse symbolic interactions, which can be understood through two mathematical criteria: monotonicity and smoothness of the ANN's average k-order output. The study demonstrates that these symbolic patterns are not merely coincidental but reflect underlying natural laws governing the behavior of ANNs. Through extensive empirical evidence, the authors show that these symbolic interactions exhibit strong transferability across different samples and models, suggesting that they encapsulate fundamental knowledge rather than being tied to specific neurons. The findings provide a theoretical foundation for symbolic explanations of ANNs and highlight the potential for communicative learning, where the inference logic can be inspected and adjusted at the symbolic level, complementing traditional end-to-end learning approaches. The results imply that similar symbolic representations may emerge in other black-box systems under certain conditions, broadening the scope of interpretability in machine learning.
Methodology
The authors utilize a logical model based on AND-OR interactions to explain the inference logic of ANNs. They analyze the properties of this model, proving its fidelity and conciseness in approximating ANN outputs across various input configurations. The study includes both theoretical proofs and extensive empirical experiments to validate the emergence and transferability of symbolic interactions.
Results
The research confirms that the logical model can accurately approximate ANN outputs across a wide range of input samples, demonstrating the universal matching property. Additionally, the sparsity of interactions indicates that only a small number of AND-OR interactions are necessary to explain the model's behavior, supporting the existence of primitive inference patterns encoded by the ANN.
Implications
The findings have significant implications for enhancing the interpretability of ANNs, suggesting new avenues for developing explainable AI systems. The concept of communicative learning could lead to more transparent models that allow for human-like reasoning and adjustments in machine learning applications. Furthermore, the potential for similar symbolic representations in other black-box systems could revolutionize the understanding and trust in complex AI models.
Equipment-centric workpiece localization in near real-time using deep learning-based vision and event-driven finite state machines
Computer Vision
Robotics
- Introduction of an equipment-centric localization framework for hot forging environments.
- Utilization of video streams and event-driven FSMs for robust workpiece tracking.
- Achieved 100% event detection accuracy and a mean localization error of 317.8 mm.
- Integration of KPGA mechanism enhances activity recognition performance.
Read more
Equipment-centric workpiece localization in near real-time using deep learning-based vision and event-driven finite state machines
Summary
This paper addresses the critical challenge of continuous workpiece localization in hot forging environments, where traditional workpiece-centric methods often fail due to extreme conditions. The authors propose an innovative equipment-centric framework that utilizes video streams from static 2D cameras to infer workpiece locations indirectly by analyzing equipment operations. This approach couples equipment behavior with workpiece-handling events, allowing for robust localization even under harsh conditions. The framework estimates 3D coordinates of handling equipment and recognizes activities such as grasping and releasing workpieces. Event-driven finite state machines (FSMs) are employed to validate and detect handling events, enabling continuous updates of workpiece states. Experimental validation in a hot forging factory demonstrated the framework's effectiveness, achieving 100% event detection accuracy and a mean localization error of 317.8 mm. The integration of a Keypoint-Guided Attention (KPGA) mechanism further improved activity recognition performance compared to traditional models. The structured outputs of the framework facilitate data-driven visualization of workpiece transfers, enhancing traceability and quality control in forging operations.
Methodology
The authors developed an equipment-centric framework that analyzes video streams from static cameras to infer workpiece locations based on equipment operations. They employed event-driven finite state machines to detect and validate handling events, continuously updating workpiece states and coordinates. The Keypoint-Guided Attention mechanism was integrated to improve activity recognition.
Results
The proposed framework achieved 100% event detection accuracy within a 33-second tolerance window, a mean localization error of 317.8 mm, and a mean system latency of 21 seconds. Ablation studies confirmed that the KPGA mechanism significantly improved activity recognition performance over baseline models.
Implications
This research has significant implications for enhancing traceability and quality control in hot forging operations. The equipment-centric approach allows for continuous localization without the need for physical tags, addressing the challenges posed by extreme conditions in manufacturing environments. It can lead to more intelligent and efficient production processes.
Quantum-Structured World Models (QSWMs) for Predictive Latent Dynamics
Theory
Reinforcement Learning
- Introduction of Quantum-Structured World Models (QSWMs) as a quantum-inspired framework for predictive modeling.
- Establishment of three foundational properties: classical inclusion, predictive sufficiency, and structured compactness.
- Demonstration of the effectiveness of ComplexQSWM over classical baselines in local predictive tasks.
- Identification of limitations in long-horizon predictions and latent interpretability.
Read more
Quantum-Structured World Models (QSWMs) for Predictive Latent Dynamics
Summary
This paper introduces Quantum-Structured World Models (QSWMs), a novel framework for predictive world modeling that leverages quantum-inspired structures for latent states. Traditional world models typically utilize classical representations such as vectors or probability distributions to summarize interaction histories and support predictions. In contrast, QSWMs employ complex-valued representations and density-matrix-like latents to capture the dynamics of latent states more effectively. The authors establish three foundational properties of QSWMs: classical inclusion, predictive sufficiency, and structured compactness, which demonstrate the potential of quantum-inspired structures to enhance predictive capabilities. The paper presents two specific implementations of QSWMsβComplexQSWM and DensityQSWMβand evaluates their performance against classical baselines using elementary cellular automata. The results indicate that ComplexQSWM outperforms traditional models in local predictive tasks, suggesting that quantum-structured representations can provide significant inductive biases for predictive dynamics. However, the study also notes limitations in long-horizon predictions and latent interpretability, highlighting areas for future research.
Methodology
The authors define QSWMs as formal objects for latent world modeling, utilizing quantum-inspired latent states, transition operators, and measurement-inspired decoding maps. They instantiate two variants, ComplexQSWM and DensityQSWM, and conduct empirical evaluations against classical models using elementary cellular automata to assess predictive performance.
Results
ComplexQSWM achieved the highest local predictive performance among all evaluated models, consistently outperforming classical baselines. The results also indicated that the benefits of QSWMs extend beyond fixed data settings, although challenges in long-horizon predictions and latent interpretability were noted.
Implications
The findings suggest that quantum-structured representations could significantly advance the field of predictive modeling, particularly in areas requiring complex latent dynamics. This work opens avenues for further exploration of quantum-inspired methodologies in machine learning and artificial intelligence.
MAUPITI: On-Device Prototype-Based Learning on a Smart Infrared Sensor
Computer Vision
Efficient ML
- MAUPITI integrates a low-power infrared sensor with on-device learning capabilities.
- The use of a prototype-based NCM classifier allows for efficient online adaptation without backpropagation.
- The system operates within tight memory and power constraints, making it suitable for embedded applications.
- Experimental results show comparable accuracy to traditional classifiers with minimal latency.
Read more
MAUPITI: On-Device Prototype-Based Learning on a Smart Infrared Sensor
Summary
This paper presents MAUPITI, a smart multi-pixel infrared (IR) sensor that integrates a 16Γ16 thermal MOSFET array with a RISC-V microcontroller, enabling on-device learning for pose and gesture recognition tasks. The system is designed to operate under strict memory and power constraints, utilizing less than 32 kB of on-chip memory and approximately 1.5 mW of power. The authors propose a prototype-based Nearest Class Mean (NCM) classifier that avoids the memory overhead associated with traditional backpropagation methods. Instead, a Convolutional Neural Network (CNN) encoder is trained offline and quantized, while class prototypes are updated on-device in a streaming manner. The approach allows for continual adaptation and online learning without the need for extensive retraining. Experimental results demonstrate that MAUPITI achieves accuracy comparable to conventional classifiers while maintaining low latency during both classification and prototype updates. This advancement enables the sensor to adapt to varying conditions and recognize new patterns efficiently, making it suitable for privacy-preserving applications in embedded systems.
Methodology
The methodology involves training a CNN encoder offline using metric learning techniques, specifically employing triplet loss for embedding generation. The Nearest Class Mean (NCM) classifier is implemented to compute class prototypes based on the embeddings. On-device updates to prototypes are performed incrementally as new samples are received, allowing for continual learning without the need for storing previous samples or performing backpropagation.
Results
The experiments conducted on two datasets indicate that the MAUPITI system achieves accuracy levels comparable to traditional CNN classifiers trained with backpropagation. The latency for both classification and prototype updates is negligible, with less than 0.29% overhead, demonstrating the system's efficiency in real-time applications.
Implications
The implications of this work extend to various applications in privacy-preserving human sensing, such as occupancy monitoring and human-computer interaction, where low power consumption and on-device learning are critical. The ability to adapt to new classes and environments without extensive retraining enhances the practicality of deploying such systems in dynamic settings.
Walkable to Whom? Capturing Subjective Variability in Walkability Perception Using Multimodal Deep Learning
Multimodal
- Introduces a new dataset linking sidewalk-view imagery with individual walkability ratings.
- Demonstrates that the viewpoint of imagery (sidewalk vs. street view) significantly affects walkability ratings.
- Proposes a multimodal deep learning framework that incorporates user attributes to model subjective variability in walkability perception.
- Achieves a 65% improvement in rank agreement over traditional image-only models.
Read more
Walkable to Whom? Capturing Subjective Variability in Walkability Perception Using Multimodal Deep Learning
Summary
This paper addresses the subjective variability in walkability perception, which varies significantly among individuals due to personal characteristics and experiences. Traditional studies often aggregate walkability ratings, assuming a uniform perception, and typically utilize vehicle-mounted street-view imagery that fails to capture the pedestrian experience. The authors introduce a new dataset comprising 29,870 walkability ratings from 1,196 respondents, linking sidewalk-view imagery from various Australian environments with individual attributes. They propose a user-conditioned multimodal deep learning framework that integrates visual features with respondent-level data to better model individual perceptions of walkability. A viewpoint-comparison study reveals that sidewalk-view images receive higher walkability ratings than street-view images, highlighting the importance of imagery source in perception surveys. The user-conditioned model significantly improves rank agreement with observed ratings, demonstrating that individual evaluator characteristics provide predictive power beyond mere image content. This research advocates for a shift from aggregated walkability scores to models that account for diverse user perspectives, thereby enabling a more inclusive assessment of pedestrian environments.
Methodology
The authors developed a multimodal deep learning framework that fuses visual features from sidewalk-view images with individual respondent attributes. They conducted a viewpoint-comparison study to assess the impact of imagery source on walkability ratings and evaluated the performance of their user-conditioned model against an image-only baseline using quadratic weighted kappa as a metric.
Results
The study found that sidewalk-view images received significantly higher walkability ratings compared to street-view images. The user-conditioned model demonstrated a 65% improvement in rank agreement with observed ratings (quadratic weighted kappa of 0.47) compared to the image-only baseline (kappa of 0.29), indicating that individual characteristics are crucial for accurately predicting walkability perceptions.
Implications
The findings suggest that urban planners and designers should consider subjective variability in walkability perception when assessing pedestrian environments. The proposed framework could enhance personalized pedestrian routing systems, allowing for tailored navigation that reflects individual preferences and perceptions, similar to existing multi-criteria routing approaches.
SkillTFM: Gated Skill Evolution for Training-Free Adaptation of Tabular Foundation Models
Time Series
Optimization
Efficient ML
- SkillTFM is the first skill-based adaptation system for training-free tabular foundation models.
- It employs a gated skill evolution mechanism that couples selective repairs with safe fallbacks.
- The system demonstrates significant improvements in prediction accuracy, particularly in boundary scenarios.
- SkillTFM's learned skill state is transferable across different TFM backbones and optimizer settings.
Read more
SkillTFM: Gated Skill Evolution for Training-Free Adaptation of Tabular Foundation Models
Summary
The paper introduces SkillTFM, a novel training-free adaptation system designed for Tabular Foundation Models (TFMs) that addresses challenges associated with distribution shifts and heterogeneous feature semantics in tabular data. SkillTFM shifts the adaptation process from traditional parameter updates to a gated evolution of agentic skills, utilizing a verifiable skill bank that identifies boundary evidence and retrieves reusable skills for task-specific adaptations. The system operates by extracting evidence from the current task, which informs the selection of candidate repairs from the skill bank. These repairs are validated through a runtime certificate, ensuring that only reliable adaptations are executed. The authors demonstrate the effectiveness of SkillTFM through experiments on simulated boundary settings and real-world applications, such as electricity-price forecasting, showing significant improvements in prediction accuracy while maintaining the integrity of the base model. SkillTFM's gated skill evolution mechanism allows for selective interventions, enhancing the adaptability of TFMs across various datasets without the need for extensive retraining.
Methodology
SkillTFM utilizes a gated external skill state to adapt TFMs without retraining. It extracts boundary evidence from the current task, including feature distributions and prediction biases, to inform the selection of candidate repairs from a skill bank. Each candidate repair undergoes validation through a runtime certificate, ensuring that only evidence-supported repairs are executed. The system also incorporates a validation-gated skill evolution process to expand its skill state.
Results
SkillTFM achieved an improvement in AUC by 0.128β0.142 across various boundary evaluations, with a notable increase from 0.699 to 0.898 on unsupported nonlinear boundaries. In electricity-price forecasting, the mean absolute error (MAE) was reduced significantly, demonstrating the effectiveness of selective repairs.
Implications
The proposed SkillTFM framework has potential applications in various fields that rely on tabular data, such as finance, healthcare, and public services. Its ability to adapt without retraining could streamline processes in environments with frequent distribution shifts, making it a valuable tool for real-time decision-making.
Align-RAG: Alignment Is All You Need for TSFM In-Context Learning
Time Series
- Align-RAG is a training-free method that enhances frozen TSFMs for time series forecasting.
- It outperforms the state-of-the-art TS-RAG method across multiple datasets without requiring learned parameters.
- The method applies amplitude rescaling and phase shifting to align retrieved data with the query.
- Align-RAG improves zero-shot forecasting accuracy significantly across various TSFM architectures.
Read more
Align-RAG: Alignment Is All You Need for TSFM In-Context Learning
Summary
The paper introduces Align-RAG, a novel method for retrieval-augmented forecasting that enhances frozen Time Series Foundation Models (TSFMs) without the need for fine-tuning or learned fusion modules. Previous methods relied on trained adapters to merge retrieved examples into forecasts, assuming that frozen models could not dynamically incorporate context. Align-RAG challenges this assumption by applying a closed-form amplitude rescaling and integer-lag phase shift to retrieved past-future windows, aligning them with the query before they are processed by the TSFM. The authors demonstrate that this approach outperforms the state-of-the-art trained retrieval adapter, TS-RAG, across all seven datasets in the benchmark, achieving an average reduction of 3.75% in Mean Squared Error (MSE). Additionally, Align-RAG shows improvements in zero-shot MSE on four other frozen TSFMs, with gains ranging from 2.5% to 13.7% without any tuning. The paper also investigates the mechanism behind these improvements, revealing that aligned demonstrations induce prediction shifts that closely track a closed-form ridge predictor, suggesting that frozen TSFMs can effectively utilize retrieved context dynamically. The findings advocate for closed-form alignment as a baseline for retrieval-augmented forecasting.
Methodology
Align-RAG employs a closed-form approach to align retrieved past-future windows with the query by applying amplitude rescaling and phase shifts. This is done without introducing any learned parameters, allowing the frozen TSFM to process the aligned data as in-context demonstrations.
Results
Align-RAG achieved an average reduction of 3.75% in MSE compared to TS-RAG across seven datasets. It also improved zero-shot MSE on four additional frozen TSFMs by 2.5% to 13.7%, demonstrating significant performance gains without any per-backbone tuning.
Implications
The findings suggest that retrieval-augmented forecasting can be effectively implemented without complex learned fusion modules, making it easier to adapt frozen TSFMs to new domains. This could lead to more efficient forecasting methods in various fields such as finance, healthcare, and energy management.
A Unified Risk View of Uncertainty: Posterior Risk for Disentanglement and Evaluation Beyond Proxies
Theory
Interpretability
- Introduces a unified definition of uncertainty as pointwise posterior risk.
- Develops a benchmark for direct computation of oracle epistemic and aleatoric uncertainty.
- Demonstrates that accurate predictions do not guarantee reliable uncertainty estimates.
- Highlights the limitations of existing proxy evaluations for uncertainty assessment.
Read more
A Unified Risk View of Uncertainty: Posterior Risk for Disentanglement and Evaluation Beyond Proxies
Summary
This paper addresses the challenge of reliable uncertainty estimation in safety-sensitive applications, emphasizing the need to disentangle epistemic and aleatoric uncertainty. The authors propose a unified definition of uncertainty as pointwise posterior risk, which represents the expected loss of a predictor under plausible ground-truth functions given the data. This approach integrates Bayesian uncertainty with estimator-dependent deviations, allowing for a more comprehensive understanding of uncertainty. The authors introduce a theory-backed benchmark that enables direct computation of oracle epistemic and aleatoric uncertainty using semi-synthetic datasets, which avoids reliance on proxy evaluations. Empirical findings reveal that accurate predictions do not necessarily correlate with reliable uncertainty estimates, highlighting the importance of method selection and the impact of dataset and modeling choices on uncertainty alignment. The proposed framework aims to enhance the evaluation of uncertainty estimates, particularly in critical domains where understanding the nature of uncertainty is crucial for decision-making.
Methodology
The authors define uncertainty as sample-conditional posterior risk and develop a semi-synthetic evaluation framework using real covariates and synthetic targets from a Gaussian-process prior. This allows for explicit computation of oracle uncertainty targets, facilitating a detailed analysis of uncertainty estimates.
Results
The empirical analysis shows significant differences in the performance of various methods for uncertainty disentanglement, revealing that some methods align better with oracle uncertainty targets than others. The findings indicate that the quality of uncertainty estimates is sensitive to the choice of dataset and modeling techniques.
Implications
The proposed framework and benchmark can improve the reliability of uncertainty estimates in machine learning applications, particularly in safety-critical fields such as healthcare and autonomous systems. By providing a clearer understanding of uncertainty, practitioners can make more informed decisions regarding model trustworthiness and risk management.
SAGA: Score-Weighted Adaptive Generation Alignment for Low-Resource Nordic Language Models
NLP
Large Language Models
Reinforcement Learning
- SAGA utilizes dependency-parser supervision to replace costly human preference annotations.
- The framework improves grammatical quality in low-resource languages without requiring human labels.
- Parser-derived supervision effectively addresses challenges such as reward hacking and alignment tax.
- Significant improvements in grammatical accuracy were observed across Danish, Icelandic, and Norwegian BokmΓ₯l.
Read more
SAGA: Score-Weighted Adaptive Generation Alignment for Low-Resource Nordic Language Models
Summary
The paper introduces SAGA (Score-weighted Adaptive Generation Alignment), a novel framework aimed at enhancing the grammatical quality of language models for low-resource Nordic languages, specifically Danish, Icelandic, and Norwegian BokmΓ₯l. Traditional preference optimization methods, which rely on human annotations, are impractical for these languages due to the scarcity of qualified annotators. SAGA addresses this challenge by utilizing existing high-quality dependency parsers to generate preference signals without the need for human labels. The framework converts parser judgments into preference pairs for Direct Preference Optimization (DPO), combines parser quality with lexical diversity into a composite reward, and employs a reward-gap criterion to filter low-information pairs. The results demonstrate significant improvements in grammatical accuracy across the three languages, showcasing the effectiveness of parser-derived supervision as a viable alternative to human preference annotations.
Methodology
SAGA employs a parser-guided preference optimization framework that converts parser judgments into preference pairs for DPO. It combines parser quality with lexical diversity to create a composite reward, filters low-information pairs based on a reward-gap criterion, and monitors for reward hacking to ensure reliable supervision.
Results
SAGA achieved a Danish parse success rate increase from 69.0% to 93.8%, a +4.5 percentage-point improvement in Icelandic on an independent evaluation, and an improvement of +28 percentage points in Norwegian BokmΓ₯l. Native speakers preferred SAGA outputs in 80% of pairwise comparisons, confirming the framework's effectiveness.
Implications
The findings suggest that parser-derived supervision can serve as a practical and cost-effective method for improving language model alignment in low-resource languages, potentially enabling better natural language processing applications in these linguistic communities.
The Tamed Subgradient Unadjusted Langevin Algorithm beyond Convexity
Optimization
Theory
Large Language Models
- Introduction of SG-TULA, a novel algorithm for sampling from non-convex distributions with non-smooth potentials.
- Derivation of non-asymptotic convergence bounds in Wasserstein-2 distance with explicit constants.
- Demonstration of SG-TULA's effectiveness in pretraining LLMs, achieving competitive results against established optimization methods.
- Addressing the challenges of superlinear gradient growth and non-convexity in optimization problems.
Read more
The Tamed Subgradient Unadjusted Langevin Algorithm beyond Convexity
Summary
This paper addresses the challenge of sampling from target distributions characterized by non-smooth potentials with superlinear gradient growth and non-convexity. The authors introduce the Subgradient Tamed Unadjusted Langevin Algorithm (SG-TULA), which is a discretization of Langevin diffusion that utilizes subgradients directly, avoiding the need for computationally intensive smoothing techniques. The algorithm employs taming techniques to ensure stability in the superlinear regime, and the authors derive non-asymptotic convergence bounds in Wasserstein-2 distance, providing explicit constants that improve upon existing rates for subgradient-based Langevin algorithms. Additionally, the paper presents excess risk estimates for the associated optimization problem, demonstrating the effectiveness of SG-TULA in pretraining large language models (LLMs) like GPT-2, showing competitive performance against fine-tuned methods such as AdamW and Muon, which lack comparable non-asymptotic guarantees.
Methodology
The authors develop SG-TULA as an explicit EulerβMaruyama discretization of Langevin diffusion that operates on subgradients. The algorithm stabilizes the drift using taming techniques, which rescale the subgradient based on its magnitude, avoiding the need for fixed thresholds. The paper also includes a coordinate-wise variant of SG-TULA and analyzes its performance under specific assumptions about the potential functions involved.
Results
The main results include non-asymptotic convergence bounds for SG-TULA in terms of Wasserstein-2 distance, with constants that are explicitly defined in relation to the problem's dimensionality and inverse temperature. The algorithm's complexity is shown to be O(Ο΅β4), with specific guarantees on the excess risk associated with the optimization problem, demonstrating that SG-TULA effectively samples from the Gibbs measure and minimizes the associated potential.
Implications
The findings suggest that SG-TULA can be a powerful tool for sampling and optimization in scenarios where traditional assumptions of Lipschitz continuity and convexity do not hold, particularly in the context of training deep learning models. This could lead to more robust training methods for complex models, such as transformers, that exhibit irregular gradient behavior.
Beyond Post-Hoc Temperature Scaling: Bilevel Optimization for LLM Calibration
NLP
Large Language Models
Optimization
- Identifies the limitations of post-hoc temperature scaling in LLM calibration.
- Proposes a bilevel optimization framework for training-time calibration adjustments.
- Introduces an entropy-maximization objective to mitigate overconfidence in LLMs.
- Demonstrates strong calibration performance across various tasks and datasets.
Read more
Beyond Post-Hoc Temperature Scaling: Bilevel Optimization for LLM Calibration
Summary
This paper addresses the issue of calibration in large language models (LLMs), which often become overconfident due to preference alignment techniques like Reinforcement Learning from Human Feedback (RLHF). Traditional calibration methods, such as post-hoc temperature scaling, are limited by their dataset dependence, failing to generalize across different domains. To overcome this, the authors propose a novel bilevel optimization framework that modifies model parameters during training to enhance calibration. The lower level of this framework fine-tunes the model under a parametric loss, while the upper level optimizes hyperparameters to maximize the entropy of predictive distributions, effectively discouraging overconfidence. The authors introduce an efficient first-order approximation to make this approach feasible for large-scale LLMs, avoiding the computational burden of second-order derivatives. Experimental results demonstrate that their method significantly improves calibration performance in both multiple-choice and open-ended generative question answering tasks, particularly excelling in out-of-domain scenarios.
Methodology
The authors formulate a bilevel optimization problem where the lower level fine-tunes the LLM using a parametric loss, while the upper level selects hyperparameters to maximize entropy in predictive distributions. They employ an efficient first-order approximation to avoid the computational complexity associated with second-order derivatives, making the method scalable for large models.
Results
The proposed method shows significant improvements in calibration performance compared to traditional calibration techniques, particularly in out-of-domain settings. Experiments reveal that the entropy-maximization approach leads to better-calibrated predictions in both multiple-choice and open-ended generative tasks.
Implications
The findings suggest that training-time calibration strategies can enhance the reliability of LLMs, making them more suitable for high-stakes applications where accurate confidence estimation is critical. This approach could lead to more robust models in fields such as healthcare and legal analysis, where miscalibrated predictions can have serious consequences.
When Do Corrective Features Help? An Agent for Corrective Feature Discovery on Black-Box Forecasters
Time Series
Interpretability
Large Language Models
- CRAFTER introduces a source-blind framework for corrective feature discovery, focusing on model-failure processes.
- The framework combines two feature generators: a compositional search and a large language model.
- CRAFTER significantly outperforms existing feature-engineering systems across multiple datasets and models.
- The effectiveness of corrective features is regime-dependent, highlighting the need for careful evaluation.
Read more
When Do Corrective Features Help? An Agent for Corrective Feature Discovery on Black-Box Forecasters
Summary
This paper addresses the challenge of correcting the predictions of frozen pretrained forecasters, which often exhibit systematic errors that are costly to fix through fine-tuning. The authors introduce CRAFTER (Corrective Residual Agent with Feature-based Temporal Exploration and Reasoning), a novel framework for corrective feature discovery that focuses on mining interpretable features from the residuals of these forecasters. Unlike traditional automated feature engineering that models the data-generating process, CRAFTER models the model-failure process by identifying features that capture the structured errors of the frozen model. The framework employs two generators: a compositional search over raw input channels and a large language model (LLM) that suggests named feature combinations and executable code. A validation-grounded gate evaluates candidate features without bias towards their origin, while a validation-selected corrector applies the accepted features or maintains the original forecast. The authors evaluate CRAFTER across six public datasets and six frozen backbone models, demonstrating that it consistently outperforms dedicated feature-engineering systems, effectively reducing errors by up to 27% on weaker models. The findings highlight the importance of understanding when corrective features are beneficial, inert, or harmful, providing insights into the conditions under which corrective feature discovery is most effective.
Methodology
The methodology involves a source-blind framework for corrective feature discovery that utilizes two generators: a compositional search over raw input channels and a large language model (LLM) for proposing named feature combinations. A validation-grounded gate evaluates candidate features, and a validation-selected corrector applies the accepted features or retains the original forecast. The evaluation is conducted across multiple datasets and frozen backbone models using a controlled comparison.
Results
CRAFTER consistently outperformed three dedicated feature-engineering systems across six public datasets and six frozen backbone families. The method demonstrated a significant reduction in forecasting errors, with improvements of up to 27% on weaker models. The results also indicated that the benefits of corrective features are robust across different LLM backends and persist even when applied to fine-tuned models.
Implications
The findings suggest that CRAFTER can serve as a valuable tool for practitioners looking to enhance the performance of frozen forecasters without the need for extensive retraining. The insights into the conditions under which corrective features are effective can guide future research and applications in automated feature engineering and model correction.
A Rate Separation for Agnostic Direct Sums
Theory
- The single-instance learning rate does not determine the direct-sum learning rate.
- Agnostic learning curves for constant functions and zero/identity functions are of order n^(-1/2).
- The paper provides a rate separation theorem demonstrating differing rates for direct sums as the number of factors increases.
- The results challenge existing assumptions about the relationship between single-instance and direct-sum learning rates.
Read more
A Rate Separation for Agnostic Direct Sums
Summary
This paper investigates the relationship between the agnostic PAC learning curves of direct sums of concept classes and their single-instance learning curves. The authors demonstrate that the single-instance learning rate does not dictate the direct-sum rate, providing a negative answer to a question posed by Hanneke, Moran, and Waknine regarding the dependence of the agnostic learning curve on the number of factors in direct sums. The study focuses on two specific classes of functions: constant binary functions and a combination of the zero function and the identity function, both exhibiting an agnostic learning curve of order n^(-1/2). The authors establish a rate separation theorem, indicating that while the learning rates of the two classes are comparable when considered individually, they diverge significantly when analyzed as direct sums. This finding highlights the complexity of learning in multi-instance settings and suggests that the structure of the concept classes plays a crucial role in determining learning rates.
Methodology
The authors utilize theoretical analysis and proofs to establish the relationships between the agnostic PAC learning curves of concept classes and their direct sums. They apply established inequalities and lemmas from statistical learning theory to derive bounds on the learning rates for both single-instance and direct-sum scenarios.
Results
The main results include the establishment of upper and lower bounds for the agnostic learning curves of the direct sums of the two classes. Specifically, the authors show that for the constant class, the learning rate is approximately 1/sqrt(n), while for the zero and identity function class, the learning rate is bounded by a function that depends on both the number of factors and the sample size. This leads to the conclusion that the rates diverge as the number of factors increases.
Implications
The findings have significant implications for understanding the complexities of agnostic learning in multi-instance settings. They suggest that practitioners should be cautious when generalizing single-instance learning rates to direct sums, as the underlying structure of the concept classes can lead to unexpected differences in learning performance.
Accelerating nanodrug development in continuous flow systems using informed prediction models based on low-cost surrogate nanoparticles
Optimization
- Introduction of a predictive modeling approach based on shape constraints for nanoparticle development.
- Utilization of controlled microfluidic methods to systematically prepare liposomes and lipid nanoparticles.
- Validation of the model with minimal empirical data, showcasing its effectiveness in predicting nanoparticle characteristics.
- Reduction of experimental workflows, leading to cost and time efficiency in nanodrug development.
Read more
Accelerating nanodrug development in continuous flow systems using informed prediction models based on low-cost surrogate nanoparticles
Summary
This paper addresses the challenges in the development of nanotherapeutics, particularly the extensive empirical optimization required due to the sensitivity of nanoparticle properties to process parameters. The authors propose a predictive modeling approach that utilizes shape constraints to enhance the estimation of nanoparticle characteristics across various production conditions. By employing controlled microfluidic methods, the study systematically prepares liposomes and lipid nanoparticles under varying conditions and validates the shape-constrained model with minimal empirical data. The results demonstrate that this modeling framework significantly reduces the need for extensive experimental workflows, facilitating accurate predictions of nanoparticle size and dispersity. This approach supports a more rational and efficient process development for manufacturing nanomedicine systems, ultimately aiming to accelerate the development of nanodrugs and improve their clinical efficacy.
Methodology
The authors employed controlled microfluidic methods to prepare liposomes and lipid nanoparticles under varying lipid concentrations, flow rates, and aqueous-to-organic mixing ratios. A shape-constrained predictive modeling approach was developed, informed by experimental data and expert knowledge, to estimate nanoparticle characteristics.
Results
The shape-constrained modeling approach successfully predicted nanoparticle size and dispersity with high accuracy, significantly reducing the need for extensive empirical testing. This validation indicates that the model can effectively guide the development of nanotherapeutics with improved efficiency.
Implications
The proposed modeling framework has the potential to streamline the development process of nanodrugs, making it more efficient and cost-effective. It could lead to faster clinical applications of nanomedicine and enhance the ability to tailor nanoparticle properties for specific therapeutic needs.
Beyond Foundation Models: Dimension-Aware Neural Architecture Search with Small-Data Representation Models for Cryocooler Lifetime Prediction
Time Series
Optimization
Efficient ML
- Introduction of the FSD-RM paradigm for small-data representation learning.
- Use of dimension-aware neural architecture search to optimize model design.
- Demonstration of competitive predictive performance on cryocooler telemetry data.
- Focus on capacity-controlled representation learning without large-scale pretraining.
Read more
Beyond Foundation Models: Dimension-Aware Neural Architecture Search with Small-Data Representation Models for Cryocooler Lifetime Prediction
Summary
This paper addresses the challenge of predicting the lifetime of cryocoolers using limited telemetry data, proposing a novel framework called the Family of Small-Data Representation Models (FSD-RM). Unlike traditional large-scale pretrained models that require abundant data, the authors focus on capacity-controlled representation learning using established encoder architectures such as CNN1D, LSTM, GRU, and Transformer. The approach integrates unsupervised representation learning with a two-stage pipeline for downstream lifetime prediction, employing dimension-aware neural architecture search (NAS) to optimize model capacity and input dimensionality. The experiments conducted on cryocooler telemetry data demonstrate that the proposed method achieves competitive predictive performance while minimizing training costs and model complexity. The findings suggest that effective representation learning can be accomplished without large-scale pretraining by applying appropriate inductive biases and controlling model capacity, making this approach particularly relevant for industrial applications where data is often scarce.
Methodology
The methodology involves unsupervised representation learning on multivariate telemetry data using encoder architectures (CNN1D, LSTM, GRU, Transformer). A two-stage pipeline is established for representation learning and downstream prediction, with dimension-aware NAS employed to systematically explore architectural trade-offs and optimize model capacity and input dimensionality.
Results
The proposed approach achieved competitive predictive performance in cryocooler lifetime prediction, demonstrating reduced training costs and model complexity compared to traditional large-scale pretrained models. The results indicate that effective representation learning is feasible in small-data settings when appropriate inductive biases and capacity controls are applied.
Implications
The findings have significant implications for predictive maintenance and reliability engineering in aerospace and other industries where telemetry data is limited. The framework can be adapted for various applications requiring lifetime predictions and anomaly detection in small-data environments.
BioM-JEPA: joint-embedding prediction of graph-connected gene blocks in single cells
Graph Learning
- BioM-JEPA predicts aggregate representations of graph-connected gene blocks instead of individual genes.
- The model employs a student-teacher framework to enhance representation learning efficiency.
- Linear attention is used to manage gene interactions, improving computational efficiency.
- BioM-JEPA outperforms existing models in retaining biological information and reducing errors in perturbation-response tasks.
Read more
BioM-JEPA: joint-embedding prediction of graph-connected gene blocks in single cells
Summary
The paper introduces BioM-JEPA, a novel joint-embedding predictive architecture designed to enhance the representation learning of single-cell transcriptomes by predicting aggregate representations of graph-connected gene blocks. Unlike traditional self-supervised models that focus on individual gene reconstruction, BioM-JEPA utilizes a student-teacher framework where the student predicts block representations from complementary genes, while the teacher provides target representations based on the full gene set. This approach leverages a binary gene graph constructed from protein associations and coexpression data, allowing for the definition of candidate target blocks. The model employs linear attention to efficiently handle gene interactions without the computational burden of a full attention matrix. The authors demonstrate that BioM-JEPA achieves superior performance in retaining biological information and reducing perturbation-response errors compared to existing models. The results indicate that graph-connected gene blocks are effective units for representation learning in single-cell biology, supporting the model's potential for broader applications in transcriptomic analysis.
Methodology
BioM-JEPA utilizes a joint-embedding predictive architecture where a student network predicts aggregate representations of gene blocks based on the remaining genes in a cell, while a slowly updated teacher network provides the corresponding target representations from the full observed gene set. The model employs linear attention to efficiently compute interactions among genes, avoiding the quadratic complexity of traditional attention mechanisms.
Results
The model demonstrated higher effective rank and weaker association with detected-gene depth in diagnostics compared to token-level predictions and reconstruction controls. In evaluations across CellBench tasks, BioM-JEPA achieved the lowest aggregate perturbation-response error and retained essential expression, pathway, and neighborhood information. Additionally, it provided significantly higher fine-tuning and embedding throughput compared to scFoundation.
Implications
The findings suggest that BioM-JEPA could be a valuable tool for analyzing single-cell transcriptomic data, potentially leading to better understanding of cellular processes and disease mechanisms. The model's efficiency and effectiveness in representation learning may facilitate advancements in personalized medicine and biological research.
Theoretical Foundations of Communication-Efficient, Robust, and Practical Distributed and Federated Optimization
Optimization
Federated Learning
Theory
- Introduction of ProxSkip for communication-efficient local gradient updates.
- Development of Variance Reduced ProxSkip to balance communication and computation costs.
- Demonstration of scalability under partial client participation.
- Establishment of a mechanism for Byzantine robustness through gradient difference clipping.
Read more
Theoretical Foundations of Communication-Efficient, Robust, and Practical Distributed and Federated Optimization
Summary
This dissertation addresses critical challenges in distributed and federated optimization, focusing on communication efficiency, robustness, and practical applicability. The author identifies seven key challenges at the intersection of theory and practice, particularly in federated learning. The first major contribution is the introduction of ProxSkip, which enhances communication efficiency by proving that local gradient steps can accelerate convergence. The framework is further extended with Variance Reduced ProxSkip, which mitigates neighborhood errors in stochastic updates while balancing communication and computation costs. The scalability of these methods is demonstrated under partial client participation, showing that local steps can still achieve communication acceleration. Additionally, the research explores server-side dynamics and data selection, proving that appropriate step sizes and sampling methods can improve convergence in heterogeneous environments. The study also highlights the benefits of compressing gradient differences over raw gradients in the context of Random Reshuffling. Furthermore, it establishes a mechanism for achieving Byzantine robustness alongside partial participation through gradient difference clipping. Finally, the dissertation introduces a theoretical framework for low-rank adaptation via randomized asymmetric chains, providing insights into fine-tuning large-scale models. The results are supported by numerical experiments that validate the theoretical findings, offering tighter analyses and more realistic assumptions than previous works.
Methodology
The author employs a theoretical approach to develop novel algorithmic frameworks addressing communication efficiency and robustness in federated learning. This includes proving the effectiveness of local gradient updates, variance reduction techniques, and exploring server-side dynamics and data selection strategies. The research also involves numerical experiments to validate theoretical claims.
Results
The findings indicate that local gradient updates can significantly enhance communication efficiency, even in heterogeneous settings. The proposed frameworks provide provable guarantees for communication acceleration and robustness against Byzantine attacks, while the low-rank adaptation framework offers new insights for fine-tuning large models.
Implications
The contributions of this dissertation have significant implications for improving the efficiency and robustness of federated learning systems, making them more practical for real-world applications. The theoretical advancements can guide future research in distributed optimization and enhance the deployment of machine learning models in decentralized environments.
Unmasking Removal-Budget Confounding: A Matched Operating-Point Evaluation Framework for Adaptive Data Cleaning
Computer Vision
Theory
Efficient ML
- Identification of granularity bias as a factor influencing removal budgets in adaptive data cleaning.
- Development of a matched operating-point evaluation framework to separate true detection capability from removal count effects.
- Demonstration that naive evaluations can misattribute performance improvements to better corruption discrimination.
- Findings indicate that performance gaps largely disappear when operating points are equalized, especially at low-to-moderate corruption rates.
Read more
Unmasking Removal-Budget Confounding: A Matched Operating-Point Evaluation Framework for Adaptive Data Cleaning
Summary
This paper addresses the issue of removal-budget confounding in adaptive data-cleaning methods, which can lead to misleading evaluations of performance improvements. The authors propose a novel evaluation framework that utilizes matched-budget and matched-recall controls, along with threshold-independent metrics like AUROC and AUPRC, to accurately assess the effectiveness of data-cleaning configurations. They demonstrate that naive evaluations can overstate performance gains due to variations in partition granularity, which alters the number of samples removed. Through experiments on CIFAR-10 and ImageNet-100, the authors reveal that performance differences largely diminish when operating points are matched, indicating that true improvements in corruption discrimination are often masked by changes in removal budgets. The study emphasizes the necessity of benchmarking adaptive cleaning methods at consistent operating points to ensure that observed performance gains are genuine and not artifacts of evaluation bias.
Methodology
The authors introduce a matched operating-point evaluation framework that controls for removal budget and recall while employing threshold-independent metrics (AUROC and AUPRC) to assess the ranking quality of corruption discrimination. They conduct experiments using a multi-cue adaptive cleaner with varying partition granularities and analyze performance across different corruption levels on CIFAR-10 and ImageNet-100 datasets.
Results
The results show that naive evaluations suggest significant performance improvements for the adaptive cleaner redesign, but these gains largely vanish when operating points are matched. The analysis reveals that clean-but-difficult samples contribute to false positives primarily at low corruption rates, and that their impact is heavily influenced by the chosen difficulty threshold. True ranking advantages are found only in specific low-prevalence settings and high-recall regions under severe corruption.
Implications
The findings underscore the importance of using matched operating points in evaluating adaptive data-cleaning methods to ensure that performance improvements are reflective of true corruption discrimination capabilities. This has significant implications for the reliability of AI systems in high-stakes applications, where data integrity is crucial.
Beyond Attention: Signed Integrated Gradients Attribution in a BiomeGPT-Style Microbiome Transformer
Interpretability
- Introduces signed Integrated Gradients as a solution to the limitations of [CLS] attention weights in BiomeGPT.
- Establishes a fusion-aware baseline that preserves species identity while isolating abundance effects.
- Demonstrates how the proposed method reveals directional relationships between microbial species and health outcomes.
- Recommends the use of second-order Integrated Hessians for understanding interactions among microbiome members.
Read more
Beyond Attention: Signed Integrated Gradients Attribution in a BiomeGPT-Style Microbiome Transformer
Summary
This paper addresses the limitations of using [CLS] attention weights for interpreting model predictions in feature-tokenized transformers, specifically in the context of BiomeGPT, a model designed for microbiome data classification. The authors highlight that [CLS] attention weights are nonnegative and cannot differentiate between contributions that support disease versus health classifications. To overcome these challenges, the paper introduces a signed, fusion-aware attribution method called Integrated Gradients (IG), which provides directional insights into how species abundance affects model outputs. The authors propose a new baseline for IG that preserves species identity while isolating the effects of abundance variation. This approach allows for a clearer understanding of the relationship between microbial species and health outcomes, revealing how certain species can be pathogenic or protective based on their abundance. The paper also suggests using second-order Integrated Hessians to analyze interactions among microbiome community members, further enhancing the interpretability of the model. Overall, this work provides a more principled framework for explainability in BiomeGPT and similar models, with implications for broader applications in biological data interpretation.
Methodology
The authors utilize Integrated Gradients (IG) to provide signed attributions for the IBD-health decision margin in BiomeGPT. They define a new baseline that maintains species identity while allowing for the analysis of abundance variations. The methodology includes empirical comparisons of IG with traditional [CLS] attention weights and the application of second-order Integrated Hessians to study community interactions.
Results
The application of signed Integrated Gradients revealed previously obscured relationships between species abundance and health outcomes, allowing for the identification of pathogenic versus protective microbial signals. The results demonstrated that IG outperformed [CLS] attention weights in providing meaningful insights into model predictions.
Implications
This work enhances the interpretability of microbiome models, which could lead to better understanding and treatment of diseases linked to microbial communities. The proposed methods can be applied to other biological data and models, improving explainability in various domains.
Hidden Gauge Controls Feature Specialization in ReLU Networks
Theory
Optimization
- Demonstrates a Ξ(DΒ²) separation in specialization times for functionally identical neurons under different hidden gauges.
- Establishes that a favorable gauge can deterministically assign feature ownership to one neuron while rendering others redundant.
- Introduces a mechanism that separates changes in functional coefficients from changes in feature direction.
- Validates results through population and finite-sample training, showing robustness to visible perturbations.
Read more
Hidden Gauge Controls Feature Specialization in ReLU Networks
Summary
This paper investigates the phenomenon of feature ownership in overparameterized ReLU networks, where multiple neurons may initially perform the same function but eventually specialize. The author introduces a Gaussian teacher-student model to explore how a hidden scaling gauge can influence which neuron becomes the owner of a task-relevant feature. By varying only the scaling gauge while keeping other parameters constant, the study demonstrates a significant separation in specialization times (Ξ(DΒ²)) between neurons with opposite gauges. The findings reveal that a favorable gauge can deterministically select one neuron as the owner, while the others become functionally redundant. The paper also establishes a mechanism that differentiates the mobilities of changing a feature's coefficient versus its direction, leading to global selection and functional pruning. The results are validated through various training scenarios, showing that the initial predictor does not dictate which neuron learns the feature or when it is learned. This work provides insights into the dynamics of feature specialization in neural networks and highlights the importance of parameter choices that may not be visible at initialization.
Methodology
The study employs a Gaussian teacher-student model to analyze the dynamics of feature specialization in ReLU networks. It varies a hidden scaling gauge while keeping the initial function and other parameters constant to observe the effects on feature ownership and specialization times. The analysis includes deriving an exact decomposition of functional changes and validating results through various training scenarios.
Results
The paper finds that opposite gauges lead to distinct feature trajectories and a sharp separation in specialization times. A favorable gauge can deterministically select one neuron as the owner of a feature, while redundant neurons lose their functional contributions. The results are consistent across different training conditions, confirming the robustness of the findings.
Implications
The findings suggest that the choice of parameters in neural network training can significantly influence feature specialization and ownership, which may have implications for designing more efficient neural architectures and training protocols. Understanding these dynamics could enhance the interpretability of neural networks and inform strategies for optimizing their performance.
EpiFlow: A framework for improving the utility of wastewater signals for disease forecasting
Time Series
- EpiFlow improves the forecasting accuracy of disease burden indicators using wastewater viral load signals.
- The framework incorporates advanced data preprocessing and signal analysis techniques to handle noise and variability in wastewater data.
- Forecasting models demonstrate a 20 percentage point improvement in forecast coverage during critical phases.
- The utility of wastewater signals is maintained even during low-prevalence periods and reporting delays.
Read more
EpiFlow: A framework for improving the utility of wastewater signals for disease forecasting
Summary
The paper presents EpiFlow, a novel framework designed to enhance the utility of wastewater viral load (WVL) signals for real-time disease forecasting, particularly in the context of COVID-19. Wastewater-based surveillance has emerged as a crucial tool for monitoring infectious diseases, providing insights into both symptomatic and asymptomatic populations. However, the inherent noise and variability in WVL data pose challenges for accurate forecasting of disease burden indicators such as hospital admissions. EpiFlow addresses these challenges through a systematic approach that includes data preprocessing, signal quality analysis, and dynamic relationship assessment between WVL and burden indicators. The framework employs permutation entropy to evaluate signal predictability and rolling-window Granger causality tests to analyze temporal dependencies. A time-varying vector autoregressive (VAR) model is utilized for probabilistic forecasting, incorporating the evolving relationship between wastewater signals and disease burden. The effectiveness of EpiFlow is demonstrated through simulations forecasting COVID-19 hospital admissions across Virginia, showing significant improvements in forecast accuracy, particularly during critical surge periods and even under conditions of reporting delays. The findings underscore the importance of WVL signals in enhancing infectious disease forecasting, especially in low-prevalence scenarios.
Methodology
EpiFlow consists of three main modules: data preprocessing (alignment and denoising of wastewater signals), signal analysis (using permutation entropy and rolling-window Granger causality), and forecasting (employing a time-varying vector autoregressive model). The framework evaluates the predictability of wastewater signals and their dynamic relationships with burden indicators to generate probabilistic forecasts.
Results
The application of EpiFlow to SARS-CoV-2 wastewater data and COVID-19 hospital admissions across Virginia demonstrated that incorporating wastewater-derived indicators significantly improved probabilistic forecasts, particularly during surge periods. The framework achieved a notable 20 percentage point increase in forecast coverage compared to baseline models.
Implications
The findings suggest that wastewater-based surveillance can play a critical role in enhancing real-time forecasting of infectious diseases, particularly in scenarios where traditional clinical data may be limited or unreliable. EpiFlow could be adapted for use in various public health contexts, potentially improving response strategies during outbreaks.
Surv-IPTB: An Attention-Based Model for Estimating Individual Probability of Treatment Benefit with Survival Data
Theory
- Surv-IPTB reformulates IPTB estimation as a binary classification problem, enhancing individual treatment benefit assessments.
- The model incorporates an attention mechanism to effectively aggregate pairwise patient comparisons and handle censored data.
- Extensive experiments show superior performance of Surv-IPTB over traditional meta-learner baselines in complex nonlinear scenarios.
- The approach provides a principled method for estimating treatment benefits tailored to individual patients, addressing limitations of average treatment effect assessments.
Read more
Surv-IPTB: An Attention-Based Model for Estimating Individual Probability of Treatment Benefit with Survival Data
Summary
This paper introduces Surv-IPTB, an innovative attention-based framework designed to estimate the Individual Probability of Treatment Benefit (IPTB) in survival analysis. The model aims to quantify the likelihood that a specific patient will experience extended survival time under treatment compared to control. By reformulating IPTB estimation as a binary classification problem, the authors leverage pairwise patient comparisons across treatment and control groups. The framework adeptly handles right-censored observations through imprecise probability representations, characterizing uncertain treatment effects with interval-valued probabilities. An attention mechanism with learnable query-key transformations facilitates flexible aggregation of pairwise comparisons while learning soft class probabilities for censored cases. Extensive experiments on synthetic datasets with complex nonlinear structures demonstrate that Surv-IPTB maintains robust performance across varying censoring rates and treatment effect strengths. The model consistently outperforms established meta-learner baselines, particularly in challenging nonlinear scenarios where traditional methods falter. The findings position Surv-IPTB as a scalable and statistically principled solution for personalized treatment benefit assessment in survival contexts, with the implementation code made publicly available.
Methodology
The authors recast the IPTB estimation as a binary classification problem, analyzing paired observations of potential survival times under treatment and control. They utilize attention mechanisms for flexible aggregation of data and employ interval-based probability distributions to accommodate uncertainties, particularly for censored observations. The methodology includes rigorous comparative analysis against various meta-learner baselines.
Results
The Surv-IPTB model demonstrated robust performance across diverse synthetic datasets with varying complexities, outperforming traditional methods like T-learner and S-learner combined with random survival forests, Cox proportional hazards, and Beran estimators. The model maintained superior accuracy even in scenarios with high censoring rates and complex treatment effects.
Implications
The proposed framework has significant implications for personalized medicine, allowing clinicians to better assess individual treatment benefits in survival analysis contexts. It addresses the limitations of traditional average treatment effect approaches, enabling more tailored and effective treatment decisions based on patient-specific data.
IFlowNets: Extending Generative Samplers to Learn Strategies in Incomplete Information Games
Reinforcement Learning
Generative Models
Theory
- IFlowNets generalize AFlowNets to handle incomplete information games effectively.
- The paper proves that existing generative flow network constraints are inadequate for incomplete information settings.
- IFlowNets maintain essential properties like flow matching, crucial for achieving valid player strategies.
- Preliminary results show IFlowNets outperform or match the performance of established methods in standard game environments.
Read more
IFlowNets: Extending Generative Samplers to Learn Strategies in Incomplete Information Games
Summary
This paper presents IFlowNets, an extension of generative flow networks (specifically Adversarial Flow Networks, AFlowNets) to address the challenges posed by incomplete information games. The authors highlight the inadequacies of existing generative sampling frameworks when applied to such games, particularly the failure of established constraints to yield valid densities and training objectives. Through theoretical proofs, they demonstrate that their proposed IFlowNets generalize AFlowNets while preserving essential properties like flow matching. The methodology involves aggregating flows over information states to ensure valid reward-proportional sampling. Preliminary experiments in three standard game environments indicate that IFlowNets perform comparably or better than existing methods such as Outcome Sampling Monte Carlo Counterfactual Regret (OS-MCCFR) and standard reinforcement learning approaches, both in terms of performance and computational speed.
Methodology
The authors extend the generative flow network framework by introducing Information Flow Networks (IFlowNets), which incorporate an infostate aggregation property to address the unique challenges of incomplete information games. They prove the inadequacy of existing constraints and propose new formulations to ensure valid reward-proportional sampling.
Results
IFlowNets were tested in three standard incomplete information game environments, showing performance comparable to or better than Outcome Sampling Monte Carlo Counterfactual Regret (OS-MCCFR) and standard reinforcement learning methods, demonstrating both efficiency and effectiveness.
Implications
The development of IFlowNets could significantly enhance the modeling and strategy learning in incomplete information games, with potential applications in various strategic decision-making scenarios, including economics, competitive gaming, and multi-agent systems.
PPDL: LLM-Based Flows as Probabilistic Programs
Large Language Models
NLP
Theory
- Introduction of PPDL, the first probabilistic programming language for LLM-based flows.
- Decoupling of inference scaling from core program logic, enhancing flexibility and usability.
- Formal semantics that clarify the interaction between prompt-based sampling and probabilistic factors.
- Empirical results demonstrating the effectiveness of PPDL with various inference engines.
Read more
PPDL: LLM-Based Flows as Probabilistic Programs
Summary
The paper addresses the challenges of building reliable applications that utilize large language models (LLMs), particularly the issues of uncertainty and lack of confidence in outputs when multiple LLM calls are involved. To tackle this, the authors introduce PPDL, a probabilistic programming language designed specifically for LLM-based flows. PPDL allows developers to quantify and propagate uncertainty throughout their applications without altering the core logic of their flows. The language integrates probabilistic constructs, enabling users to specify soft and hard constraints that influence the probability of execution traces. PPDL decouples inference scaling from program logic, allowing for experimentation with various probabilistic inference engines seamlessly. The paper includes an experimental study demonstrating PPDL's capabilities and a case study where it is applied to create a theorem proving agent for the Rocq theorem prover. The authors formalize the semantics of PPDL, showcasing its dual role as both a prompting and probabilistic language, and present empirical results that highlight its versatility across different benchmarks.
Methodology
The authors developed PPDL by extending a high-level prompting language with probabilistic constructs. They formalized its semantics and implemented a runtime that manages the exploration of multiple execution traces. The paper also details the integration of different probabilistic inference engines, such as majority voting and particle filtering, to assess output distributions.
Results
The empirical results indicate that PPDL is versatile and effective for experimenting with different inference scaling algorithms across multiple benchmarks. The case study demonstrated successful implementation of a theorem proving agent using PPDL, highlighting its practical applicability.
Implications
PPDL has the potential to enhance the reliability of applications that rely on LLMs by providing a structured way to manage uncertainty. It can be particularly useful for developers and researchers working on multi-step workflows involving LLMs, enabling them to create more trustworthy and interpretable applications.
QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding
NLP
Large Language Models
Efficient ML
- QEvict introduces a recoverable eviction strategy that allows for dynamic management of KV cache, addressing the limitations of traditional irreversible eviction methods.
- The method categorizes token windows into three tiers, enabling the retention of important contexts while maintaining a fixed memory budget.
- QEvict effectively reduces missed attention and improves information retention in long-context decoding tasks.
- The proposed diagnostics, Future Missed Mass and Global LIR, provide insights into the importance of cached states over time.
Read more
QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding
Summary
The paper introduces QEvict, a novel three-tier Key-Value (KV) cache management scheme designed to enhance the performance of autoregressive large language models (LLMs) during long-context decoding. Traditional eviction policies in KV caches make irreversible decisions, discarding tokens that may later gain importance as the decoding process evolves. This paper identifies the limitations of such policies, particularly the issue of 'attention drift' where the relevance of tokens fluctuates over time. To address this, QEvict employs a recoverable eviction strategy that categorizes windows of tokens into three tiers: high-confidence windows are retained in full precision, intermediate windows are stored in a quantized recoverable format, and the least important windows are evicted. This dynamic approach allows for the reactivation of previously evicted windows when their importance increases, thereby preserving contextual coherence and improving information retention. The authors validate QEvict against existing methods across various long-context benchmarks, demonstrating its superior ability to reduce missed attention and enhance overall performance.
Methodology
QEvict employs a three-tier KV-cache hierarchy that ranks historical windows based on cumulative attention scores. High-importance windows are stored in full precision, while intermediate windows are quantized for recoverability. The method allows for dynamic transitions between these tiers, promoting windows back to full precision when their importance re-emerges, thus facilitating a more flexible and context-aware cache management system.
Results
The experimental results show that QEvict consistently outperforms existing eviction and quantization baselines across long-context understanding, retrieval, and reasoning benchmarks. It significantly reduces missed attention and enhances the retention of relevant information, demonstrating its effectiveness in managing KV caches for LLMs.
Implications
QEvict's approach to recoverable eviction could lead to more efficient memory management in large language models, potentially lowering inference costs and improving performance in applications requiring long-context understanding, such as dialogue systems, document summarization, and complex reasoning tasks.
When Does Consensus Mean Correctness? Measuring the Agreement-Accuracy Coupling with Semantics-Preserving Re-Rendering
Computer Vision
Large Language Models
Theory
- Introduces RENDEQ, a tool for generating semantically equivalent renderings for accurate measurement of model correctness.
- Demonstrates that re-rendering is superior to resampling in assessing model accuracy and reliability.
- Finds that model agreement does not always correlate with correctness, especially when errors are diffuse.
- Identifies that fine-tuning on consensus can lead to decreased accuracy, contrary to existing literature.
Read more
When Does Consensus Mean Correctness? Measuring the Agreement-Accuracy Coupling with Semantics-Preserving Re-Rendering
Summary
This paper investigates the relationship between model agreement across perturbed inputs and the accuracy of those models, challenging the assumption that agreement is a reliable indicator of correctness. The authors introduce RENDEQ, a generator that creates render-equivalence sets from scientific figures, allowing for precise measurement of correctness and concentration without confounding factors. By applying this methodology to three vision-language models (VLMs), the authors demonstrate that re-rendering outperforms resampling in terms of both accuracy and reliability. They find that agreement surpasses a baseline in two out of three models, while also revealing that the dispersion of errors is significantly influenced by the plotting library used. Furthermore, fine-tuning based on cross-render consensus leads to decreased accuracy, contradicting previous findings on natural images. The study concludes that the coupling between agreement and correctness is conditional, depending on the diffusion of model errors, and that an objective rewarding agreement can exacerbate error concentration.
Methodology
The authors developed RENDEQ to create render-equivalence sets from scientific figures, allowing for the measurement of correctness and concentration without confounding variables. They applied this methodology to three open-weight vision-language models, measuring agreement and accuracy across multiple independent instantiations.
Results
The study found that re-rendering outperformed resampling in both accuracy and reliability. Agreement metrics surpassed a baseline in two models and tied in the third. The analysis revealed that error dispersion was primarily influenced by the plotting library used, and fine-tuning based on cross-render consensus resulted in reduced accuracy across multiple runs.
Implications
The findings suggest that relying solely on model agreement as a measure of correctness can be misleading, particularly in cases of correlated errors. This has implications for self-training methods in machine learning, indicating a need for careful consideration of error diffusion when designing training objectives.
Matrix Zonotopic Attention: A Context-Adaptive Value Projection for Set Transformers
Theory
Efficient ML
Optimization
- Introduces Matrix Zonotopic Attention (MZAttn) for improved set transformer performance.
- Defines Transformation Degrees of Freedom (TDOF) to analyze the complexity of target operators.
- Demonstrates that MZAttn can represent complex targets with fewer layers compared to traditional attention mechanisms.
- Experimental results indicate significant performance improvements on high-complexity tasks.
Read more
Matrix Zonotopic Attention: A Context-Adaptive Value Projection for Set Transformers
Summary
This paper introduces Matrix Zonotopic Attention (MZAttn), a novel attention mechanism designed for set transformers that addresses the limitations of traditional multi-head attention. The authors analyze the asymmetry in multi-head attention, where the routing is input-dependent while the value projection is not, leading to inefficiencies in representing permutation-invariant set targets. They introduce the concept of Transformation Degrees of Freedom (TDOF) to measure the complexity of target operators and demonstrate that context-rigid attention requires depth proportional to TDOF, while MZAttn can represent the same targets with a single layer. MZAttn employs a context-adaptive matrix-zonotope family for value projection, allowing for a more flexible representation that adapts to the input set. The experimental results show that MZAttn outperforms standard attention mechanisms on tasks requiring high-rank, sparsely combinatorial representations, while maintaining competitive performance on simpler tasks. The findings suggest that the architectural advantages of MZAttn are selective and depend on the nature of the target operator.
Methodology
The authors propose MZAttn, which replaces the fixed value projection in standard multi-head attention with a context-adaptive matrix-zonotope family. This involves a center matrix and generator matrices weighted by input-dependent gates. The methodology includes a depth-separation analysis and empirical validation through experiments on various set-prediction tasks.
Results
MZAttn significantly outperformed standard attention mechanisms on tasks requiring high-rank, sparsely combinatorial representations, achieving R2 = 0.69 on the minimum enclosing ball radius task, while traditional methods remained below R2 = 0.025. The results validate the theoretical predictions regarding TDOF and the selective advantages of the proposed architecture.
Implications
The findings suggest that MZAttn could be applied to various domains requiring set-based representations, such as point-cloud analysis, molecular modeling, and in-context learning, potentially leading to more efficient and effective models in these areas.
Online Security Learning in Cooperative Multi-Agent Systems under Hidden Byzantine Attacks
Theory
Reinforcement Learning
Robotics
- Introduces a Byzantine team MDP framework to model multi-agent systems under hidden Byzantine attacks.
- Establishes the concept of security regret, decomposing it into return regret and response gap.
- Demonstrates that public feedback cannot certify security against unrestricted attacks.
- Develops a robust estimation-to-decisions learner with a proven regret bound.
Read more
Online Security Learning in Cooperative Multi-Agent Systems under Hidden Byzantine Attacks
Summary
This paper addresses the challenge of online cooperative control in multi-agent systems (MAS) that are vulnerable to Byzantine attacks, where a subset of agents can stealthily alter their actions after observing the planned joint action. The authors propose a Byzantine team Markov decision process (MDP) framework to analyze the impact of such attacks on team performance. They establish that the attacker's information significantly influences the decision-making process, leading to different geometric representations of the MDP based on whether the attacker observes the planned actions. The paper introduces the concept of security regret, which quantifies the performance gap between the optimal worst-case return and the actual policy's return under attack. The authors demonstrate that public feedback alone cannot guarantee security against all possible attacks, highlighting the necessity of understanding the response gap. They develop a robust estimation-to-decisions learner that achieves a regret bound, providing theoretical and algorithmic foundations for reliable MAS under Byzantine threats. The findings emphasize the importance of robustness in cooperative systems, particularly in adversarial environments.
Methodology
The authors utilize a finite-horizon cooperative team MDP approach to model the interactions among agents in the presence of Byzantine attacks. They analyze the impact of the attacker's information on the decision problem and establish a framework for security regret. The proposed learner is based on robust estimation-to-decisions (E2D) principles, incorporating stage-tied estimators to improve performance metrics.
Results
The paper proves that the expected security regret can be bounded and shows that the proposed learner achieves an expected return regret of eO(H2SβAK) over K episodes, where A represents joint actions. The findings indicate that the response gap is a necessary component for understanding security in the context of public feedback.
Implications
The results have significant implications for the design of robust multi-agent systems, particularly in applications where agents must operate in adversarial environments, such as autonomous vehicles, coordinated energy systems, and other collaborative frameworks. The insights gained can inform the development of more secure and resilient cooperative strategies.
Multi-Level Modeling of Large Language Model Inference Latency and Energy via Hybrid Analytical--Machine-Learning Predictors
Large Language Models
Efficient ML
Optimization
- Introduction of HYMELL, a hybrid framework for estimating LLM inference latency and energy.
- Three-level modeling approach combining analytical and machine learning techniques.
- High predictive accuracy achieved, with less than 5% error for LLaMA 3 8B model.
- Framework supports diverse architectures, enhancing its applicability.
Read more
Multi-Level Modeling of Large Language Model Inference Latency and Energy via Hybrid Analytical--Machine-Learning Predictors
Summary
The paper addresses the increasing computational cost, energy consumption, and inference latency associated with the rapid scaling of Large Language Models (LLMs). The authors introduce a novel framework called Hybrid Modeling for Energy and Latency of LLMs (HYMELL), which combines analytical modeling with machine learning to estimate LLM inference latency and energy consumption. HYMELL operates on a three-level hierarchy: (1) analytical estimation of primitive operations, (2) machine learning prediction of higher-level components, and (3) an end-to-end model that captures system-level overheads during both the prefill and decode phases of LLM execution. The framework is versatile, accommodating various architectures, including dense and mixture-of-experts (MoE) feedforward networks, as well as multi-head attention (MHA) and grouped-query attention (GQA) mechanisms. Evaluations on an NVIDIA H100 GPU demonstrate that HYMELL achieves high predictive accuracy, with less than 5% error for both prefill and decode phases for the LLaMA 3 8B model. This approach enables rapid, hardware-free design space exploration and energy-efficient optimization, making it a significant contribution to the field of sustainable AI deployment.
Methodology
The HYMELL framework employs a three-level hierarchy for modeling LLM inference: (1) operator-level modeling using analytical estimators for primitive operations, (2) block-level modeling where ML predicts costs of block components, and (3) system-level modeling that aggregates these predictions to estimate overall latency and energy consumption.
Results
HYMELL demonstrated high predictive accuracy on an NVIDIA H100 GPU, achieving less than 5% error in estimating both prefill and decode phases for the LLaMA 3 8B model, indicating its effectiveness in modeling LLM inference.
Implications
The HYMELL framework can significantly aid in the sustainable deployment of LLMs by providing accurate estimations of energy and latency, facilitating hardware-aware design decisions, and optimizing energy efficiency in various applications of AI.
CertBind from Multimodal Connectivity to Certifiable Retrieval Decisions
Multimodal
Theory
Graph Learning
- CertBind enables certifiable composition of multimodal connectors, enhancing decision-making capabilities.
- The framework operates at multiple scales, addressing ambiguities in retrieval outputs.
- Different deployment actions are defined based on the support of retrieval routes.
- The methodology includes contract-aware conformal ranks and a majority path-transversal budget for robustness.
Read more
CertBind from Multimodal Connectivity to Certifiable Retrieval Decisions
Summary
The paper introduces CertBind, a framework designed to enhance the composability of frozen multimodal encoders by ensuring certifiable retrieval decisions. CertBind addresses the challenges posed by the deployment of multimodal connectors, which can alter established retrieval capabilities. The authors propose a multiscale theory that encompasses various levels of decision-making: node scale for task identification, edge scale for error control, path scale for recovery budgeting, and query scale for generating a top-k candidate set. The framework allows for different deployment actions based on the support of routes, categorizing them as Direct, Certified, or Abstain. The results demonstrate that while the C-MCR shared route reduced the native CLIP retrieval accuracy, the production fallback achieved a high recovery rate, indicating the effectiveness of CertBind in maintaining retrieval integrity across multimodal tasks.
Methodology
CertBind employs a multiscale approach to certifiable composition, analyzing graph structures at the node, edge, path, and query levels. It utilizes contract-aware conformal ranks for error control and a majority path-transversal budget to manage shared failures. The framework categorizes retrieval outputs based on their support and certifiability, allowing for a structured decision-making process.
Results
The evaluation of CertBind showed that the C-MCR shared route reduced the native CLIP retrieval accuracy from 0.524 to 0.290. However, the production fallback mechanism recovered 0.963 Β± 0.002 of clean retrieval, while the passing branch maintained a no-harm value of 1.000, demonstrating the framework's effectiveness in preserving retrieval integrity.
Implications
CertBind has the potential to improve the deployment of multimodal systems by ensuring reliable retrieval decisions, which is crucial for applications in areas such as image and text retrieval, automated reasoning, and interactive AI systems. Its structured approach to handling multimodal data can lead to more robust and interpretable AI models.
When GNNs Fail: Quantifying and Overcoming Temporal Correlation Volatility in Time Series
Graph Learning
Time Series
- Introduces Temporal Correlation Volatility (TCV) as a metric for quantifying instability in time series correlations.
- Demonstrates that existing GNN models struggle in high-TCV environments, leading to significant performance degradation.
- Proposes GLIDE, a novel GNN layer that effectively addresses dynamic graph structures through innovative design mechanisms.
- Shows that GLIDE outperforms existing models by up to 45.6% on average, with gains reaching 85.7% in specific scenarios.
Read more
When GNNs Fail: Quantifying and Overcoming Temporal Correlation Volatility in Time Series
Summary
This paper investigates the limitations of Graph Neural Networks (GNNs) in forecasting multivariate time series, particularly under conditions of dynamic temporal correlations. The authors introduce a new metric, Temporal Correlation Volatility (TCV), to quantify the instability of pairwise dependencies in time series data. They demonstrate that many existing models, including popular architectures like Transformers, perform poorly in high-TCV environments, often being outperformed by simpler models. To address these challenges, the authors propose a novel GNN layer called Graph Layer for Inference in Dynamic Environments (GLIDE), which incorporates two key mechanisms: Path-based Message Passing and Static and Dynamic Propagation Separation. These enhancements allow GLIDE to better capture temporal variability in graph structures while maintaining robustness in static scenarios. Extensive experiments reveal that GLIDE significantly improves forecasting performance across both static and dynamic settings, achieving performance gains of up to 85.7%.
Methodology
The authors conducted a theoretical analysis of correlation matrices under dynamic conditions and introduced the TCV metric to measure fluctuations in dependencies. They empirically evaluated 18 baseline models against TCV to assess performance degradation. The proposed GLIDE layer was developed with two main components: Path-based Message Passing and Static and Dynamic Propagation Separation, which were tested in extensive experiments on both synthetic and real-world datasets.
Results
GLIDE demonstrated significant improvements in forecasting accuracy, with average performance enhancements of up to 45.6% across various settings. In specific cases, the performance gains reached as high as 85.7%, indicating a substantial advantage over existing models in dynamic environments.
Implications
The findings suggest that GNNs need to adapt to dynamic temporal correlations for effective multivariate time series forecasting. The introduction of TCV provides a valuable diagnostic tool for assessing model performance in varying conditions, which could influence future research and applications in fields such as finance, energy management, and weather prediction.
Evaluating Machine Learning Models for Post-Wildfire Debris-Flow Prediction
Interpretability
- TabPFN outperforms other models in predicting post-wildfire debris flows.
- Short-duration rainfall intensity and storm accumulation are the most important features for prediction.
- Synthetic data augmentation significantly improves model performance.
- The study provides a systematic evaluation framework for machine learning models in hazard prediction.
Read more
Evaluating Machine Learning Models for Post-Wildfire Debris-Flow Prediction
Summary
This paper addresses the critical challenge of predicting post-wildfire debris flows, which pose significant risks to communities and infrastructure during intense rainfall in recently burned areas. The authors systematically evaluate a broad range of machine learning models, including the newly introduced Tabular Prior-Data Fitted Network (TabPFN), to determine their performance, feature importance, and response to synthetic data augmentation. Using basin-scale observations from the western United States, the study compares 15 different models. The results from repeated stratified cross-validation reveal that TabPFN achieves the highest unaugmented performance with a threat score of 0.637, closely followed by leading tree-based models. Feature importance analysis using SHAP indicates that short-duration rainfall intensity and storm accumulation are the most critical features for predicting debris flows, while burn severity and terrain features are less significant. The paper also explores the utility of synthetic data augmentation, demonstrating that it enhances the performance of all models except for CNN, with the greatest improvement seen in deep learning models. This comprehensive evaluation provides a framework for improving the accuracy and reliability of debris-flow predictions, which is essential for effective hazard mitigation.
Methodology
The authors conducted a systematic evaluation of 15 machine learning models, including the TabPFN, using basin-scale observations of post-wildfire debris flows. They employed repeated stratified cross-validation to assess model performance and used SHAP for feature importance evaluation. Additionally, they explored synthetic data augmentation to address the scarcity of training data.
Results
The TabPFN model achieved the highest threat score of 0.637 in unaugmented performance. Feature importance analysis revealed that rainfall intensity and storm accumulation were the most influential features. Synthetic data augmentation improved the performance of all models except CNN, with the largest mean gain in threat score observed among deep learning models.
Implications
The findings of this study have significant implications for disaster risk management and hazard mitigation strategies in post-wildfire scenarios. By enhancing the accuracy of debris-flow predictions, communities can better prepare for and respond to potential hazards, ultimately saving lives and protecting infrastructure.
Quantization Damage Is Multiplicative, Not Additive
NLP
Large Language Models
Theory
- Quantization damage is characterized as a multiplicative loss of decision margin rather than additive noise.
- The concept of 'margin shrinkage' explains how quantization affects model decisions, particularly in safety-critical applications.
- A predictive model for decision flip probabilities is developed, showing high accuracy without relying on fitted flip data.
- The findings challenge existing strategies that focus on protecting specific weights during quantization.
Read more
Quantization Damage Is Multiplicative, Not Additive
Summary
This paper investigates the effects of quantization on large language models, specifically challenging the prevailing assumption that quantization damage is additive. The authors introduce the concept of 'margin shrinkage,' demonstrating that quantization reduces the decision margin multiplicatively rather than adding fixed noise. Through extensive experiments involving 16 models across 8 families and various quantization methods, they measure the decision margins before and after quantization. The findings reveal that the margin shrinks as the bit-width decreases, leading to a higher likelihood of decision failures. The authors propose a predictive model for flip probabilities based on this multiplicative damage, achieving accurate predictions on held-out decisions. This work has significant implications for understanding model robustness and guiding future quantization strategies.
Methodology
The authors conducted a series of experiments measuring decision margins for binary choices across various models and quantization levels. They compared the margins before and after quantization to establish the relationship between margin shrinkage and bit-width, using statistical methods to fit and validate their predictive model for decision flips.
Results
The study found that quantization leads to a multiplicative reduction in decision margins, with median survival fractions of 0.86 at 4 bits, 0.33 at 3 bits, and 0.00 at 2 bits. The predictive model for flip probabilities was accurate within a median of 1.8 percentage points on held-out decisions, demonstrating the effectiveness of the proposed framework.
Implications
The findings suggest that current strategies for mitigating quantization damage may be misguided, as protecting specific weights does not yield better outcomes. This research could inform the development of more effective quantization techniques and enhance the robustness of large language models in practical applications.
Threshold-Based Early Stopping of Accumulations in Neural Networks with Binary Activation
Efficient ML
- Introduces a post-training early-stopping mechanism for binary neural networks.
- Demonstrates significant reductions in computational operations without retraining model parameters.
- Achieves 86.6% reduction in accumulation terms with minimal accuracy drop.
- Focuses on the efficiency of AI in constrained environments, addressing both accuracy and resource usage.
Read more
Threshold-Based Early Stopping of Accumulations in Neural Networks with Binary Activation
Summary
This paper addresses the inefficiencies in binary neural networks, particularly in the context of binary activations where the accumulation of inputs can lead to unnecessary computations. The authors propose a post-training early-stopping mechanism that leverages the predictable nature of running partial sums to determine the final output sign before all contributions are evaluated. By monitoring these accumulations, the method allows for the skipping of terms that do not affect the final output, thus reducing computational overhead. The approach is evaluated on the VGG11 architecture using the CIFAR-10 dataset, demonstrating significant reductions in arithmetic operations while maintaining acceptable accuracy levels. The study emphasizes the importance of efficiency in AI, particularly in resource-constrained environments, and presents a novel algorithmic solution that is hardware-agnostic.
Methodology
The authors characterize the behavior of running accumulations in binary-activation layers and develop a mechanism to predict the final output sign based on partial sums. They implement threshold calibration and checkpoint scheduling to optimize when to stop computations. The method is evaluated on VGG11 with CIFAR-10, measuring arithmetic savings relative to dense accumulation.
Results
The proposed method achieved an 86.6% reduction in accumulation terms for the deepest convolutional layer with only a 0.37-point drop in accuracy. When applied to the three deepest convolutions simultaneously, it resulted in a 25% reduction in full-network arithmetic with a 1.36-point accuracy drop.
Implications
This work has significant implications for deploying binary neural networks in resource-constrained environments, such as mobile devices and edge computing, where efficiency is critical. The proposed early-stopping mechanism can enhance the performance of AI systems by reducing computational costs while maintaining acceptable accuracy levels.
Every Cache Entry Earns Its Place: Global Allocation of Resolution and Coverage for KV Cache Compression
Large Language Models
Optimization
Efficient ML
- Introduces GraceKV, a global approach for KV cache compression that optimally allocates resources across layers and context slots.
- Utilizes a prototype tree structure to represent KV entries, allowing for adaptive balancing of resolution and coverage.
- Demonstrates superior performance compared to existing methods, achieving high compression ratios without additional training.
- Validates the approach through systematic experiments across diverse long-context tasks, showing effectiveness in resource allocation.
Read more
Every Cache Entry Earns Its Place: Global Allocation of Resolution and Coverage for KV Cache Compression
Summary
The paper addresses the bottleneck in Key-Value (KV) cache storage and access in large language models (LLMs) as they process increasingly long contexts. Existing KV cache compression methods are limited by predefined rules and local allocation strategies that do not allow for efficient resource distribution across layers and context slots. To overcome these limitations, the authors propose GraceKV, a global approach that formulates the KV cache compression process as a global resource allocation problem under a fixed budget. GraceKV constructs a prototype tree for each layer-KV head-slot combination, where each node in the tree represents a different level of resolution and coverage. The method allows for adaptive allocation of cache resources, balancing local resolution and information coverage. The authors demonstrate that GraceKV does not require additional training and can be executed entirely on GPUs. Experimental results show that GraceKV outperforms existing methods, achieving first place in 24 out of 32 settings and maintaining robustness up to 128Γ compression, validating the effectiveness of global budget allocation.
Methodology
GraceKV constructs a prototype tree for each atomic unit defined by layer-KV head-slot combinations. It employs two operations, Add and Split, to manage information coverage and local resolution. The method uses a bottom-up value flow to aggregate contextual values and a utility-guided budget flow to prioritize operations based on their marginal utility, enabling global resource allocation.
Results
GraceKV ranks first in 24 out of 32 experimental settings and demonstrates robustness at compression ratios up to 128Γ. The results confirm the effectiveness of the global budget allocation strategy in balancing resolution and coverage.
Implications
The proposed method can significantly enhance the efficiency of long-context inference in LLMs, making it applicable in various domains requiring extensive context processing, such as long-document understanding and complex reasoning tasks.
EvoHarness-RL: Learning Self-Evolving Runtime Harness for Long-Horizon LLM Agents
Large Language Models
Reinforcement Learning
Robotics
- EvoHarness-RL introduces a self-evolving runtime harness for long-horizon LLM agents.
- The framework abstracts harness components into a unified Belief, Progress, and Experience (BPE) state.
- A two-stage training process enhances the agent's ability to construct and utilize external state effectively.
- The approach significantly improves task success rates and efficiency in long-horizon interactions.
Read more
EvoHarness-RL: Learning Self-Evolving Runtime Harness for Long-Horizon LLM Agents
Summary
The paper presents EvoHarness-RL, a novel framework designed to enhance long-horizon LLM agents' capabilities by enabling them to learn and adapt their runtime harness policies. Long-horizon tasks require agents to maintain state, track progress, and utilize external tools effectively. Traditional methods rely heavily on manual engineering of prompts and heuristics, which limits the agents' ability to dynamically adapt to their environments. EvoHarness-RL addresses this by introducing a unified policy-facing harness state, encapsulated in three components: Belief (environment state), Progress (execution status), and Experience (reusable knowledge). The framework employs a two-stage training process: first, supervised fine-tuning to familiarize the agent with the harness action space, followed by cost-aware GRPO to optimize the agent's decision-making regarding when to access and update the external state. The implementation on ALFWorld with a Qwen3-8B LLM demonstrates that EvoHarness-RL achieves a success rate of 96.9% on seen tasks and 86.6% on unseen tasks, showcasing significant improvements in task execution efficiency and adaptability. The study reveals two critical dynamics: harness annealing, where the agent learns to use the harness more selectively, and harness evolution, where the agent refines its external state management through ongoing interactions.
Methodology
The methodology consists of a two-stage training process. The first stage involves supervised harness fine-tuning, which teaches the agent the semantics of the BPE action protocol and how to externalize useful states from interaction traces. The second stage employs cost-aware GRPO to optimize the agent's policy, focusing on when to read, update, or consolidate harness state while considering the interaction budget.
Results
EvoHarness-RL achieved a success rate of 96.9% on the seen split and 86.6% on the unseen split of the ALFWorld environment. The analysis revealed two dynamics: harness annealing, where the agent shifts from frequent to selective harness calls, and harness evolution, where the agent refines its external state management through ongoing updates.
Implications
The findings suggest that long-horizon LLM agents can significantly benefit from trainable policies for managing external harness workspaces, leading to more efficient and adaptable agents in complex interactive environments. This has potential applications in various fields such as robotics, web navigation, and workflow automation.
How Molecular Generative Models Organize Molecular Identity
Generative Models
Graph Learning
Theory
- Molecular generative models exhibit a structured internal organization of chemical identities.
- The arrangement of identities is influenced by representation type, identity conventions, and stochasticity in decoding.
- Local chemical cohesiveness stabilizes during training, but the number of distinct identities in neighborhoods changes.
- Understanding the internal organization is crucial for treating generative spaces as chemically navigable.
Read more
How Molecular Generative Models Organize Molecular Identity
Summary
This paper investigates how molecular generative models organize and represent molecular identities within their latent spaces. The authors emphasize the importance of understanding the internal arrangement of discrete chemical identities produced by generative models, which is often overlooked in evaluations focused on output validity and diversity. By making molecular identity explicit and analyzing the generative process, the authors reveal that the internal organization of molecular identities is structured into piecewise-constant regions, influenced by factors such as representation type, identity conventions, and stochastic decoding processes. The study employs three different molecular generative architecturesβMolMiner, HierVAE, and GDSSβto explore these properties. The findings indicate that while local chemical organization stabilizes during training, the number of distinct molecular identities in each neighborhood evolves. This highlights the need for a thorough characterization of internal organization before considering a generative space as navigable for chemical exploration.
Methodology
The authors analyze molecular generative models by making molecular identity explicit as an equivalence relation on output representation spaces. They pull back these identities through the generative process and examine the resulting partitions across three architectures: MolMiner, HierVAE, and GDSS. The study employs various identity conventions to assess the structural properties of the generated representations.
Results
The analysis reveals that the internal organization of molecular identities is arranged into piecewise-constant regions, with local neighborhoods exhibiting chemical cohesiveness. The study finds that while some models maintain identity persistence under varying stochastic conditions, others show less stability. The results underscore the importance of characterizing the internal organization of generative models to understand their navigability in chemical space.
Implications
The findings have significant implications for the design and evaluation of molecular generative models, suggesting that a deeper understanding of their internal organization can enhance their utility in chemical discovery and optimization tasks. This work could inform future research on improving the navigability of generative spaces for molecular design.
Fluid-DiT: Graph-Free Diffusion Transformers for Fluid Flow Simulations Learning
Generative Models
Graph Learning
Efficient ML
- Fluid-DiT is the first graph-free diffusion transformer for fluid flow simulations, enhancing local precision and long-range coupling.
- The model employs a latent-space diffusion formulation that improves training efficiency and reduces high-frequency artifacts.
- Fluid-DiT consistently outperforms state-of-the-art graph-based models in distributional fidelity and scalability.
- The framework generalizes effectively across diverse geometries and Reynolds numbers without requiring mesh-specific tuning.
Read more
Fluid-DiT: Graph-Free Diffusion Transformers for Fluid Flow Simulations Learning
Summary
Fluid-DiT introduces a novel approach to simulating complex fluid flows by utilizing a Graph-Free Diffusion Transformer, which replaces traditional graph-based message passing with attention-based denoising. This innovation allows for the modeling of equilibrium distributions without the constraints of handcrafted graph architectures, enabling better scalability and generalization across various geometries and Reynolds numbers. The framework employs a latent-space formulation that separates geometric fidelity from distributional learning, effectively reducing high-frequency artifacts and enhancing sampling efficiency. Experimental results demonstrate that Fluid-DiT outperforms existing graph-based diffusion models in terms of sample quality and distributional accuracy across several benchmarks, including laminar cylinder wakes and turbulent wing experiments. The model shows robust performance even with incomplete trajectory data, indicating its potential for broader applications in computational physics and fluid dynamics.
Methodology
Fluid-DiT utilizes transformer attention layers to couple nodes in fluid flow simulations directly, eliminating the need for graph-based message passing. It introduces a latent-space representation that disentangles geometric details from distributional learning, allowing for efficient sampling and reduced noise amplification.
Results
Fluid-DiT demonstrated superior performance on canonical benchmarks, achieving higher R2 correlations and lower Wasserstein distances compared to graph-based diffusion models. The model's ability to generalize from short, incomplete trajectories to unseen conditions was also validated, showcasing its robustness and efficiency.
Implications
The advancements presented by Fluid-DiT have significant implications for computational fluid dynamics, enabling faster and more accurate simulations in various applications such as aerodynamics, weather forecasting, and biomedical flows. Its scalability and generalization capabilities could accelerate design processes and scientific discovery in fluid dynamics.
Does Latent Context Help? A Controlled Evaluation of Inverse Reinforcement Learning in Arctic Shipping
Reinforcement Learning
Robotics
Theory
- Latent context in IRL does not necessarily improve performance and may reduce it in certain scenarios.
- Observable route and environmental conditions explain most behavioral variations in Arctic shipping.
- Nonlinear reward models significantly outperform linear models in predicting vessel behavior.
- A context-need diagnostic is proposed to evaluate the necessity of latent context in decision-making.
Read more
Does Latent Context Help? A Controlled Evaluation of Inverse Reinforcement Learning in Arctic Shipping
Summary
This paper investigates the effectiveness of latent context in Inverse Reinforcement Learning (IRL) for Arctic shipping navigation. The authors highlight the challenges posed by rapidly changing sea-ice conditions and the need for robust, interpretable reward models. They conduct a controlled evaluation using 3,186 voyages from 202 vessels over nine Arctic shipping seasons, comparing three reward models: a linear shared reward model (MCE-IRL), a nonlinear shared reward model (AIRL), and a latent-context model (PEMIRL). The study reveals that while the nonlinear reward model significantly outperforms the linear model, the addition of vessel-specific latent context actually reduces performance. The authors argue that the observed behavioral variations can largely be explained by observable route and environmental factors rather than hidden preferences. They also introduce a context-need diagnostic to assess when latent context may be beneficial. The findings emphasize the importance of careful evaluation of entity-specific adaptive AI systems, particularly in safety-critical applications like maritime navigation.
Methodology
The authors employed a controlled empirical evaluation framework, analyzing 3,186 AIS-derived voyages using three different reward models: MCE-IRL for linear rewards, AIRL for nonlinear rewards, and PEMIRL for nonlinear rewards with latent context. They utilized behavioral analysis, context probes, and feature-hiding ablation studies to assess the contributions of each model.
Results
The nonlinear shared reward model (AIRL) improved held-out likelihood by 50.9% compared to the linear model (MCE-IRL). However, the introduction of vessel-specific latent context (PEMIRL) resulted in a 16.5% reduction in performance. The analysis indicated that most behavioral variations were attributable to observable factors rather than hidden preferences.
Implications
The findings suggest that AI systems in safety-critical applications, such as maritime navigation, should prioritize observable features over latent context to enhance reliability and interpretability. This could lead to more effective AI-assisted navigation systems that adapt to changing environmental conditions without unnecessary complexity.
SNI-GNN: SmartNIC-Assisted Full-Graph GNN Training with In-Network Embedding Prediction
Graph Learning
Efficient ML
Theory
- Introduction of SNI-GNN, a SmartNIC-assisted system for full-graph GNN training.
- Reduction of inter-node communication by 21-45% through in-network embedding prediction.
- Achieved end-to-end speedups of 1.3-3.6x over existing full-graph training systems.
- Maintained accuracy loss within 0.01 while scaling to 16 GPUs on large graphs.
Read more
SNI-GNN: SmartNIC-Assisted Full-Graph GNN Training with In-Network Embedding Prediction
Summary
The paper presents SNI-GNN, a novel SmartNIC-assisted system designed to enhance the training of full-graph Graph Neural Networks (GNNs) in multi-server clusters. Full-graph training is known for its high accuracy but suffers from significant scalability issues due to the heavy communication overhead associated with embedding exchanges between nodes. SNI-GNN addresses this challenge by implementing in-network embedding prediction, which reduces the need for frequent inter-node communication. The system utilizes a lightweight linear-trend predictor deployed on SmartNICs to refine cached historical embeddings, thereby mitigating staleness without introducing new synchronization requirements. Additionally, SNI-GNN employs an importance-based boundary-node sampling policy and an asynchronous DPU-GPU data pipeline to optimize data transfer and computation overlap. Theoretical bounds are provided to demonstrate that the predictor's bias is controlled, ensuring convergence despite the use of inexact gradients. The implementation on NVIDIA BlueField-3 shows that SNI-GNN can significantly cut communication by 21-45% and achieve end-to-end speedups of 1.3-3.6x over existing systems while maintaining accuracy loss below 0.01. This work highlights the potential of SmartNICs in enhancing communication efficiency for large-scale GNN training.
Methodology
SNI-GNN employs a SmartNIC-based architecture that integrates a lightweight linear-trend predictor for in-network embedding prediction, boundary-node importance sampling for efficient neighbor selection, and an asynchronous DPU-GPU pipeline to optimize data transfer and computation. The system is designed to operate on NVIDIA BlueField-3 DPUs, allowing for active staleness mitigation directly on the data path.
Results
The implementation of SNI-GNN demonstrated a reduction in inter-node communication volume by 21-45%, achieving end-to-end speedups of 1.3-3.6x compared to BNS-GCN and up to 1.29x over the SANCUS baseline. The system was tested on large datasets with tens of millions of nodes and maintained model accuracy within 1% of ideal synchronous training.
Implications
SNI-GNN's approach to leveraging SmartNICs for in-network prediction has significant implications for the scalability and efficiency of GNN training in distributed environments, making it a practical solution for handling large-scale graph data in various applications such as social networks, recommendation systems, and molecular biology.