AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
57
Papers today
8h
Update frequency
7
Days of history
CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation
Multimodal
Time Series
- Introduces a unified cardiac representation model leveraging ECG, PPG, and PCG data.
- Employs a joint-embedding predictive architecture to learn shared latent cardiac states.
- Incorporates a delay-aware alignment strategy to handle temporal offsets between modalities.
- Demonstrates superior performance across multiple downstream tasks compared to modality-specific models.
Read more
CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation
Summary
The paper introduces CardioState-JEPA, a novel cardiac foundation model designed to learn a unified representation of cardiac activity from three complementary sensing modalities: electrocardiography (ECG), photoplethysmography (PPG), and phonocardiography (PCG). Traditional models have focused on individual modalities, neglecting the shared physiological insights that can be gained from a cross-modal approach. CardioState-JEPA employs a physiology-aware joint-embedding predictive architecture that maps heterogeneous waveforms into a common token space and utilizes a single shared Transformer encoder. The model is trained to predict masked latent cardiac states, emphasizing shared physiological characteristics rather than modality-specific features. To address the temporal misalignment of the different modalities, a learned delay aligner is implemented, allowing for accurate cross-modal predictions. The training process is structured in two stages: first, the model learns from abundant unimodal data, followed by alignment using paired recordings. The efficacy of CardioState-JEPA is demonstrated across 25 downstream tasks, where it significantly outperforms existing modality-specific models, showcasing its potential for improved cardiac signal analysis.
Methodology
CardioState-JEPA utilizes a two-stage training approach. Initially, it learns the structure of each modality from abundant unimodal data through intra-modal masked latent prediction. Subsequently, it aligns the modalities using paired recordings and delay-aware cross-modal prediction. The model employs a shared Transformer encoder to process the mapped waveforms in a common token space, focusing on predicting masked latent states rather than reconstructing raw signals.
Results
The model was evaluated as a frozen encoder across 25 downstream tasks, achieving an average improvement of 8.2 AUROC points in PPG classification, 18.8 AUROC points in PCG murmur detection, and 15.5 AUROC points in ECG classification compared to the best self-supervised signal baseline. CardioState-JEPA also matched or exceeded the performance of cardiac models trained with clinical text or supervised labels on several ECG benchmarks.
Implications
The findings suggest that a unified approach to cardiac signal representation can enhance the performance of various cardiac analysis tasks. This model could facilitate better clinical diagnostics, monitoring, and wearable health technologies by leveraging the complementary information from multiple sensing modalities.
MAG: MAnifold Guided Semi-Supervised Multi-modal In-Context Learning
Multimodal
Large Language Models
Graph Learning
- MAG effectively addresses label scarcity in few-shot multi-modal ICL by utilizing unlabeled data.
- The framework employs a two-stage strategy for demonstration selection, enhancing efficiency and relevance.
- Textual representations are crucial for initial relevance propagation, while both modalities are needed for final selection.
- MAG demonstrates substantial performance gains across diverse benchmarks, particularly in reasoning-intensive tasks.
Read more
MAG: MAnifold Guided Semi-Supervised Multi-modal In-Context Learning
Summary
The paper introduces MAG (MAnifold Guided semi-supervised in-context demonstration selection), a novel framework aimed at enhancing few-shot in-context learning (ICL) with multi-modal large language models (MLLMs). The authors address the challenge of label scarcity in ICL, which limits the effectiveness of demonstration selection. MAG leverages abundant unlabeled multi-modal data through a semi-supervised approach, formulating demonstration selection as a propagation problem on a multi-modal graph. The framework consists of two stages: the first stage employs relevance score propagation to identify a compact set of high-impact unlabeled samples for pseudo-labeling, thereby reducing inference costs; the second stage selects the final demonstrations using multi-modal relevance. The authors find that textual representations are more effective for relevance propagation, while both visual and textual modalities are essential for high-quality demonstration selection. Experiments across eight multi-modal benchmarks show that MAG outperforms strong baselines in label-scarce scenarios, achieving significant improvements with limited pseudo-labeling budgets.
Methodology
MAG employs a two-stage framework that utilizes graph-based relevance score propagation to filter unlabeled samples for pseudo-labeling. The first stage identifies relevant unlabeled samples, while the second stage selects the most effective demonstrations from both labeled and pseudo-labeled data. This method balances scalability with demonstration quality, allowing efficient use of large unlabeled datasets.
Results
The experiments conducted on eight multi-modal benchmarks reveal that MAG consistently outperforms existing methods that rely solely on labeled data. The framework shows particularly pronounced improvements in tasks requiring complex cross-modal reasoning, demonstrating the effective incorporation of unlabeled data.
Implications
The findings suggest that leveraging unlabeled multi-modal data can significantly enhance the performance of ICL in scenarios with limited labeled data. This has potential applications in various fields, including visual emotion recognition, scene text understanding, and visual question answering, where labeled data is often scarce.
Vero: Can AI Agents Build Formally Verified Software Repositories?
Theory
- Vero is the first benchmark for evaluating joint implementation and proof synthesis at the repository level.
- It includes 43 multi-module instances from real-world repositories, enhancing the scope of verified code generation.
- An audit mechanism is implemented to identify and correct errors in specifications and reference implementations.
- Current AI agents show limitations in solving complex verification tasks, particularly those requiring cross-module reasoning.
Read more
Vero: Can AI Agents Build Formally Verified Software Repositories?
Summary
The paper introduces Vero, a novel benchmark designed to evaluate AI agents' capabilities in generating formally verified software repositories. Unlike existing benchmarks that focus on individual functions or proof generation with fixed implementations, Vero assesses the joint implementation and proof synthesis at the repository level. It comprises 43 multi-module instances sourced from real-world repositories in various programming languages, including Python and Lean 4, covering diverse domains such as cryptographic protocols and distributed systems. Each instance includes curated formal specifications and reference implementations, allowing for both proof-only and code-and-proof evaluation modes. A unique audit mechanism is integrated to identify and correct latent errors in specifications and implementations, enhancing benchmark reliability. The evaluation of frontier coding-agent configurations reveals that the strongest agent only fully solves 27 out of 43 instances, highlighting the challenges in achieving coherent implementation and proof choices across complex codebases. The findings indicate that current agents struggle with specifications that require cross-module invariants and custom mathematical theories, suggesting directions for future advancements in repository-scale verified software synthesis.
Methodology
The authors developed Vero through a multi-stage curation process, creating benchmark instances that consist of Lean 4 projects with multiple modules, formal specifications, and reference implementations. The evaluation involved testing frontier coding-agent configurations under two task modes: code-and-proof and proof-only. An audit mechanism was also integrated to address potential errors in the benchmark.
Results
The evaluation revealed that the best-performing agent fully solved only 27 of the 43 instances, with 10 instances resisting all configurations. The analysis indicated that unsolved instances were primarily those requiring complex cross-module proofs and custom mathematical theories, demonstrating the limitations of current AI agents in repository-scale verification.
Implications
The findings from Vero can guide future research in developing more capable AI agents for verified software synthesis, emphasizing the need for agents to reason globally across codebases. The benchmark itself serves as a valuable resource for measuring progress in this area and improving the reliability of AI-generated software.
Sampling Luck Masquerades as Allocation Gain: Auditing Test-Time Budget Allocation for Neural Combinatorial Optimization
Optimization
- First measurement of allocation value for NCO test-time sampling, revealing no detectable gain in in-distribution workloads.
- Quantification of in-sample selection bias, showing that traditional measurement methods can produce misleading gains.
- Demonstration of a significant allocation gain (11-12%) under distribution shift conditions with a pre-registered confirmatory experiment.
- Introduction of a budget-accounted policy that retains performance gains while managing sample costs.
Read more
Sampling Luck Masquerades as Allocation Gain: Auditing Test-Time Budget Allocation for Neural Combinatorial Optimization
Summary
This paper investigates the effectiveness of non-uniform allocation of sample budgets in Neural Combinatorial Optimization (NCO) solvers, which traditionally use a uniform sampling strategy. The author measures the potential gains from instance-wise allocation compared to the conventional uniform approach. The study reveals two significant findings: first, while in-sample evaluations suggest a small allocation gain of 2.2-2.6% across three pretrained solvers (POMO, AM, SymNCO), this gain is not detectable out-of-sample, indicating a bias in the in-sample measurement. Second, under distribution shift conditions, a pre-registered experiment demonstrates that allocation guided by held-out sample statistics can improve performance by 11.5% to 12.0% without charging for the guiding signal. The paper also introduces a budget-accounted policy for deployment that retains gains while charging for sample acquisition. Furthermore, it provides a correction procedure for selection bias and a comprehensive pre-registration record, ensuring transparency in the research process.
Methodology
The study employs a comparative analysis of allocation strategies in NCO solvers, measuring the performance of uniform versus instance-wise sample allocation. It uses both in-sample and out-of-sample evaluations to quantify allocation gains and biases, alongside a pre-registered experimental design to validate findings under distribution shifts.
Results
The results indicate that while in-sample evaluations suggest a small allocation gain, out-of-sample measurements show no detectable gain. However, under distribution shifts, the allocation guided by held-out sample statistics leads to significant performance improvements of 11.5% to 12.0%. A budget-accounted policy also retains gains while managing sample acquisition costs.
Implications
The findings suggest that NCO solvers may benefit from tailored sample allocation strategies, particularly under varying workload distributions. The quantification of selection bias emphasizes the need for careful evaluation methodologies in machine learning research, potentially influencing future practices in resource allocation and performance measurement.
EGRL: Edge generation-guided relation-aware learning for RNA-protein interaction prediction
Graph Learning
- EGRL introduces implicit meta-path learning to capture relational semantics dynamically.
- The framework includes a graph generator that predicts soft edges for cold-start nodes.
- A multi-relation-aware attention mechanism enhances the fusion of interaction patterns.
- EGRL demonstrates superior performance in cold-start scenarios, outperforming previous methods.
Read more
EGRL: Edge generation-guided relation-aware learning for RNA-protein interaction prediction
Summary
The paper presents EGRL (Edge Generation-guided Relation-aware Learning), a novel framework designed to enhance RNA-protein interaction prediction (RPIP) using Graph Neural Networks (GNNs). Traditional methods for RPIP often rely on costly wet-lab experiments or simplistic machine learning models that struggle with the complexity of biological interactions. EGRL addresses these challenges by introducing several innovative components: implicit meta-path learning that captures relational semantics without predefined paths, a multi-relation-aware attention mechanism for adaptive interaction pattern fusion, and a graph generator that predicts potential edges for cold-start scenarios involving unknown molecules. The framework is jointly trained with a primary task loss and an auxiliary generator loss, allowing it to effectively learn from sparse data. Evaluations on four benchmark datasets reveal that EGRL not only achieves competitive performance compared to existing state-of-the-art methods but also excels in cold-start settings, significantly improving the Area Under the Receiver Operating Characteristic curve (AUROC) and Area Under the Precision-Recall curve (AUPR) metrics. This work highlights the potential of GNNs in modeling complex biological systems and offers a promising computational alternative to traditional experimental methods.
Methodology
EGRL employs a heterogeneous GNN framework that integrates implicit meta-path modeling, a graph generator for predicting interactions involving unseen molecules, and multi-relational Graph Attention Network (GAT) layers. The model is trained using a combination of primary and auxiliary losses to optimize both the interaction prediction and the edge generation tasks.
Results
EGRL achieved an AUROC of 0.867 and an AUPR of 0.861 in cold-start scenarios, marking improvements of 8.6% in AUROC and 5.0% in AUPR over existing state-of-the-art methods. Comprehensive evaluations across four benchmark datasets confirmed its competitive performance.
Implications
The findings suggest that EGRL can significantly improve the efficiency and accuracy of RNA-protein interaction predictions, which are crucial for understanding cellular functions and advancing drug design. The framework's ability to generalize in cold-start situations opens new avenues for exploring uncharacterized biological molecules.
The Time Value of Evolution
Reinforcement Learning
Optimization
Theory
- Introduces the concept of the time value of evolution, distinguishing between immediate and long-term mutation benefits.
- Presents Lineage-Value Policy Gradients (LVPG) as a new actor-critic framework for evolutionary search.
- Demonstrates that long-horizon credit assignment improves search efficiency and performance metrics.
- Shows that LVPG results in fewer temporary regressions and better recovery from them compared to traditional methods.
Read more
The Time Value of Evolution
Summary
This paper addresses the limitations of traditional evolutionary search methods that often undervalue mutations leading to weak offspring, which may nonetheless open pathways to high-fitness regions. The authors introduce the concept of the 'time value of evolution,' formalizing it within a finite-horizon Markov decision process. They propose a novel framework called Lineage-Value Policy Gradients (LVPG), which separates search control into specialized policy heads over a shared generative backbone. This architecture includes a bootstrapped critic head that estimates the value of potential future lineages from multi-step mutation trees and an actor head that modulates mutation intensity based on the remaining search budget. The authors conduct extensive empirical evaluations, demonstrating that LVPG significantly enhances search efficiency and performance compared to immediate-return optimization methods. The findings indicate that long-horizon credit assignment accelerates finite-budget searches, yielding stronger policies and fewer temporary regressions, thus providing a more nuanced approach to evolutionary search.
Methodology
The authors formalize the time value of evolution within a finite-horizon Markov decision process and introduce LVPG, which employs a bootstrapped critic to assess lineage potential and an actor to control mutation intensity. They conduct paired experiments comparing LVPG with traditional immediate-return optimization methods across various parameters.
Results
The empirical results indicate that LVPG improves validation best-so-far AUC by 0.394 Sharpe units, accelerates search efficiency, and produces fewer temporary regressions, with better recovery rates compared to immediate-return optimization.
Implications
The findings suggest that incorporating long-horizon credit assignment in evolutionary algorithms can lead to more effective search strategies, particularly in complex environments like automated trading, where the temporal dynamics of mutations are critical.
Doubly Robust Estimation of Causal Effect on CVR with Targeted Regularization
Theory
- Introduction of a new doubly robust estimator for CVR causal effect estimation.
- Theoretical guarantees provided through semiparametric theory and von Mises expansion.
- Development of a targeted regularization framework to improve numerical stability.
- Extensive validation through experiments on synthetic and real-world datasets.
Read more
Doubly Robust Estimation of Causal Effect on CVR with Targeted Regularization
Summary
This paper addresses the challenge of estimating the causal effect on post-click conversion rate (CVR), a crucial metric in e-commerce and advertising. Traditional causal inference methods applied to clicked samples suffer from sample selection bias and increased variance due to the exclusion of non-click data. The authors propose a new doubly robust causal effect estimator tailored for chain-structured outcomes like CVR, derived from semiparametric theory. This estimator achieves faster convergence rates compared to nuisance parameter estimation and is more robust when using flexible nonparametric estimators, such as neural networks. Furthermore, the authors introduce a framework based on targeted regularization to enhance numerical stability and practical applicability. Extensive experiments conducted on synthetic and real-world data validate the effectiveness and robustness of the proposed method, demonstrating that naive combinations of loss debiasing with standard causal estimators underperform compared to their tailored approach.
Methodology
The authors develop a doubly robust estimator based on semiparametric theory, focusing on chain-structured outcomes. They derive the influence function and apply targeted regularization to enhance the stability of the estimation process. The method is validated through extensive experiments on various datasets.
Results
The proposed method demonstrates strong performance and robustness across synthetic, semi-synthetic, and real-world data, outperforming naive combinations of loss debiasing with standard causal estimators.
Implications
This work has significant implications for industries relying on accurate CVR estimation, such as e-commerce and advertising, by providing a more reliable method for causal effect estimation that can enhance decision-making processes.
The data geometry of masking diffusion: Certified-optimal schedules via unmasking growth complexity
Theory
Generative Models
Efficient ML
- Introduction of unmasking growth complexity (UGC) as a measure of data geometry in masking diffusion.
- Establishment of a unified analysis framework for Bernoulli-subset and fixed-cardinality unmasking schemes.
- Development of certified-optimal samplers with high-probability guarantees on KL error.
- Demonstration of significant dimension-dependent improvements in sampling efficiency.
Read more
The data geometry of masking diffusion: Certified-optimal schedules via unmasking growth complexity
Summary
This paper investigates the concept of masking diffusion for discrete sampling, introducing a novel measure of data geometry termed unmasking growth complexity (UGC). The UGC provides a path-resolved analysis that directly influences the Kullback-Leibler (KL) discretization error, facilitating a unified examination of both Bernoulli-subset and fixed-cardinality unmasking schemes. By employing log-reveal-odds coordinates, the paper derives optimized sampling schedules that adapt computational resources according to the underlying data geometry. A significant contribution is the ability to estimate UGC increments from samples, leading to the development of certified-optimal samplers that achieve a specified KL error with high probability and maintain iteration complexity close to that of oracle procedures. The paper also connects the UGC mass to classical multivariate dependence measures and previous complexity analyses of discrete diffusion. In the fine-partition limit, the squared integral of the square-root UGC density is shown to dictate the leading-order optimal Euler discretization error. The findings demonstrate substantial improvements in sampling efficiency, particularly in high-dimensional settings, highlighting the potential for dimension-dependent gains over traditional coarse schedules.
Methodology
The methodology involves analyzing the unmasking growth complexity (UGC) through a path-resolved approach, estimating local increments from samples along coupled reveal trajectories. The paper derives optimized sampling schedules based on UGC, employing log-reveal-odds coordinates to adapt computational efforts to the data geometry.
Results
The main results include the derivation of certified-optimal sampling algorithms that achieve a specified KL error with high probability, as well as the identification of substantial improvements in sampling efficiency, particularly in high-dimensional spaces. The UGC framework also provides a refined understanding of the discretization error in relation to classical dependence measures.
Implications
The findings have significant implications for the design of efficient sampling algorithms in various applications, including machine translation, image synthesis, and generative AI. The ability to adaptively optimize sampling schedules based on data geometry can enhance performance in high-dimensional settings.
GENADA: efficient generative time series adversarial attack framework
Time Series
Generative Models
Efficient ML
- GENADA generates adversarial perturbations in a single forward pass, eliminating the need for iterative optimization.
- The framework is trained on a frozen target model, allowing for efficient gradient-free inference.
- Empirical results show that GENADA achieves competitive attack quality while significantly reducing generation time.
- The approach is validated across multiple datasets and neural architectures, including recurrent, convolutional, and transformer models.
Read more
GENADA: efficient generative time series adversarial attack framework
Summary
The paper introduces GENADA (GENerative ADversarial Attack), a novel framework for generating adversarial attacks specifically designed for time series data. Traditional gradient-based adversarial attacks are computationally intensive, requiring multiple iterations and backpropagation through the target model to compute input gradients. In contrast, GENADA employs a generative model that learns to create adversarial perturbations in a single forward pass, significantly reducing the computational burden during inference. The authors validate their approach on various neural network architectures and datasets, demonstrating that GENADA achieves comparable attack quality to established methods like FGSM and iFGSM while being more efficient. The framework is trained using a frozen target model, allowing for gradient-free inference, and includes options for fine-tuning training parameters. The empirical evaluation shows that GENADA not only maintains strong performance across different time series classification tasks but also excels in terms of computational efficiency, making it a promising tool for adversarial attacks in time series analysis.
Methodology
GENADA utilizes a generative model to produce adversarial perturbations that maximize the loss of the target classifier. The training process involves a frozen target model, which allows the generator to learn effective perturbations without requiring gradient computations during inference. The framework includes both single-step and iterative generative attack schemes, providing flexibility in its application.
Results
The empirical evaluation of GENADA on datasets such as PowerCons, GunPoint, and Strawberry demonstrates that it achieves comparable attack quality to strong baselines like FGSM and iFGSM, while significantly reducing the time required to generate adversarial examples. The results indicate that GENADA is effective across various neural network architectures and maintains high performance in time series classification tasks.
Implications
The development of GENADA has significant implications for enhancing the robustness of time series models against adversarial attacks. Its efficiency makes it suitable for real-time applications in critical domains such as healthcare and finance, where timely and accurate predictions are essential. Furthermore, the framework can serve as a foundation for future research in adversarial learning and security in sequential data analysis.
Exponential Convex Calibration Dimension for the Multi-Label Jaccard Measure
Theory
- The Jaccard score's loss matrices are nonsingular with affine dimension 2s - 1.
- Exact calibration of the Jaccard score requires exponentially many prediction coordinates.
- Two polynomial-dimensional approximation methods are provided, including a transformation from F1 to Jaccard.
- The paper establishes bounds for the convex calibration dimension of the Jaccard loss.
Read more
Exponential Convex Calibration Dimension for the Multi-Label Jaccard Measure
Summary
This paper investigates the calibration dimension of the Jaccard score, a critical measure in multi-label classification and binary segmentation. The author establishes that the loss matrices associated with the Jaccard score, shifted-loss, and ordinary loss are nonsingular, with affine dimensions of 2s - 1, where s is the number of labels. The paper proves that the convex calibration dimension (CCdim) for the Jaccard loss lies between 2s - 1 and 2s, indicating that exact calibration requires exponentially many prediction coordinates. Additionally, the author presents two polynomial-dimensional approximation guarantees, including a transformation from F1 to Jaccard that yields a polynomial-time rule with controlled regret. The findings highlight the complexity of achieving exact calibration in multi-label settings and provide practical approximation methods with manageable dimensions.
Methodology
The author employs a combination of finite MinHash Gram representations and Boolean Möbius inversion to analyze the Jaccard score and its associated loss matrices. The study includes proofs of the nonsingularity of loss matrices and the derivation of bounds for the convex calibration dimension. Approximation methods are developed using transformations and MinHash feature maps to achieve polynomial-dimensional surrogates.
Results
The paper establishes that the convex calibration dimension for the Jaccard loss is bounded by 2s - 1 and 2s, confirming that exact calibration necessitates exponential dimensions. It also presents two polynomial-dimensional approximation methods, one using a transformation from F1 to Jaccard and another utilizing MinHash for uniform approximation of the Jaccard score matrix.
Implications
The findings suggest that while exact calibration in multi-label classification is complex and requires high-dimensional representations, practical approximation techniques can be employed to achieve manageable dimensions. This has implications for designing efficient algorithms in multi-label classification tasks, particularly in fields like computer vision and information retrieval.
The Boolean Power of ReLU
Graph Learning
Theory
- ReLU-MPLang is strictly more expressive than TrReLU-MPLang for Boolean queries.
- The study resolves an open problem regarding the expressiveness of different activation functions in GNNs.
- Boolean queries derived from ReLU activations can express properties that are not expressible by truncated ReLU activations.
- The findings emphasize the importance of activation function selection in GNN architectures.
Read more
The Boolean Power of ReLU
Summary
This paper investigates the expressive power of graph neural networks (GNNs) using different activation functions, specifically focusing on the ReLU (Rectified Linear Unit) and truncated ReLU (TrReLU) functions. The authors prove that the Boolean queries expressible in the Σ-MPLang (a message-passing language for GNNs) are a strict subset of those expressible in ReLU-MPLang, thereby resolving an open question regarding the comparative expressiveness of these two languages in the context of Boolean queries. The study establishes that ReLU-GNNs are strictly more expressive than GNNs using TrReLU or identity activations when it comes to Boolean queries on graphs with Boolean features. The findings highlight the impact of activation function choice on the ability of GNNs to recognize properties of nodes after thresholding their outputs to Boolean values, demonstrating that the unbounded nature of ReLU not only influences numerical embeddings but also the Boolean characteristics that can be derived from them.
Methodology
The authors utilize a formal approach to define and compare the expressive powers of Σ-MPLang and ReLU-MPLang. They construct specific Boolean queries and demonstrate their non-definability in Σ-MPLang through a series of logical proofs and structural inductions, focusing on finite simple undirected graphs with a single Boolean node feature.
Results
The main result shows that there exists a Boolean ReLU query that cannot be expressed in Σ-MPLang for any collection of eventually constant activation functions. This establishes that bool(TrReLU-MPLang) is a strict subset of bool(ReLU-MPLang), confirming that ReLU-GNNs have superior expressive capabilities for Boolean queries.
Implications
These findings have significant implications for the design of GNN architectures, particularly in applications where Boolean properties of graphs are critical. The results suggest that using ReLU activations can enhance the ability of GNNs to capture and express complex node properties, which may be beneficial in various domains such as social network analysis, recommendation systems, and biological network modeling.
Scaling Automatic Research Agents via World Models
Reinforcement Learning
Large Language Models
Efficient ML
- Introduction of World Model RL (WMRL) to replace costly environment execution in AutoResearch agents.
- Implementation of Online Debiasing and Inverse-Variance Denoising to enhance the reliability of the world model.
- Theoretical proof of improved convergence guarantees with the proposed mechanisms.
- Empirical validation showing 3-4x training acceleration and superior performance compared to larger models.
Read more
Scaling Automatic Research Agents via World Models
Summary
This paper addresses the challenge of scaling Automatic Research (AutoResearch) agents, which utilize reinforcement learning (RL) to conduct empirical research autonomously. The authors identify a critical bottleneck in the training process: while agent generation can be efficiently batched, environment execution incurs significant costs as each execution requires isolated resources. To mitigate this issue, they introduce World Model RL (WMRL), which substitutes real environment execution with a world model that simulates outcomes based on agent-generated solutions. This approach allows for scalable training without the limitations of real execution. The authors also propose two mechanisms—Online Debiasing and Inverse-Variance Denoising—to correct biases and noise in the world model's outputs, thereby improving convergence guarantees. Empirical results demonstrate that WMRL accelerates training by 3-4 times across various tasks and agent scales, outperforming larger baseline models. The method also shows versatility by extending to post-training embodied VLA policies, indicating its broader applicability beyond AutoResearch.
Methodology
The authors propose a framework where a world model simulates the execution outcomes of AutoResearch agents, allowing for efficient training. They introduce two correction mechanisms to address biases and noise in the world model's predictions, and they provide theoretical grounding for the convergence improvements achieved through these methods.
Results
WMRL accelerates training by 3-4 times across various tasks and agent scales, while outperforming larger models (48B and 120B parameters) with smaller post-trained agents (4B and 9B parameters). The proposed mechanisms also ensure that the performance matches or exceeds that of traditional RL training with real environments.
Implications
The findings suggest that WMRL can significantly enhance the efficiency of training AutoResearch agents, making it feasible to scale their capabilities. The generalizability of the method to other domains indicates potential applications in various fields requiring autonomous empirical research and decision-making.
I-SDPO: Instance-Level Adaptive Self-Distillation Policy Optimization
Reinforcement Learning
Large Language Models
Optimization
- I-SDPO addresses the degenerate gradient problem in GRPO by adapting the use of self-distillation based on the success of responses.
- The routing decision for self-distillation is made at the instance level, allowing for more effective learning from both successful and unsuccessful trajectories.
- I-SDPO achieves state-of-the-art performance on the SciKnowEval benchmark across multiple scientific domains.
- The method automatically adjusts the expected distillation rate as the model's performance improves, reducing reliance on the teacher over time.
Read more
I-SDPO: Instance-Level Adaptive Self-Distillation Policy Optimization
Summary
This paper introduces I-SDPO (Instance-Level Adaptive Self-Distillation Policy Optimization), a novel approach to address the limitations of Group Relative Policy Optimization (GRPO) in reinforcement learning for large language models (LLMs). GRPO struggles with the degenerate gradient problem, where all sampled responses are incorrect, leading to negligible policy gradients. To mitigate this, I-SDPO employs a capability-dependent routing mechanism that determines whether to use privileged self-distillation or GRPO based on the success of responses in a rollout group. Specifically, if all responses are incorrect, the model utilizes self-distillation to provide dense supervision; if at least one response is correct, it retains the GRPO framework. This method reduces the expected distillation rate as the model's success probability increases, effectively balancing the influence of the teacher and the reward signal. The authors demonstrate that I-SDPO significantly improves performance on the SciKnowEval benchmark, achieving an average mean@16 accuracy of 70.31%, compared to 56.67% with GRPO, highlighting its effectiveness in enhancing reasoning capabilities in LLMs.
Methodology
I-SDPO employs a routing mechanism that decides whether to apply privileged self-distillation or retain GRPO based on the correctness of responses in a rollout group. This instance-level decision-making allows the model to leverage self-distillation only when necessary, thus optimizing the learning process without introducing excessive bias from the teacher.
Results
I-SDPO outperforms GRPO and other methods, achieving an average mean@16 accuracy of 70.31% on the SciKnowEval benchmark, marking a significant improvement over the 56.67% accuracy achieved with GRPO. The maximum domain gain observed was 18.24 points, demonstrating the effectiveness of the proposed method.
Implications
The findings suggest that adaptive self-distillation can enhance the training of large language models, particularly in scenarios where correct responses are rare. This approach could be applied to various reinforcement learning tasks, improving the efficiency and effectiveness of model training.
Diagnosing JEPA World Models with Action-Conditioned Predictive Consistency
Reinforcement Learning
Robotics
Theory
- Introduction of Action-Conditioned Predictive Consistency (ACPC) for diagnosing visual perturbation effects in JEPAs.
- Establishment of bounds on multi-step prediction error and planning costs due to visual perturbations.
- Development of Invariance Radius (IR) and Separation Rate (SR) as metrics for assessing model robustness.
- Empirical validation of ACPC across multiple tasks and perturbation types, showing its predictive power.
Read more
Diagnosing JEPA World Models with Action-Conditioned Predictive Consistency
Summary
This paper introduces Action-Conditioned Predictive Consistency (ACPC), a diagnostic tool for evaluating the robustness of Joint-Embedding Predictive Architectures (JEPAs) against visual perturbations. JEPAs learn world models that predict outcomes in a compact latent space, but they can still be affected by visual changes. ACPC measures the divergence between predicted trajectories of a clean history and its perturbed counterpart when rolled forward under the same action sequence. The authors establish that this divergence bounds the perturbation-induced changes in multi-step prediction error and planning costs. Two complementary measures, Invariance Radius (IR) and Separation Rate (SR), are defined to summarize the sensitivity to visual perturbations and the distinguishability of different states, respectively. The paper validates ACPC through experiments on four visual control tasks, demonstrating that it effectively predicts changes in prediction error and planning costs due to visual perturbations. The findings indicate that lower IR and higher SR correlate with better performance under perturbations, providing a framework for diagnosing and improving world models.
Methodology
The authors propose ACPC, which involves rolling forward a clean history and its perturbed view under the same action sequence to measure the distance between their predicted trajectories. They define IR and SR to summarize the results of ACPC and prove theoretical bounds on prediction error changes. Experiments are conducted on four visual control tasks to evaluate the effectiveness of these diagnostics.
Results
The experiments reveal that ACPC can predict perturbation-induced changes in multi-step prediction error and planning costs. The IR and SR metrics effectively identify models that maintain performance under visual perturbations, with lower IR and higher SR indicating better robustness. The results are consistent across different tasks and architectures, confirming the utility of ACPC as a diagnostic tool.
Implications
The findings suggest that ACPC can be used to enhance the robustness of world models in reinforcement learning and control applications, potentially leading to more reliable AI systems in dynamic environments. The metrics IR and SR can guide the development of models that are less sensitive to visual noise, improving their applicability in real-world scenarios.
Beyond Simulated Benchmarks: Evaluating Motion Representations for Fall Detection Under Real-World Data Scarcity
Time Series
- Real-world fall detection is hindered by the extreme scarcity of actual fall data.
- Simulated datasets often lead to overestimated performance in laboratory settings.
- Interval-based representations achieve the best real-world performance, while symbolic representations with impact descriptors show robustness under data scarcity.
- The study emphasizes the importance of representation choice for generalization from simulated to real-world conditions.
Read more
Beyond Simulated Benchmarks: Evaluating Motion Representations for Fall Detection Under Real-World Data Scarcity
Summary
This paper addresses the critical issue of fall detection for older adults, emphasizing the challenges posed by the scarcity of real-world fall data. While wearable sensors have been explored for detecting falls, the rarity of actual fall events necessitates reliance on simulated datasets, which often lead to overestimated performance in controlled environments. The authors systematically evaluate various motion representations—interval-based, kernel-based, symbolic, and foundation model representations—using accelerometer signals from both simulated and clinically verified real-world datasets. They introduce a lightweight symbolic representation, FallLM, which combines symbolic tokens with physically-grounded impact descriptors to enhance interpretability. The study employs a streaming evaluation pipeline to assess the robustness of these representations under data scarcity and cross-dataset transfer. Results indicate that while complex models perform well on simulated data, they struggle with real-world applicability. The interval-based representation shows the best real-world performance, but the augmented symbolic representation demonstrates resilience against domain shifts and retains detection sensitivity under extreme data scarcity. This work highlights the necessity of evaluating motion representations beyond simulated benchmarks and underscores the importance of representation choice for effective fall detection in real-world scenarios.
Methodology
The authors conducted a systematic evaluation of different motion representations for fall detection using accelerometer data. They compared interval-based, kernel-based, symbolic, and foundation model representations through a unified streaming evaluation pipeline, employing subject-wise splits for realistic testing. The study utilized both simulated and clinically verified real-world datasets, focusing on cross-dataset transfer and robustness under data scarcity.
Results
The findings reveal that highly parameterized kernel and foundation models perform well on simulated data but degrade significantly when faced with real-world data scarcity and domain shifts. The interval-based representation achieved the highest absolute performance in real-world scenarios, while the augmented symbolic representation maintained detection sensitivity under extreme data scarcity, albeit with lower precision.
Implications
This research has significant implications for the development of fall detection systems, particularly in geriatric care and remote health monitoring. By emphasizing the importance of motion representation choice, the findings can guide future research and technology development to improve the reliability and effectiveness of fall detection in real-world applications.
Unifying Generative Models with Path Integrals
Generative Models
Theory
- Generative modeling is unified under a single master action using path integrals.
- Various generative models are shown to be limits or special cases of this master action.
- A one-loop correction method is introduced that significantly reduces sampling errors.
- Imperfect learned scores are treated as diagrammatic insertions, leading to a new score-matching objective.
Read more
Unifying Generative Models with Path Integrals
Summary
This paper presents a novel framework for generative modeling by formulating it as a path integral, where various generative models such as flow-based, diffusion-based, variational, and adversarial models are derived as different evaluation principles from a single master action. The author employs the Martin-Siggia-Rose-Janssen-de Dominicis (MSRJD) formalism to separate free from interacting probability flows, allowing for a diagrammatic perturbation theory approach. This leads to a one-loop correction for deterministic samplers, significantly reducing errors from 53% at tree-level to 1.6% in practical applications. The paper also introduces a response-weighted score-matching objective that incorporates imperfect learned scores and discusses symmetry-equivariant drift design through an effective field theory (EFT) power counting approach. The theoretical emphasis of the work is validated through numerical studies on solvable and nonlinear drift models, demonstrating the framework's robustness and potential for improving generative modeling in high-dimensional probability distributions.
Methodology
The paper employs a path integral formulation of generative models, utilizing the MSRJD representation to analyze probability flows. It develops a diagrammatic perturbation theory to derive corrections for deterministic samplers and introduces a response-weighted score-matching objective for handling imperfect scores.
Results
The proposed framework yields a one-loop correction that reduces sampling errors from 53% to 1.6%. The numerical validation confirms the effectiveness of the loop formulas on synthetic drifts, showcasing the framework's applicability in generative modeling.
Implications
This work has significant implications for the development of more accurate and efficient generative models in various fields, including high-energy physics, where probabilistic modeling is crucial for simulations and uncertainty quantification. The unification of different generative approaches may lead to more robust algorithms and architectures in machine learning.
Into the ORBIT for Time Series: Training Regimes for Foundation Models
Time Series
- Introduction of ORBIT, a training paradigm for TSFMs that controls effective pre-training distribution.
- Bootstrap Multi-Level Sampling and Omni-Range Incremental Training are key components of ORBIT.
- Falcon-2.0, trained under ORBIT, shows strong zero-shot forecasting capabilities.
- Rank-Guided Cross-Depth Alignment improves representation alignment across Transformer depths.
Read more
Into the ORBIT for Time Series: Training Regimes for Foundation Models
Summary
This paper introduces ORBIT (Omni-Range Bootstrap Incremental Training), a novel training paradigm for Time Series Foundation Models (TSFMs) that aims to enhance the effective pre-training distribution of heterogeneous time series data. The authors identify that existing training regimes inadequately address challenges such as cross-domain imbalance, varying context requirements, prediction horizons, and handling of missing data. ORBIT consists of two main components: Bootstrap Multi-Level Sampling, which hierarchically controls data exposure by creating a global training stream and stochastic sample indices for time series records, target variables, context windows, and prediction horizons; and Omni-Range Incremental Training, which allows for the simultaneous consumption of variable-length examples during training without the need for stage-specific schedules. The authors implement ORBIT in training Falcon-2.0, a simple univariate encoder-only Transformer model that incorporates missingness-aware tokenization and quantile prediction. They also introduce Rank-Guided Cross-Depth Alignment, a training objective that aligns representations across different depths of the Transformer. The results demonstrate that Falcon-2.0 achieves strong zero-shot forecasting performance across various domains and frequencies, with ablation studies highlighting the significance of stochastic sample construction and diverse context exposure.
Methodology
The methodology involves two main components: Bootstrap Multi-Level Sampling for hierarchical data exposure control, and Omni-Range Incremental Training for consuming variable-length examples during a single training run. The training process utilizes stochastic sampling techniques to create a diverse set of training examples while maintaining reproducibility.
Results
Falcon-2.0, trained using ORBIT, achieved strong zero-shot forecasting performance across diverse domains and frequencies. The ablation studies confirmed the effectiveness of the stochastic sample construction and the simultaneous exposure to various context lengths and prediction horizons.
Implications
The findings suggest that ORBIT can significantly improve the training of TSFMs, leading to better forecasting capabilities in real-world applications. The approach can be beneficial for industries relying on time series data, such as finance, healthcare, and supply chain management.
Which Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods
Interpretability
- Utilizes free satellite data to predict glacial lake outburst floods and landslides in the Nepal Himalaya.
- Develops models to identify susceptible sites and timing of triggers for various hazards.
- Achieves significant predictive performance for large bursts and landslides using weather data.
- Finds that terrain susceptibility is influenced by historical failure patterns rather than inherent risk.
Read more
Which Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods
Summary
This paper investigates the prediction of glacial lake outburst floods (GLOFs), landslides, and smaller floods in the Nepal Himalaya using free satellite data. The authors utilize radar interferometry to monitor moraine dam deformations and satellite weather data to assess when lakes are under stress. They propose predictive models to identify which sites are susceptible to hazards and when triggers may occur. The study employs a dataset of 589 dated outbursts and thousands of landslides, comparing events against similar unfailed sites while adhering to a strong baseline model. The results indicate that antecedent weather can effectively time triggers for large bursts, landslides, and small floods, achieving area under the curve (AUC) scores of 0.73, 0.83, and 0.82, respectively. Terrain analysis shows that while susceptibility appears high, it is largely influenced by historical failure clustering in wetter regions. The study concludes that while deep learning models did not outperform simpler models, a decision tree based on ruggedness and rainfall provided useful insights. The authors present a ranked watchlist for Nepal, emphasizing the limitations of free data in hazard prediction.
Methodology
The authors conducted a predictive analysis using free satellite data, including radar interferometry for monitoring moraine dam deformations and weather data for assessing lake stress. They compared 589 historical outburst events against similar unfailed sites, employing spatial cross-validation to ensure model robustness. The study tested multiple models, including deep learning approaches, against a baseline model based on gradient boosting.
Results
The study found that antecedent weather data effectively predicts the timing of triggers for large moraine-dam bursts (AUC 0.73), landslides (AUC 0.83), and smaller floods (AUC 0.82). Terrain analysis indicated that while susceptibility scores appeared high, they were largely influenced by historical clustering of failures in wetter areas. The baseline model outperformed deep learning models in predicting lake hazards, with a decision tree model providing useful insights.
Implications
The findings suggest that free satellite data can be effectively utilized for hazard monitoring and prediction in remote regions, potentially aiding in disaster preparedness and risk management. The ranked watchlist can serve as a prioritization tool for monitoring glacial lakes, contributing to national efforts in mitigating flood risks.
Neural Quadratic Forms: A Unified Minimal Model for Sudden Learning and Scaling Laws
Theory
- Introduces Neural Quadratic Forms (NQF) as a minimal model for understanding neural network learning dynamics.
- Demonstrates that abrupt learning steps and smooth power-law scaling can be unified through symmetry considerations.
- Establishes a universal quadratic form for training dynamics applicable across various neural architectures.
- Identifies the structure matrix A(x) as a key component that encapsulates architectural details.
Read more
Neural Quadratic Forms: A Unified Minimal Model for Sudden Learning and Scaling Laws
Summary
This paper introduces a novel framework termed Neural Quadratic Forms (NQF), which unifies various neural network architectures under a minimal model that captures their learning dynamics. The authors observe that neural networks exhibit two distinct learning behaviors: abrupt drops in training loss after long plateaus and smooth power-law scaling across different model sizes and training conditions. By leveraging the symmetry inherent in neural architectures, the authors derive a universal quadratic form that describes the training dynamics of various architectures, including perceptrons, attention layers, and convolutions. The key insight is that the training dynamics can be characterized by a single structure matrix A(x), which encodes architectural details, while the overall behavior is governed by an order parameter M = WW⊺. The authors demonstrate that the dynamics can be reduced to a Lotka-Volterra equation, revealing that the timing of feature acquisition is influenced by the initial weight scale. This framework not only provides a theoretical basis for understanding the observed learning behaviors but also offers predictive capabilities regarding the scaling laws of neural networks.
Methodology
The authors employ a theoretical approach grounded in symmetry principles to derive a universal quadratic form for neural network training dynamics. They analyze the implications of this form across different architectures and validate their findings through numerical simulations.
Results
The study confirms that the derived neural quadratic form accurately captures the training dynamics of various neural architectures, demonstrating that the abrupt learning steps and smooth power-law behaviors are manifestations of a unified underlying mechanism. The predictions regarding the scaling laws are validated numerically, showing consistency across different training methods and architectures.
Implications
This work has significant implications for the design and understanding of neural networks, suggesting that a minimal model can effectively predict learning behaviors and scaling laws. It may guide future research in optimizing neural architectures and training strategies, as well as inform large-scale deployment decisions in machine learning applications.
Training Under Challenge: Executable Certificates and Challenge-Closed Optimality for Neural Networks
Theory
Optimization
- Introduces an executable-certificate framework for assessing neural network training.
- Defines a challenge-power modulus to quantify optimality gaps.
- Demonstrates the framework's effectiveness on a ResNet-18 distillation problem.
- Establishes the importance of coverage mechanisms in determining global optimality.
Read more
Training Under Challenge: Executable Certificates and Challenge-Closed Optimality for Neural Networks
Summary
This paper introduces 'Training Under Challenge', an innovative framework designed to assess the global optimality of neural networks during training. The authors argue that a flat training curve does not necessarily indicate that a neural network has achieved a global optimum, as it could also signify local traps or representation limitations. The proposed framework utilizes executable certificates that allow for the construction of alternative models within the same certified class, enabling a reevaluation of the objective function. The authors define a challenge-power modulus that quantifies the largest optimality gap compatible with successful passage of a challenge suite. They demonstrate the effectiveness of their approach through a case study involving a ResNet-18 model, where they successfully covered all output directions and established bounds on the empirical global-optimality gap. The paper also discusses the importance of coverage mechanisms and presents a theoretical foundation for the framework, including the relationship between challenge power and detectable suboptimality. Overall, the framework provides a systematic method for auditing neural network training and ensuring that models are not only trained effectively but also verified for optimality.
Methodology
The authors developed an executable-certificate framework that constructs alternative models through architecture-valid procedures. They defined a challenge suite to reevaluate the same objective and established a resource-indexed challenge-power modulus to characterize optimality gaps. The methodology includes a systematic audit of checkpoints and the use of theoretical claims to validate the results.
Results
The framework was applied to a ResNet-18 distillation problem, where eight internal challenges successfully covered all 240 output directions. The realized-residual bounds were found to be within factors of 1.74 to 3.02 of the true optimality gap, demonstrating the framework's capability to assess and certify model performance effectively.
Implications
This framework has significant implications for improving the reliability and transparency of neural network training. By providing a structured method for auditing and certifying models, it can enhance the understanding of model performance and guide practitioners in achieving true optimality in their neural networks.
Robust data-driven discovery of fractional differential equations via weak formulations and Pareto-based subset selection
Theory
- Weak-Pareto combines weak formulations with Pareto-based subset selection for discovering fractional differential equations.
- The method effectively mitigates noise amplification issues associated with fractional differentiation.
- Weak-Pareto demonstrates superior robustness and accuracy in recovering equations from noisy data compared to traditional methods.
- The framework allows for continuous-order optimization, avoiding the pitfalls of fixed-order dictionaries.
Read more
Robust data-driven discovery of fractional differential equations via weak formulations and Pareto-based subset selection
Summary
This paper presents a novel framework called Weak-Pareto for the data-driven discovery of fractional differential equations (FDEs) from noisy measurements. Traditional methods struggle with FDEs due to the amplification of high-frequency noise and the challenge of unknown derivative orders. Weak-Pareto addresses these issues by combining an adjoint-consistent weak formulation of fractional terms with a Pareto-based subset selection approach. The framework effectively transfers fractional operators from measured fields to smooth test functions, minimizing noise sensitivity. Coefficients are fitted using ridge regression within a differential-evolution search that optimizes the orders of the fractional derivatives. The method demonstrates robustness against noise, successfully recovering the correct support across various benchmarks, including fractional advection-diffusion and Burgers equations, even under significant noise levels. The results indicate that Weak-Pareto outperforms traditional strong-form methods, particularly in noisy environments, and achieves more consistent operator recovery with lower computational costs compared to neural network baselines. Additionally, a two-dimensional extension of Weak-Pareto is capable of recovering coordinate-dependent orders, showcasing its versatility in complex systems.
Methodology
The Weak-Pareto framework employs an adjoint-consistent weak formulation to derive fractional terms, replacing pointwise differentiation with smoothing integration. It utilizes ridge regression for coefficient fitting and a differential-evolution search for optimizing fractional orders. The Pareto-based subset selection ensures that only the most relevant terms are included, enhancing model parsimony and robustness.
Results
Weak-Pareto successfully recovers the correct structure of fractional differential equations from both clean and noisy measurements across various benchmarks. It retains accuracy even under high levels of multiplicative and additive noise, outperforming traditional strong-form methods and neural network approaches in terms of operator recovery consistency and computational efficiency.
Implications
The findings suggest that Weak-Pareto can be applied to various fields where fractional dynamics are relevant, such as physics, biology, and finance. Its robustness against noise makes it particularly useful for real-world applications where data is often imperfect. The ability to discover nonlocal dynamics could lead to better modeling of complex systems.
Demand Transfer Estimation at Scale via Restricted Logit Modeling
Optimization
- Introduces a scalable method for estimating Demand Transfer coefficients in large item universes.
- Combines independent demand forecasting with adjustments for item relationships to improve accuracy.
- Utilizes a modified Markov Chain model to address scaling issues in demand estimation.
- Demonstrates improved demand forecasting through experiments on historical transaction data.
Read more
Demand Transfer Estimation at Scale via Restricted Logit Modeling
Summary
This paper addresses the challenge of item demand forecasting in the context of store assortment optimization (SAO), particularly for large item universes with over a million items. The authors propose a novel approach to estimate Demand Transfer (DT) coefficients, which quantify how much demand for a target item is redirected to other items when the target is unavailable. Traditional methods often require separate demand forecasts for each assortment, which is inefficient. The proposed method combines independent demand forecasting with adjustments for item relationships, allowing for scalable computation of DT coefficients. The authors utilize a modified Markov Chain model to estimate these coefficients, overcoming scaling and stability issues present in previous approaches. Experiments conducted on historical transaction data demonstrate that the proposed method accurately estimates DT coefficients and enhances demand forecasting accuracy, leading to better assortment decisions for retailers.
Methodology
The authors propose a modified Markov Chain model to estimate Demand Transfer coefficients, which allows for scalable computation across large item universes. This approach integrates independent demand forecasts with adjustments based on item substitution behavior, enabling accurate estimation of demand redirection when items are removed from the shelf.
Results
The experiments show that the proposed method effectively estimates DT coefficients, leading to significant improvements in demand forecasting accuracy. This enhancement allows retailers to make more informed assortment decisions, optimizing shelf space and inventory management.
Implications
The findings suggest that retailers can leverage the proposed method to enhance their assortment optimization strategies, ultimately improving customer satisfaction and operational efficiency. The scalability of the approach makes it applicable to a wide range of retail contexts, from small businesses to large enterprises.
MARCH: Scaling Recurrent Memory with Content-Routed State Anchors
NLP
Large Language Models
Efficient ML
- MARCH introduces a scalable memory architecture that enhances recurrent models' long-context retrieval capabilities.
- The architecture employs periodic caching of recurrent states as state anchors, allowing for efficient historical information retrieval.
- MARCH outperforms existing linear attention variants in various tasks, indicating its effectiveness in recall-intensive scenarios.
- The model maintains a balance between memory cost and historical resolution, providing a flexible approach to memory management.
Read more
MARCH: Scaling Recurrent Memory with Content-Routed State Anchors
Summary
The paper introduces MARCH (Memory-Anchor Routing across Context History), a novel architecture designed to enhance the efficiency and effectiveness of recurrent memory systems in handling long-context sequences. Traditional recurrent models often struggle with recall-intensive tasks due to their tendency to overwrite earlier information with new updates, leading to a loss of historical context. MARCH addresses this limitation by periodically caching cumulative recurrent states as 'state anchors' and associating each anchor with a compact, content-conditioned key. This allows the model to maintain a growing memory bank that can be accessed during token processing, enabling a balance between historical resolution and memory cost. The architecture employs an anchor query mechanism to attend to all causally available state anchors, facilitating an attention-style aggregation of historical information. Experimental results demonstrate that MARCH consistently outperforms various linear attention models across tasks such as commonsense reasoning and in-context retrieval, showcasing its ability to strengthen long-range memory while maintaining computational efficiency.
Methodology
MARCH employs a memory-augmented recurrent architecture that periodically caches cumulative states as state anchors. Each anchor is linked to a compact descriptor, enabling the model to route queries to relevant historical states. The architecture integrates an anchor query mechanism that allows for attention-based aggregation of information from the memory bank, facilitating effective retrieval of past states without altering the underlying recurrence structure.
Results
MARCH demonstrated superior performance compared to multiple linear attention variants across several benchmarks, including commonsense reasoning and in-context retrieval tasks. The results indicate a significant improvement in long-context retrieval capabilities, validating the effectiveness of content-routed state caching in enhancing recurrent memory.
Implications
The advancements presented in MARCH could lead to improved performance in applications requiring long-document understanding, multi-turn interactions, and in-context learning. By enhancing memory management in recurrent models, MARCH may facilitate more effective information retrieval in real-world scenarios, potentially benefiting various NLP tasks.
Defensive Boosting for Online Probabilistic Forecasting
Theory
Efficient ML
Optimization
- Introduces the Defensive Booster algorithm for online probabilistic forecasting.
- Achieves dual guarantees: competitive Brier scores and reduced classification error under weak-learning conditions.
- Utilizes a single weak-class learner for efficiency, unlike previous methods requiring multiple learners.
- Provides local hard-core certificates for weak-learning conditions, enhancing robustness.
Read more
Defensive Boosting for Online Probabilistic Forecasting
Summary
This paper addresses the challenge of online probabilistic forecasting of binary outcomes in the presence of an adaptive adversary. The authors propose a novel algorithm called the Defensive Booster, which combines the strengths of existing online boosting techniques to achieve two key guarantees: competitive Brier scores with the best predictor in the span of weak hypotheses and the ability to drive classification error to zero under a smooth weak-learning condition. The Defensive Booster operates efficiently by utilizing a single weak-class learner, contrasting with previous methods that require maintaining large ensembles of weak learners. The paper demonstrates that the Defensive Booster not only meets the performance benchmarks of online gradient boosting but also excels in scenarios where the weak-learning condition is satisfied, providing a robust framework for probabilistic forecasting in dynamic environments. Experimental results on synthetic and real data streams indicate that the Defensive Booster significantly outperforms prior baselines in predictive performance while achieving faster runtimes.
Methodology
The Defensive Booster algorithm is developed as a black-box reduction from an online learning algorithm for weak hypothesis classes. It operationalizes the dual view of boosting, utilizing mistake weights to certify the failure of the weak-learning condition when necessary. The algorithm is designed to be strongly adaptive, allowing it to maintain performance guarantees over contiguous intervals of time.
Results
The Defensive Booster achieves a Brier score competitive with the best predictor in the span of weak hypotheses on every adaptive sequence. When the smooth weak-learning condition holds, it guarantees both Brier score and randomized classification error to be minimized effectively. The algorithm also provides local hard-core certificates indicating when the weak-learning condition fails, enhancing its reliability. Experimental results show substantial improvements in predictive performance and runtime efficiency compared to existing methods.
Implications
The findings suggest that the Defensive Booster can be effectively applied in various online learning scenarios where adversarial conditions are present, such as financial forecasting, real-time decision-making systems, and adaptive risk management. Its efficiency and robustness make it a valuable tool for practitioners in fields requiring accurate probabilistic predictions.
H-VAEP and H-xT: Valuing Offensive On-the-Ball Actions in Handball by Estimating Probabilities
Theory
Optimization
- Introduction of H-xT and H-VAEP frameworks tailored for handball analytics.
- Development of a handball-native court zoning layout for improved action valuation.
- Demonstration of the robustness of H-xT compared to traditional rectangular grids.
- H-VAEP provides stable and intuitive player ratings that emphasize build-up play.
Read more
H-VAEP and H-xT: Valuing Offensive On-the-Ball Actions in Handball by Estimating Probabilities
Summary
This paper presents the first comprehensive adaptation of Expected Threat (xT) and Valuing Actions by Estimating Probabilities (VAEP) frameworks for handball, addressing the limitations of traditional player evaluation methods that rely on basic box-score metrics. The authors utilize five seasons of tracking-derived event data from the Handball Bundesliga to develop Handball-xT (H-xT) and Handball-VAEP (H-VAEP). H-xT employs a handball-specific court zoning layout, demonstrating improved robustness over standard rectangular grids through simulations. H-VAEP is optimized for the sport's dynamics by refining its feature space and context length to mitigate team-identity leakage. The evaluation of H-VAEP reveals stable and intuitive player ratings that effectively highlight contributions to build-up play. The authors also provide a complete code repository to facilitate the practical application of these models by professional clubs.
Methodology
The authors adapted the xT and VAEP frameworks for handball by utilizing tracking data from five seasons of the Handball Bundesliga. They developed a custom court zoning layout that respects the sport's geometry and optimized the feature space and context length for H-VAEP. Simulations were conducted to validate the robustness of H-xT against standard models.
Results
The evaluation of H-VAEP showed strong face validity, reliability, and discrimination compared to traditional performance indicators. The models successfully highlighted the contributions of players in build-up play, which are often undervalued in conventional metrics.
Implications
The findings suggest that adopting these advanced action valuation frameworks can significantly enhance player evaluation and recruitment strategies in professional handball, allowing teams to better understand the contributions of individual players beyond basic statistics.
Simulation-to-real transfer learning for infrared spectroscopic chemical sensing and analysis from molecules to complex samples
Efficient ML
- Introduction of UltraIR, a foundation model for IR spectroscopy with over 100 million parameters.
- Utilization of simulation-to-real transfer learning to enhance data efficiency and reliability in chemical inference.
- Demonstrated strong performance across multiple analytical tasks and real-world applications.
- Pretraining on simulated IR spectra allows for effective adaptation to downstream tasks with limited labeled data.
Read more
Simulation-to-real transfer learning for infrared spectroscopic chemical sensing and analysis from molecules to complex samples
Summary
This paper presents UltraIR, a foundation model designed for infrared (IR) spectroscopy aimed at enhancing chemical sensing and analysis from molecular to complex samples. Traditional methods for interpreting IR spectra are labor-intensive and often rely on expert knowledge, making them difficult to scale and less reliable for unfamiliar compounds. The authors propose a simulation-to-real transfer learning approach that leverages a large dataset of simulated IR spectra (approximately 60 million) to pretrain a model with over 100 million parameters. This model learns a shared spectral representation through three pretraining objectives: spectral reconstruction, molecular fingerprint similarity alignment, and functional-group prediction. The pretrained model can then be adapted to specific tasks using limited labeled experimental spectra. The paper demonstrates UltraIR's effectiveness across various benchmark evaluations, including functional-group prediction and molecular structure elucidation, as well as real-world applications such as bacterial classification and soil property prediction. The results indicate that UltraIR outperforms conventional machine-learning methods and shows strong performance in scenarios with limited labeled data, highlighting its potential for scalable and reliable chemical analysis.
Methodology
The authors developed UltraIR by pretraining on a large dataset of simulated IR spectra using three complementary objectives. The model learns a shared spectral representation, which is then fine-tuned for specific tasks using limited labeled experimental data. This approach enables effective transfer learning from simulated to real-world scenarios.
Results
UltraIR demonstrated superior performance in various benchmark tasks, including functional-group prediction and molecular structure elucidation, as well as in real-world applications such as bacterial classification and soil property prediction. The model effectively adapted to tasks with limited labeled data, outperforming conventional machine-learning and task-specific deep-learning baselines.
Implications
The findings suggest that UltraIR can significantly enhance the reliability and scalability of chemical sensing systems, making it applicable in diverse fields such as environmental monitoring, clinical analysis, and materials characterization. Its ability to perform well with limited labeled data may facilitate broader adoption of machine learning in chemical analysis.
A Probe Direction Is a Property of Its Prompt
NLP
Large Language Models
Theory
- The choice of prompt significantly influences model evaluation scores.
- Reported scores can vary widely based on prompt wording, affecting perceived model performance.
- A single-prompt design is insufficient for reliable comparisons across models.
- The paper advocates for a multi-prompt evaluation framework to enhance reliability.
Read more
A Probe Direction Is a Property of Its Prompt
Summary
This paper investigates the reliability of model evaluations in machine learning, particularly focusing on how the choice of prompt influences the reported scores of models. The author critiques the standard method of evaluating models by contrasting activations from prompts that announce an evaluation against those that do not. The study reveals that the reported scores are significantly affected by the specific wording of the prompts used, rather than solely reflecting the model's capabilities. By conducting factorial experiments that vary prompt framing while keeping task text constant, the author demonstrates that the direction and magnitude of scores can change based on prompt choice. This variability indicates that the model's sense of being evaluated is not the sole determinant of the scores, as the prompt itself introduces a substantial amount of variance. The findings suggest that a single-prompt design is inadequate for reliable model comparisons, and the paper provides recommendations for a more robust evaluation framework that includes multiple prompts. The study also emphasizes the importance of transparency in reporting prompt choices in model evaluations.
Methodology
The author employs a factorial experimental design, varying prompt framing while keeping the task text constant. This approach allows for the isolation of the effects of prompt choice on model evaluation scores. The analysis utilizes generalizability theory to assess the variance components related to model performance and prompt framing.
Results
The results indicate that the reported scores are heavily influenced by the choice of prompt, with different framings leading to different trends in model performance. The correlation between model size and evaluation scores can change sign based on prompt selection, demonstrating that the prompt is a significant factor in the evaluation process. Additionally, the study finds that the variance attributed to the model itself is small compared to the variance introduced by prompt interactions.
Implications
The findings suggest that researchers should be cautious when interpreting model evaluation scores, as they may not accurately reflect model capabilities due to prompt influence. The proposed multi-prompt evaluation framework could lead to more reliable assessments of model performance, fostering better comparisons across different models and contributing to the development of more robust machine learning systems.
Dual Spatial-Temporal Attribution: Architecture-Aligned Post-Hoc Explainability for Recurrent Graph Anomaly Detection
Graph Learning
Interpretability
Time Series
- Introduction of X-AddGraph, the first explainability framework for AddGraph and GCN+GRU models.
- Development of Dual Spatial-Temporal Attribution (DSTA) to align explainability with the architecture of the anomaly detector.
- Preservation of detection performance (AUC) while providing post-hoc explanations.
- Long-term attribution identifies more informative historical snapshots compared to random selection.
Read more
Dual Spatial-Temporal Attribution: Architecture-Aligned Post-Hoc Explainability for Recurrent Graph Anomaly Detection
Summary
This paper addresses the challenge of explainability in dynamic graph anomaly detection, specifically focusing on the AddGraph framework, which combines graph convolutional networks (GCNs) and Gated Recurrent Units (GRUs). Despite achieving high accuracy in detecting anomalies, AddGraph lacks transparency, providing only anomaly scores without explanations. The authors introduce X-AddGraph, a post-hoc explainability framework that employs a Dual Spatial-Temporal Attribution (DSTA) mechanism. This mechanism consists of three components aligned with AddGraph's architecture: spatial relevance attribution based on the current adjacency structure, short-term attribution utilizing contextual attention weights, and long-term attribution through gradient rollback in recurrent states. The framework preserves detection performance while providing meaningful explanations. Evaluations on the UCI Message benchmark demonstrate that X-AddGraph maintains the original detection performance (AUC = 0.8705) while offering insights into the factors influencing anomaly scores. The long-term attribution component significantly outperforms random selection in identifying historical snapshots with counterfactual signals, enhancing the interpretability of the anomaly detection process. The implementation is made publicly available for reproducibility.
Methodology
The authors developed the X-AddGraph framework, which utilizes the DSTA mechanism comprising three components: (1) gradient-based relevance attribution over the adjacency structure for spatial insights, (2) direct reading of contextual attention weights for short-term insights, and (3) gradient rollback through GRU hidden states for long-term insights. This approach allows for a comprehensive explanation of anomaly detection without altering the original AddGraph model.
Results
The X-AddGraph framework successfully reproduces the original detection performance of AddGraph (AUC = 0.8705) while providing explanations. The long-term attribution component was shown to identify historical snapshots with significantly more counterfactual signals compared to random selections, indicating its effectiveness in enhancing interpretability.
Implications
The findings suggest that incorporating explainability into anomaly detection frameworks can improve trust and auditability in regulated environments. The ability to provide clear explanations for detected anomalies can facilitate better decision-making in cooperative information systems, such as financial networks and enterprise communication platforms.
Exploring Oversmoothing with Householder Matrices
Graph Learning
Theory
- Introduces Householder matrices as a method to combat oversmoothing in GNNs.
- Proves that HouseGNN preserves the Euclidean norm of node representations.
- Demonstrates scale and sign invariance of the Householder reflector.
- Shows that pairwise distances between nodes can vary with orthogonal transformations.
Read more
Exploring Oversmoothing with Householder Matrices
Summary
This paper addresses the issue of oversmoothing in deep graph neural networks (GNNs), where node representations converge towards a low-information subspace as the network depth increases. The author introduces the Householder Graph Neural Network (HouseGNN), which updates node embeddings using Householder reflections instead of the standard graph convolutional network (GCN) approach. By employing aggregated neighborhood messages to estimate a reflection direction, HouseGNN preserves the Euclidean norm at each node and depth, thereby mitigating oversmoothing. The paper proves three key properties of HouseGNN: (i) preservation of node-wise Euclidean norm at every internal layer, (ii) scale and sign invariance of the Householder reflector in the message, and (iii) the ability for pairwise distances between nodes to change due to mismatches in node-wise orthogonal operators. This novel approach offers a new perspective on combating oversmoothing in GNNs and enhances the discriminative power of node representations.
Methodology
The Householder Graph Neural Network (HouseGNN) employs Householder reflections to update node embeddings. Instead of directly modifying hidden states, it uses neighborhood messages to define a reflection direction, followed by a norm-preserving nonlinearity called GroupSort. This approach results in a piecewise orthogonal layer that maintains the Euclidean norm across all nodes and depths.
Results
The paper demonstrates that HouseGNN effectively mitigates the oversmoothing problem commonly faced by deep GNNs. The theoretical properties established show that the method maintains the discriminative variation of node representations, which is crucial for tasks such as node classification and link prediction.
Implications
The findings suggest that Householder transformations can be a powerful tool in enhancing the performance of GNNs, particularly in applications where maintaining the distinctiveness of node representations is critical. This could lead to improved outcomes in various graph-based learning tasks, including social network analysis, molecular property prediction, and more.
Training AI Scientists to Replicate Research
Large Language Models
Reinforcement Learning
Theory
- Introduction of Replica, a task space for replicating research papers.
- Development of Faraday, an AI Scientist agent that surpasses existing models in replication tasks.
- Implementation of a rubric-based judge for evaluating replication quality.
- Demonstration of Faraday's ability to adopt a scientifically rigorous approach.
Read more
Training AI Scientists to Replicate Research
Summary
This paper addresses the replication crisis in scientific research, particularly in machine learning, by developing a scalable task space called Replica for AI agents to replicate research papers. The authors introduce Faraday, a 27B-parameter AI Scientist agent that utilizes coding agents as tools to perform replication tasks. The study highlights the challenges faced by existing AI agents in replicating research due to the underspecified nature of papers and the need for open-ended exploration. To evaluate replication quality, the authors create an auto-generated rubric-based judge that aligns closely with human assessments. Faraday demonstrates superior performance compared to existing models, such as Claude Opus 4.8 and GPT-5.5, achieving significant improvements in replication tasks. The results indicate that Faraday not only excels quantitatively but also exhibits a more scientifically principled approach in its methodology, paving the way for future AI agents capable of long-term scientific innovation without complex harnesses.
Methodology
The authors constructed the Replica task space by curating 310 figure-replication tasks from 100 machine learning and AI-for-science papers. They trained Faraday using a modified version of GRPO with a per-task rubric-based judge that aggregates multiple samples to provide a low-noise reward signal. Faraday employs Codex as a coding tool, allowing it to navigate the complexities of replication tasks effectively.
Results
Faraday outperformed Claude Opus 4.8 and GPT-5.5 on 73% of in-distribution ML tasks and 60% of held-out AI-for-science tasks, achieving an average improvement of 6% over Claude and 8% over Codex. Human experts rated Faraday as more effective than its competitors in tasks where it had an advantage according to the rubric judge.
Implications
The findings suggest that AI agents like Faraday can significantly contribute to addressing the replication crisis in science, enhancing the reliability of research findings. This work lays the groundwork for future AI systems that can autonomously engage in scientific exploration and innovation.
The Objective Is the Bottleneck: Latent World Models Encode What Their Planners Cannot Use
Reinforcement Learning
Robotics
Optimization
- The predictor's capability is not the limiting factor in long-horizon planning.
- The planning objective can saturate and lead to counterintuitive results, such as moving away from the goal lowering costs.
- A learned cost function must be trained on the distribution relevant to the planner's scoring.
- Reachability is a more effective objective than proximity for planning tasks.
Read more
The Objective Is the Bottleneck: Latent World Models Encode What Their Planners Cannot Use
Summary
This paper investigates the limitations of latent world models in planning tasks, specifically focusing on the LeWorldModel in the TwoRoom environment. The author argues that the bottleneck in planning is not the predictive capability of the model but rather the planning objective itself. Through experiments, it is shown that the planner's objective saturates and can even lead to suboptimal planning decisions, where moving away from the goal can reduce the planner's cost. The study reveals that the predictor remains informative over long horizons, while the planner only utilizes a short horizon of twenty-five steps. The paper also discusses the dissociation between prediction accuracy and long-horizon planning success, demonstrating that a learned cost function must be aligned with the planner's scoring distribution. The findings suggest that reachability, rather than proximity, should be the focus of the planning objective, leading to significant improvements in goal achievement without retraining the model. Overall, the research highlights the importance of the planning objective in latent world models and provides insights into enhancing their performance.
Methodology
The study involved reproducing the LeWorldModel on the TwoRoom environment, using a joint-embedding predictive architecture with a cross-entropy method for planning. The experiments were conducted using the released checkpoints and evaluation protocols, focusing on the relationship between prediction accuracy and planning success over varying horizons.
Results
The results showed that the predictor maintained a low error rate over long horizons, while the planner's effectiveness diminished beyond twenty-five steps. By replacing the planning objective, the success rate of reaching goals at a distance of 100 steps improved dramatically from 26.0% to 98.0%. Additionally, the study found that a learned cost function performed poorly when not aligned with the planner's scoring distribution.
Implications
These findings suggest that improving latent world models requires a reevaluation of the planning objectives used. By focusing on reachability rather than proximity, models can achieve better performance in planning tasks. This research could influence future work in reinforcement learning and robotics, where planning and decision-making are critical.
CAKE: Compiler-Agent Co-Design for Frontier Kernel Evolution
Optimization
Efficient ML
- CAKE integrates compiler and agent co-design to improve GPU kernel evolution.
- The framework uses a typed intermediate representation (Cake IR) for better hardware control.
- Localized correctness and performance diagnostics are provided to agents for informed decision-making.
- Significant performance improvements were observed in benchmarks compared to traditional CUDA/PTX implementations.
Read more
CAKE: Compiler-Agent Co-Design for Frontier Kernel Evolution
Summary
The paper introduces CAKE, a novel framework that integrates compiler and agent co-design to enhance the evolution of GPU kernels. Traditional approaches treat the compiler as a black box, limiting the feedback available to agents working on kernel optimization. CAKE addresses this by allowing agents to work with a typed intermediate representation (Cake IR) that provides detailed insights into hardware-specific scheduling and performance diagnostics. This enables agents to make informed decisions about kernel modifications, leading to improved performance and correctness. The framework evolves both the kernels and the compiler, allowing for continuous improvement based on feedback from the execution environment. The authors demonstrate CAKE's effectiveness through various benchmarks, showing significant performance gains over traditional methods, particularly in the context of complex workloads and kernel families. The results indicate that CAKE can bridge the gap between expert kernel programming and automated kernel generation, making it a valuable tool for optimizing GPU performance.
Methodology
The authors developed Cake IR, a hardware-specific intermediate representation that allows agents to edit and optimize GPU kernels with fine-grained control. The framework incorporates structured feedback from the compiler, enabling agents to receive detailed diagnostics on correctness and performance. The methodology includes a bottom-up evolution of Cake IR based on production kernels, allowing for iterative improvements and validation against established benchmarks.
Results
In experiments, CAKE achieved a median performance of 1.144× over the tuned FlashML baseline with Cake IR, compared to 0.928× for direct CUDA/PTX. Additionally, the Kimi Delta Attention implementation showed a 2.05× geometric-mean speedup over the official FlashKDA. Dispatcher-backed KNN and KMeans families demonstrated performance improvements ranging from 1.42× to 2.12× across various shapes, with four kernel changes made available as upstream pull requests.
Implications
CAKE has the potential to significantly enhance the efficiency of GPU programming by automating kernel optimization while maintaining high performance and correctness. This framework could be applied in various domains requiring intensive computation, such as machine learning and scientific computing, where optimized GPU performance is critical.
Learning Under Treatment-Induced Label Indeterminacy with Expert Annotations of Counterfactual Outcomes: A Case Study in Neurological Prognostication
Theory
- Introduces the concept of treatment-induced label indeterminacy in clinical prediction models.
- Proposes a framework for evaluating prediction models that separates certain and uncertain cases.
- Develops a prediction model that balances accuracy on certain cases with alignment to expert estimates for uncertain cases.
- Demonstrates that traditional evaluation metrics can miss important insights in uncertain cases.
Read more
Learning Under Treatment-Induced Label Indeterminacy with Expert Annotations of Counterfactual Outcomes: A Case Study in Neurological Prognostication
Summary
This paper addresses the challenge of developing clinical prediction models in scenarios where treatment decisions render outcomes unobservable, specifically in post-cardiac-arrest neurological prognostication. The authors analyze a cohort of 2,497 patients, among whom 1,429 had indeterminate outcomes due to treatment limitations. Independent clinical experts provided estimates of counterfactual outcomes for these uncertain cases. The study proposes a novel framework for evaluating prediction models that distinguishes between certain cases (with observable outcomes) and uncertain cases (with expert estimates). A prediction model is developed that incorporates both types of target labels, allowing for a trade-off between accuracy on certain cases and alignment with expert estimates for uncertain cases. The findings reveal that conventional evaluation metrics may overlook critical failure modes in uncertain cases, emphasizing the need for tailored evaluation strategies in clinical settings where treatment decisions impact outcome observability.
Methodology
The authors employed a binary classification approach to predict favorable versus unfavorable outcomes in post-cardiac-arrest patients. They utilized expert assessments as imperfect reference signals for uncertain cases and developed a model that integrates both certain and uncertain case labels. The evaluation framework explicitly separates the performance metrics for both types of cases.
Results
The proposed model achieved similar AUROC scores for certain cases compared to baseline models but showed significant differences in Brier scores and probability estimates for uncertain cases. The results indicated a trade-off between improving alignment with uncertain case labels and accuracy on certain cases, revealing that standard evaluation metrics could obscure critical performance issues.
Implications
This research has significant implications for clinical decision-making, particularly in high-stakes environments like post-cardiac arrest care. By improving the evaluation of prediction models in the presence of treatment-induced label indeterminacy, clinicians can better understand the risks and benefits of treatment options, ultimately leading to improved patient outcomes.
Represent, Then Generate: Multimodal-Conditioned Time-Series Generation under Irregular Missingness
Generative Models
Time Series
Multimodal
- ReCoGen effectively handles multimodal physiological data with irregular missingness.
- The framework decouples condition representation from target generation, enhancing performance.
- Ablation studies reveal the importance of learnable cross-attention and dual token-plus-AdaLN routes.
- ReCoGen surpasses existing methods in generating physiological signals, achieving high utility.
Read more
Represent, Then Generate: Multimodal-Conditioned Time-Series Generation under Irregular Missingness
Summary
The paper addresses the challenge of generating continuous physiological time series in clinical settings where many informative signals are either invasive, expensive, or unavailable. Existing conditional generation methods struggle with multimodal data characterized by irregular missingness and heterogeneous modalities. The authors propose a novel framework called ReCoGen (Represent Conditions, then Generate), which consists of two stages: (1) a masked autoencoder for each modality that transforms time-variant conditions into compact, missingness-tolerant token sequences, and (2) a flow-matching generator that synthesizes the target signal by integrating these tokens with static conditions. The proposed method is evaluated across three physiological benchmarks, demonstrating superior performance compared to six existing conditional generators. Notably, ReCoGen achieves or exceeds the utility of real signals in thirteen out of sixteen evaluation settings. The findings indicate that effective representation of conditions is crucial for successful target generation, paving the way for less invasive and lower-cost continuous clinical monitoring.
Methodology
The methodology involves a two-stage framework: Stage I employs masked autoencoders for each modality to create robust token sequences from time-variant conditions, while Stage II utilizes a flow-matching generator to synthesize the target signal by integrating these tokens with static conditions. This design allows for effective handling of missingness and modality-specific characteristics.
Results
ReCoGen outperformed six representative conditional generators across three physiological benchmarks, achieving the best downstream utility in all sixteen (dataset, task, metric) settings. In thirteen of these settings, the utility reached or exceeded that of real signals, indicating the effectiveness of the proposed method.
Implications
The findings suggest that ReCoGen can significantly enhance clinical monitoring by providing informative surrogates for invasive or unavailable signals. This could lead to improved patient care through more accessible and cost-effective monitoring solutions.
When Local Variance Optimality Is Not Enough: RoPE-Aligned Q/K Rotations for Dynamic 4-Bit Quantisation
NLP
Large Language Models
Optimization
- The only commuting orthogonal maps for distinct RoPE frequencies are independent pairwise rotations.
- The derived rotation angle minimizes channel variance but does not improve quantization accuracy in practice.
- The head-shared pairwise configuration results in higher perplexity compared to full-head mixing.
- Estimating the shared angle from K alone improves performance but does not close the gap with full-head mixing.
Read more
When Local Variance Optimality Is Not Enough: RoPE-Aligned Q/K Rotations for Dynamic 4-Bit Quantisation
Summary
This paper investigates the effectiveness of rotation-based post-training quantization methods, specifically focusing on RoPE (Rotary Position Embedding) aligned Q/K rotations in the context of dynamic 4-bit quantization. The authors explore whether a transformation that respects the frequency pair decomposition of RoPE can enhance quantization accuracy compared to full-head mixing. They establish that for distinct frequencies, the only commuting orthogonal maps are independent pairwise rotations. The authors derive a rotation angle that minimizes channel variance under a pooled-covariance surrogate and verify that their implementation achieves this minimum. However, their experiments reveal that this head-shared pairwise configuration does not improve accuracy and, in fact, increases perplexity across various checkpoints. The findings suggest that optimality in a structured surrogate does not necessarily lead to reduced quantization error when misaligned with the quantizer's scale-setting statistic.
Methodology
The authors derive a closed-form rotation angle that minimizes the larger channel variance using a pooled-covariance, position-averaged surrogate. They conduct experiments on various checkpoints, comparing the performance of head-shared pairwise configurations against full-head Hadamard mixing under dynamic W4A4KV4 quantization.
Results
The experiments show that the head-shared pairwise configuration consistently increases perplexity across all evaluated contexts compared to full-head mixing. While estimating the shared angle from K improves performance, it does not eliminate the performance gap relative to full-head mixing. Additionally, broader mixing support is associated with lower quantization error and perplexity degradation.
Implications
These findings highlight the limitations of local variance optimality in quantization methods and suggest that careful consideration of mixing support and alignment with quantization statistics is crucial for improving model performance in low-bit quantization scenarios.
Perturbation-based Regional Interpretability through Subtraction Mapping (PRISM): naming-error dissociations in language models and post-stroke aphasia
Interpretability
Large Language Models
NLP
- PRISM adapts subtraction analysis from neuroimaging to interpret LLMs, providing a structured framework for mechanistic interpretability.
- The framework demonstrates that perturbation-induced error profiles in LLMs can be compared to lesion patterns in patients with aphasia.
- Both LLMs and aphasia patients show a robust phonemic-favoring dissociation, indicating shared cognitive processing patterns.
- The methodology allows for spatially resolved testing of functional specialization claims in LLMs.
Read more
Perturbation-based Regional Interpretability through Subtraction Mapping (PRISM): naming-error dissociations in language models and post-stroke aphasia
Summary
This paper introduces PRISM (Perturbation-based Regional Interpretability through Subtraction Mapping), a novel framework for interpreting large language models (LLMs) by applying subtraction analysis, a method commonly used in human neuroimaging, to perturbed transformers. The authors aim to bridge the interpretability gap in LLMs by providing a spatially resolved and falsifiable approach to test whether internal components of LLMs are specialized for distinct cognitive operations. Building on the Brain-LLM Unified Model (BLUM), which demonstrated that error profiles from a perturbed LLM correspond to lesion patterns in aphasic patients, PRISM maps the Philadelphia Naming Test categories and conducts pairwise subtractions of error classes. This analysis is run in parallel with a study of 213 chronic post-stroke aphasia patients, using correlation-difference lesion-symptom mapping. The results reveal a robust phonemic-favoring dissociation in both LLMs and human patients, indicating that deep layer clusters in the LLM and frontal-perisylvian cortical clusters in patients replicate consistently. The semantic-favoring direction shows a non-significant trend. PRISM thus offers a new method for testing functional specialization in transformer models, with further confirmatory interventions suggested for future work.
Methodology
The authors developed PRISM by applying subtraction analysis to LLM perturbation data and conducting parallel analyses on aphasia patients. They utilized the Philadelphia Naming Test to categorize errors and performed pairwise subtractions of these error classes. The analysis involved correlation-difference lesion-symptom mapping on patient data, with both LLM and patient analyses structured to match in subject and spatial dimensions.
Results
The study found a significant phonemic-favoring dissociation in both the LLM and the patient cohort, with consistent replication of deep layer clusters in the LLM and frontal-perisylvian cortical clusters in patients. The semantic-favoring direction was observed as a non-significant trend, suggesting further exploration is needed.
Implications
PRISM provides a new avenue for understanding the cognitive operations underlying LLMs, potentially leading to improved interpretability and trust in AI systems. It may also facilitate better alignment between AI models and human cognitive processes, with applications in both AI development and clinical neuroscience.
Federated Compositional Muon Optimizer for Matrix-Wise Models
Federated Learning
Optimization
- Introduction of FedCoMuon and FedCoMuon-VR optimizers for matrix-wise compositional optimization.
- Theoretical convergence analysis under non-convex and non-i.i.d. settings.
- FedCoMuon-VR achieves lower sample complexity than existing FedMuon algorithms.
- Extensive experiments demonstrate competitive performance and improved accuracy.
Read more
Federated Compositional Muon Optimizer for Matrix-Wise Models
Summary
This paper introduces the Federated Compositional Muon (FedCoMuon) optimizer, designed to address the challenges of distributed matrix-wise compositional optimization in federated learning settings. The authors identify a gap in existing Muon-based optimizers, which are primarily tailored for single-level stochastic objectives, and propose FedCoMuon to effectively handle hierarchical structured problems. The FedCoMuon optimizer integrates compositional gradient tracking with orthogonalized momentum, while a variance-reduced variant, FedCoMuon-VR, is also introduced, leveraging a momentum-based variance reduction technique. The paper provides a theoretical analysis of the convergence properties of both algorithms under non-i.i.d. and non-convex conditions, demonstrating that FedCoMuon-VR achieves a lower sample complexity of O(ϵ−3) for finding an ϵ-stationary solution compared to existing FedMuon algorithms. Empirical evaluations on robust federated learning and task-distributed risk-sensitive meta learning show that the proposed methods outperform existing compositional baselines, achieving superior accuracy in various scenarios.
Methodology
The FedCoMuon optimizer combines compositional gradient tracking with orthogonalized momentum to optimize matrix-wise models in a federated learning context. The FedCoMuon-VR variant incorporates a momentum-based variance reduction technique to enhance performance. The algorithms are analyzed theoretically for their convergence properties, focusing on sample and communication complexities.
Results
The proposed FedCoMuon and FedCoMuon-VR optimizers show significant improvements in sample complexity, with FedCoMuon-VR achieving O(ϵ−3) complexity for finding ϵ-stationary solutions. Numerical experiments indicate that these optimizers outperform existing compositional baselines, achieving the best reported accuracy in several settings, particularly in robust federated learning and task-distributed meta learning.
Implications
The development of FedCoMuon and FedCoMuon-VR has potential applications in privacy-sensitive and resource-constrained environments, such as mobile intelligence, healthcare, and recommendation systems, where federated learning is crucial. The ability to effectively optimize matrix-wise models in these contexts can lead to more robust and efficient AI systems.
Towards Socially Compliant Navigation in Deep Reinforcement Learning via Proxemics-Based Reward Modeling
Reinforcement Learning
Robotics
- Introduction of a proxemics-based reward model for DRL navigation.
- Validation of the model across multiple DRL methods and crowd scenarios.
- Demonstrated improvements in social metrics without sacrificing navigation efficiency.
- Emphasis on the importance of comfort-aware navigation assessment.
Read more
Towards Socially Compliant Navigation in Deep Reinforcement Learning via Proxemics-Based Reward Modeling
Summary
This paper addresses the challenge of robot navigation in crowded environments, emphasizing the need for social compliance alongside task efficiency. Traditional deep reinforcement learning (DRL) methods often prioritize navigation tasks without adequately considering social interactions, leading to behaviors that may be collision-free but socially inappropriate. The authors propose a novel proxemics-based reward model that incorporates Hall's proxemics theory to define personal space as a radial Gaussian-mixture field. This model generates a robot-centric local cost that encourages robots to minimize proxemic intrusion while navigating efficiently. The proposed reward is integrated into established DRL navigation methods and evaluated through simulations across various crowd scenarios. The results demonstrate that the proxemics-based reward consistently enhances social metrics while maintaining competitive navigation performance compared to existing reward models. The paper also provides insights into reward design factors critical for improving social navigation in robotic systems.
Methodology
The authors developed a proxemics-based reward model grounded in Hall's theory of interpersonal distance, representing personal space as a radial Gaussian-mixture field. This model computes a local proxemic cost within the robot's field of view and is integrated into existing DRL navigation frameworks. The evaluation involved simulations across various crowd densities and scenarios, comparing the proposed reward against established baselines using both navigation and social metrics.
Results
The proposed proxemics-based reward model consistently improved social metrics in simulation experiments while maintaining competitive navigation performance compared to other reward formulations. The results highlighted the effectiveness of the model in promoting smoother and more socially compliant robot behaviors in dense crowds.
Implications
The findings suggest that incorporating social compliance into robot navigation systems can enhance human-robot interactions in crowded environments, making robots more acceptable and effective in real-world applications such as malls, hospitals, and airports. This work lays the groundwork for future research in socially aware robotic navigation.
Incremental Evaluation and Training in Relational Deep Learning
Graph Learning
Time Series
Theory
- Introduces an incremental, multi-episode evaluation paradigm for RDL.
- Demonstrates the prevalence of temporal concept drift in RDL tasks.
- Presents multiple effective incremental training strategies for model fine-tuning.
- Proposes a new temporal evaluation metric focusing on near-future accuracy.
Read more
Incremental Evaluation and Training in Relational Deep Learning
Summary
This paper addresses the limitations of current evaluation practices in Relational Deep Learning (RDL), which typically rely on static, single-episode dataset snapshots. The authors propose an incremental, multi-episode evaluation and training paradigm that reflects the dynamic nature of real-world databases. By modeling multi-tabular databases as temporal heterogeneous graphs, the paper demonstrates that model performance can degrade over time due to temporal concept drift. The authors empirically validate their approach using large-scale datasets, revealing that temporal concept drift is prevalent in RDL tasks. They introduce multiple incremental training regimes for fine-tuning models and show that transfer learning is effective in this context. Additionally, a new temporal evaluation metric prioritizing near-future accuracy is proposed, leading to consistently better performance of incrementally fine-tuned models compared to traditional from-scratch training methods. This work aims to enhance the robustness and adaptability of RDL models in real-world applications.
Methodology
The authors developed an incremental evaluation framework that simulates the continuous growth of relational databases. They employed established large-scale datasets to analyze data evolution and model training dynamics, testing various incremental fine-tuning strategies and introducing a new evaluation metric that emphasizes near-future predictive accuracy.
Results
The study found that temporal concept drift significantly affects model performance in RDL tasks. The proposed incremental training regimes demonstrated that models could be effectively fine-tuned with new data, leading to better performance than models trained from scratch. The new evaluation metric also provided a more relevant assessment of model accuracy in practical applications.
Implications
This work has significant implications for the deployment of RDL models in real-world scenarios, where databases are continuously evolving. The proposed methodologies can enhance the adaptability and robustness of predictive models, making them more suitable for dynamic environments across various industries.
Exemplar-based objective classification of gust-induced loads across multiple flight conditions
Robotics
Theory
Interpretability
- Introduces an exemplar-based classification method for gust-induced loads in UAVs.
- Demonstrates the identification of nine fundamental response types from a large dataset.
- Compares the new classification approach with traditional parameter-based methods.
- Enhances understanding of fluid mechanics related to gust-wing interactions.
Read more
Exemplar-based objective classification of gust-induced loads across multiple flight conditions
Summary
This paper addresses the challenge of classifying gust-induced loads on small uncrewed aerial vehicles (UAVs) across various flight conditions. Traditional methods of classifying these loads based on human-defined parameters, such as flight attitude, may not adequately capture the complexity of gust interactions. The authors propose an exemplar-based classification approach that utilizes machine learning to encode a large dataset of experimental observations. By applying a summarization procedure, they identify a minimal subset of significant exemplars that serve as a basis for classification. The study analyzes a database of 3480 pressure-load measurements from a flying-wing model subjected to random gusts across six flight attitudes. The authors identify nine fundamental response types that recur across different attitudes, providing insights into the underlying fluid mechanics. This approach not only enhances the interpretability of gust-load classifications but also aids in guiding further experimental investigations.
Methodology
The authors employed a machine learning approach to analyze a large dataset of gust-induced pressure-load measurements. They utilized a summarization procedure to extract a minimal set of significant exemplars, which were then used to classify the gust-load responses objectively. The classification was compared to traditional methods based on flight attitude.
Results
The analysis revealed nine distinct fundamental response types that were consistent across multiple flight attitudes. The exemplar-based classification provided a more interpretable and organized criterion for understanding gust-induced loads compared to traditional parameter-based classifications.
Implications
The findings suggest that the exemplar-based classification method can improve the understanding of gust-induced loads in UAVs, potentially leading to better design and operational strategies for small aircraft in turbulent environments. This approach may also inform future research in fluid mechanics and aerodynamics.
Understanding Backdoor Vulnerabilities in Vertical Federated Learning: The Gap Between Research and Practice
Federated Learning
- Identifies a gap between theoretical research on backdoor attacks in VFL and practical applications.
- Highlights unrealistic assumptions in existing methodologies that lead to overestimated attack success rates.
- Introduces BVBench, a benchmark for fair evaluation of backdoor vulnerabilities in VFL.
- Recommends redefining threat models to align with realistic operational constraints.
Read more
Understanding Backdoor Vulnerabilities in Vertical Federated Learning: The Gap Between Research and Practice
Summary
This paper investigates the vulnerabilities associated with backdoor attacks in Vertical Federated Learning (VFL), highlighting a significant gap between theoretical research and practical application. VFL allows organizations to collaborate on machine learning tasks without sharing sensitive data, but this information asymmetry can be exploited by malicious participants. The authors argue that existing studies on backdoor attacks often rely on unrealistic assumptions and poorly designed evaluation practices, leading to an overestimation of the effectiveness of current defenses. To address these issues, the paper proposes a redefined threat model that accounts for realistic constraints and introduces BVBench, a benchmark for evaluating backdoor vulnerabilities in VFL. The findings reveal that many existing attack strategies may not be as effective in real-world scenarios, emphasizing the need for more robust defenses and a better understanding of practical vulnerabilities in VFL systems.
Methodology
The authors conducted a systematic review of existing literature on backdoor attacks in VFL, identifying methodological and evaluation design flaws. They proposed new threat models and developed BVBench to facilitate comprehensive and practical evaluations of backdoor vulnerabilities.
Results
The study found that existing attack strategies often rely on unrealistic assumptions, leading to inflated success rates in controlled environments. The introduction of BVBench revealed the fragility of current understandings of VFL backdoor risks, demonstrating that many attacks are less effective under realistic conditions.
Implications
The findings suggest that practitioners may be underestimating the risks associated with VFL systems, potentially leading to inadequate defenses. The proposed BVBench can serve as a foundation for future research, guiding the development of more effective security measures in federated learning environments.
Large-scale Testing Global Optimization Methods with Black-box Adversarial Attacks
Optimization
- Existing global optimization benchmarks are outdated and limited in scope.
- Black-box adversarial attacks can serve as effective benchmarks for global optimization methods.
- The study evaluates various evolutionary algorithms and metaheuristics for solving BBAA problems.
- The findings highlight the need for modern benchmarks that reflect real-world optimization challenges in machine learning.
Read more
Large-scale Testing Global Optimization Methods with Black-box Adversarial Attacks
Summary
This paper addresses the limitations of existing global optimization benchmark suites, which are often small and based on outdated analytical functions. The authors propose that black-box adversarial attack (BBAA) tasks can serve as a valuable benchmark for global optimization in high-dimensional spaces. They evaluate the effectiveness of various evolutionary algorithms and metaheuristics in solving BBAA problems, demonstrating a significant step towards aligning global optimization methods with the current challenges in machine learning. The paper discusses the evolution of global optimization benchmarks, highlighting the inadequacies of traditional suites and the need for more diverse and modern testing scenarios. The authors emphasize the importance of adversarial attacks in assessing the robustness of machine learning models and suggest that these attacks can provide a rich set of optimization problems that reflect real-world complexities.
Methodology
The authors conducted experiments using various evolutionary algorithms and metaheuristics to solve black-box adversarial attack problems. They analyzed the performance of these methods in high-dimensional optimization scenarios, comparing their effectiveness against traditional benchmarks.
Results
The results indicate that evolutionary algorithms and metaheuristics can effectively tackle BBAA problems, showcasing their potential as robust optimization methods. The study provides evidence that these approaches can adapt to the complexities of modern machine learning tasks, suggesting a shift in how global optimization is benchmarked.
Implications
The findings suggest that integrating black-box adversarial attacks into global optimization benchmarks can enhance the development of more resilient machine learning models. This approach may lead to improved methodologies for hyperparameter tuning and robustness testing in various applications, particularly in computer vision and other fields reliant on machine learning.
Symmetry-Breaking De Novo Crystal Generation via Markovian Jump Diffusion
Generative Models
- Introduces a novel framework (SbCD) for generating complete crystallographic structures.
- Utilizes Markovian jump-diffusion to model symmetry-breaking dynamics.
- Outperforms existing models in generating crystals with full structural specifications.
- Addresses limitations of traditional methods that rely on empirical sampling of space groups.
Read more
Symmetry-Breaking De Novo Crystal Generation via Markovian Jump Diffusion
Summary
This paper presents a novel framework for generating crystalline structures, addressing limitations in existing generative models that fail to produce complete crystallographic specifications. The authors introduce the Symmetry-breaking Crystal Diffusion (SbCD) model, which utilizes a Markovian jump-diffusion process to facilitate symmetry-breaking dynamics in crystal generation. By reversing from low-symmetry priors, SbCD effectively captures global symmetry and structural dependencies, which are often overlooked in traditional methods that sample space groups from empirical distributions. The framework is inspired by spontaneous symmetry breaking in physics and is designed to model inter-space-group transitions. The authors demonstrate that their approach significantly outperforms existing symmetry-preserving models in de novo generation experiments, showcasing its potential for advancing the field of materials science and accelerating crystal discovery.
Methodology
The authors develop a theoretical variational bound objective that unifies structural dependencies among crystal components. They leverage a Markovian jump-diffusion process to model space-group distributions, allowing for the simulation of symmetry-breaking processes. The framework adapts to both continuous and discrete state spaces, enforcing space-group constraints and enabling the generation of stable crystal structures.
Results
In de novo generation experiments on datasets MP20 and MPTS-52, the SbCD model demonstrated substantial improvements over symmetry-preserving counterparts, effectively generating complete crystallographic specifications, including Wyckoff positions and space groups, which are crucial for accurate crystal representation.
Implications
The SbCD framework has significant implications for materials science, potentially accelerating the discovery of new crystalline materials with desired properties. By providing a more accurate generative modeling approach, it could lead to advancements in various applications, including pharmaceuticals, semiconductor design, and energy storage technologies.
Decentralized Multi-Player Q-Learning in Episodic Markov Decision Processes with Information Asymmetry
Reinforcement Learning
Theory
Robotics
- The paper addresses decentralized multi-player reinforcement learning in episodic MDPs with information asymmetry.
- Three forms of information asymmetry are explored, leading to the development of tailored algorithms for each scenario.
- The proposed algorithms achieve regret bounds that are competitive with centralized benchmarks, indicating effective coordination without communication.
- The results highlight the exponential growth of regret bounds with the number of players, emphasizing the challenges in multi-agent settings.
Read more
Decentralized Multi-Player Q-Learning in Episodic Markov Decision Processes with Information Asymmetry
Summary
This paper investigates decentralized multi-player reinforcement learning within episodic tabular Markov Decision Processes (MDPs) that exhibit information asymmetry. The authors categorize the problem into three distinct scenarios: (A) unobserved actions with common rewards, (B) observed actions with independent rewards, and (C) unobserved actions with independent rewards. The study introduces two main algorithms, mQ-learning and mQ-learning-intervals, which achieve a regret bound of ˜O(√H4SAjoint T) for Problems A and B. For Problem C, the authors propose mEXC and mEXC-Bellman, which are explore-then-commit algorithms yielding a regret of ˜O(H(SAjoint)1/3T 2/3). The results demonstrate that decentralized learning under information asymmetry can match the performance of centralized joint-action Q-learning, with regret bounds that grow exponentially with the number of players, M. The findings suggest that while asymmetry does not impose additional costs beyond the centralized rate, it remains a significant factor in the complexity of multi-agent systems.
Methodology
The authors develop decentralized algorithms for multi-player reinforcement learning that operate under three models of information asymmetry. They utilize a deterministic protocol for implicit coordination among players, allowing them to learn effectively without direct communication. The algorithms are designed to maintain upper and lower confidence bounds for action selection and to coordinate action elimination across players.
Results
The algorithms mQ-learning and mQ-learning-intervals achieve a regret of ˜O(√H4SAjoint T) for Problems A and B, while mEXC and mEXC-Bellman yield a regret of ˜O(H(SAjoint)1/3T 2/3) for Problem C. The results indicate that decentralized learning can match the performance of centralized Q-learning algorithms, with regret bounds that are logarithmically related to the joint action space size.
Implications
The findings have significant implications for multi-agent systems in various applications, such as communication networks, multi-robot coordination, and distributed resource allocation, where agents must learn and coordinate without centralized control. The results suggest that effective learning is possible even in the presence of information asymmetry, which is common in real-world scenarios.
Novel Knowledge-Guided Generative Methods for Synthetic Transcriptomic Data
Generative Models
Graph Learning
- Synthetic data generation can mitigate challenges in accessing high-quality transcriptomic data.
- MK-TGAN, a novel GAN variant, effectively integrates biological knowledge through graph neural networks.
- Incorporating prior biological knowledge improves the realism and utility of synthetic transcriptomic data.
- The study highlights the importance of structured biological relationships in generative modeling.
Read more
Novel Knowledge-Guided Generative Methods for Synthetic Transcriptomic Data
Summary
This paper addresses the challenges in generating synthetic transcriptomic data, which is crucial for biomedical research but often limited by data quality issues such as imbalances and biases. The authors present a comparative analysis of generative models that incorporate biological knowledge through gene graphs to enhance the realism and utility of synthetic data. They introduce three variants of Generative Adversarial Networks (GANs), including the novel Multi-Kernel Transcriptional Generative Adversarial Network (MK-TGAN), which utilizes multiple graph neural network kernels to explicitly integrate biological knowledge. The study evaluates these models on breast cancer transcriptomic data, focusing on Epithelial-to-Mesenchymal Transition (EMT)-related genes. Results indicate that MK-TGAN outperforms other models in generating synthetic data that is both realistic and biologically plausible, demonstrating the effectiveness of integrating prior biological knowledge into generative frameworks.
Methodology
The authors developed three GAN variants: Graph-Modulated GAN (GM-GAN), Graph-Regularized GAN (GR-GAN), and MK-TGAN. GM-GAN and GR-GAN integrate biological knowledge implicitly, while MK-TGAN explicitly utilizes multiple graph neural network kernels to enhance data generation. The models were benchmarked using breast cancer transcriptomic data focused on EMT-related genes to assess their performance in generating realistic and biologically relevant synthetic data.
Results
MK-TGAN consistently demonstrated superior performance compared to other generative models in terms of precision, recall, and downstream biological tasks. The integration of biological knowledge through graph neural networks significantly improved the fidelity, diversity, and biological coherence of the generated synthetic samples.
Implications
The findings suggest that incorporating biological knowledge into generative models can enhance the quality of synthetic transcriptomic data, which may facilitate research in genomics and personalized medicine. This approach could help overcome ethical and legal constraints associated with real patient data, enabling broader access to high-quality datasets for machine learning applications in biomedical research.
Personalized Scorer Modeling: A Learning-Based Framework for Deriving Robust Sleep Stage Labels from Multiple Experts
Time Series
Multimodal
- Introduces a framework for improving sleep stage classification by addressing inter-scorer variability.
- Utilizes multi-scored datasets to derive more reliable sleep stage labels.
- Employs confusion matrices to model scorer-specific behavior and aggregate probabilities for labeling.
- Demonstrates improved classification metrics compared to traditional hypnograms.
Read more
Personalized Scorer Modeling: A Learning-Based Framework for Deriving Robust Sleep Stage Labels from Multiple Experts
Summary
This paper addresses the challenge of sleep stage classification, which is crucial for diagnosing and treating sleep disorders. Traditional methods often rely on a single reference hypnogram, which can lead to inaccuracies due to inter-scorer variability. The authors propose a novel framework called Personalized Scorer Modeling (PSM) that leverages multi-scored datasets, where multiple experts annotate the same polysomnographic recordings. By analyzing two publicly available datasets (DOD-H and DOD-O), the authors extract 30 features from EEG and EMG signals and develop a learning-based hypnogram (LBH) that models the behavior of each scorer using confusion matrices. These matrices are normalized to estimate the probability of each sleep stage based on the labels assigned by the scorers, which are then aggregated to produce a final label for each epoch. The performance of the LBH is evaluated using random forest, support vector machine, and multilayer perceptron classifiers, demonstrating significant improvements over traditional methods. The study concludes that personalized scorer modeling enhances the reliability of sleep stage labels without disregarding individual expert contributions.
Methodology
The study analyzes two multi-scored sleep datasets (DOD-H and DOD-O) and extracts 30 features from EEG and EMG signals. A learning-based hypnogram is generated by modeling the behavior of each scorer through confusion matrices, which are normalized to estimate stage-specific probabilities. These probabilities are aggregated to determine the final sleep stage label for each epoch. The proposed method is evaluated using various classifiers under different signal configurations.
Results
The proposed LBH achieved an accuracy of 86.07%, precision of 85.46%, and F1-score of 85.29% on the DOD-H dataset, and 86.04% accuracy, 85.21% precision, and 84.70% F1-score on the DOD-O dataset. These results indicate a consistent improvement over the original dataset hypnogram and the best-scorer hypnogram.
Implications
The findings suggest that personalized scorer modeling can enhance the construction of reference hypnograms in multi-scored sleep datasets, leading to more reliable automatic sleep staging systems. This approach could improve clinical practices and research in sleep medicine by providing more accurate sleep stage classifications.
A Multispectral Framework for the Detection of Calcium Carbide-Induced Ripening and Shelf-Life Estimation in Climacteric Fruits
Multimodal
- Development of a non-invasive multispectral framework for detecting CaC2-induced ripening in fruits.
- Utilization of visible-near infrared spectroscopy to analyze spectral profiles of mango and banana.
- Integration of feature engineering and PCA for effective data representation.
- High classification accuracy achieved with XGBoost algorithms for ripening method classification.
Read more
A Multispectral Framework for the Detection of Calcium Carbide-Induced Ripening and Shelf-Life Estimation in Climacteric Fruits
Summary
This study addresses the health risks associated with the use of industrial-grade Calcium Carbide (CaC2) for ripening climacteric fruits such as mango and banana. The authors propose a non-invasive multispectral framework that utilizes visible-near infrared spectroscopy to differentiate between naturally ripened fruits and those ripened with CaC2, while also estimating their ripening progression and remaining shelf life. The research employs the AS7265x spectral triad sensor to analyze the spectral profiles of mango and banana at 18 discrete wavelengths. The findings reveal that CaC2-treated samples exhibit distinct spectral intensity drops, indicative of accelerated chlorophyll degradation. A feature engineering strategy is implemented, incorporating spectral variance, intensity ratios, and environmental parameters. Principal Component Analysis (PCA) is used for dimensionality reduction, retaining over 90% of spectral variance. The study trains three eXtreme Gradient Boosting (XGBoost) algorithms for classification and estimation tasks, achieving a classification accuracy of 95% for mango and 81% for banana, with notable recall rates for the CaC2 class. This framework demonstrates a promising approach for non-invasive detection of harmful ripening practices in fruits.
Methodology
The study employs visible-near infrared spectroscopy using the AS7265x spectral triad sensor to capture spectral profiles of mango and banana. A feature engineering strategy is applied, integrating spectral variance, intensity ratios, and environmental factors. PCA is utilized for dimensionality reduction, followed by training three XGBoost algorithms for classification and quantitative estimation of ripening progression and shelf life.
Results
The framework achieved a classification accuracy of 95% for mango samples and 81% for banana samples. The recall for the CaC2 class was 0.67 for mango and 0.74 for banana, indicating effective differentiation between ripening methods.
Implications
This research has significant implications for food safety and public health by providing a reliable method to detect harmful ripening practices in fruits, potentially reducing health risks associated with the consumption of chemically ripened produce.
History-informed Lagrangian Neural Networks
Robotics
Time Series
Theory
- HiLNN infers hidden velocities and adapts system parameters from position-only observations.
- The framework utilizes a recurrent encoder to extract a latent context from historical position data.
- HiLNN employs a differentiable RK4 rollout scheme for optimized trajectory predictions.
- Empirical results show superior accuracy and physical consistency compared to traditional LNNs and other baselines.
Read more
History-informed Lagrangian Neural Networks
Summary
The paper presents History-informed Lagrangian Neural Networks (HiLNN), a novel framework aimed at improving the long-horizon forecasting of mechanical systems using only position observations. Traditional Lagrangian Neural Networks (LNNs) require complete state inputs, including velocity, which limits their applicability in scenarios where only position data is available. HiLNN addresses this limitation by leveraging the temporal position sequences to extract a latent context through a recurrent encoder. This context not only estimates the unobserved initial velocity but also adapts the mass matrix, potential energy, and damping coefficients of the Lagrangian system. The authors employ a differentiable RK4 rollout scheme for end-to-end optimization under multi-step trajectory supervision and energy-consistency regularization. Empirical evaluations demonstrate that HiLNN significantly enhances prediction accuracy and maintains energy consistency across various mechanical systems, including conservative, dissipative, and variable-parameter systems, outperforming existing state-of-the-art methods.
Methodology
The HiLNN framework uses a recurrent encoder to derive a latent context from historical position data, which is then used to condition a structured Lagrangian dynamics model. The model is optimized using a differentiable RK4 rollout scheme, allowing for end-to-end training under multi-step trajectory supervision and energy-consistency regularization.
Results
HiLNN demonstrated improved long-term prediction accuracy and maintained precise energy profiles across conservative, dissipative, and variable-parameter systems, outperforming traditional LNNs, Hamiltonian Neural Networks (HNN), Neural ODEs, and MLP predictors in empirical evaluations.
Implications
The development of HiLNN has significant implications for fields requiring accurate mechanical forecasting from limited observations, such as robotics, control systems, and scientific modeling. Its ability to adapt to varying dynamics and maintain physical plausibility can enhance the reliability of predictive models in real-world applications.
Finding the Needle in a Haystack: Test-Time Analog Circuit Representation Adaptation for Bayesian Optimization
Optimization
- TTARO adapts circuit representations in real-time during the optimization process, improving alignment with the optimization objective.
- The framework is compatible with various acquisition functions and Gaussian-process kernels, making it versatile for different optimization scenarios.
- TTARO demonstrates significant performance improvements over traditional fixed-embedding BO methods and DKL in analog circuit topology searches.
Read more
Finding the Needle in a Haystack: Test-Time Analog Circuit Representation Adaptation for Bayesian Optimization
Summary
This paper presents Test-Time Analog Representation Adaptation for Bayesian Optimization (TTARO), a novel online framework that enhances Bayesian optimization (BO) for analog circuit topology search by adapting circuit representations during the optimization process. Traditional representation-based BO methods often use fixed embeddings, which can misalign with the optimization objective, leading to inefficient searches. TTARO addresses this by continuously updating circuit representations based on the figure of merit (FoM) of evaluated circuits. The framework jointly learns a nonlinear feature transformation and a Gaussian-process surrogate model as new FoM observations are gathered. The authors conduct extensive experiments across various circuit encoders, kernels, and acquisition functions, demonstrating that TTARO significantly outperforms conventional BO methods and Deep Kernel Learning (DKL) by aligning the search space with the optimization objective. The results indicate that TTARO reduces regret AUC by an average of 15.2% compared to standard BO and 20.7% compared to DKL, showcasing its effectiveness in optimizing analog circuit designs.
Methodology
TTARO employs an online deep-kernel Bayesian optimization approach that continuously refines circuit representations based on newly observed FoM labels. It integrates a nonlinear feature transformation with a Gaussian-process surrogate model, updating both after each evaluation to enhance the search process.
Results
In extensive experiments involving 160 configurations across two public circuit topology benchmarks, TTARO achieved an average reduction in regret AUC of 15.2% compared to conventional BO and 20.7% compared to DKL, with some configurations showing reductions as high as 46.7%. This indicates a substantial improvement in optimization efficiency.
Implications
The findings suggest that adaptive representation learning can significantly enhance the efficiency of optimization in complex design spaces, particularly in electronic design automation. This approach could be applied to other domains requiring optimization over structured, discrete spaces.
Adaptive $k$ Nearest Neighbors Classifier via Granular Ball Computing
Efficient ML
Theory
- Introduces an adaptive KNN method using granular-ball computing.
- Utilizes the Fisher criterion for effective granular ball partitioning.
- Dynamically determines the effective k value based on local neighborhood structure.
- Demonstrates improved robustness against noise and local perturbations.
Read more
Adaptive $k$ Nearest Neighbors Classifier via Granular Ball Computing
Summary
This paper presents an innovative approach to the k-Nearest Neighbor (KNN) algorithm by introducing an adaptive method based on granular-ball computing. The authors identify the critical issue of selecting the optimal k value, which significantly influences classification performance. The proposed method operates in two stages: during training, the dataset is coarsely partitioned into granular balls to simplify data distributions, and the Fisher criterion is employed to manage the splitting and stopping of these balls, resulting in a multi-granularity representation. In the prediction phase, the nearest granular ball is identified using a weighted distance mechanism, and an adaptive neighborhood is constructed around the test sample. This neighborhood dynamically determines the effective k value based on the actual number of samples present, enhancing robustness against noise and local perturbations. Experimental results demonstrate that the proposed adaptive granular ball KNN (GBKNN) outperforms existing KNN variants in terms of accuracy and efficiency across multiple datasets, showcasing its potential for practical applications in various classification tasks.
Methodology
The proposed GBKNN algorithm involves two main stages: a training stage where the dataset is partitioned into granular balls using the Fisher criterion for adaptive generation, and a prediction stage where the nearest granular ball is located, and an adaptive neighborhood is formed to determine the effective k value based on the local sample distribution.
Results
The experimental results indicate that the GBKNN method significantly enhances classification accuracy and efficiency compared to traditional KNN and its variants, validating the effectiveness of the adaptive approach in handling complex datasets.
Implications
The findings suggest that the adaptive granular ball KNN can be effectively applied in various fields requiring robust classification methods, particularly in high-dimensional and noisy environments, improving decision-making processes in applications such as pattern recognition and speech recognition.
Concept Drift Detection and Adaptive Retraining of Malware Classification Models
Efficient ML
Theory
Time Series
- Concept drift significantly impacts the performance of malware classification models.
- OCSVM outperforms MK-Means and MMD in detecting concept drift.
- Drift-aware retraining improves efficiency while maintaining high accuracy.
- Automated concept drift detection can reduce computational resource demands.
Read more
Concept Drift Detection and Adaptive Retraining of Malware Classification Models
Summary
This paper addresses the issue of concept drift in malware classification models, which occurs when the statistical properties of data change over time, leading to performance degradation. The authors propose two machine learning-based approaches for automated concept drift detection: One-Class Support Vector Machines (OCSVM) and Minibatch K-Means (MK-Means), alongside a comparison with Maximum Mean Discrepancy (MMD). The study evaluates four classification models—Multilayer Perceptron (MLP), Random Forest (RF), Support Vector Machines (SVM), and eXtreme Gradient Boosting (XGB)—across three scenarios: static (no retraining), periodic (constant retraining), and drift-aware (retraining only upon drift detection). The experiments reveal that the drift-aware approach, particularly using OCSVM, achieves classification accuracy comparable to periodic retraining while significantly reducing the number of retraining instances. The findings suggest that the proposed techniques are efficient, practical, and can be automated, providing a robust solution for maintaining malware detection performance in the face of evolving threats.
Methodology
The authors conducted a series of experiments comparing three concept drift detection techniques (OCSVM, MK-Means, and MMD) across four classification models (MLP, RF, SVM, XGB) in three scenarios: static, periodic retraining, and drift-aware retraining. The effectiveness of the models was assessed through a total of 400 distinct experiments, focusing on accuracy and training efficiency.
Results
The results indicate that all three concept drift detection techniques achieved classification accuracy similar to that of periodic retraining while requiring fewer retraining instances. The OCSVM-based drift-aware retraining generally outperformed the MK-Means and MMD approaches, demonstrating a favorable tradeoff between accuracy and efficiency.
Implications
The findings suggest that implementing drift-aware retraining in malware classification systems can enhance performance and reduce resource consumption, making it a viable approach for real-world applications in cybersecurity. The automation of the model updating process could lead to more resilient and adaptive malware detection systems.
The Impact of Temporal Context Length and Encoding Strategies on Self-Supervised ECG Representation Learning
Time Series
- Longer temporal contexts (5 and 10 minutes) improve downstream classification and patient-level consistency.
- Continuous convolutional patch embeddings outperform discretized vector-quantized tokens across all evaluated time horizons.
- Discretization may lead to loss of clinically relevant waveform details, impacting diagnostic accuracy.
- The study emphasizes the need for ECG models that integrate extended context and continuous encoding strategies.
Read more
The Impact of Temporal Context Length and Encoding Strategies on Self-Supervised ECG Representation Learning
Summary
This paper investigates the effects of temporal context length and encoding strategies on self-supervised learning (SSL) for electrocardiogram (ECG) representation. The authors conducted a controlled study using the Icentia11k dataset, varying the input horizon (16 seconds, 1 minute, 5 minutes, and 10 minutes) and the front-end representation (continuous convolutional patch embeddings vs. fixed vector-quantized tokens). The study aims to determine whether shorter training windows and discretized tokenization compromise the ability to capture essential rhythm dynamics and patient-specific characteristics. The results indicate that longer temporal contexts (5 and 10 minutes) significantly enhance transfer learning and retrieval accuracy, suggesting that these models better capture slow-varying rhythm dynamics. Additionally, continuous patch embeddings consistently outperform discretized tokens, highlighting the importance of maintaining fine-grained morphological details for clinical applications. The findings advocate for the development of ECG foundation models that prioritize extended context and continuous encoders for improved clinical prediction and patient similarity assessments.
Methodology
The authors performed a controlled study on the Icentia11k dataset, varying the input horizon and representation type while keeping the Transformer backbone and training protocol constant. They evaluated the models based on downstream abnormal rhythm detection and patient-level retrieval metrics.
Results
The study found that increasing the temporal context beyond 16 seconds led to stronger transfer learning and higher retrieval accuracy, with optimal performance at 5 and 10 minutes. Continuous patch embeddings consistently outperformed discretized tokens, indicating that quantization can obscure important waveform details.
Implications
These findings suggest that ECG foundation models should focus on longer temporal contexts and continuous encoders to enhance clinical prediction capabilities and improve patient-specific representation learning.
Prof-K: Probabilistic One-Pass Filtering for Efficient Top-k Selection
Efficient ML
Optimization
Theory
- Prof-K provides a fast, scalable, and distribution-agnostic solution for top-k selection.
- The algorithm guarantees probabilistic correctness, ensuring robustness against adversarial inputs.
- Empirical evaluations show significant speed improvements over existing top-k algorithms.
- Prof-K allows users to balance accuracy and speed through adjustable parameters.
Read more
Prof-K: Probabilistic One-Pass Filtering for Efficient Top-k Selection
Summary
The paper introduces Prof-K, a novel algorithm designed for efficient top-k selection, which is crucial in various computational fields including databases and machine learning. Traditional methods for top-k selection face challenges as data sizes increase, leading to high memory and computational costs. Prof-K addresses these issues by employing a one-pass filtering approach that utilizes a small random sample to estimate an adaptive threshold. This allows the algorithm to stream through the input data only once, retaining elements above the threshold in a compact buffer, followed by an exact top-k selection from this reduced set. The authors provide theoretical guarantees for the correctness and buffer size of Prof-K, ensuring that the algorithm performs reliably across different input distributions, including adversarial scenarios. Empirical results demonstrate that Prof-K achieves significant speedups (1.5x to 10x) over existing optimized implementations, particularly in large-scale settings with small-to-moderate k values. Additionally, the algorithm allows for flexible accuracy-speed trade-offs, enabling users to adjust parameters based on their specific needs. The paper also highlights the practical benefits of Prof-K in training BatchTopK Sparse Autoencoders, showcasing its potential to reduce training times without compromising performance.
Methodology
Prof-K employs a probabilistic one-pass filtering technique where a small random sample is used to estimate a threshold. The algorithm streams through the input data once, retaining only those elements that exceed the threshold, thus reducing the problem size before applying an exact top-k selection on the filtered candidates.
Results
Prof-K achieves speedups of 1.5x to 10x compared to highly optimized implementations like PyTorch topk and RadiK, especially in scenarios with large data sizes and small-to-moderate k values. The algorithm maintains high accuracy and robustness, even under challenging input distributions.
Implications
The introduction of Prof-K has significant implications for various applications in machine learning and data processing, particularly in scenarios involving large datasets and sparse activations. Its efficiency can lead to faster training times for models that rely heavily on top-k selection, making it a valuable tool for practitioners in the field.
Beyond Outcome Rewards: Step-Level Self-Distilled Policy Optimization for Deep Search Agents
Reinforcement Learning
NLP
Large Language Models
- Introduces Evidence Anchors as a new form of privileged information for multi-turn search tasks.
- Proposes SSPO, which uses step-level advantage weights to improve policy optimization in deep search agents.
- Demonstrates that naive self-distillation can lead to performance degradation due to information asymmetry.
- Achieves superior sample efficiency compared to GRPO, even with fewer gradient steps.
Read more
Beyond Outcome Rewards: Step-Level Self-Distilled Policy Optimization for Deep Search Agents
Summary
This paper addresses the challenge of sparse rewards in reinforcement learning (RL) for deep search agents, which operate over long trajectories but receive only a single outcome reward per trajectory. The authors propose a novel approach called Step-Level Self-Distilled Policy Optimization (SSPO) that utilizes Evidence Anchors—concise, step-level evidence snippets extracted from the web—as privileged information to guide the learning process. SSPO converts the disagreement between teacher and student models into step-level advantage weights, applied exclusively to incorrect trajectories, allowing for a more nuanced update mechanism. This decouples the direction of policy change from the magnitude of updates, preserving the diversity of correct trajectories. The proposed method is evaluated on the Qwen3-8B model across three benchmarks: BrowseComp, GAIA, and FRAMES, demonstrating consistent performance improvements over the existing GRPO method, even with significantly fewer training steps and minimal computational overhead.
Methodology
The authors developed SSPO, which employs Evidence Anchors as step-level supervision for training deep search agents. This method converts teacher-student disagreement into advantage weights for incorrect trajectories, allowing for targeted updates while preserving the diversity of correct trajectories. The approach was evaluated using the Qwen3-8B model across multiple benchmarks.
Results
SSPO consistently outperformed GRPO across all tested benchmarks (BrowseComp, GAIA, and FRAMES), achieving better performance with fewer training steps and only about 5% additional computational overhead per step. Ablation studies confirmed the effectiveness of the proposed design choices, particularly the use of step-level advantage weights over token-level signals.
Implications
The findings suggest that incorporating step-level supervision can significantly enhance the training efficiency and effectiveness of deep search agents in complex environments. This approach could be applied to various domains requiring multi-turn reasoning and exploration, such as web search, dialogue systems, and other interactive AI applications.
Balanced Adaptive Prototype Selection for Scalable TabPFN Inference on Large-Scale Tabular Data
Efficient ML
- Identifies bounded inference context as a primary scalability bottleneck for pretrained tabular foundation models.
- Proposes BAPS, an information-preserving context-construction framework that does not require model retraining.
- Demonstrates that BAPS can maintain predictive performance on large datasets with significant context compression.
- Establishes the necessity of effective context construction for scaling pretrained models to million-scale datasets.
Read more
Balanced Adaptive Prototype Selection for Scalable TabPFN Inference on Large-Scale Tabular Data
Summary
This paper addresses the scalability challenges of pretrained tabular foundation models, particularly TabPFN, when applied to large-scale datasets. The authors introduce a novel framework called Balanced Adaptive Prototype Selection (BAPS), which constructs compact inference contexts that preserve essential predictive information without modifying or retraining the pretrained model. BAPS focuses on maintaining representative structure, informative decision boundaries, local density, class balance, and feature-space diversity. The framework is evaluated on large datasets, specifically the million-row HIGGS and SUSY datasets, demonstrating that it can achieve strong predictive performance and reliable calibration using only 512 prototypes, resulting in a significant context compression of approximately 1,953-fold. This work highlights the importance of effective context construction as a means to extend the applicability of pretrained tabular foundation models to larger datasets, thereby overcoming the limitations imposed by bounded inference contexts.
Methodology
The BAPS framework is designed to construct compact inference contexts by optimizing the retention of critical information from the original training data. It jointly preserves various aspects such as representative structure, decision boundaries, local density, class balance, and feature diversity, ensuring that the selected prototypes are informative for inference without altering the pretrained model.
Results
Experiments conducted on the HIGGS and SUSY datasets show that using BAPS allows for the effective use of only 512 prototypes while achieving strong predictive performance and reliable calibration. This corresponds to an approximate 1,953-fold reduction in context size, demonstrating the framework's efficiency and effectiveness in large-scale inference scenarios.
Implications
The findings suggest that BAPS can significantly enhance the scalability of pretrained tabular foundation models, making them more applicable to large datasets in various domains. This could lead to more efficient deployment of machine learning models in real-world applications where large-scale tabular data is prevalent.
Learning Discrete Decisions for MIPs with Constraint-Aware Diffusion
Optimization
Generative Models
Graph Learning
- Introduction of Constrained Graph Diffusion (CGD) for mixed-integer optimization.
- Integration of feasibility projections in the diffusion process to ensure valid discrete decisions.
- Decomposition of the optimization problem into discrete and continuous components for efficiency.
- Demonstrated effectiveness on optimal transmission switching and portfolio optimization tasks.
Read more
Learning Discrete Decisions for MIPs with Constraint-Aware Diffusion
Summary
This paper introduces a novel learning-based approach for solving mixed-integer optimization problems (MIPs), which are known for their complexity due to the need to simultaneously determine discrete and continuous decisions while adhering to intricate combinatorial constraints. The proposed method, named Constrained Graph Diffusion (CGD), utilizes a graph-based generative diffusion model that learns the discrete components of MIPs. A key innovation is the integration of a training-free feasibility projection operator into the reverse diffusion process, which guides the generation of intermediate samples towards the feasible set. After generating the discrete decisions, the remaining problem simplifies to a continuous optimization task that can be efficiently solved using existing numerical methods. The CGD framework is versatile and can be adapted to a wide range of mixed-integer optimization problems through appropriate projection operators. The authors evaluate CGD on two specific applications: optimal transmission switching for AC optimal power flow and discrete portfolio optimization. The results show significant improvements in both feasibility and solution quality compared to existing learning-based methods, along with speed enhancements of up to 425 times over traditional numerical solvers for mixed-integer nonlinear programming (MINLP) problems.
Methodology
The methodology involves a graph-based generative diffusion model that learns the discrete decision variables of mixed-integer optimization problems. The model incorporates a feasibility projection operator during the reverse diffusion process to ensure that generated samples are feasible. Once discrete decisions are made, the continuous variables are optimized using standard numerical methods, allowing for efficient resolution of the overall problem.
Results
The CGD framework was tested on two challenging mixed-integer optimization problems, yielding significant improvements in feasibility and solution quality. The method achieved speedups of up to 425 times compared to state-of-the-art numerical solvers for MINLPs, demonstrating its effectiveness in practical applications.
Implications
The proposed CGD framework has the potential to enhance decision-making processes in various fields that rely on mixed-integer optimization, such as power systems, logistics, finance, and manufacturing. Its ability to efficiently generate feasible solutions could lead to more effective resource allocation and operational strategies in these domains.
When Can You Trust Offline Evaluation of Equal-Cost Top-k Allocation? A Controlled, Reproducible Benchmark and Practitioner's Guide
Theory
- Weak overlap is primarily determined by logger-target action alignment rather than logging sharpness.
- Cross-fitting the outcome nuisance does not eliminate reuse bias; honest policy-level splitting is necessary.
- Propensity-estimation error is the most significant factor affecting performance in offline evaluations.
- The paper provides a reproducible benchmark and practical guidance for practitioners in the field.
Read more
When Can You Trust Offline Evaluation of Equal-Cost Top-k Allocation? A Controlled, Reproducible Benchmark and Practitioner's Guide
Summary
This paper addresses the challenges of offline evaluation in equal-cost top-k allocation, particularly focusing on the implications of weak overlap in logged data. The author benchmarks six estimators across five datasets and two known-effect sweeps, validating the mechanisms against a non-simulated paired reference. Key findings indicate that weak overlap is influenced more by logger-target action alignment than by logging sharpness alone. The study reveals that the optimizer's curse persists even with cross-fitting of outcome nuisance, necessitating honest policy-level splitting to mitigate reuse bias. Additionally, propensity-estimation error emerges as a significant factor degrading performance, with out-of-fold estimates adversely affecting Inverse Propensity Score (IPS) methods. The paper provides a comprehensive guide for practitioners on when offline evaluation can be trusted and how to select appropriate estimators, emphasizing the importance of understanding the logging environment and the implications of estimator choice.
Methodology
The study employs a benchmarking approach, testing six different estimators across multiple datasets while controlling for known effects. It integrates empirical evidence with theoretical analysis to assess the impact of logging strategies and estimator choices on evaluation accuracy. The author uses a non-simulated paired reference for validation and conducts a series of experiments to analyze the effects of various factors on overlap and estimator performance.
Results
The results indicate that the effective sample size (ESS) is significantly affected by the alignment of logger and target actions, with misalignment leading to poor overlap. The study finds that honest policy-level splitting reduces bias significantly compared to traditional cross-fitting methods. Additionally, the propensity estimation error is shown to be a critical factor that can invert overlap diagnostics, emphasizing the need for accurate propensity estimates in offline evaluations.
Implications
The findings have significant implications for practitioners involved in budget-constrained allocation decisions, as they provide insights into the reliability of offline evaluation methods. The guidance on logging design and estimator selection can help organizations make more informed decisions about deploying targeting rules based on logged data. This work also contributes to the broader understanding of off-policy evaluation in machine learning, particularly in contexts where accurate estimations of treatment effects are crucial.