AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
66
Papers today
8h
Update frequency
7
Days of history
Why GPT-Style Models Do Not Directly Transfer to Symbolic Music: Compression in the Wrong Coordinate System
Theory
Generative Models
Audio & Speech
- Tokenization in GPT-style models gains predictive advantage from constructing effective coordinate systems rather than just from reusable combinations.
- The Effectiveness–Losslessness Framework establishes boundaries for tokenization, emphasizing the importance of decoupling observable facts from contextual relations.
- Controlled experiments validate that effective coordinate construction improves predictive compressibility in symbolic music.
- Higher-order musical organization can emerge without explicit structural supervision when contextual freedom is preserved.
Read more
Why GPT-Style Models Do Not Directly Transfer to Symbolic Music: Compression in the Wrong Coordinate System
Summary
This paper investigates the limitations of GPT-style models when applied to symbolic music, arguing that the success of these models in language is not directly transferable to music due to differences in tokenization and compression. The author introduces the Effectiveness–Losslessness Framework, which emphasizes the importance of constructing a coordinate system that allows for effective representation of musical facts. The framework defines two critical boundaries: the Fact–Token Boundary, which focuses on decoupling and denesting to expose stable regularities, and the Token–State Boundary, which preserves contextual relations for model computation. Through controlled experiments, the paper demonstrates that effective coordinate construction enhances predictive compressibility, while fixed relational projections hinder contextual modeling. The findings suggest that the fundamental challenge lies in discovering effective representations that reveal observable regularities while maintaining relational freedom, which is essential for contextual structure to emerge in music.
Methodology
The author formulates the Effectiveness–Losslessness Framework to define tokenization boundaries and conducts controlled symbolic-music experiments to validate the framework's principles. The experiments include matched coordinate interventions and fixed relational projections to assess the impact on predictive compressibility and contextual modeling.
Results
The results indicate that effective coordinate construction significantly enhances predictive compressibility, while fixed relational projections restrict contextual modeling capabilities. The experiments also reveal that higher-order musical structures can naturally emerge in model outputs without the need for explicit structural labels.
Implications
The findings suggest that developing domain-specific tokenization strategies is crucial for improving the performance of models in symbolic music. This could lead to advancements in music generation and analysis, as well as a better understanding of how to adapt language models to different modalities.
Conformal Prediction for Molecular Properties under Label Shift
Theory
- Introduces a conformal prediction framework tailored for label shift in molecular property prediction.
- Addresses the challenge of overconfident predictions in AI models by providing reliable prediction intervals.
- Enhances uncertainty quantification, which is crucial for regulatory compliance in drug development.
- Offers a method that does not require retraining, making it efficient for real-world applications.
Read more
Conformal Prediction for Molecular Properties under Label Shift
Summary
This paper addresses the challenges of predicting molecular properties in drug discovery, particularly under conditions of label shift, where the distribution of target properties changes while the feature distribution remains constant. Traditional AI models often provide single-value predictions that lack reliability due to overconfidence, especially when faced with novel compounds. The authors propose a conformal prediction framework that incorporates weighted conformal scores based on marginal label probability ratios, allowing for the generation of statistically rigorous prediction intervals without the need for retraining. This approach enhances uncertainty quantification, aligning AI predictions with regulatory demands for transparency and supporting more reliable decision-making in drug development. By moving beyond mere accuracy, the framework provides actionable confidence measures, addressing a critical bottleneck in the transition from preclinical to clinical trials.
Methodology
The authors develop a weighted conformal prediction method that adjusts conformal scores using marginal label probability ratios to account for label shift. This method allows for the construction of prediction intervals that maintain statistical validity even when the underlying distribution of target properties changes.
Results
The proposed framework successfully generates reliable prediction intervals for molecular properties under significant label shift, demonstrating improved uncertainty quantification compared to traditional point predictions. This method enhances the trustworthiness of AI-driven predictions in drug development.
Implications
The findings suggest that integrating uncertainty quantification into AI predictions can significantly improve decision-making processes in drug discovery, potentially reducing costs and increasing the success rate of new drug candidates. This approach also aligns with regulatory expectations for AI applications in healthcare.
Debate Training Reduces Reward Hacking in RLAIF
Reinforcement Learning
Large Language Models
Theory
- Debate training reduces reward hacking in RLAIF by maintaining judge performance.
- The method shows a 45% improvement in peak validation accuracy compared to a single-player baseline.
- Further weakening the judge accelerates hacking, but can be mitigated by adding debate rounds.
- Critique word limits help balance the game and prevent judge hacking, though they restrict expressive clarity.
Read more
Debate Training Reduces Reward Hacking in RLAIF
Summary
This paper investigates the effectiveness of debate training as a method to mitigate reward hacking in Reinforcement Learning from AI Feedback (RLAIF). Reward hacking occurs when a policy exploits systematic errors in its AI judge, leading to degraded task performance. The authors propose a two-player adversarial game setup where a generator produces solutions and a critic argues against them, with a weaker LLM judge evaluating the arguments. The study focuses on mathematical tasks with verifiable correctness, allowing for precise measurement of reward hacking dynamics. The results indicate that debate training significantly reduces reward hacking compared to a single-player RLAIF baseline, maintaining judge performance and achieving a higher peak validation accuracy. The paper also explores the effects of judge strength and the introduction of additional debate rounds, demonstrating that debate can effectively counteract misalignment and reward hacking while highlighting the importance of balancing multi-agent training. Overall, the findings suggest that debate training is a promising approach for scalable oversight in AI systems, particularly as model capabilities increase.
Methodology
The authors employed a two-player zero-sum debate game where a generator and a critic compete to convince a weaker LLM judge. They trained a Gemini 2.5 Flash-class policy using multi-agent RL self-play, comparing the debate approach against a single-player RLAIF baseline. The experimental design focused on mathematical reasoning tasks with verifiable ground-truth labels to measure reward hacking dynamics accurately.
Results
The debate training approach maintained higher accuracy and judge performance throughout training, contrasting with the baseline that exhibited increasing reward but declining accuracy. The debate method showed a 45% performance gap recovery in peak validation accuracy and demonstrated resilience against reward hacking dynamics over many RL steps.
Implications
The findings suggest that debate training could serve as a scalable solution for oversight in AI systems, particularly in scenarios where ground-truth labels are unavailable. This approach may enhance the alignment and safety of increasingly capable AI models, making it relevant for various applications in AI development.
Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL
Reinforcement Learning
Large Language Models
Multimodal
- CO-RL enables unsupervised reasoning through cooperative multi-agent training.
- Diverse model cohorts reduce correlated errors and enhance feedback quality.
- The framework consistently outperforms traditional RL methods and label-free approaches.
- Significant performance gains are observed across multiple benchmarks in text-only and multimodal domains.
Read more
Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL
Summary
The paper presents CO-RL, a novel framework for unsupervised reasoning in multi-agent reinforcement learning (RL) that mitigates the reliance on costly ground-truth supervision. Traditional RL methods often depend on verifiable rewards, which can be scarce and expensive to obtain. CO-RL addresses this by allowing multiple decoupled models to optimize simultaneously using rewards derived from their peers, rather than from self-generated feedback. This approach leverages the diversity of model architectures and training samples to reduce correlated errors, thereby enhancing reasoning performance and maintaining behavioral diversity. The authors demonstrate that CO-RL outperforms both baseline models and previous label-free methods across various text-only and multimodal benchmarks, achieving significant improvements without requiring any ground-truth labels. The results indicate that cooperative multi-agent training can effectively provide independent learning signals that enhance model reasoning capabilities.
Methodology
CO-RL employs a cooperative multi-agent RL approach where multiple decoupled models are optimized simultaneously. Each agent generates completions for a given prompt and aggregates them into a pseudo-answer using majority voting. Rewards for each agent are derived from the performance of other agents' pseudo-answers, breaking the self-reinforcing feedback loop typical of self-rewarding methods. The diversity of the cohort is enhanced by using different model architectures, sizes, and rephrased training samples.
Results
CO-RL achieved average performance improvements of 3.0–8.6% across seven text-only benchmarks for large language models (LLMs) and 2.3–7.2% across four multimodal benchmarks for vision-language models (VLMs). It outperformed the strongest self-rewarding baselines and matched or surpassed supervised methods in several settings, demonstrating the effectiveness of cross-agent supervision.
Implications
The findings suggest that cooperative multi-agent frameworks can significantly enhance reasoning capabilities in models without requiring labeled data. This could lead to more efficient training processes in various applications, including natural language processing and computer vision, where labeled data is often limited.
Backward through Time, Algebraically
Theory
Optimization
Reinforcement Learning
- Introduces an algebra-generic evaluation engine for linear temporal logic in softly-valued systems.
- Addresses the limitations of traditional boolean-valued LTL in the context of neural networks.
- Emphasizes the importance of differentiability for training signals derived from temporal logic satisfaction.
- Provides an executable specification of various algebras and their implementations.
Read more
Backward through Time, Algebraically
Summary
This paper addresses the challenges of applying linear temporal logic (LTL) in the context of softly-valued systems, such as neural networks and adaptive controllers. Traditional LTL operates over boolean values, which are inadequate for systems that require differentiable feedback for training. The author proposes an algebra-generic evaluation engine that allows for the differentiation of various logical semantics without committing to a single algebra upfront. This approach enables the evaluation of temporal logic formulas over continuous values, facilitating the integration of LTL into machine learning frameworks like PyTorch. The paper details the implementation of several algebras and their behavior in both forward and backward evaluations, emphasizing the importance of differentiability in the training signal derived from the satisfaction of temporal logic formulas. The resulting library, telos, provides a flexible and powerful tool for researchers and practitioners working with temporal logic in machine learning contexts.
Methodology
The author develops an algebra-generic framework that abstracts the semantics of various logical algebras, allowing for the evaluation of temporal logic formulas over continuous values. This involves defining abstract operations for logical connectives and implementing them in a way that supports differentiation. The framework is designed to be flexible, enabling users to choose their preferred algebra without being constrained by a single implementation.
Results
The paper demonstrates that the proposed algebra-generic approach successfully evaluates temporal logic formulas over continuous values, providing a differentiable feedback mechanism for training neural networks. Various algebras are implemented and audited for their performance, revealing that each algebra represents a different way of handling satisfaction and dissatisfaction in the context of temporal logic.
Implications
The findings suggest that integrating linear temporal logic into machine learning frameworks can enhance the expressiveness and control of neural policies and adaptive systems. This work opens up new avenues for research in combining formal logic with machine learning, potentially leading to more robust and interpretable AI systems.
Dynamic Compression in Recurrent Networks
NLP
Time Series
Efficient ML
- Dynamic compression allows RNNs to selectively revisit and update their state based on later context.
- This approach reduces the recurrent state capacity needed for accurate function reuse.
- Dynamic compression scales more favorably as the number of stored functions increases.
- The method exposes a tradeoff between recurrent-state capacity and additional computation.
Read more
Dynamic Compression in Recurrent Networks
Summary
This paper introduces a novel approach called dynamic compression for recurrent neural networks (RNNs), which addresses the limitations of traditional single-pass architectures in processing long contexts. In standard RNNs, the model compresses its history into a fixed-size state during a single left-to-right pass, which constrains its ability to adaptively manage information for future tasks. Dynamic compression allows the model to selectively revisit and update its fixed-size state based on later context, enabling it to refine its representation of past information as needed. The authors demonstrate this approach through a controlled continual function-reuse task, where the model learns multiple functions and later encounters few-shot tasks requiring the identification and reuse of these functions. By employing selective re-scanning, the model can focus on relevant past tokens, significantly reducing the recurrent state required for accurate function reuse and improving scalability as the number of functions increases. The findings reveal a computation-memory tradeoff, where the model can allocate more computation to revisit history, thus making more effective use of its fixed-size state.
Methodology
The authors utilize a controlled continual function-reuse task to evaluate dynamic compression. They implement a selective re-scanning mechanism that allows the model to identify and revisit relevant past tokens, updating its recurrent state accordingly. The study employs Gated DeltaNet, a linear-attention model, to maintain a fixed-size state while processing sequences. The model is trained with oracle supervision to establish an upper bound on the benefits of revisiting the correct functions, followed by a self-supervised procedure to derive re-scan targets based on the model's write strengths.
Results
The results indicate that dynamic compression significantly reduces the recurrent-state capacity required for accurate function reuse. The model demonstrates improved performance in identifying and reusing functions as the number of functions increases, showcasing a favorable scaling behavior compared to traditional single-pass models. The selective re-scanning mechanism allows the model to adapt its state representation dynamically, leading to better task performance.
Implications
The findings suggest that dynamic compression could enhance the efficiency of recurrent models in various applications requiring long-context inference, such as natural language processing and time series analysis. By allowing models to adaptively manage their memory and computation, this approach could lead to more effective and scalable architectures in machine learning.
Composing Flow-Matching Energies with Known Physics: Generation, OOD Detection, and Inversion on PDE Fields
Generative Models
Theory
Optimization
- Introduces a potential parameterization for flow matching that yields time-dependent energy functions.
- Develops an Energy-based Predictor-Corrector framework for physics-aware data generation and posterior sampling.
- Demonstrates the complementary nature of data energy and physical energy for improved OOD detection.
- Shows significant performance improvements on PDE datasets compared to existing methods.
Read more
Composing Flow-Matching Energies with Known Physics: Generation, OOD Detection, and Inversion on PDE Fields
Summary
This paper presents a novel approach to probabilistic modeling of physical fields by integrating flow-matching models with known physical structures, specifically governing equations. The authors propose a method that utilizes energy-based models (EBMs) to facilitate efficient data generation, out-of-distribution (OOD) detection, and inverse problem solving. By employing a potential-induced velocity parameterization, the authors derive explicit scalar energy functions that are time-dependent and can be computed directly from a matching regression objective. This approach circumvents the need for variational forms or additional MCMC steps, allowing for straightforward training and sampling. The proposed framework, termed Energy-based Predictor-Corrector (EnergyPC), combines data energy with physical energy to enhance OOD detection and sampling accuracy. The authors demonstrate the effectiveness of their method on high-dimensional PDE datasets, showing significant improvements over traditional flow ODE and other baseline methods in terms of data generation, OOD detection, and PDE solution field reconstruction.
Methodology
The authors utilize flow-matching models with a potential-induced velocity to derive explicit scalar energy functions. They construct a total energy by combining learned data energy with known physical structures, enabling efficient sampling and inference tasks. The EnergyPC framework is employed to facilitate physics-aware generation and posterior sampling.
Results
The proposed method outperforms traditional flow ODE and other baseline approaches in generating high-dimensional PDE datasets, improving OOD detection accuracy and enhancing the reconstruction of PDE solution fields and coefficient inference.
Implications
This work has potential applications in scientific machine learning, particularly in fields requiring accurate modeling of physical phenomena, such as fluid dynamics and heat transfer. The integration of known physics with data-driven approaches could lead to more robust and interpretable models in various scientific domains.
Generative Learning of Separatrices
Generative Models
- Introduces a novel framework combining supervised classification and generative modeling for approximating separatrices in dynamical systems.
- Utilizes uncertainty metrics from neural network classifiers to identify high-uncertainty regions that correspond to separatrices.
- Employs score-based generative models to generate sample densities consistent with the empirical distribution of separatrix manifolds.
- Addresses the limitations of traditional computational methods in high-dimensional systems, providing a systematic and data-driven approach.
Read more
Generative Learning of Separatrices
Summary
This paper addresses the challenge of identifying and reconstructing separatrices, the boundaries separating basins of attraction in multistable, multidimensional dynamical systems. Traditional computational methods struggle with high-dimensional systems and often require prior knowledge of the system's equations. The authors propose a novel framework that combines supervised classification with generative modeling to effectively approximate these complex structures. The approach involves training neural network classifiers on sampled initial conditions labeled by their corresponding basins of attraction. By using uncertainty metrics from the classifiers, the method identifies high-uncertainty regions that correspond to separatrices. Subsequently, score-based generative models are trained on these regions to generate sample densities consistent with the empirical distribution of the separatrix manifold. This iterative, data-driven framework leverages the strengths of both discriminative and generative models, resulting in accurate reconstructions of separatrix manifolds that are otherwise difficult to sample directly due to their low probability density in phase space.
Methodology
The methodology involves training neural network classifiers on sampled initial conditions labeled by their corresponding basins of attraction. Uncertainty metrics from these classifiers are used to identify high-uncertainty regions, which are then targeted for further sampling using score-based generative models to generate densities that reflect the structure of the separatrix manifolds.
Results
The proposed framework successfully reconstructs approximate separatrix manifolds in high-dimensional dynamical systems, demonstrating improved sampling of regions that are typically under-sampled in traditional simulations. The results indicate that the method can effectively capture the complex geometric structures of separatrices, providing a more accurate representation of transition pathways in multistable systems.
Implications
The findings have significant implications for computational dynamics, particularly in fields such as chemical kinetics, ecology, and climate science, where understanding transition thresholds and basin boundaries is crucial. The framework could enhance predictive modeling and control strategies in complex systems by providing better insights into their behavior.
Quantifying Depth Sufficiency in Residual Neural Networks: A First-Order Criterion
Theory
Optimization
- Introduces a first-order criterion for assessing depth sufficiency in residual networks.
- Establishes that residual non-degeneracy is necessary and sufficient for determining the value of additional depth.
- Demonstrates that activation-gradient norms decrease with depth, indicating diminishing returns.
- Provides experimental evidence supporting the use of low activation-gradient energy as an indicator of depth sufficiency.
Read more
Quantifying Depth Sufficiency in Residual Neural Networks: A First-Order Criterion
Summary
This paper addresses the challenge of determining when a trained residual neural network has reached sufficient depth for optimal performance. The authors introduce a first-order criterion for assessing depth sufficiency based on a fixed function-preserving residual-growth protocol. They define 'first-order residual depth saturation' as the point where no further local decrease in performance can be achieved through additional depth. The authors prove that the condition of residual non-degeneracy is both necessary and sufficient for determining the value of adding depth. Their main theorem establishes that the boundary for depth saturation is invariant across different descent-compatible first-order updates. The paper also discusses the implications of activation gradients as a diagnostic tool for assessing the remaining optimization value of residual depth. Through experiments on various models, including ResNets and GPT-2-style architectures, the authors demonstrate that the maximum activation-gradient norm decreases with depth, indicating diminishing returns on performance gains. The findings suggest that low activation-gradient energy can serve as a practical indicator of when further depth may not yield significant benefits.
Methodology
The authors utilize a fixed residual-growth protocol to study the effects of adding depth to neural networks. They define first-order residual depth saturation and derive a criterion based on the projection of activation gradients onto the residual tangent space. The methodology includes theoretical proofs and empirical experiments across various neural network architectures to validate their claims.
Results
The experiments reveal that as depth increases, the maximum activation-gradient norm decreases, indicating a transition to a low-signal regime where additional depth yields minimal performance gains. The function-preserving growth method shows competitive performance compared to training deeper networks from scratch, reinforcing the practical utility of their proposed criterion.
Implications
The findings have significant implications for model scaling and adaptive growth strategies in neural networks. By providing a clear criterion for when to stop adding depth, the research can inform practitioners on optimizing network architectures, potentially leading to more efficient training and deployment of deep learning models.
From Abductive Explanations to Global Logical Rules for Node Classification in SGCs
Graph Learning
Interpretability
- Introduces AXSGC, a logic-based framework for node classification in SGCs.
- Utilizes minimal abductive explanations (AXps) to ensure irredundancy and sufficiency in rule extraction.
- Demonstrates significant improvements in fidelity and compactness of global rules compared to existing methods.
- Provides a structured approach to explainability in GNNs, particularly for node classification tasks.
Read more
From Abductive Explanations to Global Logical Rules for Node Classification in SGCs
Summary
This paper presents a novel logic-based framework for node classification in Simple Graph Convolution (SGC) networks, addressing the need for explainability in Graph Neural Networks (GNNs). The authors propose a method called AXSGC (Abduction-based Explanations for SGC), which utilizes minimal abductive explanations (AXps) as an intermediate representation for extracting global logical rules. The framework operates in three stages: first, it computes a minimal set of node-feature pairs for each node that are sufficient to preserve the predicted class; second, these pairs are transformed into distance-indexed predicates; and third, decision trees are trained to derive global logical rules from these predicates. The approach is evaluated on benchmark datasets, demonstrating that it produces compact global rules while maintaining high fidelity to the original SGC model, outperforming existing methods like LogicXGNN in terms of rule size and fidelity.
Methodology
The methodology involves three main stages: (1) computing minimal node-feature pairs (AXps) for each node that ensure the predicted class, (2) encoding these pairs into distance-indexed predicates, and (3) training decision trees to extract global logical rules from these predicates. The use of AXps guarantees the irredundancy of the extracted rules, and the linearity of the SGC model allows for efficient computation.
Results
The experiments conducted on benchmark datasets show that AXSGC achieves up to 30.2% higher fidelity to the SGC model compared to LogicXGNN, while extracting up to 83.8% fewer rules. This indicates a significant improvement in both the quality and compactness of the generated explanations.
Implications
The proposed framework has potential applications in various domains where GNNs are used, such as social network analysis, recommendation systems, and biological network classification. By providing clearer and more concise explanations for node classifications, it enhances the interpretability of GNNs, which is crucial for trust and accountability in automated decision-making systems.
Study-Strategy Clusters from EdNet Logs Track Engagement, Not Mastery
Theory
- Unsupervised clustering of ITS logs reveals distinct study-strategy styles but does not predict mastery.
- Early engagement metrics can be forecasted from study-strategy clusters, but not later accuracy.
- The study emphasizes the need for correctness-aware models to assess mastery rather than relying solely on behavioral clusters.
- A stable hierarchy of study strategies was identified, with implications for targeted engagement support.
Read more
Study-Strategy Clusters from EdNet Logs Track Engagement, Not Mastery
Summary
This paper investigates the validity of using unsupervised clusters derived from intelligent tutoring system (ITS) logs as predictors of learning outcomes. The authors analyze data from the EdNet-KT3 corpus, focusing on study-strategy features such as resource use, revision, video watching, and problem practice from 5,000 active learners. They identify a stable hierarchy of study-strategy clusters, which include four primary styles (reading-focused, video-heavy, revision-heavy, and problem-first) and four additional finer styles. The study reveals that while early clusters can predict later engagement metrics, such as persistence and session completion, they do not reliably predict mastery or unassisted accuracy in learning outcomes. The findings challenge the assumption that behavioral clusters can serve as proxies for knowledge gains, suggesting that they are more indicative of study styles and engagement rather than mastery. The authors recommend using these clusters to inform engagement strategies rather than as standalone predictors of learning success.
Methodology
The authors utilized clustering techniques on early practice logs from EdNet-KT3 to identify study-strategy clusters. They employed silhouette scores for cluster validation and split learner timelines to prevent leakage of late behavior into early cluster features. A knowledge-tracing model (SAKT) was also applied to assess mastery prediction.
Results
The analysis yielded a hierarchy of eight study-strategy clusters, with early clusters predicting later engagement but not mastery. The knowledge-tracing model showed only modest improvements over a baseline, indicating that mastery signals are largely independent of behavioral styles.
Implications
The findings suggest that while log-based clusters can effectively describe study styles and engagement, they should not be used as proxies for knowledge gains. For mastery-focused applications, it is recommended to utilize models that incorporate correctness and external assessments.
A Pre-Specified Construction-Confirmation Test of Operation-Level Causal Transfer Across Finite Isomorphic Symbolic Domains
Theory
Interpretability
Large Language Models
- The study focuses on operation-level causal transfer in symbolic domains.
- It employs a pre-specified construction-confirmation design to test hypotheses.
- Results indicate significant evidence for operation transfer, but are limited to one candidate.
- The research emphasizes the distinction between behavioral accuracy and underlying operational structures.
Read more
A Pre-Specified Construction-Confirmation Test of Operation-Level Causal Transfer Across Finite Isomorphic Symbolic Domains
Summary
This paper investigates the operation-level causal transfer of models across finite isomorphic symbolic domains, specifically focusing on the transfer of hidden-state differences between operations. The study poses a critical question: if the hidden-state difference between two operations is estimated separately for each source input, does adding that difference to a mapped recipient input influence the model's output towards the corresponding recipient answer? The research employs a design that compares input-specific interventions against various control conditions, including wrong-operation and random controls. The experiments were conducted using a frozen Qwen2.5-7B-Instruct model, where one specific candidate operation was pre-specified and tested for its causal transfer capability. The results demonstrated statistically significant confirmation of the operation transfer, with p-values indicating strong evidence for the hypothesis. However, the findings are limited to one prompt route and one candidate, and do not imply generalization across different models or operations. The paper emphasizes the importance of distinguishing between behavioral success and the underlying operational structures that facilitate such success, contributing to the broader discourse on model interpretability and causal inference in machine learning.
Methodology
The methodology involves a pre-specified construction-confirmation test where hidden-state differences between operations are estimated for each source input. The study compares the effects of adding these differences to mapped recipient inputs against various control conditions. The experiments were conducted on a frozen model, ensuring that the candidate operations were predefined and not altered during confirmation.
Results
The main result indicates that one specific operation candidate passed both construction and confirmation tests with statistically significant p-values (0.000198 and 0.006943 after Holm adjustment). The replication of results across different implementations (PyVene and NNsight) further supports the findings, although the results are limited to a single route and candidate.
Implications
The findings have implications for understanding how models represent and transfer operations across different symbolic domains, contributing to the fields of model interpretability and causal inference. The study highlights the need for rigorous testing of operational structures to ensure that behavioral success is not mistaken for operational understanding.
Leveraging Association Context Retrieval in Knowledge Editing to Build White-Box Attacks on LLMs
NLP
Large Language Models
Theory
- Introduces a novel white-box attack method for LLMs using knowledge editing techniques.
- Enhances attack effectiveness by leveraging associative knowledge retrieval.
- Demonstrates portability and architecture-agnostic applicability across various LLMs.
- Minimizes degradation of general model performance while unlocking harmful prompt categories.
Read more
Leveraging Association Context Retrieval in Knowledge Editing to Build White-Box Attacks on LLMs
Summary
This paper presents a novel approach to white-box attacks on large language models (LLMs) by leveraging knowledge editing techniques. The authors propose a category-level attack that utilizes associative knowledge retrieval to modify model weights, thereby enhancing the model's susceptibility to harmful prompts without significantly degrading its overall performance. The method builds on the locate-then-edit paradigm, which allows for precise alterations in model behavior while minimizing the impact on general functionality. The authors demonstrate the effectiveness of their approach across various LLM architectures, including GPT-based, Llama-based, and Qwen3-MoE models, highlighting its portability, editor-agnostic nature, and architecture-agnostic applicability. The findings underscore the potential for improved attack strategies that can exploit the vulnerabilities of LLMs while maintaining their operational integrity.
Methodology
The authors modify existing knowledge editing frameworks to create a category-level white-box attack. This involves a tracing procedure to identify entities to edit within harmful prompts, allowing for the adjustment of model weights without explicit prompt injections. The method is designed to be portable and applicable to any model architecture, utilizing various knowledge editing techniques to achieve desired outcomes.
Results
Experiments conducted on multiple LLM architectures demonstrate that the proposed attack method significantly improves the effectiveness of white-box attacks compared to existing techniques. The results indicate that the approach can unlock entire categories of harmful prompts while inflicting minimal damage on the model's overall performance.
Implications
The findings of this research have significant implications for the security and robustness of LLMs, particularly as these models are increasingly deployed in sensitive applications. Understanding the vulnerabilities of LLMs can inform the development of more effective safeguards and alignment strategies, ultimately contributing to safer AI systems.
J-Miner: Recovering Executable Decision Knowledge from Language-Model Classifiers
NLP
Large Language Models
Interpretability
- J-Miner effectively extracts and organizes decision knowledge from language-model classifiers into named concepts and executable rules.
- The approach achieves high fidelity in reproducing source classifier decisions, outperforming traditional surface-level rule extraction methods.
- The mined decision knowledge can be transferred to a smaller model, maintaining high accuracy and efficiency.
- J-Miner reveals internal semantic patterns and decision structures that are not visible through final predictions alone.
Read more
J-Miner: Recovering Executable Decision Knowledge from Language-Model Classifiers
Summary
The paper introduces J-Miner, a novel approach for extracting and encoding executable decision knowledge from fine-tuned language-model classifiers. While large language models (LLMs) excel in diverse text classification tasks, they typically output only final labels, concealing the intricate decision-making knowledge acquired during fine-tuning. J-Miner addresses this by mining internal decision knowledge from the classifier, transforming it into an explicit, inspectable, and reusable format. The methodology involves aggregating vocabulary-aligned internal signals across layers and token positions to identify named concepts, followed by learning executable decision rules based on the classifier's predictions. The results demonstrate that J-Miner can reproduce up to 98.3% of the source classifier's decisions and achieve significantly higher behavioral fidelity compared to traditional rule extraction methods. Furthermore, the decision knowledge can be effectively transferred to a compact standalone model, which retains 99.8% of the source classifier's accuracy while using only a fraction of its parameters. This work highlights the potential for making task-specific decision knowledge accessible and reusable beyond the original classifier, paving the way for more interpretable and efficient machine learning applications.
Methodology
J-Miner aggregates internal signals from fine-tuned language-model classifiers across various layers and token positions to identify named concepts. It then learns executable decision rules based on the classifier's predictions, distilling local readouts into a coherent knowledge representation.
Results
J-Miner successfully reproduces up to 98.3% of the decisions made by the source classifier and achieves 6.0–29.5 percentage points higher behavioral fidelity than compact rules learned from input words. The compact student model retains 99.8% of the source classifier's mean task accuracy while using about 1/24 of its parameters.
Implications
The findings suggest that task-specific decision knowledge can be effectively extracted and reused, enhancing the interpretability of machine learning models. This has potential applications in various domains, including spam filtering, content moderation, and sentiment analysis, where understanding model decisions is crucial.
NICE: Scale-Stable Perturbations for Graph Neural Network Explanations via Noise Corruption
Graph Learning
Interpretability
- Identification of Scale Drift as a critical issue in the Perturb-Query paradigm for GNN explanations.
- Introduction of Noise Corruption (NC) as a scale-stable alternative to Element-wise Masking (EM).
- Development of NICE, a framework that utilizes NC to learn a Stochastic Restoration Boundary for generating explanations.
- Demonstrated improvements in explanation quality, with average Recall and AUC-ROC increases of 8.67% and 7.57%, respectively.
Read more
NICE: Scale-Stable Perturbations for Graph Neural Network Explanations via Noise Corruption
Summary
This paper addresses the challenges of post-hoc explanations for Graph Neural Networks (GNNs) that rely on the Perturb-Query paradigm, which can introduce significant distribution shifts that undermine the reliability of explanations. The authors identify a phenomenon termed 'Scale Drift,' which occurs due to the commonly used Element-wise Masking (EM) method that suppresses edge-induced messages toward zero, leading to deterministic scale contraction across message-passing layers. To overcome this issue, they propose a new perturbation mechanism called Noise Corruption (NC), which maintains the expected squared message norm while perturbing messages through matched-norm random-direction corruption. Building on NC, the authors introduce NICE, a framework that learns a Stochastic Restoration Boundary (SRB) to balance target-prediction restoration against compactness. The framework employs Boundary-Integrated Gradient (BIG) to convert the learned boundary into edge attributions, thereby providing more faithful explanations. Experimental results across multiple benchmarks demonstrate that NICE significantly improves explanation performance and model faithfulness, effectively reducing Scale Drift compared to traditional methods.
Methodology
The authors propose a new perturbation mechanism, Noise Corruption (NC), which replaces the traditional Element-wise Masking (EM) with matched-norm random-direction corruption. This method preserves the expected squared norm of messages while perturbing them. The NICE framework learns a Stochastic Restoration Boundary (SRB) to balance the restoration of target predictions against compactness, and utilizes Boundary-Integrated Gradient (BIG) to attribute the contributions of individual edges to the restoration process.
Results
NICE outperforms existing methods across eight benchmarks, achieving an average Recall improvement of 8.67% and an AUC-ROC improvement of 7.57% over the best baseline. The framework also demonstrates higher retention fidelity, indicating more reliable explanations.
Implications
The findings suggest that addressing the perturbation mechanism itself can lead to more reliable and interpretable explanations for GNNs, which is crucial for applications in areas requiring transparency in decision-making, such as social networks, bioinformatics, and recommendation systems.
Detecting and Discriminating Operator Misspecification in Hybrid PDE-Parameter Learning: a Reference-Free Instrument, with Discrimination Bounded In Sample
Theory
- Introduces a reference-free instrument for detecting operator misspecification in hybrid PDE learning.
- Demonstrates the ability to distinguish between wrong operators and unidentifiable parameters.
- Findings indicate that traditional accuracy checks can fail to identify significant model misspecifications.
- Shows that different models can converge to the same incorrect parameter values, regardless of complexity.
Read more
Detecting and Discriminating Operator Misspecification in Hybrid PDE-Parameter Learning: a Reference-Free Instrument, with Discrimination Bounded In Sample
Summary
This paper presents a novel instrument designed to detect and discriminate operator misspecification in hybrid partial differential equation (PDE) parameter learning without requiring an oracle. The proposed method distinguishes between a wrong operator and an unidentifiable parameter based on a single fit. The author evaluates the instrument on a self-adjoint parabolic inverse problem, demonstrating its effectiveness through an information-matrix statistic that shows significant differences in rejection rates between correct and misspecified models. The instrument remains silent under correct but unidentifiable designs, indicating its ability to differentiate between types of failures. The findings reveal that traditional accuracy checks may overlook significant misspecifications, as the misspecified estimator can yield low in-domain root mean square error (RMSE) while being substantially inaccurate in parameter recovery. The paper emphasizes the need for such detection mechanisms in learning systems, as various models converge to similar incorrect values regardless of their complexity. The results highlight the importance of understanding model specifications in hybrid learning frameworks, particularly when using physics-informed neural networks.
Methodology
The methodology involves the development of an information-matrix statistic that assesses the fit of a hybrid PDE-parameter estimator. The instrument analyzes the performance of the estimator under various conditions, including correct specification, single misspecification, and double misspecification, using statistical measures to evaluate rejection rates and parameter recovery accuracy.
Results
The results indicate that under correct specification, the median statistic is 0.19 with a rejection rate of 0.033, while under misspecifications, the statistic rises significantly to 224 and 85, indicating clear detection capabilities. The instrument remains silent (0.050) under non-identifiable designs, demonstrating its effectiveness in distinguishing between types of failures. The misspecified estimator shows a RMSE of 2.7 × 10−2, which is misleadingly low compared to the actual parameter recovery error of 29.7% at zero noise.
Implications
The implications of this research are significant for the development of robust hybrid learning systems, particularly in fields that rely on accurate parameter estimation from PDEs. The findings suggest that practitioners should incorporate such detection mechanisms to avoid the pitfalls of model misspecification, which can lead to erroneous conclusions and ineffective models in scientific and engineering applications.
REFLEX: Reflexive Equilibrium Fixed-point Learning for Endogenous eXchanges
Theory
Optimization
- Introduces REFLEX, a framework for assessing stability in OTC corporate bond markets.
- Replaces abstract stability conditions with measurable dealer behaviors.
- Demonstrates significant amplification of instability due to competition among dealers.
- Provides a market-level safety margin for evaluating automated quoting systems.
Read more
REFLEX: Reflexive Equilibrium Fixed-point Learning for Endogenous eXchanges
Summary
This paper introduces REFLEX, a framework designed to analyze and ensure stability in over-the-counter (OTC) corporate bond markets where dealers use machine learning to set bid and ask prices. The authors highlight the feedback loop created when dealers retrain their models based on the trades attracted by their quotes, raising concerns about market stability. REFLEX replaces unobservable stability conditions with measurable dealer behaviors, allowing for the estimation of a retraining modulus that predicts whether retraining will lead to convergence or instability. The framework is validated through simulations, showing a close agreement between predicted and measured stability. The authors also demonstrate that competition among dealers can significantly amplify market instability. By calibrating their model with 36 years of public market data, they establish a market-level safety margin that can be evaluated prior to deploying automated quoting systems, thus addressing systemic risk in algorithmic trading.
Methodology
The authors develop REFLEX by deriving stability constants from market microstructure primitives and validating these through simulations. They compute a retraining modulus based on dealer behaviors and use historical market data to calibrate their model. The framework incorporates a learned market-response operator and employs various instruments to assess stability.
Results
The REFLEX framework predicts stability with an accuracy of within 8% compared to measured outcomes. It identifies a critical retraining modulus threshold (1.21) beyond which ordinary retraining becomes unstable. The study finds that competition among dealers increases market instability by factors of 1.74 and 3.16 for two and three dealers, respectively. The framework also establishes finite-sample robust certificates and provides a lazy deployment strategy for optimal retraining.
Implications
The findings suggest that dealers can better manage risks associated with algorithmic trading by using REFLEX to evaluate the stability of their quoting strategies before deployment. This has significant implications for regulatory oversight and the design of automated trading systems, potentially reducing systemic risks in financial markets.
Integrating Novelty and Surprise for Experience Prioritization and Exploration in Image-Based Reinforcement Learning
Reinforcement Learning
- Introduction of NSPER, a novel experience prioritization method using intrinsic rewards.
- Extension to NSPER+R, which utilizes novelty and surprise for both prioritization and intrinsic rewards.
- Comprehensive evaluation of prioritization strategies and their impact on training efficiency.
- Ablation study isolating the contributions of novelty and surprise.
Read more
Integrating Novelty and Surprise for Experience Prioritization and Exploration in Image-Based Reinforcement Learning
Summary
This paper addresses the challenge of sample efficiency in reinforcement learning (RL), particularly in image-based domains where agents learn from high-dimensional visual inputs. Traditional sampling methods often lead to redundant updates and slow learning due to suboptimal experience selection. To tackle this, the authors propose Novelty and Surprise Prioritized Experience Replay (NSPER), which combines novelty and surprise as intrinsic signals to prioritize informative experiences and enhance exploration. Novelty captures underrepresented states, while surprise highlights gaps in the agent's understanding of the environment. The authors further extend this approach with NSPER+R, which integrates these signals as intrinsic rewards to improve both replay quality and exploration. Experiments conducted on the DeepMind Control Suite demonstrate that NSPER and NSPER+R significantly enhance training efficiency and convergence speed compared to existing methods, showcasing the effectiveness of combining intrinsic rewards with experience prioritization in image-based RL.
Methodology
The authors developed NSPER by integrating novelty and surprise as intrinsic signals within the Prioritized Experience Replay (PER) framework. They further enhanced this with NSPER+R, which uses these signals as intrinsic rewards. The methods were implemented in the PixelTD3 algorithm and evaluated on tasks from the DeepMind Control Suite.
Results
The experiments showed that both NSPER and NSPER+R improved training efficiency and convergence speed compared to traditional methods, indicating that the integration of novelty and surprise enhances the learning process in complex, high-dimensional visual environments.
Implications
The findings suggest that incorporating intrinsic rewards into experience prioritization can lead to more efficient learning in RL, particularly in image-based applications. This approach could be beneficial in various domains such as robotics, gaming, and autonomous driving, where efficient learning from visual inputs is crucial.
Domain-Adapted Molecular Language Models for Efficient Search of Make-on-Demand Libraries
Graph Learning
Efficient ML
- Pretrained molecular language models show variable performance across different molecular libraries.
- Traditional molecular fingerprints serve as a strong baseline for molecular representation.
- Domain adaptation through fine-tuning improves the performance of molecular language models.
- The quality of molecular representations is closely linked to the target domain of the virtual library.
Read more
Domain-Adapted Molecular Language Models for Efficient Search of Make-on-Demand Libraries
Summary
This paper investigates the effectiveness of pretrained molecular language models as molecular encoders for discovering new molecules from virtual libraries across various domains, including drug discovery, materials chemistry, and catalysis. The authors benchmark four molecular language models against six virtual libraries and find that native embeddings show significant variability in performance, while traditional molecular fingerprints provide a robust baseline. They identify a domain-representation mismatch as a key issue and demonstrate that explicit domain adaptation through fine-tuning on target library structures significantly enhances representation performance. The results indicate that the quality of molecular representations is highly dependent on the target domain, and that domain-adapted models can improve sample efficiency in molecular discovery tasks. This work highlights the importance of adapting molecular representations for effective decision-making in virtual screening and automated laboratories, paving the way for more efficient resource utilization in molecular design.
Methodology
The authors conducted systematic benchmarks of four molecular language models across six virtual libraries, focusing on drug discovery, materials chemistry, and catalysis. They evaluated the performance of native embeddings and traditional molecular fingerprints, and implemented domain adaptation by fine-tuning the models on structures from the target libraries without requiring property labels.
Results
The study found that native molecular language model embeddings exhibited substantial performance variability across different libraries, while molecular fingerprints provided consistent results. Domain adaptation significantly improved the performance of the models, with several adapted encoders emerging as top performers in the benchmark tasks, demonstrating that tailored representations can enhance molecular discovery efficiency.
Implications
The findings suggest that domain-adapted molecular representations can lead to more effective and efficient decision-making in virtual screening and self-driving laboratories, potentially transforming approaches to molecular discovery and resource allocation in experimental campaigns.
FinFraudBench: A Heterogeneous Graph Benchmark for Financial Fraud Detection
Graph Learning
- FinFraudBench is the first benchmark to construct large heterogeneous graphs specifically for financial fraud detection.
- The benchmark includes two datasets that preserve multiple financial entity types and relation types, reflecting real-world complexities.
- A reproducible evaluation protocol is provided, supporting controlled comparisons under limited labels and extreme class imbalance.
- Experiments show that heterogeneous graph methods significantly outperform traditional non-graph methods.
Read more
FinFraudBench: A Heterogeneous Graph Benchmark for Financial Fraud Detection
Summary
The paper introduces FinFraudBench, a benchmark designed to enhance financial fraud detection through the use of heterogeneous graphs. Traditional fraud detection methods often rely on simplified models that do not capture the complex interrelations among various financial entities. FinFraudBench addresses this by providing two large-scale heterogeneous graph datasets (CreditCard-Fraud and BankTrans-Fraud) that encompass multiple entity types and relation types, reflecting the intricate nature of real-world financial systems. The datasets include up to 8.99 million nodes and 89.23 million directed edges, preserving the multi-entity and multi-relational characteristics of financial data. The authors establish a standardized evaluation protocol that incorporates ranking and imbalance-sensitive classification metrics, allowing for a comprehensive assessment of existing methods. Extensive experiments reveal that models designed for heterogeneous graphs outperform others, highlighting the importance of relational context in fraud detection. The benchmark aims to facilitate future research and improve the practical effectiveness of fraud detection systems.
Methodology
The authors constructed two heterogeneous graph datasets from public transaction-level fraud data, preserving multiple entity types and relations. They established a unified schema and a reproducible evaluation protocol that includes labeled and unlabeled training transactions, validation splits, and standardized metrics. The evaluation involved comparing various baseline methods, including non-GNN, homogeneous GNN, multi-relation, and heterogeneous approaches.
Results
The experiments demonstrated that methods tailored for heterogeneous graphs achieved the best performance, while multi-relation and fraud-oriented methods also showed competitive results. Non-GNN baselines were limited due to their inability to leverage relational context, underscoring the advantages of using heterogeneous graph structures for fraud detection.
Implications
FinFraudBench has the potential to significantly advance the field of financial fraud detection by providing a robust framework for evaluating and developing new models. It encourages the adoption of graph-based approaches that can better capture the complexities of financial interactions, ultimately leading to more effective fraud detection systems.
OOD Detection for EEG-based Machine Learning in High-Risk Environments
Time Series
- Introduces a novel benchmark for evaluating OOD detection methods specifically for EEG data.
- Demonstrates that generative models significantly outperform discriminative models in OOD detection tasks.
- Clarifies the distinction between OOD detection and model uncertainty, which are often conflated in the literature.
- Provides actionable insights for improving the deployment of EEG-based machine learning models in clinical settings.
Read more
OOD Detection for EEG-based Machine Learning in High-Risk Environments
Summary
This paper addresses the challenges of deploying machine learning models for electroencephalography (EEG) analysis in high-risk environments, particularly due to their susceptibility to out-of-distribution (OOD) data. The authors highlight that existing OOD detection methods have not been thoroughly explored in the context of EEG, which can lead to significant predictive failures in clinical applications. To bridge this gap, the authors introduce a comprehensive benchmark for EEG OOD detection, evaluating a range of methods while linking their performance to downstream clinical prediction tasks. The study distinguishes between OOD detection and model uncertainty, revealing that generative models outperform discriminative models in detecting OOD data. However, both types of models provide valuable insights into model performance in clinical settings. The findings emphasize the importance of combining these methods to enhance the safety and reliability of EEG-based machine learning applications.
Methodology
The authors developed a systematic evaluation framework for OOD detection in EEG, which includes a suite of perturbations to create controlled OOD data. They benchmarked various OOD detection methods, categorizing them into discriminative and generative approaches, and assessed their performance in relation to downstream clinical prediction tasks.
Results
The results indicate that generative models achieved significantly better OOD detection performance compared to discriminative models, especially as the severity of perturbations increased. In clinical evaluations, both types of models effectively predicted model performance, with generative models identifying unfamiliar data and discriminative models serving as proxies for model uncertainty.
Implications
The findings suggest that integrating OOD detection and model uncertainty estimation can create a more robust framework for deploying EEG-based machine learning models in high-risk environments, potentially improving diagnostic reliability and patient safety.
Mr.Dec: Daily-Scale Longitudinal Multimodal Modeling for 30-Day Readmission Prediction
Multimodal
Time Series
Computer Vision
- Introduces Mr.Dec, a multimodal trajectory architecture for predicting 30-day readmissions.
- Utilizes a Transformer Decoder to model daily EHR and CXR data in a chronological sequence.
- Achieves state-of-the-art performance on MIMIC-IV and MIMIC-CXR datasets.
- Identifies 'Critical Days' for actionable risk stratification.
Read more
Mr.Dec: Daily-Scale Longitudinal Multimodal Modeling for 30-Day Readmission Prediction
Summary
The paper presents Mr.Dec, a novel approach for predicting 30-day hospital readmissions by modeling patient data as a chronological sequence of daily multimodal events. Traditional methods often compress longitudinal clinical histories into fixed representations, losing critical day-level signals that reflect a patient's evolving health status. Mr.Dec addresses this issue by utilizing a Transformer Decoder to integrate daily Electronic Health Record (EHR) updates and intermittent Chest X-ray (CXR) findings in a time-aligned manner. The model employs Disease-Specific Supervised Contrastive Learning to enhance robustness and ensure a diagnosis-aware latent space structure. Evaluations on the MIMIC-IV and MIMIC-CXR datasets demonstrate that Mr.Dec achieves state-of-the-art performance while maintaining the integrity of clinical sequences. Additionally, the model identifies 'Critical Days' during hospital stays, providing actionable insights for real-time risk stratification. This approach not only improves prediction accuracy but also offers clinically grounded interpretations of risk changes throughout a patient's admission.
Methodology
Mr.Dec employs a causal Transformer architecture that processes daily EHR records and sporadic CXR observations as a chronological token stream. The model uses masked self-attention to ensure causality, allowing it to predict readmission risk based on the accumulated clinical evidence throughout the hospital stay. The integration of Disease-Specific Supervised Contrastive Learning serves as an auxiliary regularization technique to enhance the model's robustness against label noise and imbalance.
Results
Mr.Dec demonstrated superior performance in predicting 30-day hospital readmissions compared to existing state-of-the-art models. The model's ability to maintain the temporal structure of clinical data resulted in improved accuracy and the identification of 'Critical Days' that correlate with significant changes in readmission risk.
Implications
The findings suggest that Mr.Dec can be effectively utilized in clinical settings to enhance patient monitoring and risk assessment, potentially leading to better resource allocation and improved patient outcomes. The model's capability to provide interpretable insights into critical periods during hospitalization may also support clinical decision-making.
MiNO: Cotangent-bundle propagator learning for PDEs
Theory
- MiNO learns the phase-space propagator of PDEs, focusing on phase and amplitude rather than the solution field.
- The method shows improved accuracy and efficiency compared to traditional PINNs and FNOs, especially for sharp discontinuities.
- MiNO can reuse a single trained generator for multiple initial conditions, enhancing its practical applicability.
- The approach provides a clear separation of trainable error from frequency truncation, leading to more reliable results.
Read more
MiNO: Cotangent-bundle propagator learning for PDEs
Summary
This paper introduces MiNO (microlocal neural operator), a novel approach to learning the propagator of partial differential equations (PDEs) by focusing on the phase and amplitude in phase space rather than the solution field itself. The authors highlight a gap in regularity where the propagator can be smoother than the solution it generates, particularly in cases of transported discontinuities. MiNO employs the eikonal equation for phase learning and the transport equation for amplitude learning, allowing for the recovery of solutions through oscillatory integrals. The methodology demonstrates significant advantages over traditional physics-informed neural networks (PINNs) and supervised Fourier neural operators (FNOs), particularly in handling sharp fronts and caustics. The experiments show that MiNO achieves a lower mean error and can effectively reuse a trained generator across multiple initial conditions without retraining, showcasing its efficiency and robustness in solving PDEs.
Methodology
MiNO utilizes a microlocal analysis framework to learn the phase and amplitude of the propagator through the eikonal and transport equations. It reconstructs solutions via oscillatory integrals, allowing for the explicit representation of sharp fronts and caustics in the propagation geometry.
Results
In benchmark tests, MiNO achieved a mean error of 3.84 × 10−3 on smooth advection, outperforming a supervised FNO with a mean error of 3.12 × 10−2. MiNO also demonstrated that it could stop improving after 10,000 optimization steps, reaching the accuracy limit of its reconstruction window, while traditional methods remained near their initial error.
Implications
The findings suggest that MiNO could be a powerful tool for solving complex PDEs in various scientific and engineering applications, particularly where sharp discontinuities and high-frequency propagation are involved. Its ability to generalize across initial conditions without retraining could streamline workflows in computational simulations.
Iterative tensor network transformations for element-wise evaluation of elementary and filtering functions
Optimization
Theory
Efficient ML
- Introduction of ITNTs for efficient evaluation of nonlinear functions on tensor trains.
- Demonstrated application in 3D reactive flow fields for high-fidelity reaction rate computation.
- Successful optimization of complex problems like Max-SAT with significant scalability.
- ITNTs operate entirely in the compressed domain, maintaining low computational costs.
Read more
Iterative tensor network transformations for element-wise evaluation of elementary and filtering functions
Summary
This paper introduces Iterative Tensor Network Transformations (ITNTs), a novel algorithmic framework designed for the element-wise evaluation of nonlinear functions on data represented as tensor trains (TTs). The authors highlight the limitations of existing tensor network methods, particularly in handling nonlinear operations, which have restricted their application in general data processing. ITNTs allow for efficient computation within the compressed domain of TTs, significantly enhancing the ability to process large-scale datasets. The framework is validated through two main applications: first, in evaluating highly nonlinear functions for a 3D reactive flow field, which improves the accuracy of reaction rate computations and region filtering; second, in solving complex optimization problems, such as Max-SAT instances, demonstrating the capability to handle up to 270 configurations. The results indicate that ITNTs can effectively address the challenges posed by nonlinear transformations, making them a foundational tool for data science and large-scale optimization tasks.
Methodology
The ITNT framework leverages iterative methods, particularly the Newton-Raphson approach, to apply arbitrary nonlinear functions directly to data encoded in tensor trains. This process involves parallel evaluations of the function, allowing for efficient computation while maintaining data in a compressed format. The methodology includes a combination of linear operations, partial integrations, and element-wise multiplications to facilitate the transformation of data.
Results
The application of ITNTs resulted in a significant improvement in the accuracy of reaction rate computations in a simulated methane/air jet flame, achieving a relative error approximately 100 times lower than traditional methods. Additionally, the framework successfully identified optimal solutions in complex optimization landscapes, demonstrating its effectiveness in handling large-scale problems.
Implications
The introduction of ITNTs has the potential to revolutionize data processing in various scientific fields, particularly those requiring the evaluation of nonlinear functions. This framework can enhance simulations in fluid dynamics, quantum physics, and machine learning, enabling more accurate modeling and optimization in complex systems.
Evaluating and improving crop-yield forecasting methods during extreme drought
Time Series
- The study focuses on crop yield forecasting during the 2012 Midwestern US drought, highlighting the impact of extreme weather on agricultural predictions.
- Feature distribution mismatch and data irregularities significantly affect the performance of forecasting models.
- Sample weighting and feature selection improve the accuracy of traditional ML models, but have limited impact on deep learning models.
- The deep learning model VITA outperforms ML models overall, despite the challenges posed by extreme drought conditions.
Read more
Evaluating and improving crop-yield forecasting methods during extreme drought
Summary
This paper investigates the challenges of crop yield forecasting during extreme drought conditions, specifically focusing on the 2012 drought in the Midwestern United States. The authors evaluate various forecasting models, including traditional machine learning (ML) and deep learning (DL) approaches, to predict county-level corn yields using 16 meteorological drivers. The study highlights the issue of feature distribution mismatch, where the training data does not adequately represent the extreme conditions present during the drought year. Additionally, the dataset suffers from spatial and temporal irregularities due to missing yield data across counties and incomplete daily measurements. To address these challenges, the authors implement sample weighting and feature selection techniques, which improve the performance of ML models. However, the deep learning model VITA, while outperforming the ML models, shows limited improvement from these modifications. The study emphasizes the importance of understanding the effects of feature distribution mismatch on forecasting accuracy and provides insights into the relative performance of ML versus DL models in the context of extreme weather events.
Methodology
The authors compared non-deep learning and deep learning forecasting models using a dataset of county-level corn yields and meteorological drivers. They addressed data irregularities through sample weighting and feature selection based on SHAP analysis to improve model performance.
Results
The modifications led to improved performance in ML models, while the deep learning model VITA maintained superior performance but showed little to no improvement from the modifications. The study revealed the significant impact of feature distribution mismatch on forecasting accuracy.
Implications
The findings suggest that while deep learning models can outperform traditional methods, they may not be as adaptable to extreme conditions without further modifications. This research can inform future crop yield forecasting efforts, particularly in the context of climate variability and extreme weather events.
MultiSigBERT: Beyond Survival Analysis through Multimodal and Sequential Modeling in Oncology
NLP
Multimodal
Time Series
- MultiSigBERT integrates multimodal data from EHRs for survival analysis in oncology.
- The framework captures higher-order temporal interactions using the Signature transform.
- It prevents temporal information leakage through a landmark-based survival modeling approach.
- The model achieved a concordance index of 0.743, indicating strong predictive performance.
Read more
MultiSigBERT: Beyond Survival Analysis through Multimodal and Sequential Modeling in Oncology
Summary
The paper introduces MultiSigBERT, a novel framework for multimodal sequential survival modeling in oncology that effectively integrates heterogeneous data from electronic health records (EHRs). Traditional survival models often rely on unimodal data and fail to account for the temporal dynamics of patient trajectories. MultiSigBERT addresses these limitations by combining narrative medical reports, structured clinical variables, and their longitudinal evolution into a unified temporal representation. The framework employs sentence embeddings for narrative texts, modality-specific PCA for dimensionality reduction, and the Signature transform from Rough Paths theory to capture higher-order temporal interactions across modalities. The model is evaluated on a large oncology cohort from the Léon Bérard Center, demonstrating its effectiveness in predicting individualized risk scores. The results indicate a concordance index of 0.743 on an independent test set, showcasing the advantages of jointly modeling multimodal temporal dynamics and patient-level geometric structures for improved survival prediction.
Methodology
The methodology involves encoding narrative medical reports into sentence embeddings, processing structured clinical variables, and aligning them temporally. These modalities are concatenated to form multimodal vectors, which are then transformed using the Signature transform. A LASSO-regularized Cox model is employed for survival analysis, allowing for individualized risk score estimation while maintaining interpretability.
Results
The MultiSigBERT model was tested on a dataset comprising over 120,000 medical reports and structured records from more than 2,500 patients. It achieved a concordance index of 0.743 (sd 0.029) on an independent test set, demonstrating its effectiveness in survival prediction through the integration of multimodal data.
Implications
The proposed framework has significant implications for clinical decision-making in oncology, as it enhances the accuracy of survival predictions by leveraging diverse data sources. This could lead to better treatment selection, improved patient outcomes, and more effective clinical trial designs.
Temporal Graph Prototype-conditioned Conformal Prediction for Fraud Detection
Graph Learning
Time Series
Theory
- ProtoCP improves calibration efficiency for fraud detection in temporal graphs.
- The framework addresses challenges of benign-dominated neighborhoods and class imbalance.
- Utilizes learned prototypes to enhance the quality of calibration contexts.
- Introduces a neighborhood-relative scoring mechanism for stable predictions.
Read more
Temporal Graph Prototype-conditioned Conformal Prediction for Fraud Detection
Summary
This paper introduces ProtoCP, a novel conformal prediction framework specifically designed for edge-level fraud detection in temporal graphs. The authors identify two significant challenges in applying existing graph conformal predictors to fraud detection: the presence of benign-dominated neighborhoods that obscure fraud signals and the extreme class imbalance that leads to conservative calibration thresholds. ProtoCP addresses these challenges by focusing calibration on fraud-relevant subgraph contexts and utilizing learned prototypes to filter out benign noise. Additionally, it implements a neighborhood-relative scoring mechanism with temporal score diffusion to enhance the stability of class-conditional calibration. Through experiments on four fraud detection benchmarks, ProtoCP demonstrates improved calibration efficiency and smaller prediction sets compared to state-of-the-art methods, thereby providing reliable coverage guarantees essential for risk-aware decision-making in fraud detection.
Methodology
ProtoCP employs a conformal prediction framework that leverages learned prototypes to suppress noise from benign interactions in the calibration context. It introduces a neighborhood-relative scoring mechanism that diffuses scores over time, allowing for more stable class-conditional calibration despite class imbalance and temporal drift.
Results
Experimental results on four fraud detection benchmarks (YelpChi, S-FFSD, FTFD, and BankSim) indicate that ProtoCP consistently achieves the desired coverage with significantly smaller prediction sets than existing state-of-the-art methods, demonstrating its effectiveness in edge-level fraud detection.
Implications
The findings suggest that ProtoCP can enhance fraud detection systems in various applications, such as telecommunications and financial networks, by providing more reliable uncertainty quantification and enabling better risk-aware decision-making.
Valid Per-Field Selective Risk Control for Document Extraction: Three Failure Modes, a Validity Ladder, and When Conditioning Pays
NLP
Large Language Models
Theory
- Identification of three failure modes in naive per-field selective risk control for document extraction.
- Introduction of a validity ladder framework to categorize solutions based on rigor and guarantees.
- Demonstration of effective risk control through fit/val split protocols and Mondrian Learn-then-Test.
- Highlighting the importance of conditioning on provenance for enhanced certification.
Read more
Valid Per-Field Selective Risk Control for Document Extraction: Three Failure Modes, a Validity Ladder, and When Conditioning Pays
Summary
This paper addresses the challenges of selective risk control in document extraction systems, particularly focusing on the per-field accept/review process. The author identifies three critical failure modes that undermine the effectiveness of naive risk control methods: document clustering, score-refit leakage, and tie-mass pathology. These issues can lead to significant violations of the expected error rates in real-world applications. To mitigate these failures, the author proposes a 'validity ladder' framework that categorizes solutions into tiers based on their rigor and guarantees. The paper presents a fit/val split protocol that restores expected selective risk control, achieving a nominal risk level while maintaining coverage. Additionally, the author introduces Mondrian Learn-then-Test, which provides PAC certificates for risk control, demonstrating its effectiveness across different document types. The results indicate that conditioning on provenance can enhance certification where pooled thresholds fail, highlighting the importance of context in risk management. The findings are validated through extensive experiments on genuine document datasets, revealing both the limitations of existing methods and the potential for improved approaches.
Methodology
The study employs a combination of empirical analysis and theoretical frameworks to diagnose failure modes in document extraction systems. It utilizes a fit/val split protocol to restore selective risk control and implements Mondrian Learn-then-Test for PAC certification. The author conducts experiments on a dataset of 13,859 fields from CORD receipts to quantify the effectiveness of proposed methods.
Results
The proposed methods achieve significant improvements in risk control, with the fit/val split protocol yielding a coverage of 0.318 at an achieved risk of 0.096, while Mondrian Learn-then-Test provides PAC certificates with varying levels of coverage and risk across different document types. The findings indicate that conditioning on provenance is crucial for effective certification, particularly in challenging scenarios.
Implications
The insights from this research can enhance the reliability of document extraction systems, making them more trustworthy for applications in various domains, including finance, healthcare, and legal documentation. The proposed frameworks and methodologies can be adopted to improve risk management practices in machine learning applications involving structured data extraction.
EMAN: Optimization-Driven Capacity Growth through Path Emergence in Multi-Task Learning
Optimization
Computer Vision
Efficient ML
- EMAN starts with a single shared path and grows additional paths based on optimization evidence.
- The framework allows for adaptive capacity allocation to meet varying task requirements.
- Extensive experiments validate EMAN's effectiveness on multiple datasets, achieving competitive performance.
- The timing of path emergence is crucial for optimizing computational resources and improving task performance.
Read more
EMAN: Optimization-Driven Capacity Growth through Path Emergence in Multi-Task Learning
Summary
The paper introduces the Emergent Modular Atomic Network (EMAN), a novel framework for multi-task learning (MTL) that addresses the limitations of existing methods which rely on fixed structures or predefined paths. EMAN proposes an optimization-driven approach that begins with a single shared path and only grows additional independent paths when persistent optimization evidence indicates a need for increased capacity. This framework allows for adaptive allocation of shared and task-specific representation capacity, thus accommodating varying task demands without incurring unnecessary computational costs. The authors demonstrate EMAN's effectiveness through extensive experiments on benchmark datasets such as PASCAL-Context and NYUv2, showing that it achieves competitive performance while optimizing resource usage. The results indicate that the timing and strength of optimization demand significantly influence the emergence of additional paths, leading to improved performance in multi-task scenarios.
Methodology
EMAN employs an optimization-driven framework that begins with one physical forward path and utilizes a latent relative phase to probe local optimization geometry. It evaluates the need for additional paths through Optimization Evidence Certification (OEC), which assesses the strength and consistency of optimization signals before materializing new paths.
Results
The experiments conducted on controlled rank settings, PASCAL-Context, and NYUv2 demonstrated that EMAN achieves improved performance compared to existing methods while using approximately 75% of the computational resources required by always-dual path models. EMAN's adaptive capacity growth led to better mean semantic segmentation and depth results, particularly on the NYUv2 dataset.
Implications
The findings suggest that EMAN can be applied to various multi-task learning scenarios where task demands fluctuate, providing a more efficient and flexible approach to capacity management. This could lead to advancements in fields such as computer vision and robotics, where multi-task learning is prevalent.
SIGMA: SHAP-Guided Implicit-Trajectory Generation for Metadata-Free LLM-Based AutoFE
Large Language Models
Optimization
Interpretability
- SIGMA replaces semantic metadata with SHAP-based importance signals for feature generation.
- The EXIT approach allows for effective long-horizon optimization without explicit trajectory representation.
- SIGMA significantly reduces duplicate feature generation rates while maintaining competitive performance.
- The framework demonstrates substantial efficiency in feature utilization compared to traditional methods.
Read more
SIGMA: SHAP-Guided Implicit-Trajectory Generation for Metadata-Free LLM-Based AutoFE
Summary
The paper introduces SIGMA, a novel framework for Automated Feature Engineering (AutoFE) that addresses two significant challenges in leveraging Large Language Models (LLMs) for feature generation: the lack of semantic metadata in practical applications and the instability caused by trajectory accumulation during long-horizon optimization. SIGMA utilizes SHAP (SHapley Additive exPlanations) values to provide task-aware signals for guiding feature generation, thereby eliminating the need for semantic descriptions. The framework employs an EXposed-feature Implicit Trajectory (EXIT) approach, which allows the optimization trajectory to be implicitly represented through visible features in the prompt, thus maintaining a constant context length. Empirical evaluations demonstrate that SIGMA achieves performance on par with state-of-the-art LLM baselines while significantly reducing the duplicate feature generation rate from 37.2% to 6.8%. Furthermore, SIGMA shows substantial efficiency gains, requiring only an average of 5.4 features to match traditional AutoFE performance.
Methodology
SIGMA employs SHAP values to categorize input features into groups based on their importance, facilitating intra-group and cross-group feature generation. The EXIT method allows for implicit representation of the optimization trajectory through visible features, thus avoiding the need for lengthy prompts and reducing the risk of exceeding context limits.
Results
SIGMA achieves comparable performance to existing LLM-based AutoFE methods while significantly lowering the duplicate feature generation rate from 37.2% to 6.8%. It also demonstrates efficiency by requiring only 5.4 features on average to match traditional state-of-the-art performance.
Implications
The proposed SIGMA framework can be applied in various domains where semantic metadata is scarce or unavailable, such as privacy-sensitive fields like healthcare. Its efficiency and reduced duplication in feature generation can enhance the robustness and interpretability of machine learning models in real-world applications.
Q-Learning With World Models
Reinforcement Learning
Robotics
Efficient ML
- QWM integrates world models with Q-learning to enhance sample efficiency.
- The framework avoids compounding model bias by training policies on real transitions only.
- QWM enables test-time search over imagined trajectories for better action selection.
- Significant performance improvements were observed on manipulation benchmarks.
Read more
Q-Learning With World Models
Summary
This paper introduces Q-Learning with World Models (QWM), a novel framework that integrates world models into standard Q-learning to enhance sample efficiency and performance in reinforcement learning (RL). Traditional model-based RL approaches often suffer from compounding biases when optimizing policies or value functions based on imagined rollouts. In contrast, QWM leverages world models to perform test-time search over imagined trajectories, allowing for improved action selection during both online rollouts and evaluations. The key innovation is that while the policy and value function are trained solely on real transitions, the world model aids in selecting high-value actions without introducing bias. The authors demonstrate the effectiveness of QWM on challenging manipulation benchmarks, Robomimic and LIBERO, where it significantly outperforms existing state-of-the-art methods in terms of both sample efficiency and overall performance. This approach opens new avenues for applying world models in online RL settings, particularly in complex, high-dimensional environments.
Methodology
QWM utilizes a learned world model to conduct test-time search over actions by imagining future trajectories. The framework operates on top of standard Q-learning, where the Q-function is used to evaluate candidate actions and select the highest-value one before execution. This method is applied during online rollouts to gather better data and during evaluation to enhance decision-making.
Results
QWM was evaluated on the Robomimic and LIBERO benchmarks, where it demonstrated significant improvements in both sample efficiency and performance compared to strong prior state-of-the-art methods. The results indicate that leveraging world models for action selection can lead to better outcomes in complex manipulation tasks.
Implications
The findings suggest that integrating world models into online RL can enhance the performance of RL agents in real-world applications, particularly in robotics and other high-dimensional environments. This approach may facilitate more efficient training and deployment of RL systems in practical scenarios.
The concentration game: Bayesian updating, regret, and information
Theory
- Introduces the concentration game as a unifying framework for Bayesian updating and exponential-weights regret.
- Establishes a decomposition of cumulative regret into three distinct components: intrinsic-time loss, retempering drift, and terminal relative-entropy transport.
- Demonstrates that various learning strategies and methods can be derived from the same underlying identity.
- Provides insights into how measurement scales and information budgets influence learning dynamics.
Read more
The concentration game: Bayesian updating, regret, and information
Summary
This paper introduces a two-player zero-sum repeated game framework termed the 'concentration game,' which unifies concepts from Bayesian updating and exponential-weights regret. The game involves a learner and nature, where the learner's strategy is to minimize regret against a comparator while nature's moves are constrained by an information budget. The paper establishes that the cumulative regret can be decomposed into three distinct components: intrinsic-time loss, retempering drift, and terminal relative-entropy transport. This decomposition provides a comprehensive ledger of the learner's performance, revealing how different strategies and measurement scales affect regret. The findings suggest that various learning and inference methods, including Thompson sampling and boosting, can be viewed as special cases of this framework, highlighting the interconnectedness of different areas in machine learning and information theory.
Methodology
The methodology involves formulating a two-player zero-sum game where the learner selects a mixture of actions while nature responds with moves constrained by an information budget. The paper employs concepts from Bayesian inference, game theory, and information theory to derive the regret decomposition and analyze the dynamics of the game.
Results
The main result is the exact decomposition of cumulative regret into three parts, which provides a clear understanding of the sources of regret in the learning process. This decomposition allows for a detailed analysis of how different strategies impact the learner's performance and connects various machine learning methods under a common theoretical framework.
Implications
The implications of this work extend to various fields within machine learning, including online learning, Bayesian inference, and algorithm design. By providing a unified approach to understanding regret and information dynamics, the findings can inform the development of more efficient learning algorithms and enhance the theoretical foundations of machine learning.
Information Geometry of Message Passing
Theory
Optimization
Graph Learning
- Introduction of natural-gradient message passing (NGMP) as a localized approach to variational inference.
- NGMP retains relevant information from exact belief-propagation messages while discarding non-representable components.
- Demonstrated improved accuracy in uncertainty calibration compared to variational message passing.
- Applicable in various scenarios including Poisson smoothing and heteroskedastic regression.
Read more
Information Geometry of Message Passing
Summary
This paper introduces a novel approach called natural-gradient message passing (NGMP) that localizes the natural-gradient stationary condition of variational inference on Forney-style factor graphs. By starting from the Bethe free energy and constraining selected edge marginals to an exponential family, the authors derive that at a stationary point, the natural parameter of an edge is the sum of two projected messages from incident factors. Each projected message is derived from the exact belief-propagation log-message, ensuring that the message sent by a factor is tailored to the receiving marginal. NGMP retains the relevant information from the exact message while discarding non-representable components, leading to improved accuracy, especially in scenarios with uncertainty, such as partially observed latent chains. Experimental results demonstrate the effectiveness of NGMP in applications like Poisson smoothing, heteroskedastic regression, and hourly ETTh forecasting, highlighting its advantages in uncertainty calibration over traditional variational message passing.
Methodology
The authors derive NGMP by starting from the Bethe free energy and imposing an exponential-family constraint on edge marginals. They project the exact belief-propagation updates onto the tangent space of the constrained family, allowing for a localized message-passing scheme that optimally uses the information from exact messages.
Results
Experiments show that NGMP outperforms variational message passing in terms of uncertainty calibration across multiple tasks, including Poisson smoothing, heteroskedastic regression, and forecasting, particularly when uncertainty is present.
Implications
The findings suggest that NGMP can enhance the performance of Bayesian inference in complex models, particularly in applications requiring accurate uncertainty quantification. This could lead to better decision-making processes in fields such as finance, healthcare, and machine learning.
Causal Local States: Scalable Simultaneous Causal Network Inference and Forecasting for Dynamical Systems
Graph Learning
Time Series
Interpretability
- Introduces Causal Local States (CLS) for simultaneous causal inference and forecasting.
- Achieves high fidelity in reconstructing causal networks from time series data.
- Forecasting accuracy is comparable to models using the true underlying network.
- Enhances interpretability by providing insights into variable interactions.
Read more
Causal Local States: Scalable Simultaneous Causal Network Inference and Forecasting for Dynamical Systems
Summary
The paper presents a novel framework called Causal Local States (CLS) that integrates causal network inference and forecasting for high-dimensional dynamical systems. Traditional machine learning methods often operate as black boxes, lacking insight into the interactions driving system dynamics. Existing causal discovery methods reconstruct interaction networks from observational data but do not consider their predictive utility. CLS addresses this by independently selecting the smallest set of neighbors for each node that optimally forecasts its future state, thus creating a Granger-causal interaction network. The authors demonstrate the effectiveness of CLS on three benchmark datasets, achieving high fidelity in network reconstruction and forecasting accuracy comparable to models using the true network. This approach not only enhances interpretability but also scales effectively to complex systems, providing a significant step towards explainable forecasting in dynamical systems.
Methodology
The CLS framework employs a hybrid approach that combines causal discovery and predictive modeling. It utilizes Granger causality to select relevant variables for each node, optimizing the predictive model's performance. The method processes candidate subsets of variables to identify the smallest necessary neighborhood for accurate forecasting, thus addressing the challenges of high-dimensional data.
Results
CLS demonstrated high fidelity in reconstructing causal networks and achieved forecasting results on par with models that had access to the true network. The method was tested on three benchmarks of increasing complexity, showing its robustness and effectiveness in both causal inference and prediction.
Implications
The CLS framework has significant implications for various fields requiring the forecasting of complex dynamical systems, such as climate modeling, power grid management, and ecological forecasting. By providing interpretable models that reveal underlying interactions, CLS can enhance decision-making processes based on data-driven insights.
Detecting Money Laundering in Rwandan Mobile Money: A Machine Learning Framework
Theory
Efficient ML
Optimization
- Development of a machine learning framework for AML tailored to Rwanda's regulatory constraints.
- Engineering of causal, account-centric behavioral features from transaction data.
- Evaluation of multiple machine learning models using operational metrics rather than traditional accuracy.
- LightGBM model shows promising results in detecting laundering cases with high precision.
Read more
Detecting Money Laundering in Rwandan Mobile Money: A Machine Learning Framework
Summary
This paper addresses the challenge of detecting money laundering and terrorism financing (ML/TF) within Rwanda's mobile money ecosystem, which has expanded financial access but also increased the risk of illicit activities. The authors develop a machine learning framework tailored to the specific constraints of Rwanda's anti-money laundering (AML) regime, which faces issues such as extreme class imbalance (with laundering prevalence around 0.1%), delayed labeling of suspicious transactions, and limited investigative resources. The study utilizes a synthetic dataset, SAML-D, comprising over 9.5 million transactions and 17 laundering typologies, to engineer causal, account-centric behavioral features. A variety of supervised classifiers and unsupervised anomaly detection methods are benchmarked, including Logistic Regression, Random Forest, LightGBM, and others. The evaluation focuses on operational metrics rather than traditional accuracy, assessing performance through PR-AUC, recall at high precision, and alert rates. The results indicate that LightGBM achieves a PR-AUC of 0.0469, identifying 64 true laundering cases with a precision of approximately 0.89. The late-fusion approach marginally reduces alert volume but does not enhance recall. The authors propose a governance-aware monitoring pipeline that aligns with the operational realities of Rwanda's financial institutions, emphasizing the importance of adapting machine learning techniques to local regulatory frameworks.
Methodology
The authors created a transaction-monitoring framework using a synthetic dataset (SAML-D) and engineered behavioral features to capture transaction patterns. They benchmarked various supervised classifiers and unsupervised anomaly detectors, focusing on operational metrics for evaluation, including PR-AUC and recall at calibrated precision levels.
Results
The LightGBM model achieved a PR-AUC of 0.0469, identifying 64 true laundering cases with a precision of approximately 0.89 and an alert rate of 0.51 per 10,000 transactions. The late-fusion approach yielded a PR-AUC of 0.0477 with a precision of around 0.91, capturing 59 true positives.
Implications
The findings suggest that machine learning can enhance the effectiveness of AML monitoring in mobile money systems, particularly in developing countries. The governance-aware pipeline can facilitate collaboration between financial institutions and regulatory bodies, improving the detection of illicit activities.
Delta2Gamma: Band-Wise Adaptive Contrastive Learning of EEG for Alzheimer's Disease Detection
Time Series
- Introduces Delta2Gamma, a self-supervised framework for EEG analysis.
- Decomposes EEG signals into five frequency bands for independent processing.
- Achieves 92.4% accuracy in detecting Alzheimer's disease from EEG data.
- Utilizes adaptive temperature scaling to balance signal statistics across bands.
Read more
Delta2Gamma: Band-Wise Adaptive Contrastive Learning of EEG for Alzheimer's Disease Detection
Summary
The paper presents Delta2Gamma, a self-supervised learning framework designed to enhance the detection of Alzheimer's disease (AD) using electroencephalography (EEG) data. Traditional imaging methods for diagnosing dementia are expensive and not easily deployable, making EEG a promising alternative due to its portability and cost-effectiveness. However, EEG data is often noisy and lacks sufficient clinical labels. Delta2Gamma addresses these challenges by decomposing EEG recordings into five canonical neural rhythms (delta, theta, alpha, beta, gamma), each processed by its own encoder and projection head. This band-wise approach allows the model to capture distinct signal characteristics associated with dementia. The framework employs adaptive temperature scaling during contrastive training to balance the different signal statistics across bands. The model was evaluated on the ADFTD cohort using a strict leave-one-subject-out protocol, achieving an impressive accuracy of 92.4% in distinguishing AD patients from cognitively normal controls, outperforming both supervised models and existing EEG methods. This work highlights the potential of self-supervised learning in leveraging unlabeled EEG data for effective dementia classification.
Methodology
Delta2Gamma employs a self-supervised learning approach based on contrastive learning. It processes EEG data by decomposing it into five frequency bands, each with its own convolutional encoder and projection head. The model uses data augmentation techniques to create multiple views of the same signal, enhancing robustness. An adaptive NT-Xent loss function is utilized to train the model, allowing it to learn meaningful representations from unlabeled data. The framework consists of two main stages: self-supervised pre-training and linear evaluation for classification tasks.
Results
The Delta2Gamma framework achieved a classification accuracy of 92.4% in distinguishing Alzheimer's disease from cognitively normal controls on the ADFTD cohort, surpassing the performance of both supervised learning methods and other dedicated EEG approaches.
Implications
The findings suggest that Delta2Gamma could serve as a scalable and effective tool for early detection of Alzheimer's disease using EEG, potentially improving access to dementia screening in various healthcare settings. This approach may also pave the way for further research into self-supervised learning applications in neurophysiological data analysis.
Real-Time State-of-Health Estimation and Online Degradation Prognosis from Partial Battery Discharge Using Physics-Informed Neural Networks
Time Series
Theory
Efficient ML
- Introduces a physics-informed deep learning approach for SOH estimation from partial battery discharge data.
- Achieves a MAPE of less than 4%, indicating high accuracy in SOH predictions.
- Develops a real-time degradation trend estimation strategy that operates without historical data.
- Applicable to a wide range of battery types, enhancing the flexibility of battery management systems.
Read more
Real-Time State-of-Health Estimation and Online Degradation Prognosis from Partial Battery Discharge Using Physics-Informed Neural Networks
Summary
This paper presents a novel approach for estimating the State of Health (SOH) of lithium-ion batteries using a physics-informed deep learning model. The method leverages incomplete discharge curves from arbitrary voltage ranges, which allows it to operate under realistic and heterogeneous conditions. By integrating data-driven learning with physically motivated degradation dynamics, the proposed model achieves a Mean Absolute Percentage Error (MAPE) of less than 4% in SOH estimation. Additionally, the authors introduce a real-time degradation trend estimation strategy that detects key aging transitions without the need for prior knowledge or historical data. This approach is particularly beneficial as it can be applied to a wide variety of batteries and overcomes limitations of previous methods that relied on fixed protocols or early predictions. The results demonstrate that the proposed method not only provides accurate SOH estimates but also facilitates continuous monitoring of battery health, making it a significant advancement in battery management systems.
Methodology
The authors employ a physics-informed neural network (PINN) framework that integrates data-driven learning with physical degradation models. This hybrid approach allows the model to learn from incomplete discharge curves while ensuring that the predictions remain consistent with the underlying physical principles governing battery degradation.
Results
The proposed method successfully estimates the SOH of lithium-ion batteries with a MAPE below 4%. It also provides real-time degradation trend estimates, enabling the detection of aging transitions without prior knowledge or historical data, thus demonstrating its effectiveness in practical applications.
Implications
This research has significant implications for the management of energy storage systems, particularly in the context of renewable energy integration and electric mobility. The ability to accurately estimate SOH and predict degradation in real-time can enhance battery lifespan, improve safety, and optimize performance in various applications.
Spectral Rank Certification for Foundation Model Adapters
Theory
Efficient ML
- Develops a finite-sample framework for rank certification in LoRA adapters.
- Introduces an empirical-null workflow for effective rank evaluation.
- Finds that calibrated effective rank is typically much smaller than nominal rank.
- Demonstrates that energy retention and statistical surprise address different questions.
Read more
Spectral Rank Certification for Foundation Model Adapters
Summary
This paper addresses the challenge of effectively certifying the rank structure of low-rank adapters (LoRA) used in foundation models. The authors develop a finite-sample framework that distinguishes between nominal and effective ranks of these adapters, which is crucial for understanding their performance. The theoretical foundation is based on an exact chi-square divergence for a Gaussian rank-one reference experiment, integrating an unknown signal direction under a rotation-invariant prior. The paper introduces a series of computational tools, including a finite-sample Le Cam bound and a Monte Carlo p-value workflow for empirical null hypothesis testing. An extensive audit of 26 public adapters reveals that the calibrated effective rank is often significantly lower than the nominal rank, indicating that traditional measures of energy retention do not adequately capture the statistical significance of the components. The findings highlight the need for careful calibration in the evaluation of adapter performance, suggesting that energy retention and statistical surprise provide different insights into model behavior.
Methodology
The authors utilize a theoretical framework based on chi-square divergence to analyze the rank structure of LoRA adapters. They employ Monte Carlo methods for p-value estimation and implement a series of statistical tests to assess the significance of singular components in the adapter spectra. The methodology includes factor reconstruction, stagewise calibration, and block testing for clustered components.
Results
The main results indicate that the calibrated effective rank of LoRA adapters is generally much lower than the nominal rank, with systematic differences observed in energy retention metrics. The empirical findings are supported by an audit of 26 public adapters, encompassing 684 modules and various architecture families, revealing that traditional metrics may not fully capture the statistical properties of the adapters.
Implications
The findings suggest that practitioners should reconsider how they evaluate the performance of low-rank adapters in foundation models. The distinction between nominal and effective ranks can inform better design choices and improve the interpretability of model behavior. This work may influence future research on low-rank adaptations and their applications in various machine learning tasks.
CORAM: Coherent Orthogonal Rotation for Model Merging
NLP
Multimodal
Large Language Models
- CORAM introduces a search-free amplification rule for model merging, maintaining performance close to the optimal coefficient.
- The method employs slice-level manifold merging, enhancing the geometric compatibility of updates.
- Spread slicing and a residual pathway are introduced to improve the distribution of updates and restore fine-tuning signals.
- CORAM demonstrates significant improvements over existing methods across diverse model families and tasks.
Read more
CORAM: Coherent Orthogonal Rotation for Model Merging
Summary
The paper introduces CORAM, a novel approach for merging finetuned models that preserves the geometric structure of weight updates. Traditional methods often rely on linear arithmetic in Euclidean weight space, which fails to account for the underlying geometry of model updates. CORAM improves upon existing techniques like Orthogonal Model Merging (OrthoMerge) by partitioning weight matrices into row slices and utilizing singular value decomposition (SVD) to represent each expert slice. This allows for merging task-specific factors on their respective manifolds. A key innovation is the introduction of a search-free amplification coefficient, λ = κˆc, which compensates for the contraction of updates during manifold averaging. The method also incorporates spread slicing to distribute updates more evenly across slices and a residual pathway for non-target layers. Evaluations across multiple model families and sizes demonstrate that CORAM outperforms OrthoMerge and matches or exceeds the best weight-space baselines, achieving improvements of 0.25 to 1.35 points across various tasks.
Methodology
CORAM partitions weight matrices into row slices and utilizes singular value decomposition (SVD) to represent each expert slice. It merges updates using geometry-compatible operations on the special orthogonal group, Euclidean spectral space, and the Stiefel manifold. The method also includes a conflict-aware variant for handling inconsistent updates and employs an amplification coefficient to counteract update contraction during merging.
Results
CORAM was evaluated across four heterogeneous model-merging suites, covering three model families and sizes ranging from 3B to 9B parameters. The results indicate that CORAM improves upon OrthoMerge by 0.25 to 1.35 points and matches or exceeds the strongest weight-space baselines, demonstrating its effectiveness in model merging tasks.
Implications
The advancements presented in CORAM have significant implications for the field of model merging, particularly in scenarios where multiple specialized models need to be combined without access to original training data. This approach can enhance the capabilities of AI systems in various applications, including natural language processing and computer vision, by allowing for more efficient and effective model integration.
SCENARIODIFF: A Scenario-level Guidance Framework for Multimodal Time Series Forecasting--Extended Version
Time Series
Multimodal
Large Language Models
- SCENARIODIFF organizes contextual information into three levels for improved interpretability.
- The framework effectively integrates textual context with numerical time series data.
- Anchor Blended Sampling allows for local refinement of forecasts without retraining.
- Experiments show strong performance in event-driven forecasting scenarios.
Read more
SCENARIODIFF: A Scenario-level Guidance Framework for Multimodal Time Series Forecasting--Extended Version
Summary
The paper introduces SCENARIODIFF, a hierarchical contextual reasoning framework designed for multimodal time series forecasting (MTSF). Traditional forecasting models primarily rely on numerical histories, often neglecting the valuable insights provided by contextual information such as news and reports. SCENARIODIFF addresses this gap by organizing contextual information into three distinct levels: a Historical Context Agent that extracts evidence from raw documents, a Scenario Agent that generates qualitative scenario descriptions for the forecast horizon, and an Anchor Guidance Agent that produces sparse anchor points for future regions likely to experience abrupt changes. This structured approach enhances the interpretability and control over the influence of contextual signals on forecasts. The framework utilizes a Multimodal Diffusion Transformer, which iteratively refines forecasts based on the structured signals from the agents. The method also incorporates Anchor Blended Sampling to locally adjust generated trajectories without the need for retraining. Experiments conducted on the Time-MMD benchmark demonstrate that SCENARIODIFF significantly improves forecasting accuracy in event-driven domains, showcasing the effectiveness of explicit hierarchical scenario guidance in multimodal time series forecasting.
Methodology
SCENARIODIFF employs a hierarchical framework consisting of three agents: a Historical Context Agent for summarizing raw documents, a Scenario Agent for creating qualitative scenario descriptions, and an Anchor Guidance Agent for generating localized anchor points. These agents guide a Multimodal Diffusion Transformer, which models future trajectories through iterative denoising based on structured signals. The method also introduces Anchor Blended Sampling to refine forecasts locally.
Results
The evaluation on the Time-MMD benchmark indicates that SCENARIODIFF outperforms existing multimodal forecasting methods, particularly in scenarios driven by external events, demonstrating its effectiveness in accurately predicting future dynamics that are not evident from numerical histories alone.
Implications
The findings suggest that incorporating structured contextual guidance can significantly enhance the accuracy of time series forecasting models, particularly in domains where external events play a crucial role. This approach could be applied in various fields such as finance, energy, and healthcare, where understanding the influence of contextual factors is essential for decision-making.
Understanding Curriculum Learning in Large Language Models via Cross-Difficulty Optimization Dynamics
NLP
Large Language Models
Optimization
- Curriculum learning's effectiveness is task-dependent and not universally optimal.
- The concept of Relative Transfer provides a theoretical framework for understanding cross-difficulty knowledge transfer.
- Transfer-aware Dynamic Curriculum Sampling (TDCS) adapts the sampling strategy based on estimated transfer benefits.
- TDCS outperforms fixed scheduling strategies across various reasoning tasks and model scales.
Read more
Understanding Curriculum Learning in Large Language Models via Cross-Difficulty Optimization Dynamics
Summary
This paper investigates the effectiveness of curriculum learning in large language models (LLMs), which organizes training data from easy to hard. The authors identify that the success of curriculum learning varies significantly across different reasoning tasks, prompting an exploration of the underlying mechanisms that determine its effectiveness. They analyze the optimization dynamics induced by various curriculum schedules and introduce the concept of 'Relative Transfer', a measure of cross-difficulty knowledge transfer that characterizes these dynamics. Based on this analysis, they propose Transfer-aware Dynamic Curriculum Sampling (TDCS), an adaptive framework that adjusts the sampling distribution according to the estimated transfer relationships during training. Extensive experiments across multiple reasoning benchmarks demonstrate that TDCS consistently outperforms traditional fixed scheduling strategies, providing a unified optimization-based explanation for the varying effectiveness of curriculum learning.
Methodology
The authors conducted a systematic empirical study across multiple reasoning benchmarks to analyze the optimization dynamics of curriculum learning. They introduced the concept of Relative Transfer to formalize the transfer relationship between different difficulty levels and developed TDCS, which dynamically adjusts the sampling distribution based on this analysis.
Results
The experiments showed that TDCS consistently outperformed existing fixed scheduling strategies across various reasoning tasks, model scales, and self-improvement settings, highlighting the importance of adaptive curriculum strategies based on optimization dynamics.
Implications
The findings suggest that understanding the optimization dynamics of curriculum learning can lead to more effective training strategies for large language models, potentially improving their performance on a wide range of reasoning tasks.
TRACE-CASH: Trial-History-Conditioned Reinforcement Learning for Adaptive Configuration Exploration in Time-Series CASH
Reinforcement Learning
Time Series
Optimization
- TRACE-CASH combines history-conditioned candidate generation with fixed search rules for effective exploration in time-series forecasting.
- The method is evaluated against six heterogeneous search strategies under a unified evaluation protocol.
- TRACE-CASH outperforms other methods in terms of MASE and WQL metrics across multiple task variants.
- The study addresses the challenges of evaluating forecasting models in a time-sensitive context.
Read more
TRACE-CASH: Trial-History-Conditioned Reinforcement Learning for Adaptive Configuration Exploration in Time-Series CASH
Summary
The paper introduces TRACE-CASH, a novel hybrid sequential optimizer designed for the combined algorithm selection and hyperparameter optimization (CASH) problem in time-series forecasting. The authors highlight the complexities of time-series forecasting, including the need for chronological validation and the computational costs associated with evaluating deep learning models. TRACE-CASH employs a task-local approach that utilizes a grouped actor-critic mechanism for candidate generation, which is conditioned on trial history. This method integrates fixed rules for model coverage, validation-guided exploitation, and exploration after stalled progress. The authors conduct a comprehensive evaluation comparing TRACE-CASH against six other search methods across 41 dataset-frequency task variants. The results demonstrate that TRACE-CASH achieves the lowest mean rank on both mean absolute scaled error (MASE) and weighted quantile loss (WQL), indicating its competitive performance in the context of time-series CASH.
Methodology
TRACE-CASH utilizes a hybrid approach that combines a grouped actor-critic mechanism for generating candidate configurations with fixed rules for model coverage and validation-guided exploitation. The method is reinitialized for each task-seed run and adapts based on accumulated trial history, allowing for dynamic exploration and exploitation of the search space.
Results
TRACE-CASH achieved the lowest mean rank on both MASE and WQL metrics compared to six alternative methods. It also demonstrated the lowest window-averaged test-MASE rank in both predefined full and late windows, indicating its effectiveness in time-series forecasting tasks.
Implications
The findings suggest that TRACE-CASH can significantly enhance the efficiency of automated machine learning systems in time-series forecasting, providing a competitive edge in selecting models and hyperparameters. This could lead to better forecasting accuracy in various applications, including finance, healthcare, and environmental monitoring.
Multi-Feature Riemannian Hypergraph for Online Test-Time Adaptation of Motor Imagery Brain-Computer Interface
Multimodal
Time Series
Graph Learning
- Introduction of the Multi-feature Riemannian Hypergraph (MRieHy) framework for MI-BCI adaptation.
- Utilization of Riemannian geometry to enhance cross-day transferability in brain signal decoding.
- Combination of covariance matrix representations and deep features for improved prediction accuracy.
- First validation of online test-time adaptation on human MI-BCI ECoG data.
Read more
Multi-Feature Riemannian Hypergraph for Online Test-Time Adaptation of Motor Imagery Brain-Computer Interface
Summary
This paper addresses the challenges of cross-day transferability and online operation in Motor Imagery Brain-Computer Interfaces (MI-BCIs). The authors propose a novel framework called Multi-feature Riemannian Hypergraph (MRieHy) that leverages Riemannian geometry to enhance the adaptability of MI-BCI systems during online testing. MRieHy computes Riemannian means of covariance matrices from training data across multiple days to align distributions, and constructs hypergraphs based on Riemannian distances and deep features. The two hypergraphs are fused using adaptively learned weights, optimizing the prediction process. During online testing, MRieHy employs a first-in-first-out buffer for recent samples, aligning buffered data with Riemannian techniques before decoding with the hypergraph. The framework is validated through extensive experiments on private and public datasets, demonstrating significant performance improvements over existing state-of-the-art methods.
Methodology
The MRieHy framework computes Riemannian means of covariance matrices from cross-day training data to align distributions. It constructs two hypergraphs: one based on Riemannian distances of covariance matrices and another on deep features using cosine similarity. These hypergraphs are fused with adaptively learned weights and optimized with label projection matrices. During online testing, a buffer of recent samples is maintained, and Riemannian alignment is performed on this data before decoding.
Results
The experiments conducted on a private four-class ECoG dataset and two public four-class EEG datasets show that MRieHy significantly outperforms state-of-the-art baselines, confirming its effectiveness in online test-time adaptation for MI-BCI systems.
Implications
The proposed MRieHy framework has the potential to improve the usability of MI-BCIs in real-world applications, particularly for individuals with severe motor impairments. By enabling effective online adaptation without the need for extensive calibration, it could facilitate more accessible and reliable brain-computer communication.
Task Specialization Fine-Tuning for Contextual Reinforcement Learning
Reinforcement Learning
Optimization
- Introduces a budget-constrained fine-tuning paradigm for CRL, shifting focus from where to train to how much fine-tuning to allocate.
- Develops TSFT, an online framework that optimizes budget allocation using integer linear programming.
- Demonstrates significant improvements in task coverage and sample efficiency across diverse decision-making domains.
- Aligns with modern machine learning practices by integrating pretraining and fine-tuning strategies.
Read more
Task Specialization Fine-Tuning for Contextual Reinforcement Learning
Summary
This paper introduces Task Specialization Fine-Tuning (TSFT), a novel approach for Contextual Reinforcement Learning (CRL) that enhances task coverage by fine-tuning a pretrained policy across various task regions. Traditional methods in CRL often involve training from scratch, which can be inefficient and limited in scalability. TSFT proposes a unified framework that first pretrains a single policy and then fine-tunes it for specialization under a constrained budget. The authors address the challenge of heterogeneous fine-tuning returns across tasks by formulating a budget-constrained fine-tuning problem. TSFT employs a parametric model to predict performance improvements and utilizes integer linear programming (ILP) to optimally allocate fine-tuning resources. The framework is designed to adaptively update its model based on new data, enhancing its performance over time. The extensive experiments conducted across various domains, including combinatorial optimization and continuous control, demonstrate that TSFT significantly outperforms existing methods, achieving up to 2-3 times better task coverage and doubling sample efficiency compared to multi-task training baselines.
Methodology
The methodology involves pretraining a single policy across a context space, followed by fine-tuning multiple specialized policies. TSFT predicts the performance of fine-tuning using a parametric model and solves the budget allocation problem through integer linear programming. The framework is designed to adaptively re-estimate the model as new data is collected, ensuring optimal allocation of resources.
Results
The empirical results show that TSFT achieves 2-3 times improvements in task coverage compared to simpler strategies and doubles the sample efficiency over multi-task training baselines. It performs comparably to an oracle policy across various settings, indicating its effectiveness in maximizing task coverage under budget constraints.
Implications
The findings suggest that TSFT can significantly enhance the efficiency and effectiveness of contextual reinforcement learning applications, making it suitable for complex decision-making tasks where resource allocation is critical. This approach could be applied in various fields, including robotics, combinatorial optimization, and large-scale machine learning tasks.
Early Cycle Charge Trajectory Generative Prediction and Full Life Cycle Health Management of Iron-Chromium Flow Batteries Based on FlowBD-E1
Time Series
Generative Models
- FlowBD-E1 predicts future charge V/I trajectories from early-cycle data, enhancing battery health management.
- The model outperformed traditional LSTM and TCN baselines, demonstrating superior accuracy in trajectory generation.
- The study emphasizes the importance of analyzing full charge trajectories rather than reducing aging to scalar metrics.
- Three deployment strategies (SLE, RLF, TFU) were compared to assess forecasting accuracy and diagnostic capabilities.
Read more
Early Cycle Charge Trajectory Generative Prediction and Full Life Cycle Health Management of Iron-Chromium Flow Batteries Based on FlowBD-E1
Summary
This paper addresses the challenge of predicting the degradation of iron-chromium redox flow batteries (RFBs) by introducing FlowBD-E1, a generative forecasting framework that predicts complete future charge voltage/current (V/I) trajectories based on early-cycle data. The authors highlight the limitations of existing prognostic studies, which often focus on lithium-ion batteries and reduce aging to scalar metrics like state-of-health (SOH). Instead, FlowBD-E1 utilizes a multi-scale convolutional encoder, a lifecycle Transformer, and an age-aware FiLM decoder to generate future charge trajectories from just the first few cycles of operation. The model was evaluated on an industrial 33 kW Fe-Cr RFB, achieving a mean absolute percentage error (MAPE) of 0.731% for V/I predictions and sub-1% MAPE for SOH estimates. The study demonstrates that early-cycle data can provide significant insights into long-term battery health, enabling better management of flow-battery systems.
Methodology
The authors developed FlowBD-E1, which integrates a multi-scale convolutional encoder, a lifecycle Transformer, and an age-aware FiLM decoder. The model was trained using data from a 33 kW industrial Fe-Cr RFB, with a focus on predicting future charge V/I trajectories based on the first few cycles of operation. Three deployment strategies were tested: single-step latent extrapolation (SLE), recursive latent forecasting (RLF), and teacher-forced updating (TFU).
Results
FlowBD-E1 achieved a joint V/I MAPE of 0.731% over the remaining lifecycle and produced SOH estimates with less than 1% MAPE. The age-aware generative architecture outperformed LSTM and TCN baselines, maintaining sub-percent errors during industrial validation. The results suggest that early-cycle trajectory generation can transform limited commissioning data into valuable long-term diagnostic signals.
Implications
The findings imply that early-cycle data can be leveraged for effective health management of iron-chromium flow batteries, potentially improving the reliability and efficiency of energy storage systems. This approach could be extended to other types of batteries and energy storage technologies, enhancing predictive maintenance and operational strategies.
Agents unlock new capabilities through Switching LoRA Adapters as a Tool (SLAaaT)
Large Language Models
NLP
Optimization
- Introduces Switching LoRA Adapters (SLAaaT) to enhance agent capabilities.
- Addresses the issue of catastrophic forgetting in small language models.
- Demonstrates significant performance improvements in synthetic coding tasks.
- Achieves up to an 18x reduction in capability tax compared to single adapter use.
Read more
Agents unlock new capabilities through Switching LoRA Adapters as a Tool (SLAaaT)
Summary
This paper introduces a novel approach for enhancing the capabilities of agents through the use of Switching LoRA Adapters (SLAaaT). The primary challenge addressed is the tradeoff between specialization and generality in machine learning models, particularly in small language models that can easily forget previously learned tasks when fine-tuned for new ones. The proposed method allows agents to autonomously switch between specialized LoRA adapters during task execution, enabling them to tackle multiple specialized tasks without suffering from catastrophic forgetting. The authors conducted experiments using synthetic coding tasks that required specific skills, demonstrating that the agent could effectively learn when to switch adapters based on a single in-context example. The results showed that this approach led to a significant reduction in capability tax—up to 18 times less than using a single specialized adapter—and outperformed traditional methods, including the use of subagents. The findings suggest that the SLAaaT method not only enhances task performance but also allows for innovative routing strategies that surpass human heuristics in certain scenarios.
Methodology
The study involved designing synthetic coding tasks that required multiple specialized skills, which were absent from the base model. The authors implemented a framework allowing agents to switch between LoRA adapters mid-task. They trained the adapters on synthetic data and evaluated the model's performance across various task configurations, comparing it to baseline methods including single adapter use and subagent invocation.
Results
The results indicated that the agent using the SLAaaT method achieved higher task success rates compared to other configurations. Specifically, the model scored significantly better in end-to-end task completion, with the autoswitch mechanism outperforming human heuristic strategies in certain conditions. The analysis revealed that the switching mechanism preserved the model's overall capacity while allowing it to effectively utilize specialized skills.
Implications
The findings suggest that the SLAaaT approach could be applied to various domains requiring flexible and adaptive learning strategies, particularly in environments where agents must handle multiple tasks without losing previously acquired knowledge. This could lead to advancements in the development of more robust AI systems capable of complex problem-solving.
Tail-Aware Top-k On-Policy Distillation
NLP
Large Language Models
Optimization
- Normalized top-k OPD neglects tail probability, leading to increased entropy and reduced accuracy.
- TA-OPD incorporates a tail token to restore the tail probability signal, improving alignment between student and teacher distributions.
- The method provides a theoretical lower bound for full-vocabulary reverse KL divergence.
- Extensive experiments show TA-OPD improves accuracy by up to 8.05 points on common benchmarks.
Read more
Tail-Aware Top-k On-Policy Distillation
Summary
This paper introduces Tail-Aware Top-k On-Policy Distillation (TA-OPD), a novel method for knowledge distillation in language models that addresses the limitations of existing on-policy distillation techniques. Traditional methods often minimize the reverse Kullback-Leibler (KL) divergence between the student and teacher's distributions over the teacher's top-k tokens, which neglects the tail probability—the probability mass outside the top-k tokens. This oversight can lead to increased tail probability and entropy in the student model, resulting in degraded performance on downstream tasks. TA-OPD rectifies this by incorporating a tail token that captures the tail probability, thereby aligning the student’s distribution more closely with the teacher’s. The authors provide theoretical backing for their approach, demonstrating that TA-OPD serves as a tight lower bound for the full-vocabulary reverse KL divergence. Extensive experiments validate the effectiveness of TA-OPD, showing significant improvements in accuracy on various benchmarks compared to normalized top-k OPD, particularly in scenarios with a large capability gap between the student and teacher models.
Methodology
The authors propose TA-OPD, which minimizes the reverse KL divergence over the top-k tokens plus a tail token that captures the tail probability. They also derive a sampled variant of TA-OPD that provides an unbiased estimate of the full-vocabulary reverse KL divergence when the sampled token's probability is known. The methodology is validated through extensive experiments on mathematical benchmarks.
Results
TA-OPD significantly outperforms normalized top-k OPD, stabilizing the distillation process and maintaining lower training entropy. For instance, when distilling the OpenThinker3-7B model into Qwen2.5-7B-Instruct, TA-OPD improved the Avg@8 accuracy on the MATH500 benchmark from 68.78% to 77.88%, while keeping training entropy below 1.5.
Implications
The findings suggest that incorporating tail probability into distillation processes can enhance the performance of smaller language models, making them more effective in practical applications. This approach could be beneficial in scenarios where model efficiency and accuracy are critical, such as in real-time language processing tasks.
Abra: Scaling Diffusion Image Training
Generative Models
Computer Vision
Optimization
- Diffusion models require approximately 200 image tokens per parameter for compute optimality, significantly more than language models.
- The training of diffusion models is robust to overtraining, allowing for smaller models trained on larger datasets.
- Generative quality metrics and representation quality scale predictably with compute, similar to language models.
- The study introduces the concept of scaling collapse in diffusion models, indicating a universal scaling behavior.
Read more
Abra: Scaling Diffusion Image Training
Summary
This paper presents a comprehensive study on the scaling laws for text-to-image diffusion models, specifically using a controlled family of flow-matching transformers called ABRA. The authors explore compute-optimal scaling laws, demonstrating that diffusion models scale predictably similar to language models but require significantly more data to achieve optimal training performance. The study reveals that compute optimality for diffusion models occurs at approximately 200 image tokens per parameter, which is ten times the requirement for large language models. Additionally, the authors find that diffusion models are robust to overtraining, suggesting that practitioners should prioritize larger datasets over larger model sizes. The paper also discusses the predictability of generative quality metrics, representation quality, and the universal shape of training curves, contributing to a deeper understanding of scaling in visual generation.
Methodology
The authors conducted a systematic scaling study using ABRA, a family of flow-matching transformers, across three orders of magnitude in compute (from 10^19 to 10^22 FLOPs). They utilized µP to derive scaling laws and analyzed the relationship between model size, dataset size, and training performance, focusing on generative quality metrics and representation quality.
Results
The findings indicate that diffusion models achieve compute optimality at 200 image tokens per parameter, with robust performance against overtraining. The study also confirms that generative metrics and representation quality scale predictably, and introduces the concept of scaling collapse, where training curves exhibit a universal form.
Implications
These insights could guide practitioners in optimizing the training of diffusion models, emphasizing the importance of data size over model size. The findings may also influence future research in visual generation and scaling laws, potentially leading to more efficient training methodologies.
GEO-Flag: Detecting and Measuring GEO-Optimized Web Content
NLP
Large Language Models
Generative Models
- Introduction of GEOFlagBench as a benchmark for evaluating GEO detection methods.
- Identification of significant weaknesses in existing detection methods, particularly reliance on authorship cues.
- Development of Intervention-Paired Training (IPT) to enhance GEO detection accuracy.
- Creation of a GEO-gated Agent system for auditing citation URL verifiability.
Read more
GEO-Flag: Detecting and Measuring GEO-Optimized Web Content
Summary
The paper addresses the emerging issue of Generative Engine Optimization (GEO), which modifies web content to enhance its visibility and citation likelihood by generative search engines. This optimization can lead to the propagation of weak or false information, as generative search engines synthesize information into direct answers, obscuring the original sources' authority. To tackle this problem, the authors introduce GEOFlagBench, a benchmark consisting of 3,200 webpages across various domains and GEO optimizer families, to systematically evaluate existing GEO detection methods. The study finds that while the best baseline achieves an F1 score of 0.880, significant weaknesses remain, particularly in relying on authorship cues. To improve detection, the authors propose Intervention-Paired Training (IPT), which enhances the F1 score to 0.944 and reduces false-positive rates. Additionally, they develop a GEO-gated Agent system for auditing citation URLs, achieving high accuracy in assessing source verifiability. The complete pipeline is deployed on real-world search results, estimating a GEO prevalence of 8.90% among analyzed pages, with a notable increase in prevalence over time. The findings highlight the necessity for systematic detection and auditing of GEO content in search ecosystems.
Methodology
The authors created GEOFlagBench to evaluate various GEO detection methods, including fine-tuned classifiers and LLMs. They proposed Intervention-Paired Training (IPT) to improve detection by supervising responses to GEO interventions. A GEO-gated Agent system was developed to assess the verifiability of citation URLs from detected GEO pages.
Results
The strongest baseline detection method achieved an F1 score of 0.880, while IPT improved this to 0.944. The GEO-gated Agent system achieved 84.75% accuracy for URL Source Tier and 83.00% for Citation URL Verifiability. The overall GEO prevalence in analyzed search results was estimated at 8.90%, with significant increases noted in subsequent years.
Implications
The findings underscore the need for robust detection and auditing mechanisms for GEO-optimized content, which can mislead users and distort information access in generative search environments. The proposed methods and benchmarks can serve as foundational tools for future research and development in this area.
PERO: Efficient Robust Post-Training Foundation Models for Encrypted Traffic Classification
Optimization
Efficient ML
- PERO introduces a lightweight pre-evaluation module for efficient risk estimation in encrypted traffic classification.
- The framework significantly reduces computational overhead while maintaining robust performance.
- Theoretical analysis provides insights into the convergence behavior and generalization of the proposed method.
- Extensive experiments validate PERO's effectiveness against existing robust post-training techniques.
Read more
PERO: Efficient Robust Post-Training Foundation Models for Encrypted Traffic Classification
Summary
The paper addresses the critical issue of encrypted traffic classification, which is essential for network security, particularly in identifying malicious traffic. The authors introduce Pre-Evaluation Robust Optimization (PERO), a novel framework designed to enhance the robustness of post-training foundation models while minimizing computational costs. Traditional methods often rely on empirical risk minimization, which can neglect high-risk tail events, leading to significant misclassification errors. PERO mitigates this by employing a lightweight proxy to estimate sample-wise risk, allowing for the selective updating of high-risk samples without the computational burden associated with full model optimization. The framework is designed to operate efficiently under limited computational and memory constraints, making it suitable for high-throughput network environments. The authors provide a theoretical analysis of PERO, detailing its convergence behavior and generalization bounds. Extensive experiments demonstrate that PERO achieves competitive or superior robustness compared to existing robust post-training methods, while significantly reducing the associated computational and memory costs.
Methodology
PERO utilizes a lightweight proxy to estimate sample-wise risk and selectively updates high-risk samples during post-training. This decouples the risk estimation process from the expensive optimization of large models, allowing for efficient and targeted improvements in model robustness.
Results
Experimental results indicate that PERO achieves competitive or superior robustness compared to leading robust post-training methods, while also incurring significantly lower computational and memory costs.
Implications
The proposed PERO framework has the potential to enhance the security of network systems by improving the accuracy of encrypted traffic classification, thereby facilitating timely detection of intrusions and malicious activities. Its efficiency makes it applicable in real-time network environments where computational resources are limited.
Policy-Invariant Reward Shaping from LLM Feedback: A Framework for Hybrid RL Agents
Reinforcement Learning
Large Language Models
Theory
- Introduces a formal framework for hybrid LLM and RL agents using GA-MDP.
- Proves that LLM-derived reward shaping does not alter optimal policies.
- Validates the theoretical guarantees through numerical experiments.
- Provides a reference implementation and full inference algorithm.
Read more
Policy-Invariant Reward Shaping from LLM Feedback: A Framework for Hybrid RL Agents
Summary
This paper presents a novel framework for integrating large language models (LLMs) with reinforcement learning (RL) agents, formalized as a Goal-Augmented Markov Decision Process (GA-MDP). The authors establish that using LLM-derived per-state progress scores as a bounded potential function ensures that the optimal policies of the augmented MDP remain unchanged, regardless of inaccuracies in the LLM's scoring. This finding offers a stronger guarantee than existing LLM-as-reward approaches. The framework is validated through numerical experiments on a small MDP and a series of MiniGrid tasks, demonstrating the feasibility of the proposed architecture. The authors also provide a reference implementation and outline a full inference algorithm, while acknowledging that empirical performance comparisons with existing baselines are planned for future work.
Methodology
The authors formalize the hybrid architecture as a GA-MDP and prove the policy invariance of the reward shaping through theoretical analysis. They conduct numerical experiments on a small MDP and evaluate the planner's performance on 20 MiniGrid tasks, while also running a pipeline-validation study to diagnose integration issues.
Results
The theoretical guarantee that LLM-derived reward shaping does not affect optimal policies was confirmed through numerical validation on a 3-state MDP under various configurations. The framework was successfully implemented and tested on MiniGrid tasks, revealing one integration failure related to vocabulary mismatches.
Implications
This work provides a theoretical foundation for the integration of LLMs in RL, ensuring that the use of LLM feedback does not compromise the learning of optimal policies. It opens avenues for more robust hybrid RL systems that leverage LLMs for task decomposition without risking suboptimal learning due to incorrect reward signals.
EMASAM: a Computationally Efficient Sharpness-Aware Minimization via EMA-Guided Perturbations
Optimization
Efficient ML
- EMASAM is a computationally efficient variant of Sharpness-Aware Minimization (SAM).
- It eliminates the need for an extra gradient computation during the perturbation step.
- EMASAM uses Exponential Moving Average (EMA) weights to guide perturbations, enhancing stability.
- The method preserves the generalization ability of SAM while reducing computational costs.
Read more
EMASAM: a Computationally Efficient Sharpness-Aware Minimization via EMA-Guided Perturbations
Summary
The paper introduces Exponential Moving Average Sharpness-Aware Minimization (EMASAM), a novel approach to enhance the efficiency of Sharpness-Aware Minimization (SAM) in training machine learning models. SAM has been recognized for its ability to improve model generalization by encouraging flatter minima in the loss landscape. However, SAM's computational cost is a significant drawback, as it requires two gradient computations per training iteration, effectively doubling the training time. EMASAM addresses this limitation by eliminating the need for an additional gradient computation during the perturbation step. Instead, it utilizes the Exponential Moving Average (EMA) of model weights to define perturbation directions, which helps in navigating towards less stable areas of the loss landscape without relying on noisy mini-batch gradients. This approach not only reduces computational overhead but also enhances training stability by mitigating gradient-induced instability. The authors conducted several experiments to validate the efficiency and robustness of EMASAM, demonstrating that it maintains the generalization capabilities of SAM while being computationally efficient. The code for EMASAM is publicly available for further research and application.
Methodology
EMASAM replaces the gradient-based perturbation in SAM with perturbations defined by the discrepancy between the main model and its EMA shadow model. This approach avoids the need for additional backpropagation and mitigates the instability caused by noisy mini-batch gradients.
Results
The experiments conducted show that EMASAM is not only computationally efficient but also maintains or improves upon the generalization performance achieved by SAM. The results indicate that EMASAM can effectively navigate the loss landscape while reducing training time.
Implications
EMASAM has the potential to be applied in large-scale machine learning tasks where computational resources are a concern. Its efficiency makes it suitable for training complex models on extensive datasets, thereby facilitating broader applications in various domains of machine learning.
Breaking the Compression Barrier: Cross-Architecture Compression Boundary Learning via Reverse Regrowth
Efficient ML
- Introduces BRIDGE, a reverse boundary-learning framework for model compression.
- Identifies compression boundary estimation as a critical yet overlooked challenge in network compression.
- Demonstrates architecture independence by effectively recovering performance in CNNs and Transformers.
- Achieves performance improvements of up to 1.49% in unstructured pruning and 4.77% in structured pruning.
Read more
Breaking the Compression Barrier: Cross-Architecture Compression Boundary Learning via Reverse Regrowth
Summary
The paper addresses the challenge of model compression for resource-constrained edge devices, focusing on the limitations of existing pruning methods that often lead to abrupt performance degradation beyond a certain sparsity threshold. The authors propose a novel framework called BRIDGE, which reformulates the compression problem as a boundary-learning task. Unlike traditional forward pruning techniques, BRIDGE first reduces the model to a highly sparse state to identify the compression boundary and then selectively regenerates critical model components to restore performance. This reverse regrowth process employs a hierarchical strategy that combines coarse-grained layer selection with fine-grained parameter recovery. The framework is shown to be architecture-independent, effectively recovering models from the brink of collapse across various architectures, including CNNs and Transformers. The proposed method achieves notable performance improvements in both unstructured and structured pruning scenarios, demonstrating its potential to extend the compression limits of neural networks while maintaining stable accuracy.
Methodology
The BRIDGE framework reformulates model compression as a boundary-learning problem. It intentionally traverses the compression boundary to expose its location and employs a reverse regrowth strategy to selectively restore the most critical parameters. The regrowth process is modeled as a Markov Decision Process, utilizing a reinforcement learning-based controller to allocate a parameter budget through a two-level hierarchical strategy: coarse-grained layer selection based on structural degradation and fine-grained parameter selection using gradient-based saliency scores.
Results
BRIDGE consistently recovers model accuracy in high-sparsity regimes across four different architectures and two pruning paradigms, with search costs maintained under 10 GPU-hours per model. The framework demonstrates significant performance improvements, achieving up to 1.49% enhancement in unstructured pruning and up to 4.77% in structured pruning.
Implications
The findings suggest that BRIDGE can be a valuable tool for deploying deep learning models on resource-constrained devices, as it allows for more aggressive compression without sacrificing performance. This has potential applications in mobile computing, embedded systems, and other areas where computational resources are limited.
Hierarchical Data Selection via Manifold Coverage and Sparse Feature Coverage in LLM Post-training
Large Language Models
Efficient ML
NLP
- MASS reformulates data selection as a coarse-to-fine hierarchical coverage problem.
- The method employs a dense autoencoder for learning principal manifold coordinates and a sparse autoencoder for fine-grained feature selection.
- MASS incorporates external quality scores to improve the selection of high-value samples.
- Experimental results show that MASS outperforms existing data selection methods under multiple conditions.
Read more
Hierarchical Data Selection via Manifold Coverage and Sparse Feature Coverage in LLM Post-training
Summary
The paper addresses the challenge of selecting high-value subsets from large pools of supervised fine-tuning data for large language models (LLMs). Traditional methods often rely on geometric metrics in the original embedding space, which can obscure important semantic distinctions and introduce noise. The authors propose a novel approach called MASS (Manifold Aware Sparse Selection), which reformulates data selection as a hierarchical coverage problem. This method first utilizes a dense autoencoder to learn low-dimensional principal manifold coordinates for coarse semantic grouping, followed by a TopK sparse autoencoder to perform quality-aware sparse feature coverage within each group. The experiments conducted on Vision Flan and LLaVA-CoT datasets demonstrate that MASS consistently outperforms existing data selection baselines across various budgets and settings, achieving performance that matches or exceeds that of full data training using only a small subset of data. The findings suggest that the hierarchical approach to data selection can significantly enhance the efficiency and effectiveness of model training.
Methodology
MASS employs a two-step approach: first, it uses a dense autoencoder to learn low-dimensional principal manifold coordinates for coarse semantic grouping. Then, it applies a TopK sparse autoencoder to extract sparse features and perform fine-grained coverage selection within each group. The method also integrates external quality scores to enhance the selection process.
Results
The experimental results indicate that MASS consistently outperforms strong baseline methods across various datasets and data budgets. In certain scenarios, it matches or surpasses the performance of full data training while utilizing only a small subset of data, demonstrating its effectiveness and robustness.
Implications
The proposed method has significant implications for reducing training costs and improving model performance in large-scale supervised learning tasks. It can be particularly beneficial in scenarios where data is abundant but computational resources are limited, enabling more efficient training of large language models.
MAGPIE-Net: Predicting short-duration heavy-rainfall events in station neighborhoods from multitemporal FY-4A AGRI observations
Time Series
- MAGPIE-Net effectively predicts short-duration heavy rainfall using satellite observations.
- The model employs a unique grid-to-station mapping that enhances prediction accuracy.
- In independent tests, MAGPIE-Net outperformed traditional gridded-output models significantly.
- The model achieved a high detection rate and lead time for early warnings of heavy rainfall events.
Read more
MAGPIE-Net: Predicting short-duration heavy-rainfall events in station neighborhoods from multitemporal FY-4A AGRI observations
Summary
The paper presents MAGPIE-Net, a novel deep learning model designed to predict short-duration heavy rainfall events in the vicinity of weather stations using multitemporal observations from the Fengyun-4A Advanced Geostationary Radiation Imager (FY-4A AGRI). The model addresses the challenge of converting satellite observations of cloud and moisture evolution into localized rainfall warnings, which is crucial for mitigating the impacts of flash floods and urban flooding in densely populated areas. Unlike traditional methods that rely on gridded precipitation predictions, MAGPIE-Net incorporates a geographically adaptive, differentiable mapping from grid to station, allowing for direct supervision of satellite-to-station learning pathways. The model combines convection-initiation features, multiscale encoding, and auxiliary gridded precipitation diagnosis to enhance prediction accuracy. In tests conducted during the 2023 warm season over central and eastern China, MAGPIE-Net demonstrated superior performance compared to baseline models, achieving a critical success index (CSI) of 0.371, 0.304, and 0.238 for 0-1, 1-2, and 2-3 hour forecasts, respectively. The model also achieved a detection rate of 65.1% with a mean lead time of 64.6 minutes, significantly outperforming the best gridded-output baseline. These findings indicate that event-oriented satellite-to-station modeling is more effective for generating local heavy-rainfall warnings than traditional gridded-precipitation approaches.
Methodology
MAGPIE-Net integrates a geographically adaptive, differentiable grid-to-station mapping with deep learning techniques that utilize convection-initiation features and multiscale encoding. The model is trained on multitemporal infrared and water-vapor observations from the FY-4A AGRI to predict heavy rainfall events within a specified radius of weather stations.
Results
MAGPIE-Net achieved critical success index (CSI) values of 0.371, 0.304, and 0.238 for 0-1, 1-2, and 2-3 hour forecasts, respectively. It recorded a detection rate of 65.1% and a mean lead time of 64.6 minutes, significantly surpassing the baseline model's performance, which had a detection rate of 23.6% and a lead time of 18.3 minutes.
Implications
The findings suggest that MAGPIE-Net can enhance operational forecasting and early warning systems for heavy rainfall, potentially reducing the risk of flash floods and improving disaster preparedness in urban areas. This model could be integrated into meteorological services to provide timely and localized rainfall warnings.
Evaluating RL Explainability Methods by How Much They Help Fix Bugs in Agents
Reinforcement Learning
Interpretability
- Introduction of EvalXRL, a benchmark for evaluating XRL methods based on their diagnostic utility.
- Focus on the practical application of XRL methods in debugging RL agents rather than traditional metrics.
- Implementation of a closed-loop evaluation process allowing iterative use of XRL methods.
- Utilization of a Large Language Model as the coding agent for diagnosis and repair tasks.
Read more
Evaluating RL Explainability Methods by How Much They Help Fix Bugs in Agents
Summary
This paper presents a novel evaluation benchmark for Explainable Reinforcement Learning (XRL) methods, termed EvalXRL. The authors critique existing evaluation metrics that focus on faithfulness and subjective ratings, proposing instead that the effectiveness of XRL methods should be assessed based on their ability to assist developers in diagnosing and fixing malfunctioning RL agents. EvalXRL utilizes a Large Language Model (LLM) coding agent to interactively diagnose and repair RL agents by employing various XRL methods. The evaluation framework is designed around tuples of environment, malfunction, and XRL method, where the success of each method is measured by the performance improvement of the repaired agent, as indicated by its reward signal. The authors emphasize a closed-loop process where the coding agent can iteratively invoke the XRL method, refine hypotheses, and adjust parameters based on intermediate results. This approach aims to provide a more practical and measurable evaluation of XRL methods, addressing the diverse needs of stakeholders in the RL community, including developers, end-users, and regulators.
Methodology
The proposed methodology involves creating deliberately malfunctioning RL agents and using a Large Language Model (LLM) coding agent to diagnose and repair these agents through various XRL methods. The evaluation is structured around a closed-loop process where the coding agent can repeatedly invoke the XRL methods, refine its approach based on feedback, and measure the performance of the repaired agent using its reward signal.
Results
The paper outlines a framework for evaluating XRL methods but does not present specific experimental results, as it is a preliminary proposal. The authors suggest that the effectiveness of each XRL method will be quantitatively assessed based on the performance improvements of the repaired RL agents.
Implications
This research has the potential to significantly enhance the evaluation of XRL methods, providing a more relevant and practical framework for developers. By focusing on the utility of explanations in debugging, it could lead to the development of more effective XRL methods that better meet the needs of various stakeholders in the RL community.
Position: Fairness Failure in Generative Models is an Evaluation Problem
Generative Models
- Fairness failures in generative models stem from evaluation problems rather than solely from model design.
- Current evaluation practices are inconsistent and hinder comparability across studies.
- The proposed 'Fairness Cards' aim to standardize reporting and improve accountability in fairness evaluations.
- Fairness should be integrated as a core design goal alongside other performance metrics in generative systems.
Read more
Position: Fairness Failure in Generative Models is an Evaluation Problem
Summary
This position paper addresses the persistent issue of fairness failures in generative models, emphasizing that these failures primarily arise from inadequate evaluation practices rather than a lack of mitigation strategies. The authors argue that current fairness evaluations are often non-comparable and lack actionable insights, which hampers progress in the field. They diagnose several empirical and conceptual failure modes in existing practices, highlighting how variations in evaluation protocols, prompt families, and scoring methods can lead to inconsistent fairness assessments. To address these challenges, the paper proposes 'Fairness Cards' as a standardized reporting tool that clarifies evaluation choices and enhances reproducibility and accountability. The authors advocate for a paradigm shift in evaluation standards to ensure fairness is treated as a critical performance dimension throughout the model lifecycle, enabling more reliable comparisons and informed deployment decisions.
Methodology
The authors conducted a comprehensive review of existing literature on fairness in generative models, identifying recurring empirical and structural obstacles in evaluation practices. They analyzed how different evaluation protocols and settings affect fairness outcomes and proposed a new reporting standard, 'Fairness Cards', to enhance transparency and comparability.
Results
The paper illustrates that the same generative model can yield vastly different fairness assessments depending on the evaluation protocol used. It highlights the instability of fairness results due to variations in prompt families, scoring methods, and system-layer behaviors. The introduction of 'Fairness Cards' is presented as a solution to improve the reliability of fairness evaluations.
Implications
The findings suggest that establishing standardized evaluation practices could lead to more reliable assessments of fairness in generative models, ultimately guiding better deployment decisions and fostering accountability in AI systems. This could also influence regulatory frameworks and encourage researchers to prioritize fairness as a fundamental aspect of model development.
Certified but Private: Scalable Zero-Knowledge Proofs for Neural Network Guarantees
Theory
Efficient ML
- PANDA combines zero-knowledge proofs with the CROWN verification algorithm to certify neural network robustness while maintaining parameter privacy.
- The system can handle neural networks with over 2.9 million parameters, scaling polynomially with the number of neurons.
- PANDA introduces a novel method for verifying linear relaxations of activation functions, making it the first privacy-preserving local robustness certification system for networks with transcendental activation functions.
- The proving time is significantly reduced to 5 minutes, with verification taking only 10 seconds, enabling practical certification for larger models.
Read more
Certified but Private: Scalable Zero-Knowledge Proofs for Neural Network Guarantees
Summary
The paper presents PANDA, a scalable system that utilizes zero-knowledge proofs (ZKPs) to certify the robustness and fairness of neural network models without disclosing their private parameters. As machine learning models are increasingly deployed in safety-critical applications, formal guarantees of their robustness are essential. However, revealing model parameters poses risks of intellectual property theft and privacy violations. PANDA addresses this challenge by integrating the CROWN verification framework with ZKPs, allowing for the certification of local robustness properties while keeping model parameters confidential. The core innovation of PANDA is a novel algorithm for proving linear relaxation bounds for non-linear activation layers, which results in lightweight proofs. PANDA can generate proofs for neural networks with over 2.9 million parameters in just 5 minutes and verify them in 10 seconds, significantly outperforming previous ZKP-based systems that relied on exponential-time algorithms. This advancement enables the certification of much larger neural networks, enhancing the practical applicability of formal verification methods in real-world scenarios.
Methodology
PANDA integrates zero-knowledge proofs with the CROWN robustness certification framework. It employs a novel algorithm for proving linear relaxation bounds for non-linear activation functions, allowing for efficient verification without exposing model parameters. The system uses a customized backend to enhance prover efficiency and scalability.
Results
PANDA successfully generates proofs for neural networks with over 2.9 million parameters in 5 minutes and verifies them in 10 seconds. This represents a significant improvement over previous ZKP-based systems, which could only handle much smaller networks and required exponential time for proving.
Implications
PANDA enables the deployment of machine learning models in safety-critical applications by providing formal robustness guarantees while preserving model confidentiality. This has potential applications in healthcare, autonomous driving, and other sectors where model reliability is crucial.
Beyond $L_2$: Generalizing Abductive Latent Explanations to Diverse Prototype-Based Architectures
Interpretability
- Generalization of Abductive Latent Explanations (ALE) to non-Euclidean prototype-based architectures.
- Development of novel bounding algorithms for diverse geometric representations.
- Validation through empirical experiments on image classifiers demonstrating the framework's effectiveness.
- Unification of various prototype-based models under a single formal framework for interpretability comparison.
Read more
Beyond $L_2$: Generalizing Abductive Latent Explanations to Diverse Prototype-Based Architectures
Summary
This paper addresses the limitations of existing Abductive Latent Explanations (ALE) that are confined to Euclidean latent spaces, which restricts their applicability to modern prototype-based neural networks that utilize non-Euclidean representations. The authors propose a generalized ALE framework that accommodates various geometric structures, including spherical metrics and Gaussian densities. By systematically deriving methods to either adapt existing bounds or create new architecture-specific bounding algorithms, the paper enables the computation of subset-minimal formal explanations across diverse prototype-based architectures. The authors validate their theoretical contributions through extensive experimentation on fully trained image classifiers, showcasing the flexibility and effectiveness of the generalized ALE framework. This work represents a significant step towards unifying different prototype-based models under a single formal framework, facilitating rigorous comparisons of interpretability across architectures.
Methodology
The authors extend the ALE framework by deriving methods to adapt existing distance bounds to non-Euclidean spaces and creating new bounding algorithms specific to different architectures. They conduct experiments to compute formal explanations on trained image classifiers, comparing the interpretability of various models.
Results
The proposed generalized ALE framework successfully computes subset-minimal formal explanations for diverse prototype-based architectures, demonstrating its flexibility and effectiveness. The empirical results indicate that the framework allows for a rigorous comparison of interpretability across different models.
Implications
This work enhances the interpretability of machine learning models, particularly in critical applications where understanding model decisions is essential, such as healthcare and autonomous systems. It paves the way for more transparent AI systems that can be trusted in high-stakes environments.
SchurQuant: Groupwise Discrete Optimization for Layer-Wise LLM Quantization
NLP
Large Language Models
Optimization
- Introduces SCHURQUANT, a new method for layer-wise quantization of LLMs.
- Utilizes SCHUROPT to optimize groupwise discrete decisions while considering the continuous suffix response.
- Achieves significant improvements in zero-shot accuracy, particularly at low bit quantization levels.
- Combines multiple techniques including teacher reconstruction and reference-weight regularization for enhanced performance.
Read more
SchurQuant: Groupwise Discrete Optimization for Layer-Wise LLM Quantization
Summary
The paper introduces SCHURQUANT, a novel approach for weight-only post-training quantization (PTQ) of large language models (LLMs) that addresses the accuracy degradation typically observed at low bit quantization levels (2-3 bits). Existing PTQ methods face limitations due to their inability to effectively utilize the continuous suffix of weights during group decisions and their reliance on fixed affine quantization grids. SCHURQUANT employs a new optimizer, SCHUROPT, which analytically determines the optimal continuous response of the suffix, allowing for a more accurate groupwise quadratic optimization. The method alternates between closed-form scale/zero-point refitting and coordinate descent over integer codes. The integration of quantized-prefix teacher reconstruction, reference-weight regularization, and other enhancements leads to significant improvements in zero-shot accuracy. The results demonstrate that SCHURQUANT outperforms existing backpropagation-free PTQ methods, achieving the highest mean zero-shot accuracy across various models, particularly excelling at 2-bit quantization.
Methodology
The methodology involves the development of SCHUROPT, which optimizes the quantization process by analytically determining the optimal continuous response of the suffix weights. This is achieved through a combination of closed-form refitting of scale and zero-point parameters and coordinate descent over quantized integer codes. The final configuration, SCHURQUANT, integrates teacher reconstruction and regularization techniques to improve the quantization objective.
Results
SCHURQUANT improves mean zero-shot accuracy on the 2-bit Qwen3-4B model by 11.88 percentage points compared to existing methods. It also outperforms the strongest baseline by 9.65 percentage points at 2 bits across eight Llama and Qwen models, although performance at higher bit levels does not consistently improve.
Implications
The findings suggest that SCHURQUANT can enable more efficient deployment of large language models on resource-constrained devices by maintaining higher accuracy at lower bit quantization levels. This has potential applications in edge computing and mobile devices where memory and bandwidth are limited.
An Omitted Mode Is a Rare Rule: The Sampling-Verification Danger Law in Continuous Code World Models
Reinforcement Learning
Theory
Robotics
- Introduces the concept of 'danger' as a measure of risk in model-based RL.
- Establishes a localization budget that quantifies model disagreement in continuous settings.
- Demonstrates that omitted modes can lead to catastrophic exploitation of planners.
- Shows that acceptance tests may not adequately capture critical mode boundaries.
Read more
An Omitted Mode Is a Rare Rule: The Sampling-Verification Danger Law in Continuous Code World Models
Summary
This paper investigates the risks associated with the acceptance of world models in continuous control settings, particularly in the context of model-based reinforcement learning (RL). The author introduces the concept of 'danger' as an expected risk, defined as the play cost multiplied by the probability of missing critical events in the model. The paper establishes a mathematical framework for understanding how errors in world models can lead to catastrophic failures in planning, particularly when the model is verified but still incorrect. A key contribution is the definition of a localization budget, which quantifies how much disagreement can exist in models that differ slightly at specific points. The author demonstrates this through three hybrid instruments that measure the consequences of omitted modes, showing that planners can be severely misled when critical events are not captured in the model. The findings reveal that even when models pass acceptance tests, they can still exploit planners, leading to significant regret and poor performance. The paper also discusses the limitations of current sampling acceptance tests, highlighting that they do not adequately control for coverage of mode boundaries, which can result in unrecognized errors.
Methodology
The author employs a theoretical framework to analyze the risks associated with world models in continuous control. This includes defining the danger law mathematically, establishing a localization budget, and conducting empirical evaluations using hybrid instruments to measure the consequences of omitted modes. The study also involves rigorous statistical analysis to assess the performance of models under various conditions.
Results
The results indicate that models can pass acceptance tests while still failing to capture critical events, leading to significant exploitation by planners. The empirical evaluations showed that planners faced substantial regret when critical modes were omitted, with a clear performance gap compared to a uniform-random baseline. The paper also found that targeted interventions aimed at recovering omitted modes were largely unsuccessful, underscoring the challenges in model recovery.
Implications
The findings suggest that reliance on sampling acceptance tests in continuous world models can be misleading, potentially leading to severe performance issues in real-world applications. This has implications for the design of more robust verification methods in model-based RL and highlights the need for improved strategies to ensure model reliability and safety.
Hybrid ML for Lightweight Pre-Route Delay Estimation in Open-Source IC Design
Optimization
Efficient ML
Interpretability
- The hybrid model combines decision trees and linear regression for improved delay estimation.
- Achieves an 80% reduction in error compared to OpenLane's estimates.
- The model is over 300 times smaller and twice as fast as traditional methods.
- Provides higher explainability and meaningful insights into timing factors.
Read more
Hybrid ML for Lightweight Pre-Route Delay Estimation in Open-Source IC Design
Summary
This paper addresses the challenges of accurate delay estimation in the design flow of digital integrated circuits, particularly during the pre-routing stage where limited physical design information is available. The authors propose a hybrid machine learning approach that combines decision trees and linear regression to enhance delay predictions generated by the open-source RTL-to-GDSII tool, OpenLane. The hybrid model significantly reduces estimation errors by 80% compared to OpenLane's predictions and achieves a 71% improvement even without specific parameters from OpenLane. The model is lightweight, being over 300 times smaller and twice as fast as traditional methods, while also providing greater explainability. By leveraging the strengths of both decision trees and linear regression, the proposed framework effectively captures non-linear relationships and enhances interpolation capabilities, making it a viable alternative to conventional delay propagation techniques.
Methodology
The authors developed a hybrid model that integrates decision trees and linear regression. The decision tree partitions the data into homogeneous regions, allowing the linear regression model to operate effectively within those regions, thereby enhancing both accuracy and interpretability.
Results
The hybrid model demonstrated an 80% reduction in delay estimation errors compared to OpenLane, and a 71% improvement was noted even without OpenLane-specific parameters. The model's lightweight nature and computational efficiency were highlighted, being over 300 times smaller and twice as fast as existing methods.
Implications
This approach can streamline the design process for digital integrated circuits by providing more accurate and efficient delay estimations, potentially reducing the number of design iterations and improving overall productivity in VLSI design workflows.
Pessimistic Meta-Induction and Its Limits: Lessons from Frequentist Statistics and Machine Learning Theory
Theory
- Challenges the pessimistic meta-inductive argument by undermining its inductive step.
- Establishes that ordinary induction can achieve everywhere convergence, while meta-induction cannot.
- Introduces a mathematical model to illustrate the differences between ordinary induction and meta-induction.
- Argues that no inference method achieves almost everywhere convergence in the context of meta-induction.
Read more
Pessimistic Meta-Induction and Its Limits: Lessons from Frequentist Statistics and Machine Learning Theory
Summary
This paper critiques the pessimistic meta-inductive argument against scientific realism by focusing on the inductive step rather than the historical premise. The author proposes a novel challenge to the meta-inductive argument, drawing from frequentist statistics and machine learning to evaluate induction based on its convergence to the truth. The paper distinguishes between ordinary enumerative induction, which can achieve everywhere convergence, and meta-induction, which fails to achieve even almost everywhere convergence. The author presents a mathematical model to illustrate these concepts, demonstrating that in contexts where meta-induction is applied, no inference method can achieve almost everywhere convergence. This work emphasizes the importance of evaluating inference methods through their convergence properties and suggests that the epistemology developed in frequentist statistics and machine learning provides a robust framework for scientific inference.
Methodology
The paper employs a mathematical modeling approach to compare ordinary induction and meta-induction, focusing on their convergence properties. It draws on concepts from frequentist statistics and machine learning to evaluate the effectiveness of different inductive methods.
Results
The main result of the paper is that while ordinary induction can achieve everywhere convergence, meta-induction fails to achieve even almost everywhere convergence. Furthermore, in contexts defined by competing hypotheses, no inference method can achieve almost everywhere convergence.
Implications
The findings suggest that the pessimistic meta-inductive argument may be less robust than previously thought, and they encourage a reevaluation of how scientific theories are inferred and validated. The insights could influence future research in scientific methodology, epistemology, and the foundations of machine learning.
Causal Discovery in Equal Variance Linear Gaussian DAGs via SURE-Tuned Ridge Regression
Graph Learning
Optimization
Theory
- Introduction of SURE-Ridge, a closed-form estimator for causal discovery in equal variance linear Gaussian DAGs.
- Utilization of Stein’s unbiased risk estimate (SURE) for adaptive selection of regularization parameters in ridge regression.
- Demonstration of SURE-Ridge's effectiveness in small-sample regimes and compute-limited scenarios.
- Adaptive thresholding procedure for extracting valid DAGs from soft adjacency matrices.
Read more
Causal Discovery in Equal Variance Linear Gaussian DAGs via SURE-Tuned Ridge Regression
Summary
This paper addresses the challenge of recovering the directed acyclic graph (DAG) of a structural equation model (SEM) from observational data, particularly in sample-limited and compute-limited regimes. The authors propose a novel method called SURE-Ridge, which is a non-iterative, closed-form estimator designed for equal variance linear Gaussian SEMs. SURE-Ridge employs parallel node-wise ridge regressions, with regularization parameters adaptively selected using Stein’s unbiased risk estimate (SURE). This approach allows for efficient DAG recovery without the need for extensive hyperparameter tuning or iterative optimization, making it particularly suitable for scenarios where the number of samples is small relative to the number of nodes. The method also includes an adaptive thresholding procedure to derive a valid DAG from the resulting soft adjacency matrix. Through numerical experiments, SURE-Ridge demonstrates superior performance, achieving the lowest structural Hamming distance (SHD) in small-sample scenarios and the fastest runtime across various sample sizes when compared to existing methods such as NOTEARS, DAGMA, and GBNSL.
Methodology
The methodology involves a closed-form estimation approach using SURE-tuned node-wise ridge regression. The regularization parameters are selected adaptively using SURE, which provides a principled risk metric for tuning without requiring held-out data. The method performs parallel regressions for each node, allowing for efficient computation and the extraction of a DAG through an adaptive thresholding procedure.
Results
The results indicate that SURE-Ridge achieves the lowest structural Hamming distance in small-sample settings and maintains the lowest runtime across all tested sample sizes when compared to baseline methods such as NOTEARS, DAGMA, and GBNSL.
Implications
The proposed SURE-Ridge method has significant implications for causal discovery in various fields, including epidemiology, wireless networks, and financial fraud detection, particularly in scenarios where data is limited or computational resources are constrained.
p-Spin Glass Network Efficient Single-Batch Continual Learning
Efficient ML
Theory
Optimization
- Introduces the p-Spin Glass Network, enhancing memory efficiency and sample efficiency.
- Achieves 8× parameter compression through ternary quantization of internal matrices.
- Demonstrates stable convergence with a micro-batch size of 1, facilitating continual learning.
- Maintains robust performance across different input modalities.
Read more
p-Spin Glass Network Efficient Single-Batch Continual Learning
Summary
The paper introduces the p-Spin Glass Network, a novel architecture designed to address the limitations of modern sequence models that rely on large memory footprints and batch sizes. The proposed architecture enhances memory efficiency through ternary quantization, achieving an 8× reduction in parameter size while maintaining activation memory at O(B · T · D). It demonstrates sample efficiency by matching the performance of a Transformer baseline with only 12.5% of the training sequences. The network also ensures stability during single-batch training, allowing for smooth convergence even with a micro-batch size of 1. This stability is modality-agnostic, effectively managing temporal credit assignment across various input types, including discrete subword tokens and long horizon byte streams. The work establishes a foundation for continual learning and edge AI by removing the dependency on large batch sizes for stable deep learning.
Methodology
The methodology involves reframing sequence representation using implicit deep learning and thermodynamic attractors. The architecture employs ternary quantization for internal matrices, coupled with a bounded fixed point solver to dampen gradient variance. This allows for efficient state-space integration and robust temporal credit assignment across various input modalities.
Results
The p-Spin Glass Network achieves asymptotic performance parity with standard Transformers while using significantly fewer training sequences. It also demonstrates stable training dynamics at a micro-batch size of 1, effectively addressing the challenges of gradient variance and optimization instability.
Implications
The findings suggest that the p-Spin Glass Network can facilitate continual learning in resource-constrained environments, making it suitable for applications in edge AI and scenarios where memory and computational efficiency are critical.
Recirculation
NLP
Large Language Models
Theory
- Recirculation enhances state tracking in foundation models, addressing limitations of feedforward transformers.
- The method incurs minimal latency during generation and requires serial processing only in the prefill phase.
- An adaptive variant of recirculation achieves significant performance gains with light hyperparameter tuning.
- The approach leads to a 23% reduction in perplexity and a 21% increase in accuracy on specific datasets.
Read more
Recirculation
Summary
The paper introduces 'recirculation', an inference-time architectural enhancement for foundation models that significantly reduces perplexity and enhances accuracy in generation and reasoning tasks. This technique addresses the limitations of state updates in feedforward transformers by incorporating a specific form of recurrence, allowing models to function as dynamical systems that can track belief states. Unlike traditional recurrent neural networks and depth-recurrence techniques, recirculation requires minimal additional latency during generation and only necessitates serial processing during the prefill phase. The authors also propose an adaptive variant of recirculation that requires light hyperparameter tuning while keeping the original model weights frozen. The results demonstrate substantial improvements over baseline models, including a 23% reduction in perplexity and a 21% increase in accuracy on the GSM8k dataset, showcasing the potential for architectural evolution informed by the properties of trained networks.
Methodology
The authors developed the recirculation technique, which introduces recurrence into the architecture of foundation models. This method allows for better state tracking by enabling the model to act as a dynamical system. They also created an adaptive version that requires minimal tuning of hyperparameters while keeping the original model weights unchanged. The performance of these methods was evaluated on various datasets, including the Gemma3 family and GSM8k.
Results
The implementation of adaptive recirculation led to a 23% reduction in perplexity across a suite of datasets and a 21% increase in accuracy on the GSM8k dataset. The approach also yielded consistent improvements in accuracy on other downstream tasks, demonstrating its effectiveness in enhancing model performance without extensive retraining.
Implications
The findings suggest that architectural modifications can be guided by the properties of existing trained networks, potentially leading to more efficient and effective model designs. This could influence future research in model architecture and state tracking, particularly in natural language processing and reasoning tasks.