AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
55
Papers today
8h
Update frequency
7
Days of history
Minimally Invasive Steering of Language Models
NLP
Large Language Models
Optimization
- Introduction of MISVO for minimally invasive steering of language models.
- Utilization of local KL geometry to penalize deviations from reference policies.
- Analytic gradient computation through matrix-vector products with frozen model heads.
- Demonstrated superior performance in reward maximization across multiple tasks.
Read more
Minimally Invasive Steering of Language Models
Summary
This paper introduces Minimally Invasive Steering Vector Optimization (MISVO), a novel approach to adapt frozen language models at test time by optimizing steering vectors added to the model's hidden states. The authors highlight the challenges of unregularized reward optimization, which can lead to significant alterations in output distribution and degrade generation quality. To address this, MISVO incorporates a penalty based on the local Kullback-Leibler (KL) geometry of the token distribution, allowing for a more controlled adjustment of the model's outputs. The method employs a Fisher quadratic measure to assess distributional sensitivity and derives an analytic gradient for optimization, which avoids the need for backpropagation through the model. The paper provides a detailed analysis of the sequence-level KL gradient, breaking it down into components that facilitate effective steering. The experimental results demonstrate that MISVO outperforms existing methods in terms of mean reward across various tasks while maintaining high coherence and diversity in generated outputs.
Methodology
The methodology involves optimizing position-specific steering vectors added to the hidden states of a frozen language model. The optimization process uses a Fisher quadratic penalty derived from the local KL divergence to balance reward maximization against deviations from a reference policy. The authors provide an analytic gradient for the steering vectors, allowing for efficient computation without backpropagation through the model.
Results
MISVO achieved the highest mean reward in six out of seven model-task settings tested, across preference and code-generation tasks using models with parameter sizes ranging from 1B to 14B. The method also maintained diversity and coherence scores comparable to those of the Best-of-N approach, indicating its effectiveness in generating high-quality outputs.
Implications
The findings suggest that MISVO can be effectively used for real-time adaptation of language models to varying user contexts and objectives without the need for retraining. This has potential applications in personalized content generation, interactive AI systems, and scenarios where model parameters cannot be updated post-deployment.
Graph-Based Inference and Topology-Aware Multi-Agent Reinforcement Learning for Large-Scale Railway Network Management
Reinforcement Learning
Graph Learning
Optimization
- Proposes a graph-based framework for railway maintenance management.
- Utilizes a hierarchical Bayesian model for environment inference.
- Introduces a topology-aware MARL framework integrating graph neural networks.
- Demonstrates zero-shot transfer learning capabilities for scalability.
Read more
Graph-Based Inference and Topology-Aware Multi-Agent Reinforcement Learning for Large-Scale Railway Network Management
Summary
This paper addresses the complex problem of infrastructure asset management, particularly in large-scale railway networks, by proposing a novel graph-based framework that combines accurate environment modeling with scalable decision support. The authors highlight the limitations of traditional maintenance strategies that often neglect system-level interactions and propose a two-step approach: first, they utilize a hierarchical Bayesian model with a Gaussian Process on Graph kernel to infer a realistic, spatially correlated environment for railway maintenance planning using real-world data from the Swiss Federal Railways. Second, they introduce a topology-aware Multi-Agent Reinforcement Learning (MARL) framework that integrates graph neural networks and graph Transformers to optimize network-level maintenance policies. A significant contribution of this work is the demonstration of zero-shot transfer learning, where agents trained on small network segments can be effectively deployed on larger, unseen networks without retraining. The results show that this method outperforms traditional heuristics and standard MARL baselines, achieving superior performance while reducing computational training time, thus showcasing its scalability and effectiveness in managing large-scale railway networks.
Methodology
The methodology involves a two-step process: first, employing a hierarchical Bayesian model with a Gaussian Process on Graph kernel to infer a spatially correlated environment from real-world railway data. Second, integrating graph neural networks and graph Transformers within a topology-aware MARL framework to optimize maintenance policies across the network.
Results
The proposed framework significantly outperformed optimized heuristics and standard MARL baselines, demonstrating effective scalability through zero-shot transfer learning. The numerical results indicated reduced computational training time while maintaining superior performance on large-scale railway networks.
Implications
The findings suggest that the proposed graph-based and topology-aware approach can enhance decision-making processes in large-scale infrastructure management, potentially leading to more efficient maintenance strategies and better resource allocation in railway networks and similar complex systems.
Not Every Token Is Worth Distilling: Selective Supervision for Direct-OPD
Reinforcement Learning
Large Language Models
Efficient ML
- Introduces S2D-OPD to improve Direct-OPD by selectively supervising states based on divergence.
- Identifies a probability-mass mismatch issue in Direct-OPD that can lead to ineffective supervision.
- Empirical results show S2D-OPD improves accuracy in most tested scenarios without extra forward passes.
- Demonstrates that performance correlates positively with the JSD percentile of selected states.
Read more
Not Every Token Is Worth Distilling: Selective Supervision for Direct-OPD
Summary
This paper introduces Selective Supervision for Direct-OPD (S2D-OPD), a novel approach to improve the efficiency of Direct On-Policy Distillation (Direct-OPD) in reinforcement learning contexts. Direct-OPD transfers policy improvements from a smaller model to a larger one using token-level log-ratios as supervision. However, the authors identify a critical issue: the log-ratio may not reflect meaningful changes in the teacher's behavior when the probability mass assigned to candidate tokens diminishes. To address this, the paper proposes S2D-OPD, which ranks states based on their Jensen-Shannon divergence (JSD) and selectively masks supervision for states with low divergence, retaining only the top 10% of states. The empirical results demonstrate that S2D-OPD outperforms dense Direct-OPD in seven out of eight tested configurations across various model sizes, achieving improved held-out accuracy without requiring additional computational resources. This work highlights the importance of selective supervision in policy transfer and provides insights into the relationship between divergence and performance in reinforcement learning.
Methodology
The authors analyze the limitations of Direct-OPD by establishing a connection between the log-ratio reward and the probability mass assigned to candidate tokens. They propose S2D-OPD, which ranks student-sampled states by their teacher-reference JSD and masks supervision for low-divergence states. This selective approach retains only the top 10% of states per response, allowing for more effective policy transfer.
Results
S2D-OPD improves held-out accuracy over dense Direct-OPD in seven out of eight settings tested, with a mean gain of 0.95 points (95% CI 0.40–1.54). The method also yields smoother validation curves during late-stage training, indicating enhanced stability and performance.
Implications
The findings suggest that selective supervision can significantly enhance the efficiency of policy transfer in reinforcement learning, potentially reducing the computational costs associated with training large models. This approach could be applied to various reinforcement learning tasks, improving the scalability and effectiveness of model training.
Direct Message Approximation (DMA): A Consistency-Based Framework for Tractable Approximate Inference on Factor Graphs
Theory
Graph Learning
Efficient ML
- Introduction of Direct Message Approximation (DMA) framework for factor graphs.
- Establishment of a consistency condition for message construction.
- Development of a master theorem bounding marginal KL divergence from message KL divergence.
- Explicit DMA message derivation for product factors, crucial for Bayesian neural networks.
Read more
Direct Message Approximation (DMA): A Consistency-Based Framework for Tractable Approximate Inference on Factor Graphs
Summary
This paper introduces Direct Message Approximation (DMA), a novel framework for approximate inference on factor graphs that addresses limitations in existing methods like Expectation Propagation (EP) and Variational Message Passing (VMP). Traditional approaches rely on marginal approximations, which can lead to iterative schedules, invalid messages, and loss of variance in Bayesian neural networks. DMA circumvents these issues by directly approximating factor-to-variable messages without relying on marginal computations. The authors establish a consistency condition that ensures the accuracy of message construction, particularly when other incoming messages are Dirac deltas. A master theorem is presented, bounding the Kullback-Leibler divergence of marginals from message KL divergence, applicable to any graph structure. The paper also details the construction of DMA messages for product factors, which are crucial for Bayesian neural network updates. The proposed method demonstrates improved predictive uncertainty handling, particularly in data-sparse regions, and eliminates the need for iterative message updates, thus enhancing computational efficiency.
Methodology
The DMA framework is built on a consistency condition that guides the construction of factor-to-variable messages directly, avoiding the iterative and potentially invalid message updates seen in traditional methods. The authors prove a master theorem that provides bounds on KL divergence, ensuring that the constructed messages maintain accuracy across various graph structures. They also derive specific message forms for product factors, leveraging moment-matching techniques.
Results
The DMA framework successfully eliminates the need for iterative message passing, reduces the occurrence of negative-precision messages, and maintains variance in Bayesian neural networks. The derived messages for product factors exhibit O(1/r²) accuracy in concentrated-input scenarios, showcasing the method's robustness and efficiency.
Implications
The DMA framework has significant implications for probabilistic inference in complex models, particularly in Bayesian neural networks where maintaining uncertainty is crucial. It offers a more efficient alternative to existing methods, potentially leading to faster convergence and improved performance in applications involving sparse data or model mismatch.
Leakage-Safe Machine Learning for Hydrogen Embrittlement Detection in 316L Stainless Steel: A Region-Held-Out Evaluation of Texture and Deep Features in SEM Micrographs
Computer Vision
- Introduces a region-held-out evaluation protocol to prevent data leakage in machine learning models for HE detection.
- Demonstrates that simple texture-based methods can outperform complex deep learning models in specific applications.
- Achieves statistically significant results validating the effectiveness of the proposed methodology.
- Highlights the importance of using robust evaluation techniques in materials science to ensure reliable model performance.
Read more
Leakage-Safe Machine Learning for Hydrogen Embrittlement Detection in 316L Stainless Steel: A Region-Held-Out Evaluation of Texture and Deep Features in SEM Micrographs
Summary
This paper addresses the challenge of detecting hydrogen embrittlement (HE) in 316L stainless steel using machine learning techniques applied to scanning electron microscopy (SEM) micrographs. Traditional machine learning evaluations often suffer from data leakage when multiple images originate from the same specimen region, leading to biased performance metrics. To mitigate this, the authors propose a region-held-out evaluation protocol, employing Leave-One-Region-Out (LORO) cross-validation across 14 spatial regions (8 as-received and 6 hydrogen-charged) with a total of 31 images. The study compares various feature-classifier combinations, including local binary patterns (LBP), grey-level co-occurrence matrices (GLCM), self-supervised convolutional embeddings, and convolutional neural networks (CNNs). The results indicate that the simplest approach, LBP combined with a support vector machine (SVM), achieved the highest balanced accuracy of 0.79, with a hydrogen-charged recall of 0.69 and precision of 0.82, outperforming more complex deep learning models. A permutation test confirmed the statistical significance of these results (p = 0.008), indicating that the findings are not due to chance. The study concludes that texture descriptors can effectively identify hydrogen-charging signatures in SEM micrographs, even with limited samples, and suggests that this methodology can be applied to larger studies in other alloy systems.
Methodology
The study employs a region-held-out evaluation protocol using Leave-One-Region-Out (LORO) cross-validation. It compares multiple feature-classifier combinations, including local binary patterns (LBP), grey-level co-occurrence matrices (GLCM), self-supervised embeddings, and convolutional neural networks (CNNs), to classify SEM micrographs of 316L stainless steel.
Results
The LBP combined with SVM approach achieved the best performance with a balanced accuracy of 0.79, hydrogen-charged recall of 0.69, and precision of 0.82. Statistical validation through a permutation test yielded a p-value of 0.008, confirming the robustness of the results.
Implications
The findings suggest that texture descriptors can effectively capture critical microstructural changes due to hydrogen embrittlement, enabling automated diagnostics. This methodology can be extended to other materials and degradation studies, enhancing the reliability of assessments in safety-critical applications.
Residual Correlation as a Diagnostic for Joint-Uncertainty Gains from GP Coregionalisation
Theory
- Residual correlation is a better predictor of joint-uncertainty gains from coregionalisation than raw target correlation.
- The proposed diagnostic Dlogdet can be computed from independent GPs and predicts NLL improvements effectively.
- Coregionalisation primarily benefits joint uncertainty estimation rather than point prediction accuracy.
- Residual-ICM enhances independent GP predictions by incorporating residual correlation, achieving the best average joint NLL.
Read more
Residual Correlation as a Diagnostic for Joint-Uncertainty Gains from GP Coregionalisation
Summary
This paper investigates the effectiveness of multi-output Gaussian processes (GPs) with an intrinsic model of coregionalisation (GP-ICM) for multi-target regression tasks. The authors find that while correlated targets are often modeled jointly to improve prediction and uncertainty quantification, the actual benefits are inconsistent across different datasets. The study reveals that the primary advantage of coregionalisation lies in joint uncertainty quantification rather than point prediction. The authors introduce a novel diagnostic tool, Dlogdet, which quantifies the residual correlation left unexplained by independent predictors and serves as a better predictor of joint-uncertainty gains than raw target correlation. Through extensive empirical validation, including synthetic experiments and real-world datasets, the paper demonstrates that the proposed diagnostic significantly outperforms traditional heuristics. Additionally, the authors present Residual-ICM, a new readout method that incorporates residual correlation while maintaining independent marginal variances, achieving superior joint negative log-likelihood (NLL) performance. The findings suggest that the residual-correlation mechanism is relevant across various representation types, including frozen transformer and CNN models for keypoint regression.
Methodology
The authors employ a unified GP readout framework to evaluate the effects of target coupling on representation selection, point prediction, and uncertainty quantification. They introduce the diagnostic Dlogdet to quantify residual correlation and validate their findings through synthetic studies and real-world datasets, comparing the performance of various methods including Residual-ICM.
Results
The results indicate that point prediction accuracy remains largely unchanged across methods, while Dlogdet strongly predicts observed NLL improvements (correlation of -0.83, p < 0.001). Residual-ICM achieves the best average joint NLL across 16 datasets, demonstrating the effectiveness of incorporating residual correlation into multi-output models.
Implications
The findings have significant implications for multi-target regression tasks in various scientific fields, suggesting that practitioners should focus on residual correlation to assess the potential benefits of coregionalisation. The proposed diagnostic can guide model selection and improve uncertainty quantification in applications such as environmental monitoring and plant phenotyping.
Active Client Selection in Federated Trajectory Prediction with Uncertainty-Awareness and Heterogeneous Complexity
Federated Learning
Robotics
Time Series
- Introduces a federated learning framework for trajectory prediction that addresses scene uncertainty and complexity heterogeneity.
- Develops active client selection methods that prioritize clients based on uncertainty and complexity metrics.
- Demonstrates improved convergence and predictive performance on the Argoverse dataset compared to traditional methods.
- Provides a reproducible setup for heterogeneous federated partitions of trajectory datasets.
Read more
Active Client Selection in Federated Trajectory Prediction with Uncertainty-Awareness and Heterogeneous Complexity
Summary
This paper addresses the challenges of trajectory prediction for autonomous vehicles using Federated Learning (FL), particularly focusing on scene uncertainty and complexity heterogeneity. The authors propose a framework called Federated Trajectory Prediction (FLTP), which incorporates uncertainty-aware objectives and active client selection methods to enhance model training without the need for centralized data. The framework includes various selectors that prioritize clients based on their informativeness, specifically targeting those with high uncertainty and complex scene characteristics. The study demonstrates that FLTP and its variants significantly improve convergence rates and predictive performance metrics compared to traditional methods. The experiments conducted on the Argoverse dataset show that even basic federated models outperform locally trained counterparts, and the proposed active selection methods lead to better generalization, particularly in heterogeneous datasets. Additionally, the authors provide a reproducible setup for creating diverse federated partitions of trajectory datasets, which could be beneficial for future research.
Methodology
The authors propose a family of active client selection methods within the FL framework, including FLTP, FLTP-AU, FLTP-NLL, and FLTP-DR. These methods utilize negative log-likelihood objectives and aleatoric uncertainty to guide client selection. The complexity-aware scoring system ranks clients based on interpretable features related to scene complexity, enhancing the selection process.
Results
Experiments reveal that the proposed FLTP framework and its variants outperform locally trained models, achieving better core metrics such as minimum Average Displacement Error (minADE), minimum Final Displacement Error (minFDE), and Miss Rate (MR). The active client selection methods accelerate convergence and improve generalization, particularly in datasets with significant scene complexity heterogeneity.
Implications
The findings suggest that federated learning can effectively leverage decentralized data for trajectory prediction in autonomous vehicles, potentially leading to safer and more efficient driving systems. The active client selection methods can be applied to other domains facing similar challenges of data fragmentation and heterogeneity.
Decoupling Knowledge and Privacy: Post-Task Self-Distillation Replay for LLM Continual Learning
NLP
Large Language Models
Theory
- Introduces SPARK, a framework for privacy-preserving continual learning in LLMs.
- Decouples knowledge retention and privacy correction to address sensitive content reproduction.
- Utilizes Self-Distillation Replay to create a stable checkpoint for effective learning.
- Demonstrates effective selective suppression of PII while maintaining model utility.
Read more
Decoupling Knowledge and Privacy: Post-Task Self-Distillation Replay for LLM Continual Learning
Summary
This paper addresses the challenge of privacy-preserving continual learning (PPCL) in large language models (LLMs), focusing on the dual objectives of retaining useful knowledge from previous tasks while minimizing the reproduction of sensitive content. The authors introduce a novel framework called SPARK, which decouples knowledge retention from privacy correction. SPARK operates in two phases: first, it employs Self-Distillation Replay to learn the current task while preserving behaviors from previous tasks, creating a stable post-task checkpoint. Second, it applies Post-Task Privacy Correction to selectively reduce the likelihood of personally identifiable information (PII) outputs while anchoring non-PII behaviors. The framework is designed to maintain continual learning utility and knowledge retention across diverse settings. The authors demonstrate through extensive evaluations that SPARK effectively suppresses sensitive outputs while preserving model performance, thus addressing the critical balance between knowledge retention and privacy in continual learning scenarios.
Methodology
The methodology involves a two-phase approach: (1) Self-Distillation Replay, which learns the current task while distilling behaviors from previous tasks to create a stable post-task distribution, and (2) Post-Task Privacy Correction, which selectively reduces the likelihood of annotated PII outputs while anchoring non-PII behaviors to the checkpoint established in the first phase. This approach allows for the preservation of useful knowledge while controlling sensitive outputs.
Results
The evaluations show that SPARK achieves significant reductions in the likelihood of sensitive content being reproduced while maintaining strong continual learning performance. The framework effectively balances the need for privacy with the retention of knowledge across tasks, outperforming existing methods in terms of both privacy control and task acquisition stability.
Implications
The findings have important implications for the development of privacy-preserving machine learning systems, particularly in applications involving sensitive data such as healthcare, finance, and personal communications. The SPARK framework can be adapted for various continual learning scenarios where privacy is a concern, potentially leading to safer and more reliable AI systems.
Learning to Discover Interesting Mathematics
Large Language Models
Theory
Generative Models
- Introduces a formal definition of mathematical interestingness based on proof length relative to statement length.
- Demonstrates a strong correlation between interestingness and utility in mathematical theorems.
- Develops a 27B parameter model that accurately predicts proof difficulty, outperforming existing models.
- Achieves a significant reduction in overlap with existing mathematical libraries, indicating the generation of novel theorems.
Read more
Learning to Discover Interesting Mathematics
Summary
This paper explores the capabilities of Large Language Models (LLMs) in generating and discovering new mathematical theorems, focusing on the concept of 'interestingness' in mathematics. The authors define interestingness as the ratio of the length of a theorem's proof to the length of its statement, and they demonstrate a strong correlation between this intrinsic measure and the extrinsic utility of theorems. A 27B parameter model is trained to predict proof difficulty conditioned on existing premises, outperforming existing models. The authors show that optimizing for interestingness leads to the generation of more novel theorems, significantly reducing overlap with existing mathematical libraries. The framework developed allows for the iterative generation and selection of interesting mathematical statements, paving the way for self-expanding mathematical libraries that can autonomously identify valuable conjectures without human input.
Methodology
The authors post-train a large language model to predict the difficulty of proving theorems based on existing premises. They define interestingness and utility metrics and use these to guide the generation of new mathematical statements. The model is evaluated against existing mathematical libraries to measure novelty and interestingness.
Results
The model successfully quadruples the mean interestingness of generated theorems and reduces the overlap with existing mathematical libraries from 91.9% to 30.6%. The framework allows for the recursive discovery of nontrivial mathematical statements.
Implications
This work has significant implications for the field of automated theorem proving and mathematical discovery, suggesting that LLMs can not only solve existing problems but also generate new, valuable mathematical knowledge autonomously. It opens avenues for creating self-sustaining mathematical libraries that can evolve without human intervention.
When Explanations Cannot Be Read: Measuring and Correcting SHAP and LIME Rendering for Right-to-Left Languages
Interpretability
NLP
- Characterizes rendering failures in SHAP and LIME visualizations for RTL languages.
- Introduces a quantitative measure of rendering correctness using OCR.
- Demonstrates that common workarounds for rendering issues are ineffective for Urdu.
- Presents SHAP-RTL, a rendering layer that corrects visualization issues while preserving attribution values.
Read more
When Explanations Cannot Be Read: Measuring and Correcting SHAP and LIME Rendering for Right-to-Left Languages
Summary
This paper addresses the challenges faced by post hoc explanation methods, specifically SHAP and LIME, when applied to right-to-left (RTL) languages such as Urdu, Arabic, Persian, and Hebrew. While the mathematical validity of attribution values remains intact, the visual representation fails due to incorrect token sequencing, broken cursive letterforms, and layouts that do not align with the natural reading direction. The authors propose SHAP-RTL, a rendering layer that corrects these visualization issues by implementing complex text layout and per-language font selection while preserving the original attribution values and feature ordering. The effectiveness of SHAP-RTL is evaluated using hate and offensive language datasets across the four RTL languages, measuring rendering correctness through an OCR round trip over 200 feature words per language. The study reveals that common workarounds for rendering issues are ineffective, particularly for Urdu, and highlights the importance of language-aware visualization in enhancing the accessibility of post hoc explainability methods across diverse writing systems.
Methodology
The authors developed SHAP-RTL, a rendering layer that applies complex text layout and per-language font selection to SHAP and LIME visualizations. They quantitatively evaluated rendering correctness by conducting an OCR round trip over 200 feature words for each of the four RTL languages, comparing the results against existing rendering methods.
Results
The evaluation showed that default rendering methods resulted in character error rates ranging from 0.820 to 0.979, with the common reshape-and-reorder workaround failing for Urdu at a rate of 0.998. In contrast, SHAP-RTL maintained rendering correctness across both library versions used in the study, demonstrating its effectiveness in addressing the visualization challenges for RTL languages.
Implications
The findings underscore the necessity for developing language-aware visualization tools in the field of explainable AI, particularly for non-Latin scripts. This work can enhance the interpretability of machine learning models in multilingual contexts, making them more accessible to speakers of RTL languages.
LabFactory: Building and Evaluating Executable AI Labs
Theory
Large Language Models
Optimization
- LABFACTORY enables the creation of executable AI labs from scientific briefs.
- The framework separates the builder's development process from the evaluation of the final product.
- 28 AI labs were constructed, exceeding reference values in various scientific tasks.
- The methodology emphasizes operational evaluation through independent execution of delivered artifacts.
Read more
LabFactory: Building and Evaluating Executable AI Labs
Summary
The paper introduces LABFACTORY, a framework designed to transform scientific briefs into executable AI labs, which are task-specific solvers integrating models, knowledge resources, tools, and controllers. The framework allows an AI builder to develop and package a lab in a controlled workspace, which can then be executed by a separate host on held-out inputs. This approach emphasizes the evaluation of the delivered system rather than the builder's progress, ensuring that the lab remains operational post-construction. The authors document 28 constructions across seven scientific task categories, demonstrating that the delivered labs surpassed their configured reference values across 33 subtests. The framework's contributions include a construction and delivery protocol, detailed accounts of constructed AI labs, and execution-grounded case records, showcasing the potential for AI agents to create functional scientific tools that can be invoked and inspected after their development.
Methodology
The LABFACTORY framework employs a task-conditioned construction process where a builder receives a scientific brief and resources, develops models or tools, and delivers an artifact that adheres to a fixed interface. The construction lifecycle includes stages of inspection, resource acquisition, model development, validation, and packaging, with independent evaluation conducted by a separate host on the delivered lab.
Results
The study reports that all 28 constructed labs exceeded their configured reference values across 33 subtests during host-side execution. The results include comparisons with raw platform LLMs and detailed construction traces, demonstrating the effectiveness of the framework in producing functional AI labs.
Implications
LABFACTORY has the potential to streamline the development of AI systems for scientific tasks, allowing researchers to efficiently create and evaluate task-specific solvers. This could enhance the reproducibility and reliability of AI applications in various scientific domains, from genomics to clinical decision support.
Safety-oriented pedestrian trajectory prediction at urban intersections using time-to-collision and crossing-zone context
Robotics
Time Series
Optimization
- The framework combines pedestrian motion history with Time-to-Collision (TTC) and crossing-zone context for improved trajectory prediction.
- A weighted loss function emphasizes larger prediction errors, enhancing safety-oriented training.
- The final model achieved significant reductions in prediction errors compared to traditional methods.
- Results highlight the importance of integrating contextual information in pedestrian trajectory prediction.
Read more
Safety-oriented pedestrian trajectory prediction at urban intersections using time-to-collision and crossing-zone context
Summary
This paper addresses the critical issue of pedestrian safety at urban intersections by proposing a safety-oriented framework for predicting pedestrian trajectories. The authors emphasize the importance of accurately forecasting pedestrian movements, particularly in complex environments where interactions with vehicles and crossing contexts play a significant role. The proposed framework integrates pedestrian motion history with Time-to-Collision (TTC) data and crossing-zone indicators, utilizing naturalistic trajectory data from the inD dataset. The study evaluates various neural network architectures, particularly a pooled Long Short-Term Memory (LSTM) model, which separately encodes TTC histories and crossing-zone contexts before merging them with pedestrian position data. The authors introduce a weighted loss function that prioritizes larger prediction errors, enhancing the model's focus on safety-critical scenarios. Results demonstrate that the safety-oriented model significantly reduces Average Displacement Error (ADE) and Final Displacement Error (FDE), indicating improved prediction accuracy. The findings suggest that incorporating interaction dynamics and contextual information can effectively mitigate large trajectory prediction errors, although the authors call for broader validation across diverse datasets and pedestrian scenarios.
Methodology
The study employs a pooled Long Short-Term Memory (LSTM) architecture to encode pedestrian motion, Time-to-Collision (TTC) histories, and crossing-zone contexts. The model is trained using naturalistic trajectory data, with a focus on minimizing both average prediction errors and the frequency of large errors through a weighted loss function.
Results
The safety-oriented LSTM model achieved an Average Displacement Error (ADE) of 0.184 m and a Final Displacement Error (FDE) of 0.491 m. The introduction of the weighted loss function reduced ADE from 0.210 to 0.190 m and FDE from 0.550 to 0.503 m, alongside a 34.8% reduction in ADE exceedance counts and a 19.8% reduction in FDE exceedance counts.
Implications
The findings suggest that integrating safety-oriented training and contextual information can significantly enhance pedestrian trajectory prediction models, potentially improving the effectiveness of Advanced Driver Assistance Systems (ADAS) and automated driving technologies in urban environments.
The Impossible Trinity of Time-Series Validation: A Conservation Law among Training Sufficiency, Test Coverage, and Temporal Causality
Time Series
Theory
- Establishes a mathematical framework for time-series validation that quantifies the trade-offs between training sufficiency, test coverage, and temporal causality.
- Proves that it is impossible to satisfy all three validation requirements simultaneously, leading to a conservation law among them.
- Introduces key metrics (α, β, Λ) that characterize validation schemes and their trade-offs.
- Demonstrates that expanding walk-forward validation is optimal for causal validation, while k-fold cross-validation is effective for maximizing future data usage.
Read more
The Impossible Trinity of Time-Series Validation: A Conservation Law among Training Sufficiency, Test Coverage, and Temporal Causality
Summary
This paper addresses the challenges of validating time-series models by introducing a mathematical framework that formalizes the trade-offs between three critical requirements: training sufficiency, test coverage, and temporal causality. The author proves that it is impossible to satisfy all three simultaneously, establishing a conservation law that quantifies the relationship between these demands. The paper defines key metrics such as the worst-fold training fraction (α), test coverage (β), and anti-causal mass (Λ), leading to the main theorem that describes the constraints governing these metrics. The findings reveal that expanding walk-forward validation achieves the optimal balance of these requirements, while k-fold cross-validation maximizes future data usage. The paper also discusses the implications of these findings on model evaluation, highlighting the biases introduced by violating causality and the importance of understanding the memory characteristics of the underlying process. Overall, the work provides a theoretical foundation for time-series validation, offering practical guidance for practitioners in the field.
Methodology
The author develops a formal framework that defines validation schemes and quantifies the three key metrics (training sufficiency, test coverage, and temporal causality). The paper employs mathematical proofs to establish the relationships between these metrics and derives inequalities that govern their interactions. Theoretical results are supported by numerical illustrations and comparisons of standard validation schemes.
Results
The main results include the establishment of the inequalities α + β ≤ 1 + Λ and α + min{β, δ/T} ≤ 1, which describe the limits of training sufficiency, test coverage, and causality. The paper shows that expanding walk-forward validation achieves the best balance of these metrics, while k-fold cross-validation provides the most future data. Additionally, it quantifies the biases introduced by insufficient training and coverage, emphasizing the significance of the distance to future training data in causal violations.
Implications
The findings have significant implications for practitioners in time-series modeling, providing a theoretical basis for selecting appropriate validation schemes. Understanding the trade-offs between training sufficiency, test coverage, and causality can lead to more robust model evaluations and improved predictive performance. The insights gained from this work can inform best practices in time-series analysis across various domains, including finance, economics, and environmental science.
SMILESGNN: Interpretable Clinical Toxicity Prediction via SMILES-Graph Cross-Attention Fusion
Graph Learning
Multimodal
Interpretability
- Introduces SMILESGNN, a multimodal architecture for drug toxicity prediction that combines SMILES and graph representations.
- Achieves high predictive performance on ClinTox and Tox21 datasets while addressing class imbalance and interpretability.
- Demonstrates that cross-attention fusion is a viable alternative to concatenation, preserving model interpretability.
- Utilizes GNNExplainer for analyzing substructures associated with toxic predictions, enhancing interpretability for toxicologists.
Read more
SMILESGNN: Interpretable Clinical Toxicity Prediction via SMILES-Graph Cross-Attention Fusion
Summary
The paper presents SMILESGNN, a novel multimodal architecture designed for predicting clinical toxicity of drugs by fusing SMILES (Simplified Molecular Input Line Entry System) representations with graph neural network (GNN) embeddings through cross-attention. This approach addresses key challenges in drug toxicity prediction, including severe class imbalance, scaffold-based generalization, and the need for interpretable predictions. The architecture consists of a SMILES Transformer encoder and a GATv2 graph encoder, allowing for the extraction of complementary information from both modalities. The authors also introduce SMILESGNN-PT, a variant that utilizes a pretrained ChemBERTa-2 backbone. The model is trained using focal loss to mitigate class imbalance and supports GNNExplainer-based analysis for interpretability. The results demonstrate that SMILESGNN achieves competitive performance on the ClinTox dataset with an AUC-ROC of 0.987 and an F1 score of 0.906, while maintaining a low parameter count of 0.4M. On the Tox21 dataset, SMILESGNN-PT shows comparable performance to existing models, confirming the effectiveness of cross-attention fusion in enhancing predictive accuracy while providing interpretability.
Methodology
The methodology involves a dual-pathway architecture that integrates a SMILES Transformer encoder and a GATv2 graph encoder through cross-attention. The model is trained on the ClinTox and Tox21 datasets, employing focal loss to address class imbalance. Data processing includes custom tokenization for SMILES sequences and feature extraction for molecular graphs, with a focus on maintaining structural details.
Results
SMILESGNN achieves an AUC-ROC of 0.987 ± 0.012 and an F1 score of 0.906 ± 0.039 on the ClinTox dataset with only 0.4M parameters. For the Tox21 dataset, SMILESGNN-PT obtains a mean AUC-ROC of 0.750 ± 0.002, demonstrating competitive performance compared to existing models.
Implications
The findings suggest that SMILESGNN can significantly improve the early prediction of drug toxicity, potentially reducing late-stage attrition in drug development. The model's interpretability features also provide valuable insights for toxicologists, facilitating safer drug candidate selection.
Learning from Mixed-Quality Deployment Experience for Robot Manipulation
Robotics
Reinforcement Learning
- PACL effectively utilizes mixed-quality deployment experiences for robot policy improvement.
- The method combines a Q-conditioned diffusion actor with a chunk-level critic for enhanced learning.
- Future latent dynamics prediction serves as auxiliary supervision, improving value learning under sparse rewards.
- PACL consistently outperforms strong baselines in both simulated and real-world robot manipulation tasks.
Read more
Learning from Mixed-Quality Deployment Experience for Robot Manipulation
Summary
This paper introduces Predictive Action Chunk Learning (PACL), a novel approach for improving robot manipulation policies using mixed-quality deployment experiences. The authors highlight the challenges of directly incorporating these experiences into imitation learning due to the risk of reinforcing undesirable behaviors. PACL addresses this by employing a predictive chunk-level critic that evaluates sequences of actions, enhancing temporal difference learning with future latent predictions. This method allows for richer supervision in long-horizon value estimation. The critic categorizes action sequences into discrete quality conditions, which guide a diffusion actor to learn from both successful and failed experiences without treating them equally. The experimental results demonstrate that PACL significantly enhances pretrained policies in both simulated and real-world tasks, outperforming traditional imitation learning and offline reinforcement learning methods. The findings suggest that leveraging naturally accumulated experiences can lead to substantial improvements in robotic performance.
Methodology
PACL employs a predictive chunk-level critic to evaluate temporally extended action sequences, enhancing traditional Q-learning methods. It integrates future latent dynamics prediction to provide dense supervision for value estimation, allowing the actor to learn from both successful and failed experiences. The actor generates multiple action chunks, from which the critic selects the highest-valued candidate during inference.
Results
The experiments reveal that PACL improves pretrained policies across various robot manipulation tasks, significantly outperforming established imitation learning and offline reinforcement learning baselines. Ablation studies confirm the positive contributions of both the Q-conditioned actor and the predictive critic to the overall performance enhancement.
Implications
The findings suggest that PACL can be applied to enhance robotic learning in real-world environments, enabling robots to autonomously improve their performance based on their operational experiences. This approach could lead to more robust and adaptable robotic systems capable of handling diverse tasks and conditions.
When Temporal Perturbations Act Like Sensor Biases: Label-Free Auditing of Wearable Activity Recognizers
Time Series
- Introduction of SpectrumAudit for label-free auditing of HAR models.
- Demonstration that temporal perturbations can act like sensor biases.
- Significant accuracy losses observed across multiple datasets and architectures.
- Separation of waveform selection from diagnosis to identify the source of failures.
Read more
When Temporal Perturbations Act Like Sensor Biases: Label-Free Auditing of Wearable Activity Recognizers
Summary
This paper introduces SpectrumAudit, a novel label-free auditing method for wearable human-activity recognition (HAR) models. The authors investigate how temporal perturbations can mimic sensor biases, affecting the reliability of HAR systems. The methodology involves selecting waveforms from unlabeled calibration windows of subjects not included in training or testing, and then replaying these waveforms to assess their impact on model accuracy. The study spans three datasets and three backbone architectures, revealing that selected waveforms can lead to significant accuracy losses, primarily due to persistent sensor offsets rather than temporal variations. The findings highlight the importance of distinguishing between these two types of perturbations in HAR systems, providing a framework for diagnosing vulnerabilities in wearable activity recognizers without relying on labeled data.
Methodology
The authors employed a two-stage audit process where waveforms were selected from unlabeled calibration windows of held-out subjects. The selected waveforms were then replayed on the same subjects to evaluate their effects on model performance. The analysis included both the direct DC component and a budget-constrained zero-mean residual to determine the contributions of persistent offsets versus temporal variations.
Results
The results indicated that the selected waveforms caused accuracy losses ranging from 2.87 to 40.83 points across 27 victims from three datasets. The DC component was found to be more damaging than the AC component in most cases, with significant recovery of accuracy observed when the DC component was replayed. In a separate evaluation on the UTD-MHAD dataset, the selected waveform resulted in a 13.49 percentage point accuracy drop and an 11.68 percentage point macro-F1 score loss, compared to negligible changes from matched random perturbations.
Implications
The findings suggest that understanding the nature of perturbations in wearable HAR systems is crucial for enhancing their robustness. The proposed auditing framework can be applied to improve the reliability of wearable devices in mobile health and assistive technologies, potentially leading to better user experiences and outcomes.
SGA: Uncertainty Quantification for Multi-Step Forecasting in Time Series Foundation Models
Time Series
- SGA effectively quantifies uncertainty in multi-step forecasting for TSFMs.
- The method utilizes a directed acyclic graph to characterize forecast branches.
- SGA outperforms existing methods in ranking predictive errors with uncertainty estimates.
- Larger TSFM models tend to exhibit lower uncertainty estimates.
Read more
SGA: Uncertainty Quantification for Multi-Step Forecasting in Time Series Foundation Models
Summary
This paper addresses the challenge of uncertainty quantification in multi-step forecasting using Time Series Foundation Models (TSFMs). While TSFMs have improved forecasting accuracy, they often exhibit significant uncertainty, which can undermine the reliability of their predictions. The authors propose a novel method called Slicing-Graphing-Alignment (SGA) to quantify this uncertainty. SGA characterizes the potential forecast branches through a directed acyclic graph (DAG), integrating topological information and inherent stochasticity of TSFMs to measure graph complexity. The method involves three stages: slicing, graphing, and alignment. In the slicing stage, multiple forecasts are sampled and segmented into slices. The graphing stage constructs a DAG to represent the temporal dependencies of these slices, while the alignment stage uses Dynamic Time Warping (DTW) to measure slice similarity and reveal the forecasting space's topology. Experimental results demonstrate that SGA outperforms existing uncertainty quantification methods, provides more extensive sampling coverage, and shows that larger TSFM models correlate with lower uncertainty estimates. This work highlights the importance of quantifying uncertainty in multi-step forecasting, particularly in high-stakes applications.
Methodology
The SGA method consists of three main stages: (1) Slicing - sampling multiple forecasts and segmenting them into slices; (2) Graphing - constructing a directed acyclic graph (DAG) to represent the temporal dependencies of these slices; (3) Alignment - using Dynamic Time Warping (DTW) to measure the similarity between slices and merging similar ones to reveal the intrinsic topology of the forecasting space. The uncertainty is quantified by measuring the complexity of the forecasting space, approximated through graph complexity.
Results
The experimental validation across 11 TSFMs and 27 datasets shows that SGA significantly outperforms existing methods in ranking predictive errors with uncertainty estimates. It provides a more extensive and precise sampling coverage, leading to a fundamentally different quantification mechanism. Additionally, the results indicate that larger model scales correlate with lower uncertainty estimates in multi-step forecasts.
Implications
The findings suggest that incorporating uncertainty quantification into multi-step forecasting can enhance the credibility of TSFM predictions, which is crucial for applications in high-stakes scenarios such as finance and industrial forecasting. This work lays the groundwork for future research on uncertainty in forecasting models.
TAM-Chain: Multi-Scale Thyroid Cytology Classification via Absorbing Markov Chains and Shannon Entropy Uncertainty Quantification for False-Negative Suppression and Domain-Shift Adaptation
Theory
Computer Vision
Efficient ML
- TAM-Chain effectively suppresses false negatives in thyroid cytology classification.
- The framework integrates multi-scale feature extraction with uncertainty quantification.
- Achieved a 0.00% false-negative rate on an internal test set.
- Demonstrated adaptability under domain shifts with robust performance on external validation.
Read more
TAM-Chain: Multi-Scale Thyroid Cytology Classification via Absorbing Markov Chains and Shannon Entropy Uncertainty Quantification for False-Negative Suppression and Domain-Shift Adaptation
Summary
The paper presents TAM-Chain, a novel framework for thyroid cytology classification that addresses significant challenges in deep learning applications, particularly high false-negative rates and overconfidence in clinical settings. By utilizing Absorbing Markov Chain theory and Shannon Entropy for uncertainty quantification, TAM-Chain dynamically models multi-magnification feature extraction (10×, 20×, 40×) as an absorbing stochastic process. This allows for optimal stopping criteria and a human-in-the-loop referral mechanism to minimize critical diagnostic errors. The framework was rigorously evaluated on an internal test set, achieving a Macro F1 score of 0.9741 with a 0.00% false-negative rate. Furthermore, on an external validation set with severe domain shifts, TAM-Chain maintained a Macro F1 score of 0.7026, demonstrating its adaptability and robustness compared to single-magnification models. The study highlights the importance of integrating automated diagnostic efficiency with stringent biological safety, making TAM-Chain a promising solution for digital pathology workflows.
Methodology
The TAM-Chain framework employs Absorbing Markov Chains to model the classification process across multiple magnifications, utilizing Shannon Entropy for uncertainty quantification. It features a dynamic early stopping mechanism that optimizes the stopping step based on uncertainty levels and a referral mechanism for cases requiring expert evaluation.
Results
On an internal test set of 235 samples, TAM-Chain achieved a Macro F1 score of 0.9741 and a false-negative rate of 0.00%. In external validation on a dataset of 1015 samples with significant domain shifts, it maintained a Macro F1 score of 0.7026, outperforming traditional single-magnification models.
Implications
The TAM-Chain framework has the potential to enhance the safety and accuracy of automated diagnostic systems in digital pathology, particularly in thyroid cancer detection, by effectively managing uncertainty and adapting to domain shifts.
Upholding Robustness in Federated Learning: Trends, Emerging Strategies, and Research Opportunities
Federated Learning
- Federated Learning is increasingly adopted for privacy protection but faces significant robustness challenges.
- The paper presents a threat-centric view categorizing various attack surfaces in FL.
- A structured taxonomy of robust aggregation strategies is proposed, differentiating between outcome-centric and security-centric approaches.
- Current evaluation practices for FL robustness are inconsistent, necessitating a unified framework for assessment.
Read more
Upholding Robustness in Federated Learning: Trends, Emerging Strategies, and Research Opportunities
Summary
This paper provides a comprehensive overview of the robustness challenges faced in Federated Learning (FL), a decentralized machine learning paradigm that enhances user privacy. The authors categorize the various threats to FL, including performance degradation, information theft, and vulnerabilities in aggregation processes. They propose a structured taxonomy that distinguishes between outcome-centric and security-centric robust aggregation strategies, as well as a layered taxonomy of defensive measures. The paper critically examines current evaluation practices for FL robustness and highlights significant applications and open research challenges. The authors emphasize the need for multi-faceted solutions to address the diverse threats and advocate for a consistent evaluation framework to assess the effectiveness of different strategies in real-world scenarios. Through this synthesis, the paper aims to guide future research towards enhancing the robustness of FL systems.
Methodology
The authors conducted a systematic review of existing literature on Federated Learning, analyzing various threats, aggregation strategies, and defensive mechanisms. They developed taxonomies to categorize these elements and critically assessed current evaluation practices.
Results
The paper identifies key factors affecting FL robustness, including adversarial threats, aggregation processes, and defensive mechanisms. It highlights the need for adaptive and customized approaches to improve robustness and calls for a consolidated discussion on defensive strategies to steer future research.
Implications
The findings of this paper have significant implications for the design and deployment of Federated Learning systems across various sectors such as finance, healthcare, and IoT. By addressing robustness challenges, the research can enhance the reliability and security of FL applications, fostering greater trust and adoption.
On the SoS Certifiability of Log-Concave Distributions
Theory
Efficient ML
Optimization
- Establishes SoS certifiability for isotropic log-concave distributions.
- Removes dependence on the Poincaré constant for moment bounds.
- Introduces efficient algorithms for high-dimensional statistical estimation.
- Utilizes stochastic localization and variance inequalities in proofs.
Read more
On the SoS Certifiability of Log-Concave Distributions
Summary
This paper investigates the certifiability of isotropic log-concave distributions through the lens of sum-of-squares (SoS) proofs. The author establishes that for any isotropic log-concave distribution P on R^d, the polynomial (Cm)m∥v∥m^2 − E[X∼P ⟨X, v⟩^m is a sum of squares for every even m ≥ 2, where C is a universal constant. This result eliminates the dependence on the Poincaré constant, thereby recovering optimal moment bounds for log-concave distributions. The methodology employs stochastic localization to express P as an average of random strongly log-concave measures, allowing for the derivation of fourth-moment certificates using Letwin’s variance inequality. The findings lead to computationally efficient algorithms with dimension-free error guarantees applicable to various high-dimensional statistical estimation problems, particularly in robust mean estimation scenarios. The paper highlights the significance of certifiability in enhancing statistical guarantees for algorithms reliant on moment bounds.
Methodology
The author employs stochastic localization to decompose isotropic log-concave distributions into averages of random strongly log-concave measures. This approach facilitates the derivation of moment certificates using Letwin’s variance inequality, allowing for the establishment of SoS proofs for moment bounds.
Results
The main result confirms that every isotropic log-concave distribution is (Cm, m)-certifiably bounded for all even m ≥ 2, achieving optimal moment bounds without dependence on the dimension. This leads to the development of efficient algorithms with dimension-free error guarantees for statistical estimation problems.
Implications
The findings have significant implications for statistical estimation in high-dimensional settings, particularly in robust mean estimation. The established SoS certifiability can enhance the performance of algorithms that rely on moment bounds, making them more efficient and reliable in practice.
Sample-Weighted End-to-End Trace-Norm Geometry for Multitask Learning
Theory
- Introduces a sample-size-weighted trace norm for multitask learning that captures the intrinsic complexity of shared representations and task-specific outputs.
- Derives the exact empirical Rademacher complexity for the proposed method, providing a theoretical foundation for its effectiveness.
- Demonstrates through experiments that weighted joint nuclear regularization outperforms traditional methods, improving multitask learning performance.
- Highlights the importance of considering the geometric properties of shared maps in understanding task relationships and their impact on learning.
Read more
Sample-Weighted End-to-End Trace-Norm Geometry for Multitask Learning
Summary
This paper addresses the complexities of multitask learning by proposing a novel approach that utilizes the sample-size-weighted trace norm of the end-to-end mapping from task coefficients to input-space predictors. The author critiques traditional methods that treat shared representations and task-specific outputs separately, arguing that this can lead to loss of information regarding the relationships between tasks. The study derives the exact empirical Rademacher complexity for the proposed fixed-radius class and characterizes it through the optimization of a positive-definite task covariance. The paper also discusses the implications of geometric properties of shared maps on the task Gram matrix and the effectiveness of a convex regularizer in enhancing multitask learning performance. Experimental results demonstrate that weighted joint nuclear regularization significantly improves average population excess compared to unweighted methods, highlighting the advantages of the proposed approach in multitask settings.
Methodology
The author employs a theoretical framework based on the sample-size-weighted trace norm to analyze the end-to-end mapping in multitask learning. The methodology includes deriving empirical Rademacher complexity, optimizing task covariance, and conducting experiments to evaluate the performance of the proposed regularization techniques across various multitask learning scenarios.
Results
The results indicate that weighted joint nuclear regularization improves average population excess by 0.00764 compared to unweighted nuclear regularization, with a 95% confidence interval of [0.00465, 0.01110]. Additionally, correct task counts enhance performance metrics significantly, with improvements of 0.01072 and 0.02847 over shifted counts in various settings. The proposed method consistently outperformed other regularization techniques, confirming its efficacy in multitask learning.
Implications
The findings suggest that incorporating a sample-size-weighted approach in multitask learning can lead to better generalization and performance. This has potential applications in various fields where multitask learning is relevant, such as natural language processing, computer vision, and other domains requiring the integration of multiple tasks.
Where Does Exactly-Once Live? Model, Harness, and Tool-Contract Effects on Duplicate Side Effects in LLM Agents
Large Language Models
- The responsibility for ensuring exactly-once behavior in LLM agents depends on the type of fault encountered.
- Frontier models can significantly reduce duplication rates when they can verify outcomes before retrying.
- Tool contracts, particularly those that include idempotency keys, are crucial for managing duplicates in uncertain scenarios.
- Client-side middleware (harnesses) has limited effectiveness in achieving exactly-once success.
Read more
Where Does Exactly-Once Live? Model, Harness, and Tool-Contract Effects on Duplicate Side Effects in LLM Agents
Summary
This paper investigates the challenge of ensuring exactly-once behavior in tool-using agents, particularly in the context of Large Language Models (LLMs) that interact with external systems. The author introduces Limbo, a deterministic sandbox designed to evaluate the effects of various factors—model, harness, and tool contract—on the occurrence of duplicate side effects when agents retry actions after failures. The study spans 25,930 episodes across nine models and various configurations, revealing that the responsibility for ensuring exactly-once behavior shifts depending on the type of fault encountered. When a read-back can clarify the outcome of a request, the model's capabilities play a crucial role, with frontier models achieving a low duplication rate of 0.5%. However, in cases where the request is still in flight or redelivered, the tool contract becomes the primary determinant, explaining 81% of the variance in duplicates. The findings highlight the limitations of verification-only policies and emphasize the importance of idempotency keys in reducing duplication rates significantly. The paper concludes that while harnesses have minimal impact, agents often remain unaware of their duplicate actions, reporting successful completions in 90% of such cases.
Methodology
The study employs a factorial experimental design using the Limbo sandbox to simulate various fault conditions and evaluate the performance of different models, harnesses, and tool contracts. The experiments involve injecting twelve fault modes and grading each episode against a ledger of committed effects.
Results
The results indicate that frontier models achieve a duplication rate of only 0.5% when read-backs are possible, while duplication rates rise to 56% and 74% under certain fault conditions without verification. The introduction of idempotency keys reduces duplication rates from 28% to 4%. The harnesses used in the experiments show minimal variation in performance, and agents report successful completions in 90% of duplicate cases.
Implications
The findings suggest that improving the design of tool contracts and incorporating idempotency keys can significantly enhance the reliability of LLM agents in executing actions without unintended duplicates. This has implications for the development of more robust agent frameworks and better handling of external system interactions.
On the second-order optimization for spiking neural networks
Optimization
Efficient ML
Theory
- Introduction of SpiKFAX, a second-order optimization method for SNNs.
- Development of a computationally tractable Kronecker-factored approximation of the Fisher information matrix.
- Demonstrated improvements in test accuracy and training stability across multiple architectures and datasets.
Read more
On the second-order optimization for spiking neural networks
Summary
This paper addresses the challenges associated with training Spiking Neural Networks (SNNs), which are known for their energy efficiency but suffer from a sharp loss landscape due to their spiking activations. Traditional first-order optimizers, such as those in the Adam family, struggle with the unique geometry of SNNs, leading to slow convergence and poor generalization. To overcome these limitations, the authors propose SpiKFAX, a second-order optimization method that utilizes a Kronecker-factored approximation of the Fisher information matrix tailored for SNNs. This method captures the second-order curvature information essential for navigating the complex loss landscape of SNNs. The authors conduct extensive empirical evaluations across five different architectures and seven datasets, demonstrating that SpiKFAX consistently outperforms popular optimizers like SGD, Adam, and AdamW in terms of both accuracy and training stability.
Methodology
The authors formulate the optimization problem for SNNs, focusing on minimizing a non-convex, non-differentiable loss function. They derive a preconditioned update rule that incorporates a curvature approximation based on the Fisher information matrix. The Kronecker-factored approximation is specifically designed to accommodate the time-recurrent dynamics of SNNs, making it computationally feasible.
Results
The empirical evaluation shows that SpiKFAX leads to significant improvements in test accuracy and training stability compared to traditional optimizers like SGD, Adam, and AdamW. The method was tested across five different SNN architectures and seven datasets, including both static and neuromorphic vision benchmarks.
Implications
The development of SpiKFAX has the potential to enhance the training of SNNs, making them more viable for practical applications in energy-efficient computing and neuromorphic systems. This could lead to advancements in areas such as robotics, real-time processing, and other domains where SNNs can be leveraged for their efficiency.
Generative Atmospheric Super-Resolution from Heterogeneous In Situ Observations through Composable Interfaces
Generative Models
Time Series
Multimodal
- Introduces composable observation interfaces for integrating heterogeneous atmospheric observations.
- Demonstrates a modular approach to conditioning a pretrained atmospheric diffusion model.
- Achieves a 9.24% reduction in RMSE by combining radiosonde, aircraft, and surface station data.
- Validates the method using extensive evaluation across different atmospheric variables.
Read more
Generative Atmospheric Super-Resolution from Heterogeneous In Situ Observations through Composable Interfaces
Summary
This paper addresses the challenge of reconstructing atmospheric states from sparse and heterogeneous observations using a generative atmospheric super-resolution approach. The authors propose a novel framework that utilizes composable observation interfaces to condition a pretrained atmospheric diffusion model on various in situ observation sources, including radiosonde, aircraft, and surface station data. The interfaces transform these diverse observations into source-specific likelihood factors, which facilitate the integration of observations with different geometries and sampling densities into a unified inference framework. The study evaluates the performance of the proposed method using data from 2019 and 2020, demonstrating significant improvements in reconstruction accuracy. The results indicate that combining observations from multiple sources leads to a reduction in root mean square error (RMSE) and continuous ranked probability score (CRPS), showcasing the effectiveness of the composable interfaces in enhancing atmospheric state estimation without the need for retraining the underlying model.
Methodology
The authors formulated the atmospheric state reconstruction as a generative super-resolution problem, utilizing a pretrained atmospheric diffusion model. They developed composable observation interfaces that convert observations from different sources into likelihood factors, which are then integrated into a Bayesian inference framework. The methodology involved evaluating the interfaces using historical data and assessing their performance through RMSE and CRPS metrics.
Results
The R+A+S interface, which combines radiosonde, aircraft, and surface station observations, resulted in a 9.24% reduction in RMSE compared to using only radiosonde data. Additionally, the combined approach lowered the CRPS and demonstrated reduced prediction errors in held-out observations, indicating complementary improvements in both upper-air and surface variables.
Implications
This research has significant implications for atmospheric modeling and forecasting, as it provides a robust method for integrating diverse observational data into generative models. The composable interface approach can enhance the accuracy of atmospheric state reconstructions, which is crucial for weather prediction, climate modeling, and environmental monitoring.
Growth-Inspired Graph Generation and Inverse Design of Mechanical Lattices via Dot Matrices Database Augmentation and GCNN
Graph Learning
Optimization
Generative Models
- Introduces a morphogenetic framework for generating mechanical lattices based on natural growth processes.
- Utilizes a graph convolutional neural network to predict effective compressive stiffness from lattice topology.
- Enables inverse design for mechanical lattices, achieving target stiffness with high accuracy.
- Extends the framework to include nonlinear materials and complex geometries, enhancing design flexibility.
Read more
Growth-Inspired Graph Generation and Inverse Design of Mechanical Lattices via Dot Matrices Database Augmentation and GCNN
Summary
This paper presents a novel morphogenetic graph-generation framework for mechanical lattices, inspired by natural growth processes. The authors utilize a discrete dot matrix to define potential nodes, allowing for a sequential growth process that mimics biological systems like leaf venation. The framework is implemented in both two and three dimensions, with a focus on a 3x3x3 nodal matrix. The generated lattices are evaluated using beam-based finite element analysis (FEA) and represented as graphs. A graph convolutional neural network (GCNN) is employed to learn the relationship between the lattice topology and its effective compressive stiffness. This approach facilitates inverse design, enabling the prediction of lattice structures that meet specific stiffness requirements. The authors extend their framework to include parameterized horseshoe-shaped beams made from nonlinear materials, allowing for more complex design variables and programmable deformation shapes. Overall, this work bridges biological morphogenesis, graph learning, and nonlinear shape programming, providing a comprehensive generative design framework for architected materials.
Methodology
The authors developed a growth-based structural representation method using a discrete dot matrix, allowing for sequential cross-layer and intra-layer growth. A graph convolutional neural network (GCNN) with three convolutional layers and dual global pooling was trained to learn the topology-property mapping. Finite element analysis was used to evaluate the generated lattices.
Results
The framework successfully predicted a lattice design with a target stiffness of 1000 MPa, achieving a predicted stiffness of 1042.43 MPa and validating it with finite element analysis at 1027.49 MPa. The extension to parameterized horseshoe-shaped beams demonstrated the capability of designing for specific deformation shapes.
Implications
This research provides a new paradigm for the design of mechanical metamaterials, integrating biological growth principles with machine learning. It opens avenues for creating advanced materials with tailored mechanical properties and programmable behaviors, which could have applications in various fields such as aerospace, robotics, and biomedical engineering.
MF-SCBO : Multi-fidelity Scalable Constrained Bayesian Optimization
Optimization
- MF-SCBO effectively integrates multi-fidelity optimization with constrained Bayesian optimization.
- The method addresses high-dimensional optimization problems with arbitrary fidelity levels.
- Two trust-region center selection strategies are proposed to enhance optimization stability.
- Experimental results show improved convergence over existing methods in challenging scenarios.
Read more
MF-SCBO : Multi-fidelity Scalable Constrained Bayesian Optimization
Summary
The paper introduces MF-SCBO, a novel method for multi-fidelity scalable constrained Bayesian optimization, addressing the challenges of optimizing high-dimensional black-box functions with constraints. Existing methods have struggled to simultaneously manage high-dimensionality, multiple fidelity levels, and non-nested sampling. MF-SCBO extends the Scalable Constrained Bayesian Optimization (SCBO) framework to multi-fidelity settings, allowing for the efficient use of both high-fidelity and low-fidelity evaluations. The authors propose two variants for determining the trust-region center based on high-fidelity evaluations and predictions, and they introduce a leave-one-out criterion to enhance predictive accuracy. The method is evaluated against standard benchmark functions and more complex real-world problems, demonstrating superior convergence performance compared to both single-fidelity SCBO and other multi-fidelity methods. The results indicate that MF-SCBO is effective in reducing the computational budget while maintaining optimization quality, making it suitable for applications in engineering, machine learning, and control.
Methodology
The MF-SCBO method builds on the SCBO framework by incorporating multi-fidelity evaluations. It utilizes Gaussian processes for surrogate modeling and introduces two strategies for selecting the trust-region center: one based on high-fidelity evaluations and another on predictions. The method employs a leave-one-out criterion to improve predictive accuracy and manages constraints that are independent of fidelity levels.
Results
The experimental evaluation of MF-SCBO on benchmark and real-world problems demonstrates that it consistently outperforms both the single-fidelity SCBO and the Multi-Fidelity Max-value Entropy Search (MF-MES) in terms of convergence speed and optimization efficiency, particularly in high-dimensional and constrained settings.
Implications
MF-SCBO has significant implications for fields that rely on expensive simulations or experiments, such as aerospace engineering, machine learning hyperparameter tuning, and robotics. Its ability to efficiently navigate high-dimensional spaces while managing constraints can lead to more effective optimization strategies in various applications.
When Identical Rows Disagree: From Benchmark Identifiability to Replication-Robust Anomaly Detection
Theory
- Identical rows in datasets can lead to significant ambiguities in anomaly detection.
- The paper introduces SCOUT, a framework that improves anomaly detection robustness by focusing on support rather than rows.
- An extensive audit of 690 datasets reveals critical issues with train-test overlap and conflicting labels.
- SCOUT maintains comparable performance to traditional methods while enhancing replication-invariant AUROC.
Read more
When Identical Rows Disagree: From Benchmark Identifiability to Replication-Robust Anomaly Detection
Summary
This paper addresses the challenges posed by identical rows in datasets used for anomaly detection, which are often treated as independent samples despite potentially encoding business frequency, repeated entities, or errors. The author identifies three main consequences of this ambiguity: an evaluation ceiling imposed by feature-identical rows, sensitivity of row-weighted AUROC to replication, and the learning of a biased law by row-trained detectors. An audit of 690 OddBench datasets reveals significant issues, including train-test overlap and conflicting labels among identical features. The paper introduces SCOUT (Support–Count Orthogonalized Unsupervised Testing), a novel anomaly detection framework that shifts the focus from row-based to support-based weighting, improving robustness against replication. SCOUT employs a one-class model for distinct support patterns and a conditional count model when applicable, achieving marginal false-positive control through factorwise split-conformal calibration. The results demonstrate that SCOUT is comparable to traditional methods like Isolation Forest in AUROC while enhancing replication-invariant performance. The findings underscore the importance of addressing multiplicity in anomaly detection and provide a comprehensive framework for understanding the implications of repeated rows in datasets.
Methodology
The methodology involves an audit of 690 OddBench datasets to identify issues related to identical rows, followed by the development of SCOUT, which utilizes a one-class model for distinct support patterns and a conditional count model. The framework employs factorwise split-conformal calibration to control false-positive rates and evaluates performance through AUROC metrics across various datasets.
Results
The analysis found that 355 datasets had train-test overlap, 147 featured identical rows with conflicting labels, and 137 had test anomalies identical to training normals. SCOUT demonstrated non-inferior performance to row-wise Isolation Forest in raw AUROC, with a slight improvement in replication-invariant AUROC. The results indicate that changing the evaluation unit from rows to support can significantly affect AUROC metrics.
Implications
The findings suggest that anomaly detection frameworks must account for the implications of repeated rows in datasets, which can affect model performance and evaluation metrics. The SCOUT framework provides a robust alternative for practitioners dealing with real-world data that may contain such ambiguities.
Canopy: Exploiting Piecewise Smooth Tree Priors for Multi-Fidelity Bandits
Large Language Models
Optimization
Efficient ML
- CANOPY learns the validity of smoothness assumptions from data rather than relying on predefined schedules.
- The framework employs random-path probes to identify regions needing expensive evaluations.
- The method demonstrates significant performance improvements across multiple LLM inference tasks.
- Theoretical guarantees are provided for fixed-budget and regret performance.
Read more
Canopy: Exploiting Piecewise Smooth Tree Priors for Multi-Fidelity Bandits
Summary
The paper introduces CANOPY, a novel multi-fidelity tree bandit framework designed to optimize large language model (LLM) inference problems by exploiting the hierarchical structure of candidate evaluations. Traditional hierarchical bandit methods rely on a predefined smoothness schedule, which can be challenging to determine in practice, especially when objectives are piecewise smooth with sharp boundaries. CANOPY addresses this limitation by learning the validity of the smoothness prior through cheap random-path probes that assess local aggregation bias. This approach allows the algorithm to direct expensive leaf evaluations toward regions where smoothness violations are detected. The authors provide theoretical guarantees for fixed-budget and regret performance, demonstrating that CANOPY can recover optimal rates under smooth conditions while adapting to dense violations. Experimental results across various benchmarks, including model routing and caching, show that CANOPY significantly outperforms existing methods, achieving higher recall rates and faster response times while maintaining budget constraints.
Methodology
CANOPY utilizes a combination of online local-Lipschitz certificates to measure aggregation bias and a discontinuity-guided adaptive sampling technique. This allows the algorithm to efficiently allocate resources by focusing evaluations on regions where smoothness assumptions fail, thus optimizing the overall performance in multi-fidelity settings.
Results
In experiments, CANOPY achieved 2.9 times higher top-10 recall on a 1000-model pool compared to the best structure-blind baseline. It also resolved 1.6 times more SWE-bench Verified issues than the best-of-N approach and reduced median time-to-first-token by 3.6 times in prefix caching scenarios. The method consistently outperformed existing techniques across various benchmarks, demonstrating its effectiveness in real-world applications.
Implications
The findings suggest that CANOPY can be applied to various LLM inference tasks, including model routing, caching, and prompt trimming, leading to more efficient resource utilization and improved performance in practical applications. This framework could enhance the efficiency of LLM deployments in production environments.
Spatio-temporally complementary feature propagation on graphs for longitudinal AADT estimation
Graph Learning
Optimization
Time Series
- Introduces a spatio-temporally complementary feature propagation framework for AADT estimation.
- Combines temporally dense loop detector data with a spatially complete transportation model.
- Utilizes Poisson energy minimization and flow ratio matrices for improved accuracy.
- Achieves convergence in minutes and a normalized mean absolute error below 10% in Zurich.
Read more
Spatio-temporally complementary feature propagation on graphs for longitudinal AADT estimation
Summary
This paper addresses the challenge of estimating Annual Average Daily Traffic (AADT) across urban networks, which is crucial for transportation planning but often hampered by the high costs and sparse distribution of physical sensors. The authors propose a novel framework that integrates two data sources: temporally dense loop detector data, which is spatially sparse, and a spatially complete but temporally sparse macroscopic transportation model. The methodology employs a feature propagation algorithm on directed graphs, formulated as a Poisson energy minimization problem that accounts for residues. A key innovation is the use of flow ratio matrices instead of standard binary adjacency matrices to better represent vehicle turn ratios at intersections. The proposed algorithm was validated in Zurich, demonstrating high computational efficiency with convergence achieved in minutes. The results indicate that the framework effectively combines theoretical models with empirical data, achieving a normalized mean absolute error below 10%. This scalable approach offers a practical solution for network-wide AADT estimation, leveraging limited sensor coverage alongside traffic models.
Methodology
The methodology involves a feature propagation algorithm on directed graphs, framed as a Poisson energy minimization problem. It replaces the standard binary adjacency matrix with flow ratio matrices to accurately capture vehicle turn ratios at intersections, enhancing the estimation process for unmonitored road segments.
Results
The proposed framework was validated in Zurich, achieving high computational efficiency with convergence in minutes and a normalized mean absolute error below 10%. This indicates a strong alignment between the theoretical models and empirical data, showcasing the effectiveness of the approach.
Implications
The findings suggest that the proposed framework can significantly improve AADT estimation across urban networks, facilitating better transportation planning and infrastructure maintenance. It demonstrates the potential for integrating sparse sensor data with comprehensive traffic models to enhance urban mobility analysis.
Tracking States or Tracking Cosets? An Algebraic Account of Learned State Tracking
Theory
- Introduces a quantitative account of partial tracking based on class size and output uncertainty.
- Establishes a connection between Transformer quotient solutions and normal/non-normal coset stages in recurrent networks.
- Identifies the geometry of coset mean vectors in hidden states and their role in subsequent predictions.
- Demonstrates that sequential updates can track non-normal cosets, expanding the understanding of state tracking in neural networks.
Read more
Tracking States or Tracking Cosets? An Algebraic Account of Learned State Tracking
Summary
This paper investigates the concept of state tracking in neural networks, particularly focusing on how models predict the running product of group elements. The authors identify that while models can achieve high accuracy, this does not necessarily indicate a faithful representation of the underlying state space. They explore the notion of quotient solutions in Transformers, where models can recover the quotient class of states while making nearly uniform predictions among its members. The study introduces a novel metric for partial accuracy based on the reciprocal of class size, extending previous parity-based accounts to non-parity quotients. The authors demonstrate that baseline Transformers exhibit minimal changes in predictions under prefix reordering, which is significant beyond the exact-tracking frontier. They prove that for finite groups with uniform i.i.d. inputs, the optimal order-blind accuracy converges to the reciprocal of the abelianization class size as the prefix length increases, aligning with observed abelianization plateaus. Furthermore, they show that sequential updates allow for tracking partitions into right cosets of subgroups, regardless of whether the subgroup is normal. The paper also highlights that recurrent networks can navigate through both normal and non-normal coset stages during training, revealing low-dimensional subspaces that encode these non-normal cosets. The findings connect the concepts of partial accuracy, learning stages, and internal computation through the subgroup cosets that models learn to track.
Methodology
The authors conducted experiments using Transformers and recurrent networks to analyze their ability to track states and cosets. They employed cross-entropy loss for training on sequences of group elements and examined the effects of input order on model predictions. Theoretical proofs were provided to support their findings regarding accuracy and recoverable information.
Results
The study found that class size predicts partial accuracy when models identify the class but predict uniformly within it. It was shown that optimal order-blind accuracy converges to the reciprocal of the abelianization class size as prefix length increases. Additionally, recurrent networks were found to utilize non-normal coset stages, with specific subspaces of the state encoding these cosets.
Implications
The findings have implications for understanding how neural networks learn and represent complex state spaces, particularly in applications involving sequential data processing, such as natural language processing and program execution. The insights into coset tracking could inform the design of more robust models that better capture the underlying structures of the data.
RLVR landscapes for iterated multiplications can be benign: Insights from spin-glass theory
Reinforcement Learning
Theory
Large Language Models
- The RLVR landscape for certain algorithmic tasks is benign, lacking local minima that trap training.
- Challenges in RLVR arise from diffusive barriers and gradient estimation errors rather than landscape ruggedness.
- Entropy regulators can help mitigate obstacles in RLVR optimization.
- A transformer model can learn complex reasoning tasks effectively when properly regulated.
Read more
RLVR landscapes for iterated multiplications can be benign: Insights from spin-glass theory
Summary
This paper investigates the optimization landscape of Reinforcement Learning with Verifiable Rewards (RLVR) applied to algorithmic tasks, specifically iterated group and quasigroup multiplications. The authors establish a mapping between RLVR and spin-glass models, allowing for a rigorous characterization of the optimization landscape. They demonstrate that for a wide range of models and tasks with uncorrelated inputs, the RLVR landscape is benign, lacking local minima that could hinder training. Instead, challenges arise from diffusive barriers and gradient estimation errors. The authors propose that these obstacles can be mitigated through careful selection of entropy regulators. Experimental results show that a transformer model trained from scratch using last-token rewards can effectively learn the algorithmic reasoning required for iterated non-Abelian group multiplications, supporting the theoretical findings.
Methodology
The authors mapped entropy-regularized RLVR over myopic tabular policies onto an energy-based spin-glass model. They analyzed the landscape using theoretical proofs and numerical experiments, focusing on the behavior of RLVR in the context of iterated multiplications. The study involved deriving closed expressions for the free energy of the spin-glass model and assessing the performance of various policy update strategies.
Results
The mapping to spin-glass models revealed that the RLVR landscape for untied models has no suboptimal local maxima, and Monte Carlo sampling can solve these models without a curriculum. However, with weight tying and input correlations, local maxima can emerge. The study also found that transformers trained with strong regulation could effectively learn the tasks, while weakly regulated models struggled.
Implications
The findings suggest that RLVR can be effectively applied to complex reasoning tasks without being hindered by local minima, provided that appropriate entropy regulation is used. This has implications for the design of reinforcement learning systems, particularly in improving the training of large language models and other AI systems that require reasoning capabilities.
BridgeMem: Causal Dyadic Transition Residuals for Temporal Knowledge Graph Forecasting
Graph Learning
Time Series
Theory
- BRIDGEMEM captures dyadic transition evidence, improving the accuracy of TKG forecasting.
- The model uses a likelihood-ratio correction based on historical interactions specific to actor-candidate pairs.
- A support-adaptive empirical-Bayes reader balances reliance on exact counts and learned estimations.
- The approach shows significant performance improvements over existing methods in multiple evaluation metrics.
Read more
BridgeMem: Causal Dyadic Transition Residuals for Temporal Knowledge Graph Forecasting
Summary
The paper introduces BRIDGEMEM, a novel approach for temporal knowledge graph (TKG) forecasting that addresses the limitations of existing methods in capturing dyadic transition evidence. Traditional forecasting techniques often summarize historical interactions through entity or relation states, which overlook the specific transition dynamics between a query actor and candidate entities. BRIDGEMEM estimates the impact of prior relations as a likelihood-ratio correction added to the scores of a frozen forecaster. It retrieves relevant past events for each actor-candidate pair, encoding their relations, directions, and recency to form a dyadic state. The model employs a support-adaptive empirical-Bayes reader that utilizes exact transition counts when available and resorts to a learned attention estimator in sparse scenarios. The uncertainty of the backbone model gates the correction, ensuring that confident predictions remain largely unchanged while ambiguous queries receive more significant adjustments. The authors demonstrate that BRIDGEMEM outperforms nine baseline methods across five benchmarks, highlighting the effectiveness of explicit dyadic transition modeling in TKG forecasting.
Methodology
BRIDGEMEM retrieves prior events for each actor-candidate pair and encodes their relations, directions, and recency into a dyadic state. It combines a support-adaptive empirical-Bayes reader that uses exact transition counts when abundant and a learned attention estimator when sparse. The model's uncertainty gates the correction, allowing for nuanced adjustments based on the confidence of predictions.
Results
BRIDGEMEM achieved improvements over the best prior results in all 20 filtered MRR and Hits@{1, 3, 10} comparisons across five benchmarks, with MRR gains of 0.0213, 0.0164, 0.0216, 0.0112, and 0.0028, demonstrating the effectiveness of its dyadic transition modeling.
Implications
The findings suggest that incorporating explicit dyadic transition modeling can significantly enhance the performance of TKG forecasting systems, which could have applications in various domains such as social network analysis, event prediction, and recommendation systems.
The Sequential Price of Continual Learning
Theory
- Distribution-level forgetting converges to population loss with an O(T −1) gap, extending existing theories to conflicting tasks.
- The limiting loss can be separated into unavoidable intrinsic loss from task conflicts and an additional sequential price from learning tasks sequentially.
- The sequential price is influenced by the strength of EWC, decreasing as 1/λ while increasing the number of tasks needed to reach stationarity.
- The theoretical framework is applied to real-world data, providing accurate predictions for the sequential price and its reduction through EWC.
Read more
The Sequential Price of Continual Learning
Summary
This paper investigates the performance costs associated with sequential task updates in continual learning, particularly focusing on the recency bias that can lead to distribution-level forgetting. The authors analyze an overparameterized linear-regression model with independent and identically distributed (i.i.d.) task sampling. They demonstrate that both distribution-level forgetting and population loss converge to the same stationary limit, which can be decomposed into an intrinsic loss due to task incompatibility and an additional 'sequential price' incurred from sequential updates. The study further explores the effects of fixed-strength elastic weight consolidation (EWC) on this sequential price, revealing that under strong regularization, the sequential price decreases inversely with EWC strength while the convergence to stationarity slows. The theoretical findings are validated through experiments on the Jester joke-rating dataset, where the authors quantify the sequential price arising from conflicting user preferences and its mitigation through EWC.
Methodology
The authors utilize a theoretical framework based on an overparameterized linear-regression model to analyze the effects of sequential learning on performance. They prove convergence results for distribution-level forgetting and population loss, and derive expressions for the sequential price in relation to EWC strength. Empirical validation is conducted using the Jester dataset to illustrate the practical implications of their theoretical findings.
Results
The study establishes that distribution-level forgetting and population loss converge to the same limit, which is composed of intrinsic loss and sequential price. In homogeneous task geometries, these two components are equal, leading to a total loss that is double that of joint training. The analysis of EWC shows that increasing regularization strength reduces the sequential price while requiring more tasks to achieve steady-state performance.
Implications
The findings highlight the importance of understanding the costs associated with sequential learning in continual learning systems. The insights into the sequential price and its dependence on EWC strength can inform the design of more effective continual learning algorithms, particularly in scenarios with conflicting tasks. This research could have applications in various domains where continual learning is critical, such as robotics, personalized recommendations, and adaptive AI systems.
SwitchPFN: Shared Switching Dynamics for Frozen In-Context Time Series Classification
Time Series
- SwitchPFN addresses the challenges of temporal order preservation and feature consistency in time series classification.
- The model uses a shared projection and regime codebook to enhance comparability of features across sequences.
- SwitchPFN achieves a mean test accuracy of 82.63% on eight UEA datasets, outperforming existing methods.
- Ablation studies and sensitivity analyses validate the effectiveness of the proposed representation design.
Read more
SwitchPFN: Shared Switching Dynamics for Frozen In-Context Time Series Classification
Summary
The paper introduces SwitchPFN, a novel approach for time series classification using tabular foundation models (TFMs). The authors identify two main challenges in existing time series representations: the loss of temporal order during global aggregation and the inconsistency of feature meanings across sequences when using independently fitted coordinate systems. SwitchPFN addresses these issues by learning a shared projection and regime codebook from training sequences, allowing for local dynamic operators and transition features to be comparable across samples. The framework captures local temporal transitions while maintaining a consistent feature definition, which is crucial for effective classification. The authors evaluate SwitchPFN on eight UEA datasets, demonstrating its superior performance with a mean test accuracy of 82.63%, outperforming the strongest baseline by 4.47%. Additionally, the paper includes ablation studies and sensitivity analyses to explore the contributions of various design choices and the model's behavior with limited labeled data.
Methodology
SwitchPFN employs a shared coordinate system to ensure that operator coefficients refer to the same directions across examples. It combines local dynamics, class residuals, ordered paths, and signal summaries into a fixed-width table for TabPFN. The model is trained on labeled sequences, mapping them to a common feature space, which allows for effective classification of new sequences.
Results
The results indicate that SwitchPFN achieves a mean test accuracy of 82.63% across eight UEA datasets, which is 3.54 percentage points higher than the second-best model. The paper also discusses the outcomes of various ablation studies and parameter sensitivity analyses, confirming the robustness of the representation design.
Implications
The findings suggest that SwitchPFN can significantly improve time series classification tasks in various fields, such as healthcare (e.g., ECG analysis) and environmental monitoring, by providing a more effective representation of temporal data. The model's ability to work well with limited labeled data also opens avenues for applications in scenarios where data collection is challenging.
Image Fidelity is Not Field Fidelity: Joint Thermodynamic Reconstruction and Error Localization in Neural Tomography
Computer Vision
- CoroNeRF optimizes 3D electron density and temperature fields directly from multiview spectral line intensities.
- Image fidelity does not equate to field fidelity; low 2D image error can correspond to poor recovery of 3D fields.
- Cross-seed instability can rank local physical-field errors, but shared model errors may lead to incorrect consensus.
- The framework highlights the challenges of recovering latent fields from indirect measurements in scientific tomography.
Read more
Image Fidelity is Not Field Fidelity: Joint Thermodynamic Reconstruction and Error Localization in Neural Tomography
Summary
This paper presents CoroNeRF, a novel framework for joint thermodynamic reconstruction of 3D electron density and temperature fields from multiview, multiline intensity observations in solar coronal tomography. The authors highlight a critical distinction between image fidelity and field fidelity, demonstrating that low error in 2D image reconstructions does not guarantee accurate recovery of the underlying 3D physical fields. The study employs a differentiable atomic-emission renderer to optimize the recovery process, addressing the challenges posed by the many-to-one nature of the forward map in tomographic imaging. The authors conduct experiments to evaluate the performance of CoroNeRF, revealing that limited-channel reconstructions can yield misleadingly high image fidelity while resulting in poor field fidelity. Additionally, they explore the concept of cross-seed instability as a means of localizing physical-field errors, providing insights into the limitations of seed-based error localization in the absence of ground truth. The findings underscore the necessity of considering both image and field fidelity in scientific tomographic imaging, particularly in contexts where direct supervision of the latent field is not feasible.
Methodology
The authors developed CoroNeRF, a differentiable multiline neural tomography framework that parameterizes a neural plasma field to jointly recover 3D electron density and temperature from multiview spectral line-intensity images. The framework uses a differentiable atomic-emission renderer to synthesize multichannel spectral line intensities from the inferred fields, optimizing them through image supervision.
Results
The experiments demonstrated that while limited-channel reconstructions can fit observations well, they often lead to significantly worse physical field recoveries. The study also confirmed that cross-seed instability can effectively rank local physical-field errors, although it has limitations when shared forward-model mismatches occur.
Implications
The findings suggest that in scientific tomographic imaging, particularly in astrophysics, it is crucial to evaluate both image fidelity and the accuracy of the reconstructed physical fields. This work could influence future approaches to tomographic reconstruction in various scientific fields where indirect measurements are prevalent.
BLADE: Distilled LLM Regularization for Calibrated Knowledge Graph Completion
Graph Learning
Large Language Models
NLP
- BLADE separates latent truth from graph recording to improve probability calibration in KGC.
- The model uses a distilled LLM as a frozen teacher to regularize KGC training.
- BLADE significantly reduces expected calibration error (ECE) compared to existing methods.
- Empirical results show that BLADE maintains competitive ranking performance while enhancing calibration metrics.
Read more
BLADE: Distilled LLM Regularization for Calibrated Knowledge Graph Completion
Summary
The paper introduces BLADE, a novel approach for knowledge graph completion (KGC) that aims to provide calibrated probabilities for inferred facts. Traditional KGC models focus on ranking candidates but often fail to deliver reliable probability estimates. BLADE addresses this by employing a variational model that separates latent truth from graph recording and utilizes a distilled language model (LLM) as a regularizer. The LLM is used during training to create a frozen teacher model that informs the KGC process without being present during inference. This allows BLADE to generate predictive probabilities and quantify epistemic uncertainty effectively. The authors evaluate BLADE across five benchmarks, demonstrating its competitive performance in ranking while significantly improving calibration metrics, such as expected calibration error (ECE). The results indicate that BLADE reduces adaptive ECE by 60.1% compared to deep ensembles and 78.1% compared to temperature-scaled RotatE. Furthermore, it outperforms other calibration methods on various metrics, including Brier score and negative log-likelihood (NLL). The paper concludes that while BLADE achieves improved calibration and maintains ranking quality, it also highlights the limitations of the model, particularly regarding the assumptions of distributional shifts and the potential leakage of knowledge from the LLM.
Methodology
BLADE employs a three-stage process: first, it uses offline language model labels to train a frozen plausibility teacher. During KGC training, a relational encoder parameterizes Gaussian embeddings, and a probabilistic decoder maps samples to latent truth probabilities. The teacher provides a bounded energy regularizer. At inference, the language model is not used, and posterior samples yield ranking scores and calibrated probabilities.
Results
BLADE achieved a macro-average reduction in adaptive ECE of 60.1% relative to deep ensembles and 78.1% relative to temperature-scaled RotatE. It also improved ECE, Brier score, and NLL over validation-selected histogram binning and a matched generative ComplEx2 model. The full triage score reached a mean AUC-PR of 0.863, outperforming its strongest non-teacher variant, which scored 0.805.
Implications
The findings suggest that BLADE can be effectively applied in scenarios requiring reliable probability estimates for knowledge graph completion, such as in recommendation systems, automated reasoning, and decision-making processes that rely on uncertain information.
SPADE-DFL: Communication-Efficient Decentralized Federated Learning via Derivative-Free Linearized ADMM
Federated Learning
Optimization
Theory
- SPADE-DFL reduces communication costs in decentralized federated learning by allowing multiple local updates before exchanging information.
- The method achieves a time-averaged stationarity and consensus bound of O(T^{-1/3}) with Θ(T^{2/3}) communication rounds.
- Differential privacy is ensured through careful isolation of local updates and the use of Gaussian noise.
- The approach incorporates a component-memory single-point estimator to enhance local training dynamics.
Read more
SPADE-DFL: Communication-Efficient Decentralized Federated Learning via Derivative-Free Linearized ADMM
Summary
This paper presents SPADE-DFL, a novel approach to decentralized federated learning (DFL) that aims to reduce communication costs while maintaining convergence performance. The authors introduce a primal-dual method that allows clients to perform multiple local updates before exchanging model information, thereby addressing the challenge of communication efficiency in scenarios where gradients are unavailable. The proposed method leverages randomized function evaluations and incorporates a component-memory single-point estimator to enhance local training dynamics. The authors establish theoretical bounds on stationarity and consensus, demonstrating that their approach achieves a time-averaged stationarity and consensus bound of O(T^{-1/3}) with Θ(T^{2/3}) communication rounds, where T is the number of local updates per client. Additionally, the paper addresses privacy concerns by isolating the accumulated data-dependent increments from the graph correction, ensuring client-level differential privacy while quantifying the optimization error. Experimental results on four classification tasks indicate that SPADE-DFL outperforms existing decentralized learning methods in terms of mean test accuracy, showcasing its effectiveness in practical applications.
Methodology
The authors develop a primal-dual method based on linearized ADMM that allows for multiple local updates using randomized function evaluations. They establish a communication-efficient schedule that balances local training length and exchange frequency while ensuring convergence. The method also incorporates a component-memory single-point estimator to improve local model accuracy and utilizes Gaussian noise for privacy protection.
Results
SPADE-DFL achieves a time-averaged stationarity and consensus bound of O(T^{-1/3}) with Θ(T^{2/3}) communication rounds. Experimental evaluations on four classification tasks show that SPADE-DFL consistently outperforms existing decentralized learning methods in terms of mean test accuracy.
Implications
The findings suggest that SPADE-DFL can significantly enhance the efficiency of decentralized federated learning systems, making it suitable for applications where communication bandwidth is limited or privacy is a concern. This approach could be particularly beneficial in scenarios such as mobile device learning and IoT applications.
Time-Series Foundation Models That Understand Data Revisions
Time Series
- Introduction of VINTAGE-TS, a revision-aware time-series forecasting model.
- Differentiation between observation time and information-availability time.
- Joint predictive distribution to model uncertainty in forecasts.
- Comprehensive evaluation framework with rolling evaluations and baseline comparisons.
Read more
Time-Series Foundation Models That Understand Data Revisions
Summary
This paper introduces VINTAGE-TS, a time-series foundation model designed to account for data revisions that occur after initial publication. Traditional forecasting models often assume that historical data is fixed, which can lead to inaccuracies when older observations are revised. VINTAGE-TS differentiates between observation time and information-availability time, allowing it to predict both the first-published value of the next period and its revised value after a specified maturation interval. The model employs a joint predictive distribution to capture the uncertainty associated with these targets. The paper outlines a comprehensive evaluation framework using ALFRED for rolling evaluations and compares VINTAGE-TS against conventional and revision-aware baselines. It also includes a software package that facilitates validity-interval reconstruction and delayed-label filtering. The authors conduct a synthetic demonstration and a sensitivity analysis to explore the impact of different configurations and revision regimes on model performance. Although the empirical advantages of the model are yet to be established, the research aims to provide a robust methodology for forecasting in the presence of data revisions.
Methodology
The methodology involves the development of VINTAGE-TS, which integrates explicit timing of information availability and separate release targets into a joint uncertainty model. The evaluation framework includes rolling evaluations using ALFRED, comparisons with conventional and revision-aware baselines, and a series of automated tests to ensure the integrity of the model's predictions.
Results
The results from synthetic demonstrations and sensitivity analyses indicate that the proposed model can effectively handle the complexities introduced by data revisions. However, the paper does not claim empirical advantages over existing models until further real-world experiments are conducted.
Implications
The findings suggest that incorporating revision histories into forecasting models could enhance the accuracy and calibration of predictions, particularly in economic and financial contexts where data revisions are common. This approach may lead to more reliable decision-making based on time-series forecasts.
BranchShine-CR: Compact Multilingual IPA Transcription with Self-Conditioned CTC and Consistency Regularization
Audio & Speech
Efficient ML
NLP
- BranchShine-CR achieves a 4.47% IPA character error rate on multilingual transcription tasks.
- The model is significantly smaller (25M parameters) compared to previous models while maintaining high accuracy.
- Ablation studies confirm the effectiveness of the model's components, including self-conditioned CTC and consistency regularization.
- The model outperforms existing baselines across multiple languages, demonstrating its robustness.
Read more
BranchShine-CR: Compact Multilingual IPA Transcription with Self-Conditioned CTC and Consistency Regularization
Summary
The paper presents BranchShine-CR, a novel 25M-parameter model designed for multilingual transcription into the International Phonetic Alphabet (IPA). This model integrates log-mel features with a rotary-position E-Branchformer encoder, employs intermediate self-conditioned connectionist temporal classification (CTC), and utilizes consistency regularization across augmented views. Evaluated on 16,646 IPApack++ test utterances, BranchShine-CR achieves a 4.47% IPA character error rate, representing a 22.3% relative improvement over the previous ZIPA-CTC-NS model, while using approximately one-twelfth the number of parameters and being trained from scratch. The model also surpasses a similarly sized NeMo Conformer baseline across all 41 dataset language labels. Ablation studies reveal that the individual components of the model work synergistically to enhance performance, indicating that BranchShine-CR is capable of effective IPA recognition even with limited computational resources, making it suitable for applications in low-resource on-device pronunciation assessment.
Methodology
The methodology involves using log-mel features as input to a rotary-position E-Branchformer encoder. The model incorporates intermediate self-conditioned CTC for enhanced training supervision and applies consistency regularization to improve model robustness. The training objective combines CTC losses from two independently masked views of the same audio input, with a consistency loss based on Kullback-Leibler divergence to encourage agreement between predictions from both views.
Results
BranchShine-CR achieved a 4.47% IPA character error rate on the IPApack++ test set, marking a 22.3% relative improvement over the ZIPA-CTC-NS model. It also outperformed a similarly sized NeMo Conformer baseline across all 41 language labels, demonstrating its effectiveness in multilingual IPA transcription.
Implications
The findings suggest that compact models like BranchShine-CR can effectively perform multilingual IPA transcription, making them suitable for applications in language documentation, pronunciation analysis, and cross-lingual speech research, particularly in low-resource settings.
Not All Synthetic Data Are Equal: Expert-Committee Audit Screening for Imbalanced Crash-Injury-Severity Prediction in Automated Driving Systems
Robotics
- Introduction of Expert-Committee Audit Screening (ECAS) for evaluating synthetic data credibility.
- Demonstrated improvement in prediction accuracy for crash injury severity using ECAS.
- Local neighborhood analysis confirms the reliability of ECAS-accepted samples.
- Shift in focus from quantity-oriented data augmentation to credibility-aware approaches.
Read more
Not All Synthetic Data Are Equal: Expert-Committee Audit Screening for Imbalanced Crash-Injury-Severity Prediction in Automated Driving Systems
Summary
This paper addresses the challenge of predicting crash injury severity in automated driving systems (ADS) due to the imbalanced nature of crash data, particularly the scarcity of severe injury reports. The authors propose a novel framework called Expert-Committee Audit Screening (ECAS) that evaluates the credibility of synthetic minority samples generated to augment the dataset. The ECAS framework employs an expert committee to assess samples based on criteria such as label support and local plausibility, ensuring that only credible samples are accepted for training. The study utilizes a dataset of 1,477 ADS-related crashes from the National Highway Traffic Safety Administration. The authors demonstrate that the ECAS framework significantly improves the performance of a Tabular Prior-data Fitted Network (TabPFN) classifier, achieving higher balanced accuracy and macro-F1 scores compared to traditional augmentation methods. Additionally, local neighborhood analysis reveals that ECAS-accepted samples are more representative of real minority crashes. The findings underscore the importance of credibility in synthetic data generation for safety-critical applications, advocating for a shift from quantity-focused to credibility-aware data augmentation strategies in ADS safety prediction.
Methodology
The study employs the ECAS framework to audit synthetic minority samples generated for imbalanced crash data. The framework uses criteria such as label support, boundary separation, and committee agreement to evaluate sample credibility. The authors utilize a combination of normalizing flow augmentation and a TabPFN classifier to assess the performance of accepted samples.
Results
The best configuration of ECAS achieved the highest balanced accuracy, macro-F1 score, and minor-injury recall. Local neighborhood analysis indicated that ECAS-accepted samples were better aligned with real minority crash data compared to unscreened samples. The fitted model's analysis revealed that lower injury severity classes were influenced by crash counterpart and pre-crash movement, while moderate-plus injuries were more sensitive to posted speed limits and operating context.
Implications
The findings suggest that adopting credibility-aware data augmentation methods can enhance the predictive performance of models used in safety-critical applications like ADS. This approach may lead to better risk governance and improved public trust in automated driving technologies.
Evaluating Cross-region Generalization for Wavelet-Diffusion Precipitation Downscaling
Generative Models
Time Series
Theory
- The study evaluates the generalization of a Wavelet-Diffusion Model for precipitation downscaling across different U.S. regions.
- A WDM trained on multiple regions outperforms one trained on a single region, though improvements vary by precipitation intensity.
- Spatial organization of precipitation fields significantly influences model performance, more so than regional differences.
- The findings support the feasibility of transferring downscaling models to regions with limited local training data.
Read more
Evaluating Cross-region Generalization for Wavelet-Diffusion Precipitation Downscaling
Summary
This study investigates the cross-region and cross-event generalization capabilities of a Wavelet-Diffusion Model (WDM) for precipitation downscaling, addressing the challenge of applying machine learning models trained in one geographic region to others. The authors utilize a dataset comprising six distinct U.S. regions characterized by various precipitation regimes, including convective, winter, tropical, and atmospheric-river events. They compare the performance of a WDM trained solely on Oklahoma data against a WDM trained on all six regions, as well as traditional nearest-neighbor and bicubic interpolation methods. The evaluation employs three families of metrics to assess image-domain reconstruction, spectral fidelity, and bin-wise precipitation detection. Results indicate that while the Oklahoma-trained WDM performs competitively outside its training region, the all-region WDM achieves superior overall performance, albeit with uneven improvements across different precipitation intensities. Notably, the study finds a strong correlation between the spatial organization of precipitation fields and model performance, suggesting that spatial characteristics are more influential than regional or event-specific differences in precipitation downscaling tasks. These findings underscore the potential for transferring downscaling models to regions with limited training data, paving the way for the development of globally consistent high-resolution precipitation products.
Methodology
The authors utilized a dataset from NOAA Multi-Radar/Multi-Sensor (MRMS) composite reflectivity fields, generating low-resolution inputs through block averaging. They compared the performance of WDMs trained on Oklahoma data and all six regions against traditional interpolation methods, using a variety of metrics to evaluate model performance across different precipitation regimes.
Results
The study found that the WDM trained on Oklahoma data remained competitive when applied to other regions. The all-region WDM provided the best overall performance, particularly in image-domain reconstruction and precipitation detection, although performance gains were not uniform across different precipitation intensities. A strong correlation was observed between the spatial organization of precipitation and model performance, indicating that spatial characteristics play a crucial role in downscaling tasks.
Implications
The results suggest that machine learning models for precipitation downscaling can be effectively transferred to regions lacking extensive local training data. This has significant implications for generating high-resolution precipitation products globally, which are essential for hydrological modeling and flood risk assessment.
FlowAtom: Atom-Based Evidence Aggregation for Multi-Label Website Fingerprinting
Theory
- FlowAtom constructs shared Atoms from flow representations without website labels.
- The method aggregates evidence across flows to predict monitored website sets.
- Achieves high micro-F1 scores in both closed-world and open-world evaluations.
- Introduces a flow-level perspective for multi-label website fingerprinting.
Read more
FlowAtom: Atom-Based Evidence Aggregation for Multi-Label Website Fingerprinting
Summary
The paper presents FlowAtom, a novel approach to multi-label website fingerprinting that addresses the challenge of identifying monitored websites in mixed encrypted traffic. Traditional methods often rely on complete traffic traces, which are not feasible in realistic scenarios where multiple websites are accessed simultaneously. FlowAtom constructs shared prototypes, termed Atoms, from flow representations without requiring website labels. It employs a flow encoder pretrained on external unlabeled traffic to aggregate Atom responses across flows within an observation window, resulting in a fixed-dimensional, permutation-invariant representation for predicting the set of monitored websites. The authors evaluate FlowAtom in various settings, achieving high micro-F1 scores, thus demonstrating its effectiveness in both closed-world and open-world evaluations. This work contributes to the field by formulating a flow-level perspective for multi-label website fingerprinting and providing a benchmark for future research.
Methodology
FlowAtom consists of three main stages: flow representation pretraining, Atom construction, and window-level multi-label prediction. It uses contrastive learning for pretraining flow representations from unlabeled traffic, constructs Atoms through clustering, and aggregates Atom responses across flows to create a fixed-dimensional representation for prediction.
Results
FlowAtom achieves micro-F1 scores of 97.82%, 94.43%, and 93.92% in closed-world evaluations for Direct HTTPS, Trojan, and VMess, respectively. In open-world evaluations, it maintains scores of 92.37%, 92.64%, and 89.85% for the same settings, outperforming baseline methods.
Implications
The findings suggest that FlowAtom can significantly enhance the accuracy of website fingerprinting in encrypted traffic scenarios, which has implications for cybersecurity, privacy protection, and traffic analysis. The methodology can also be adapted for other multi-label classification tasks in different domains.
Monitoring Urban Traffic Dynamics at Fine Spatiotemporal Resolution Using Distributed Acoustic Sensing and Deep Learning
Time Series
Audio & Speech
- DAS provides a novel method for real-time urban traffic monitoring using existing fiber-optic infrastructure.
- The integration of deep learning enhances the ability to detect vehicle trajectories and infer traffic states from DAS data.
- Distinct traffic patterns are observed during different urban events, highlighting the system's sensitivity to dynamic conditions.
- The methodology supports privacy-preserving monitoring, addressing concerns associated with traditional traffic sensing methods.
Read more
Monitoring Urban Traffic Dynamics at Fine Spatiotemporal Resolution Using Distributed Acoustic Sensing and Deep Learning
Summary
This study explores the integration of Distributed Acoustic Sensing (DAS) and deep learning to monitor urban traffic dynamics at high spatiotemporal resolution. DAS utilizes existing underground fiber-optic cables as continuous sensor arrays, allowing for real-time traffic monitoring with meter-level spatial and second-level temporal resolution. The authors developed a deep learning framework that processes raw ground vibration data to detect vehicle trajectories and infer traffic states, such as volume and congestion. A hybrid training strategy combining synthetic and manually annotated data was employed to enhance vehicle detection in noisy conditions. The results indicate that urban events, such as football game days and holidays, produce distinct traffic patterns, showcasing the effectiveness of DAS in capturing nuanced traffic dynamics. This approach offers a promising solution for real-time, privacy-preserving urban traffic monitoring, particularly in scenarios where traditional data sources are limited or disrupted.
Methodology
The authors deployed DAS along a roadway network and developed a deep learning framework to convert raw vibration waveforms into spatiotemporal representations. A hybrid training strategy was utilized, combining synthetic and manually annotated images to improve vehicle detection under various conditions.
Results
The study found that different urban events produced unique traffic-state signatures, with home football game days leading to localized congestion and the Thanksgiving holiday causing sustained traffic suppression. The deep learning model effectively characterized system-level traffic dynamics from DAS data.
Implications
The findings suggest that DAS combined with deep learning can significantly enhance urban traffic monitoring capabilities, providing a scalable, efficient, and privacy-preserving alternative to traditional methods. This approach could be particularly beneficial in emergency situations or areas with limited conventional traffic data.
Stable and Faithful Explanations for Knowledge Tracing
Interpretability
- Introduces a validation protocol for assessing KT model explanations.
- Demonstrates that XGBoost can achieve competitive predictive accuracy compared to deep learning models.
- Establishes the stability of feature importance rankings across different conditions.
- Shows that removing top-ranked features degrades model performance more than random feature removal, indicating faithfulness of explanations.
Read more
Stable and Faithful Explanations for Knowledge Tracing
Summary
This paper addresses the challenge of providing interpretable explanations for Knowledge Tracing (KT) models, which predict student performance but often lack transparency. The authors propose a validation protocol that assesses predictive accuracy, explanation stability, and faithfulness of explanations. They engineered thirteen behavioral features from the ASSISTments datasets and rebuilt the 2009 dataset to mitigate data leakage issues. An Extreme Gradient Boosting (XGBoost) model was utilized and explained using Tree SHapley Additive exPlanations (TreeSHAP), and its performance was compared against four deep learning baselines (DKT, SAKT, AKT, and SimpleKT) under an information-matched protocol. The study found that while XGBoost achieved competitive AUC scores, the differences in performance were attributed to the information supplied rather than the model architecture. The stability of feature importance rankings was confirmed across various conditions, and the removal of top-ranked features significantly impacted model performance, indicating the faithfulness of the explanations provided by TreeSHAP. The findings highlight the importance of transparent and interpretable models in educational contexts, enabling instructors to make informed pedagogical decisions.
Methodology
The authors engineered behavioral features from ASSISTments datasets and rebuilt the 2009 dataset to address data leakage. They employed an Extreme Gradient Boosting (XGBoost) model explained with TreeSHAP and compared its performance against deep learning baselines under an information-matched protocol. The study included validation of predictive accuracy, explanation stability, and faithfulness through systematic testing.
Results
XGBoost achieved AUC scores of 0.777 on the 2012 dataset and 0.786 on the rebuilt 2009 dataset. The model's performance was comparable to deep learning baselines when restricted to the same information. The stability of feature rankings was confirmed with high Spearman correlation coefficients, and the removal of top-ranked features significantly affected model performance, supporting the faithfulness of TreeSHAP explanations.
Implications
The findings suggest that interpretable KT models can enhance the ability of educators to understand and act on student performance predictions. This can lead to more effective instructional strategies and improved learning outcomes. The validation protocol established in this study can serve as a framework for future research in educational data mining and explainable AI.
Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning
Reinforcement Learning
Robotics
Theory
- Introduction of TRACE, a novel attack framework for reconstructing trajectories in embodied RL.
- Utilization of temporal correlations in gradients to enhance reconstruction accuracy.
- Demonstrated high fidelity in reconstructing observations and actions from gradient streams.
- Evaluation across various architectures and environments, showing robustness and adaptability.
Read more
Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning
Summary
This paper addresses the privacy risks associated with gradient inversion attacks in distributed embodied reinforcement learning (RL) settings, where agents share policy gradients while keeping raw sensor data on-device. The authors introduce TRACE (Temporal Reconstruction Attack on Consecutive Encodings), a novel amortized temporal gradient-inversion attack that reconstructs sequences of private observation-action trajectories from per-step policy-learning gradients. TRACE leverages two key structural signals: the cross-time correlation between successive gradients and the closed-form recovery of actions from policy-head gradient structures. The method employs a gradient encoder, a causal transformer for temporal context propagation, and an autoregressive decoder for joint reconstruction of observations and actions. The experimental results demonstrate that TRACE achieves a peak signal-to-noise ratio (PSNR) of 18.8 dB with near-perfect action recovery, outperforming existing learning-based and optimization-based baselines in terms of reconstruction fidelity and speed. The findings suggest that protecting temporal gradient streams may necessitate sequence-aware privacy mechanisms.
Methodology
TRACE employs an autoregressive model consisting of a gradient encoder that maps gradients to compact latents, a causal transformer that maintains temporal context, and a decoder that reconstructs observations and predicts actions. The method is designed to exploit the temporal coherence in the gradient streams of embodied RL agents, allowing for effective trajectory-level inversion.
Results
TRACE achieved an 18.8 dB PSNR in reconstruction fidelity and demonstrated near-perfect action recovery at a speed of 3–4.5 ms per frame. It outperformed both optimization-based and learning-based baselines across all reconstruction metrics, showing strong zero-shot reconstruction capabilities on unseen environments and rapid adaptation with minimal target-domain data.
Implications
The findings highlight significant privacy vulnerabilities in embodied RL systems and suggest that existing defenses may be inadequate. The need for sequence-aware privacy mechanisms is emphasized, indicating potential avenues for future research in secure distributed learning.
Vector Bellman Theory for Multichain Robust Average-Reward Markov Decision Processes
Reinforcement Learning
Optimization
Theory
- Introduces a vector Bellman theory for robust average-reward MDPs that addresses state-dependent rewards and transition uncertainty.
- Develops a gain-first, bias-second optimization principle leading to a coupled vector gain-bias system.
- Characterizes solvability conditions and sufficient criteria for distinct recurrent-class gains.
- Proposes a robust approximately shifted Halpern planning algorithm for effective undiscounted planning.
Read more
Vector Bellman Theory for Multichain Robust Average-Reward Markov Decision Processes
Summary
This paper presents a novel vector Bellman theory tailored for robust average-reward Markov decision processes (MDPs) that account for uncertainty in transition dynamics and state-dependent rewards. The authors introduce a gain-first, bias-second optimization principle that leads to a coupled vector gain-bias system, allowing for the identification of optimal robust gains and stationary saddle strategies applicable from all initial states. The theory addresses the challenges posed by recurrent-class rewards and transition uncertainty, providing a framework that connects finite Bellman certificates to undiscounted planning. The authors characterize conditions for solvability and establish sufficient criteria for distinct recurrent-class gains. Additionally, they propose a robust approximately shifted Halpern planning algorithm that ensures convergence of gain estimates and Bellman displacements to the optimal gain vector, enabling average-optimal greedy controllers. This work significantly enhances the understanding of robust average-reward MDPs and offers practical methods for planning under uncertainty.
Methodology
The authors formulate a coupled Bellman system prioritizing continuation gain for both nature and the controller. They analyze conditions for optimal control, explore regimes allowing distinct recurrent-class gains, and adapt the Halpern iteration for robust Bellman operators to ensure convergence and optimality in planning.
Results
The paper demonstrates that the proposed vector Bellman theory provides a robust framework for average-reward MDPs, yielding optimal gain vectors and average-optimal controllers from all initial states. The results indicate that finite Bellman certificates can be established under certain conditions, leading to effective planning strategies.
Implications
This research has significant implications for decision-making in uncertain environments, particularly in fields like robotics, finance, and operations research, where robust performance under varying conditions is crucial. The findings can enhance the design of algorithms for real-world applications that require long-term performance optimization.
Beyond Static Graph World Models: Learning Stochastic Latent Dynamics over Evolving Topologies
Graph Learning
Reinforcement Learning
Robotics
- Introduction of the Graph Dynamics Model (GDM) for evolving topologies in graph-based environments.
- Development of the Graph Distribution Distance (GDD) metric for evaluating graph-based world models.
- Demonstration of GDM's superior performance in stochastic and partially observable settings.
- Evidence of zero-shot generalization capabilities to larger graphs.
Read more
Beyond Static Graph World Models: Learning Stochastic Latent Dynamics over Evolving Topologies
Summary
This paper introduces the Graph Dynamics Model (GDM), a novel approach to graph-based world modeling that addresses the limitations of existing models which typically operate under fixed topologies or deterministic environments. The GDM is designed to handle evolving topologies in stochastic and partially observable settings, utilizing a sparse recurrent adjacency matrix for topology updates and a recurrent state-space architecture for stochastic transition modeling. A significant contribution of this work is the introduction of the Graph Distribution Distance (GDD) metric, which allows for a comprehensive evaluation of graph-based world models by comparing predicted and true distributions over joint graph states. The authors demonstrate the effectiveness of the GDM across various environments, showcasing its ability to generalize to larger, unseen graphs and outperform baseline models in multi-step rollouts. The study highlights the importance of accurately modeling topology and message passing in dynamic graph environments, providing a foundation for future research in graph-based reinforcement learning and planning.
Methodology
The GDM employs a sparse recurrent adjacency matrix to model topology updates and a recurrent state-space architecture to predict stochastic transitions. The GDD metric is based on maximum mean discrepancy (MMD) with a graph kernel, allowing for a comprehensive comparison of joint next-state distributions in graph environments.
Results
The GDM outperformed state-of-the-art baseline models in various environments, demonstrating effective generalization to larger graphs not encountered during training. The GDD metric provided a principled evaluation of model predictions, capturing the joint distribution of graph topology, node features, and graph features.
Implications
The findings suggest that GDM can significantly enhance the modeling of dynamic environments in reinforcement learning and planning applications, particularly in scenarios involving complex relational structures. The introduction of the GDD metric also sets a new standard for evaluating graph-based models, promoting further research in this area.
ICE: Task-Aligned Clifford Latent Fields for Multimodal Graph Foundation Models
Graph Learning
Multimodal
- ICE introduces a novel multimodal graph foundation model that preserves entity semantics and constructs higher-order relations.
- The model utilizes a node-indexed Clifford latent field to effectively manage multimodal data and interactions.
- ICE achieves state-of-the-art performance across multiple node classification and link prediction datasets.
- The study emphasizes the significance of task-specific access and structured representations in multimodal graph learning.
Read more
ICE: Task-Aligned Clifford Latent Fields for Multimodal Graph Foundation Models
Summary
This paper introduces ICE (Interaction-aware Clifford Encoder), a novel multimodal graph foundation model designed to effectively learn representations across multimodal attributed graphs (MAGs). The authors argue that traditional methods, which typically fuse node representations into a single Euclidean vector, fail to adequately capture the complex interactions and semantics inherent in MAGs. ICE addresses this by utilizing a node-indexed Clifford latent field that incorporates topology, text, and image data into explicit Cl(3) addresses. The model employs edge-aware geometric products to transform these inputs into various relational forms, ensuring that entity semantics are preserved while enabling the construction of higher-order relations. The paper presents a comprehensive empirical evaluation across multiple datasets, demonstrating that ICE outperforms existing methods in both node classification and link prediction tasks. The findings highlight the importance of maintaining a structured representation that allows for task-specific access and interaction construction, ultimately leading to superior performance in multimodal graph learning tasks.
Methodology
ICE employs a node-indexed Clifford latent field to encode multimodal data (topology, text, images) into explicit addresses. It utilizes edge-aware geometric products to transform these inputs into scalar, bivector, and trivector relations, preserving entity semantics while allowing for higher-order relational construction. The model features fresh supervised and few-shot heads that query the latent field, ensuring task alignment and effective representation learning.
Results
ICE ranks first in all evaluated tasks, including six node classification and three link prediction datasets, demonstrating superior performance in both supervised and few-shot settings. The results indicate that component removals negatively impact task performance, linking improvements to the model's geometric transport capabilities and the retention of semantic information.
Implications
The findings suggest that ICE can be applied to various domains requiring multimodal graph learning, such as recommendation systems, social network analysis, and knowledge graph construction. The model's ability to maintain semantic integrity while facilitating complex interactions could enhance the performance of applications in these areas.
UO-FIE: Combining Exact-Label Supervision with Graded Utility for Factivity Inference
NLP
Large Language Models
Efficient ML
- Introduces UO-FIE, a parameter-efficient system for factivity inference that combines exact-label supervision with graded utility.
- Identifies two empirical patterns: majority-label collapse and regression toward nearby but inexact intervals.
- Utilizes a score-matrix-guided categorical learning approach to improve prediction accuracy while accounting for class imbalance.
- Achieves top rankings in both fine-tuning and non-fine-tuning tracks of the FIE2026 evaluation.
Read more
UO-FIE: Combining Exact-Label Supervision with Graded Utility for Factivity Inference
Summary
This paper presents UO-FIE, a novel approach for factivity inference in Chinese context-hypothesis pairs, addressing the challenges posed by the Factivity Inference Evaluation 2026 (FIE2026). The task involves classifying pairs into nine ordered factivity intervals, where both exact predictions and proximity to the correct interval are rewarded. The dataset is imbalanced, with a significant majority of examples belonging to a single class. Preliminary experiments revealed that existing models, such as mDeBERTa, tended to predict the dominant class, while a Huber-regression baseline improved proximity predictions but reduced exact matches. UO-FIE combines exact-label supervision with a utility-based approach, predicting a distribution over the nine classes and utilizing hard-label supervision, utility-based soft targets, scheduled class weights, and an ordinal loss. The model is based on Qwen3.5-9B with LoRA and achieved first place in the fine-tuning track with a macro utility of 0.8316, while a prompt-based ensemble ranked third in the non-fine-tuning track with a macro utility of 0.8450. The paper contributes to the understanding of empirical patterns in classification tasks and introduces a score-matrix-guided categorical learning approach that retains exact-label supervision while considering neighborhood utility and ordinal structure.
Methodology
UO-FIE employs a parameter-efficient architecture that predicts a distribution over nine factivity classes. It integrates hard-label supervision with utility-based soft targets and uses scheduled class weights alongside an ordinal loss function. The model leverages Qwen3.5-9B with LoRA for fine-tuning and applies ordinal calibration for decision-making.
Results
UO-FIE achieved a macro utility score of 0.8316, ranking first in the fine-tuning track of FIE2026. Additionally, a separate prompt-based ensemble ranked third in the non-fine-tuning track with a macro utility of 0.8450, showcasing the model's effectiveness in handling the task's challenges.
Implications
The findings suggest that combining exact-label supervision with graded utility can significantly enhance performance in classification tasks with imbalanced datasets. This approach may be applicable to other NLP tasks requiring nuanced predictions and could inform future research on utility-aligned learning.
Auditability Is Not One Property: Rule Overlap, Behavioural Agreement, and Composition in Reinforcement Learning
Reinforcement Learning
- Auditability in RL is decomposed into six independently testable predicates.
- High rule coverage does not guarantee behavioral agreement among policies.
- Fused policies are limited by the quality of the rule bank rather than the knowledge contained within it.
- The paper presents a failure-driven design history to enhance transparency in the evaluation process.
Read more
Auditability Is Not One Property: Rule Overlap, Behavioural Agreement, and Composition in Reinforcement Learning
Summary
This paper investigates the auditability of reinforcement learning (RL) policies, which are often opaque and lack clear explanations of their learned behaviors. The author proposes a decomposition of auditability into six distinct predicates: trace integrity, lossless coding, rule coverage, behavioral agreement, composition quality, and value-model reliability. Each predicate is tested independently, revealing that passing one does not imply passing another. The study introduces a protocol involving agents that share a symbolizer, a passive observer that logs transitions into condition-action rules, and a ledger that certifies the integrity of the recorded traces. The findings indicate that while rule sets may overlap across independently trained policies, this does not guarantee behavioral agreement, which is a significant negative result. Additionally, the paper discusses the limitations of rule-level fusion compared to value-based composition, emphasizing that the correctness of fused policies is constrained by the quality of the underlying rule bank. The research aims to provide a clearer understanding of how to create auditable and composable RL policies, highlighting the challenges and limitations in achieving this goal.
Methodology
The study employs a protocol where agents share a frozen symbolizer, and a passive observer logs transitions into confidence-scored condition-action rules. An append-only ledger certifies the integrity of the recorded traces, and rules from multiple seeds are fused using a confidence-ranked arbitration policy.
Results
The research finds that symbolic overlap does not lead to behavioral agreement, with independently trained policies showing agreement rates indistinguishable from chance. The correctness of fused policies is limited by the quality of the rule bank, and the apparent failure of the protocol in certain tasks is attributed to mismatches in evaluation methods rather than task limitations.
Implications
The findings suggest that achieving auditability in RL requires careful consideration of the distinct properties involved. This work has implications for the development of more interpretable and composable RL systems, which are crucial for applications in safety-critical domains.
WeatherDiagFlow: Evidence-Grounded Radar Nowcasting with Diagnostic Flow Refinement
Time Series
- Introduces WeatherDiagFlow (WDF) for radar nowcasting, integrating forecasting, bulletin generation, and auditing.
- Formulates radar nowcasting as an evidence-grounded task with a clear boundary between forecast and verification.
- Employs diagnostic-conditioned rolling flow refinement to improve prediction accuracy and preserve strong echoes.
- Demonstrates competitive performance on FJRADAR, particularly in strong-echo event skill.
Read more
WeatherDiagFlow: Evidence-Grounded Radar Nowcasting with Diagnostic Flow Refinement
Summary
The paper presents WeatherDiagFlow (WDF), a novel framework for radar nowcasting that addresses the limitations of conventional systems by integrating evidence-grounded forecasting, operational bulletins, and post-event audits. Traditional radar nowcasting primarily focuses on predicting future radar fields, which often leads to the loss of critical strong echoes and lacks structured evidence for operational communication. WDF reformulates radar nowcasting as an evidence-grounded forecast-bulletin-audit task, ensuring a clear boundary between forecast-time guidance and post-event verification. The framework predicts motion, growth and decay, heavy-echo risk, and uncertainty, which condition rolling flow refinement. Additionally, a frozen-scaffold residual calibration technique is employed to enhance the preservation of strong echoes. The system utilizes a multi-agent layer to convert structured evidence into operational bulletins while independently generating verification audits. Experiments conducted on the FJRADAR dataset demonstrate that WDF achieves competitive performance, particularly in preserving strong-echo events, and provides consistent evidence-based bulletin generation and auditing.
Methodology
WDF utilizes a coarse forecaster to generate a deterministic scaffold for future radar fields, followed by a diagnostic encoder that predicts motion, growth/decay, heavy-echo risk, and uncertainty. The framework incorporates rolling flow refinement to enhance predictions while maintaining a strict boundary between forecast-time bulletins and post-event audits. The frozen-scaffold residual calibration is applied to preserve strong echoes over longer lead times.
Results
The experiments on the FJRADAR dataset indicate that WeatherDiagFlow achieves competitive overall performance and significantly improves the skill in predicting strong-echo events. The structured evidence generated by WDF allows for reliable operational bulletins and effective post-event audits.
Implications
The proposed framework has significant implications for improving short-term weather forecasting, enhancing operational communication during severe weather events, and providing a more accountable verification process for radar nowcasting systems.
Learnable Time-Frequency Masks for Explaining Time-Series Classifiers
Time Series
Interpretability
- XACT provides a transform-agnostic framework for generating explanations in time-frequency representations.
- The method incorporates a time-frequency objective that enhances the sparsity and smoothness of the learned masks.
- XACT extends the application of Layer-wise Relevance Propagation (LRP) to wavelet transforms, improving interpretability.
- The framework demonstrates improved robustness against spurious features compared to existing attribution methods.
Read more
Learnable Time-Frequency Masks for Explaining Time-Series Classifiers
Summary
This paper addresses the challenge of explainability in time-series classifiers, where critical information is often encoded in latent frequency or time-frequency features rather than in the raw signal. The authors propose a novel framework called eXplainable Attributions from Coefficients of Transforms (XACT), which learns sparse attribution masks over coefficients derived from arbitrary invertible time-frequency transforms. The framework is evaluated using three different transforms: the Short-Time Discrete Fourier Transform (STDFT), the Continuous Wavelet Transform (CWT), and the Discrete Wavelet Transform (DWT). XACT extends the virtual inspection layer approach to these transforms, enabling Layer-wise Relevance Propagation (LRP) to generate explanations in the time-frequency domain. The results indicate that XACT produces precise and structured explanations on both synthetic and real-world datasets, outperforming existing methods in terms of robustness to spurious features, although no single method excels across all evaluation metrics. This work highlights the potential of learning explanations directly in time-frequency representations for better interpretability of deep learning models applied to time series data.
Methodology
The authors developed the XACT framework, which learns sparse attribution masks over coefficients from various time-frequency transforms. The methodology includes extending the stationary smoothness term to two dimensions and incorporating an energy-aware penalty to maintain relevance in meaningful regions of the explanation domain. The framework was tested on synthetic and real-world datasets, utilizing LRP for generating explanations in the time-frequency domain.
Results
XACT produced precise explanations on a synthetic dataset and demonstrated less susceptibility to highlighting irrelevant features compared to baseline methods. On real-world datasets, it generated sparse and structured explanations, although performance varied across different quantitative evaluation criteria, indicating that no single method was the best across all metrics.
Implications
The findings suggest that learning explanations in time-frequency representations can significantly enhance the interpretability of deep learning models for time series data, which is crucial in domains like finance, healthcare, and climate science where understanding model decisions is essential.
Downside-Controlled Online Forecast Combination under Delayed and Revised Outcomes
Time Series
- Introduces a novel method for downside-controlled online forecast combination using frozen models.
- Combines static and online correctors to minimize forecast deterioration while maximizing gains.
- Achieves significant improvements in mean squared error across multiple benchmarks and real-world applications.
- Demonstrates the effectiveness of learning from provisional outcomes to enhance settled forecasts.
Read more
Downside-Controlled Online Forecast Combination under Delayed and Revised Outcomes
Summary
This paper addresses the challenge of improving forecasts from frozen models, which cannot be retrained, by introducing a method for downside-controlled online forecast combination. The authors propose a framework that combines a frozen forecaster with both static and online correctors, focusing on minimizing the worst-case deterioration of forecasts while allowing for gains. The methodology leverages the structure of residuals from the frozen models and employs a gate mechanism to dynamically adjust the weights of the combined forecasts based on their performance. The approach is evaluated across seven benchmarks and four base models, including foundation models, demonstrating that it effectively lowers mean mean squared error (MSE) in various scenarios. The results indicate that the proposed method achieves a maximum deterioration of only 0.15% while realizing gains of up to 11.5%. Furthermore, the application to day-ahead load forecasts for European transmission system operators shows consistent improvements across all bidding zones, highlighting the practical utility of the approach in real-time forecasting contexts.
Methodology
The authors develop a combination framework that integrates a frozen forecaster with a static corrector and an online corrector. The weights of these components are dynamically adjusted based on their performance using a gate mechanism, which is informed by the realized losses of each expert. The methodology emphasizes downside control, ensuring that the combined forecast does not significantly worsen compared to the initial forecast.
Results
The proposed method shows a worst-case deterioration of only 0.15% across 28 dataset-model pairs, while achieving gains of up to 11.5%. In practical applications, such as day-ahead load forecasting for seven European bidding zones, the method consistently lowers mean MSE, outperforming single correctors that can increase MSE by up to 102%.
Implications
The findings suggest that the proposed framework can be effectively utilized in environments where retraining of models is impractical, such as in regulated industries. It offers a robust solution for improving forecast accuracy in real-time applications, particularly in time-sensitive domains like electricity load forecasting.
From Graphs to Feeders: Constraint-Guided Diffusion for Rule-Compliant Feeder Generation
Generative Models
Graph Learning
- PG-DiGress formulates feeder synthesis as a constraint-guided discrete graph generation problem.
- The model incorporates electrical compatibility rules into the graph generation process.
- A significant increase in the strict feeder pass rate from 13.7% to 96.8% was achieved.
- The generated feeder models are usable for downstream power system analysis.
Read more
From Graphs to Feeders: Constraint-Guided Diffusion for Rule-Compliant Feeder Generation
Summary
This paper addresses the challenge of generating distribution feeder models that comply with specific electrical rules, which are essential for accurate power system analysis. The authors propose a novel generative model called Power-Grid-constrained Discrete Denoising Diffusion (PG-DiGress) that integrates domain-specific constraints into the graph generation process. Unlike traditional graph generators that focus solely on statistical properties, PG-DiGress ensures that the generated feeder topologies adhere to electrical compatibility and radiality requirements. The model employs soft masks during the reverse diffusion process to suppress incompatible edge classes, followed by a projection step that constructs a valid feeder graph. The evaluation of PG-DiGress demonstrates a significant improvement in generating rule-compliant feeders, with the strict feeder pass rate increasing from 13.7% to 96.8% compared to an unconstrained baseline. The generated graphs can also be converted into executable feeder models for downstream analysis, highlighting the practical utility of the proposed approach.
Methodology
The authors developed PG-DiGress, a discrete diffusion model that learns categorical node and edge patterns from feeder data. The model integrates soft masks during the reverse diffusion process to ensure local compatibility of generated edges and nodes, followed by a projection step to ensure the global validity of the generated feeder graph.
Results
PG-DiGress significantly outperformed the unconstrained baseline in generating rule-compliant feeders, achieving a strict feeder pass rate of 96.8%. The model was also able to convert the generated graphs into executable feeder models for further analysis.
Implications
The proposed method has the potential to enhance the generation of realistic distribution feeder models, which are crucial for power system studies and simulations. It addresses the limitations posed by the lack of detailed feeder models and can facilitate better planning and operational studies in power distribution systems.
FB-GDM: Fully-Bayesian Guided Diffusion Models for High-Dimensional Linear Inverse Problems via Unsupervised Variational Inference
Generative Models
Computer Vision
Efficient ML
- FB-GDM eliminates the need for task-specific hyperparameter tuning in guided diffusion models.
- The model infers precision parameters as latent variables using variational inference, enhancing flexibility and robustness.
- FB-GDM outperforms traditional methods like Î GDM, even when the latter is calibrated with true noise levels.
- The approach maintains computational efficiency, scaling linearly with image resolution.
Read more
FB-GDM: Fully-Bayesian Guided Diffusion Models for High-Dimensional Linear Inverse Problems via Unsupervised Variational Inference
Summary
The paper introduces FB-GDM, a fully-Bayesian guided diffusion model designed to address high-dimensional linear inverse problems without the need for task-specific hyperparameter tuning. Traditional methods like Diffusion Posterior Sampling (DPS) and Pseudoinverse-Guided Diffusion Models (Î GDM) require calibration against ground truth data, which is often unavailable in practical applications. FB-GDM overcomes this limitation by treating two precision parameters as latent variables, which are inferred through variational inference at each reverse step of the diffusion process. This approach allows for a closed-form conditional score that scales linearly with the number of pixels, maintaining tractability at full image resolution. The model only requires the observation and the forward operator as inputs, eliminating the need for noise level estimation or ground truth. Experimental results on CelebA-HQ inverse problems demonstrate that FB-GDM outperforms Î GDM by up to 14 dB and closely matches the performance of a ground-truth-calibrated Î GDM oracle. Additionally, FB-GDM shows robustness to changes in the forward operator, noise level, and image distribution, effectively mitigating hallucinations that can occur with other methods.
Methodology
FB-GDM employs a fully-Bayesian framework where precision parameters related to denoising and observation likelihood are treated as latent variables. These parameters are inferred at each reverse step using variational inference, allowing the model to adaptively balance the contributions of the learned diffusion prior and the observed data without manual tuning.
Results
FB-GDM was tested on CelebA-HQ inverse problems, showing performance improvements of up to 14 dB over Î GDM at its nominal settings. It also matched the performance of a calibrated Î GDM oracle within 0.1 dB. The model maintained robustness across varying conditions, outperforming DPS and exhibiting fewer hallucinations.
Implications
The development of FB-GDM has significant implications for fields requiring high-dimensional image reconstruction, such as medical imaging and astronomical imaging, where ground truth data is often unavailable. Its ability to operate without extensive calibration makes it a practical tool for real-world applications.