AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
69
Papers today
8h
Update frequency
7
Days of history
Rethinking Personalized Reward Modeling for LLMs under Preference Heterogeneity via Group-Debiased Federated Learning
Large Language Models
Federated Learning
Optimization
- A single shared model can outperform group-specific models in personalized reward modeling under balanced preference groups.
- FedGD introduces group-debiased sampling to address group imbalance in federated learning.
- The shared initialization from FedAvg provides a better starting point for personalization due to its adaptability.
- Preference learning consists of shared knowledge and client-specific knowledge, necessitating a focus on initialization rather than final model training.
Read more
Rethinking Personalized Reward Modeling for LLMs under Preference Heterogeneity via Group-Debiased Federated Learning
Summary
This paper addresses the challenge of aligning large language models (LLMs) with human preferences in the context of federated learning (FL), particularly under preference heterogeneity. Traditional federated methods often cluster clients with similar preferences and train separate reward models for each group, which the authors argue is unnecessary. They demonstrate that a single FedAvg model can outperform group-specific models after minimal local optimization, as it learns richer shared representations while mitigating conflicting preferences. The study introduces FedGD (Federated Learning with Group Debiasing), a method that identifies latent preference groups during training and employs group-debiased client sampling to counteract group imbalance. This approach allows for effective personalization of the reward model without prior knowledge of the underlying preference groups. The authors find that the shared initialization from FedAvg has a lower Gradient Quotient (GQ), indicating better adaptability for personalization compared to centralized models. The results suggest that balancing client sampling across preference groups enhances the model's ability to serve minority clients, ultimately leading to improved personalization outcomes.
Methodology
The authors propose FedGD, which involves two phases: first, discovering latent preference groups during federated training, and second, training a single reward model using group-debiased client sampling. The effectiveness of the model is evaluated based on its adaptability and performance in personalization tasks.
Results
The experiments show that the FedAvg model, despite starting with low accuracy, quickly surpasses both centralized training and separate models for each preference group after a few local optimization steps. FedGD effectively mitigates the negative impact of group imbalance, allowing for better personalization for minority clients.
Implications
The findings suggest that federated learning can be effectively utilized for personalized reward modeling in LLMs, enabling better alignment with user preferences while maintaining data privacy. This approach has potential applications in various human-facing AI systems where user satisfaction is critical.
BRiG-AFA: Bellman Risk-to-Go Learning for Non-Myopic Active Feature Acquisition
Reinforcement Learning
Optimization
Theory
- BRiG-AFA offers a supervised alternative to traditional greedy and reinforcement learning methods for Active Feature Acquisition.
- The method learns a budget-specific risk-to-go function, enabling non-myopic decision-making in feature acquisition.
- Controlled benchmarks show significant accuracy improvements over one-step ablation methods.
- The approach is deployable without requiring online reinforcement learning or generative models.
Read more
BRiG-AFA: Bellman Risk-to-Go Learning for Non-Myopic Active Feature Acquisition
Summary
The paper introduces BRiG-AFA, a novel approach to Active Feature Acquisition (AFA) that addresses the limitations of greedy methods and complex reinforcement learning techniques. AFA is concerned with determining which unobserved feature to measure next under a budget constraint. Traditional greedy methods may overlook context features that only reveal their value through subsequent acquisitions. In contrast, BRiG-AFA employs a supervised learning framework that learns a candidate-conditioned risk-to-go function for each remaining budget. This function is derived from the one-step terminal classification risk and is fitted backward using Bellman targets. The inference process minimizes the learned terminal risk based solely on observed values, the mask, candidate identity, and remaining budget. The authors conducted controlled benchmarks demonstrating that BRiG-AFA significantly improves accuracy over one-step ablation methods, achieving notable gains on datasets like Fashion-MNIST and MiniBooNE. The results indicate that BRiG-AFA effectively captures the non-myopic nature of feature acquisition, establishing a reproducible mechanism-level case for direct Bellman risk regression.
Methodology
BRiG-AFA utilizes a supervised learning framework to learn budget-specific action-value networks that predict terminal loss after acquiring features. The method employs a backward fitting process based on Bellman targets, allowing it to derive a risk-to-go function for each remaining budget. The inference process is designed to operate without requiring labels or unobserved values, focusing solely on observed data and the remaining budget.
Results
BRiG-AFA demonstrated an average accuracy improvement of 4.84 ± 2.17 and 4.39 ± 1.10 percentage points over one-step methods at budgets two and three, respectively. On the Fashion-MNIST dataset, it achieved an average accuracy gain of 10.20 ± 0.74 points at four acquisitions and a mean paired gain of 3.50 ± 0.37 points across various budgets. The MiniBooNE study showed mixed results at smaller budgets but positive outcomes at larger budgets of 8 and 16 acquisitions.
Implications
The findings suggest that BRiG-AFA can be effectively applied in scenarios where sequential feature acquisition is critical, such as medical diagnostics, sensor networks, and adaptive learning systems. The method's ability to learn from observed data without requiring extensive metadata makes it suitable for real-world applications where data acquisition is costly or time-consuming.
End-to-End Fairness Optimization with Fair Decision-Focused Learning
Optimization
Theory
- E2EFO provides a comprehensive framework for integrating fairness in predictive modeling and decision-making.
- FDFL optimizes three objectives: prediction accuracy, prediction fairness, and decision regret, using multi-task learning techniques.
- The decision Jacobian is derived for fair allocation, enabling efficient computation in the optimization process.
- Numerical experiments validate the effectiveness of jointly addressing prediction and decision fairness.
Read more
End-to-End Fairness Optimization with Fair Decision-Focused Learning
Summary
This paper introduces End-to-End Fairness Optimization (E2EFO), a framework that integrates fairness considerations throughout the prediction-to-decision pipeline, particularly in resource allocation scenarios. The authors emphasize the importance of addressing both prediction fairness—minimizing accuracy disparities across groups—and decision fairness—ensuring equitable outcomes in resource distribution. They propose Fair Decision-Focused Learning (FDFL), a training paradigm that jointly optimizes prediction accuracy, prediction fairness, and decision regret. The methodology involves multi-task learning techniques to balance these objectives, with a focus on deriving a decision Jacobian for fair allocation. The paper presents numerical experiments in healthcare and synthetic resource allocation contexts, demonstrating the advantages of jointly considering prediction and decision fairness in improving decision-making outcomes.
Methodology
The authors develop Fair Decision-Focused Learning (FDFL) algorithms that utilize gradient-based training combined with multi-task learning (MTL) techniques. They derive a closed-form decision Jacobian for specific fair allocation scenarios and apply differentiable optimization layers for broader cases, allowing for efficient training of predictors that account for multiple fairness objectives.
Results
The experiments conducted on healthcare resource allocation and synthetic scenarios show that the E2EFO framework, through FDFL, significantly improves both prediction and decision fairness compared to traditional methods that focus on either prediction or decision fairness in isolation.
Implications
The proposed framework can be applied in various fields where predictive models influence critical decisions, such as healthcare resource allocation, loan processing in finance, and social services, ensuring that fairness is maintained throughout the decision-making process.
FedChronos: Federated Fine-Tuning of Time-Series Foundation Models for Privacy-Preserving Commodity Price Forecasting
Time Series
Federated Learning
Efficient ML
- FedChronos enables federated fine-tuning of time-series models without centralizing sensitive data.
- The use of Low-Rank Adaptation (LoRA) allows for efficient model updates, reducing communication overhead.
- Differential privacy acts as a regularization method, improving model performance on small datasets.
- The framework is tested on real-world agricultural price data, showcasing its practical applicability.
Read more
FedChronos: Federated Fine-Tuning of Time-Series Foundation Models for Privacy-Preserving Commodity Price Forecasting
Summary
The paper introduces FedChronos, a novel framework designed for federated fine-tuning of time-series foundation models (TSFMs) like Chronos, specifically aimed at addressing the challenges of privacy and data fragmentation in commodity price forecasting. Traditional methods of fine-tuning TSFMs require centralized data, which is often not feasible due to regulatory and competitive constraints. FedChronos utilizes Low-Rank Adaptation (LoRA) to efficiently fine-tune a pre-trained TSFM across distributed clients, allowing them to share only lightweight adapter weights instead of full model parameters. The framework is evaluated using daily commodity prices from 15 agricultural markets in India, demonstrating that naive LoRA fine-tuning can lead to overfitting on small datasets. However, the introduction of differential privacy (DP) noise serves as an implicit regularization technique, significantly improving forecasting accuracy. The results indicate that FedChronos can effectively balance privacy and accuracy, making it suitable for edge AI applications where both network and device constraints exist.
Methodology
FedChronos employs Low-Rank Adaptation (LoRA) to fine-tune the Chronos-T5 backbone model across distributed clients. Each client trains on its local data and shares only the adapter weights with a central server, which aggregates these updates using FedAvg or FedProx. The model's base parameters remain fixed and are never transmitted, ensuring data privacy.
Results
The evaluation of FedChronos on daily commodity prices revealed that naive LoRA fine-tuning led to significant overfitting, dropping below zero-shot performance. However, incorporating differential privacy with an ε value of 5 resulted in a 31% reduction in mean absolute percentage error (MAPE) compared to zero-shot performance, and a 26% improvement over the best traditional baseline.
Implications
FedChronos has significant implications for sectors where data privacy is paramount, such as agriculture, finance, and healthcare. It enables institutions to collaboratively improve forecasting models without compromising sensitive data, thus enhancing decision-making processes in fragmented data environments.
OnlineCache: Learning Dynamic Caching Policies with Error Correction for Efficient Diffusion Inference
Generative Models
Efficient ML
Optimization
- OnlineCache dynamically adapts caching policies based on input complexity and error sensitivity.
- The framework employs a policy gradient method for optimizing resource allocation during inference.
- A bilevel optimization approach is used to jointly train the caching policy and error corrector.
- OnlineCache achieves substantial speed improvements while preserving generation quality.
Read more
OnlineCache: Learning Dynamic Caching Policies with Error Correction for Efficient Diffusion Inference
Summary
The paper introduces OnlineCache, a dynamic caching framework designed to enhance the efficiency of diffusion inference in generative models. Traditional caching strategies are static and do not account for the varying difficulty of generation tasks or the fluctuating error sensitivity across timesteps. OnlineCache addresses these limitations by employing a policy gradient approach to learn adaptive caching policies that optimize resource allocation based on the complexity of the input prompts and the error sensitivity of the denoising process. The framework consists of a policy network that decides when to cache and a corrector that mitigates approximation errors. This dual approach is optimized through a bilevel optimization framework, allowing for improved speed-quality trade-offs. Extensive experiments demonstrate that OnlineCache achieves significant speedups—up to nearly 3×—while maintaining high generation fidelity across various models, outperforming existing cache-based methods.
Methodology
The authors propose a dynamic caching framework called OnlineCache, which utilizes a lightweight policy network to evaluate latent states at each denoising timestep. The policy network decides whether to reuse cached representations or perform full computation. The optimization is performed using policy gradient methods, and a bilevel optimization framework is employed to jointly train the caching policy and an error corrector, ensuring that the policy learns to skip computations based on the corrector's ability to handle induced errors.
Results
OnlineCache demonstrates a nearly 3× speedup on the FLUX.1-dev model while preserving generation fidelity. On other models like DiT and CogVideoX, it also provides competitive acceleration without compromising quality. The method consistently outperforms existing cache-based acceleration baselines, achieving significant reductions in error metrics such as L1 error and perceptual quality.
Implications
The findings suggest that adaptive caching strategies can significantly enhance the efficiency of generative models, making them more suitable for real-time applications. This approach could be applied to various domains where diffusion models are used, potentially improving performance in image, video, and audio generation tasks.
SEDR-Seq2P: A Lightweight Dilated Residual Sequence-to-Point Network for Multi-Task Industrial NILM
Time Series
Efficient ML
Optimization
- Introduces a multi-task industrial NILM framework that disaggregates multiple machine loads from aggregate power.
- Proposes SEDR-Seq2P, a lightweight Seq2Point model that improves accuracy and reduces inference latency.
- Establishes a unified benchmarking protocol for evaluating various NILM models under identical conditions.
- Implements an accuracy-delay criterion to assess the trade-off between predictive performance and computational efficiency.
Read more
SEDR-Seq2P: A Lightweight Dilated Residual Sequence-to-Point Network for Multi-Task Industrial NILM
Summary
This paper addresses the challenges of Non-Intrusive Load Monitoring (NILM) in industrial settings, where measurement noise and concurrent machine operations complicate load disaggregation. The authors propose a multi-task disaggregation approach using a single network to estimate multiple industrial machine loads from aggregate power data. They benchmark various models including Seq2Seq, Seq2SubSeq, Seq2Point, GRU, and WaveNet on the IMDELD dataset, focusing on energy estimation metrics and an accuracy-delay criterion. The proposed SEDR-Seq2P model enhances the Seq2Point architecture with dilated residual blocks and squeeze-and-excitation attention, achieving a better balance between accuracy and computational efficiency. The results demonstrate that SEDR-Seq2P reduces mean absolute error (MAE) by approximately 7% and inference latency by around 58% compared to WaveNet, making it suitable for scalable industrial applications.
Methodology
The authors adopt a one-to-many multi-task learning approach for industrial NILM, utilizing a lightweight Seq2Point architecture enhanced with dilated residual blocks and squeeze-and-excitation attention. They benchmark several models on the IMDELD dataset, employing energy estimation metrics and an accuracy-delay criterion to evaluate performance.
Results
SEDR-Seq2P achieves a 7% reduction in MAE compared to the Seq2Point baseline, a 1% improvement in the coefficient of determination, and a 0.8% increase in match rate. Additionally, it reduces inference latency by approximately 58% compared to WaveNet, indicating a favorable accuracy-delay trade-off.
Implications
The proposed SEDR-Seq2P model offers a scalable solution for industrial NILM, enabling efficient energy management and predictive maintenance in industrial settings. Its lightweight architecture makes it suitable for real-time applications where computational resources are limited.
Toward Plasticity-Preserving KL Regularization for Capability Retention in LLM Reinforcement Learning
Large Language Models
Reinforcement Learning
Optimization
- Introduction of Correctness-Conditioned KL Regularization (CoKL) for LLM post-training.
- CoKL narrows the preservation constraint to correctness-conditioned distributions, avoiding unnecessary restrictions on incorrect outputs.
- Decouples total correctness probability from correctness-conditioned distributions, addressing limitations of full-policy KL regularization.
- Demonstrated superior performance in balancing target-task improvement and capability retention in experiments.
Read more
Toward Plasticity-Preserving KL Regularization for Capability Retention in LLM Reinforcement Learning
Summary
This paper addresses the challenge of capability retention in large language models (LLMs) during reinforcement learning (RL) post-training. Traditional KL regularization methods, while effective in mitigating forgetting of previously learned capabilities, often impose constraints on the entire response distribution, which can hinder exploration and adaptation to new tasks. The authors propose a novel approach called Correctness-Conditioned KL Regularization (CoKL), which focuses on constraining only the correctness-conditioned response distributions. By doing so, CoKL allows for the preservation of existing capabilities while minimizing interference with the learning of new tasks. The method is instantiated using forward KL divergence and is shown to decouple the total probability assigned to correct responses from their correctness-conditioned distribution. This results in a more favorable balance between improving target-task performance and retaining prior capabilities. Experimental results demonstrate that CoKL outperforms existing regularization methods in various controlled environments and continual post-training settings, indicating its effectiveness in enhancing the adaptability of LLMs without compromising their foundational skills.
Methodology
The authors developed CoKL as a conditional regularization framework that applies forward KL divergence to correctness-conditioned response distributions. They derived a finite-group training objective for RL-based LLM post-training and conducted population-level analyses to evaluate the decoupling of total correctness probability from the correctness-conditioned response distribution.
Results
Experiments conducted in controlled multi-solution environments and continual post-training settings across various model scales showed that CoKL achieved a better balance between target-task improvement and retention of prior capabilities compared to existing regularization methods. The results indicate that CoKL effectively mitigates the risk of capability forgetting while allowing for the adaptation to new objectives.
Implications
The findings suggest that CoKL can be a valuable tool for enhancing the post-training process of LLMs, enabling them to adapt to new tasks without losing previously acquired capabilities. This has significant implications for the development of more robust and versatile language models that can maintain performance across diverse applications.
Interpretable machine learning for predicting splitting strength of asphalt concrete: insights from SHAP analysis
Interpretability
- Developed an interpretable machine learning framework for predicting asphalt concrete splitting strength.
- Compared six machine learning models, with TabPFN showing superior performance metrics.
- Identified nine key variables influencing splitting strength through SHAP analysis.
- Quantified optimal parameter ranges for improving splitting strength.
Read more
Interpretable machine learning for predicting splitting strength of asphalt concrete: insights from SHAP analysis
Summary
This paper introduces an interpretable machine learning framework aimed at predicting the splitting strength (ST) of asphalt concrete, which is crucial for evaluating and optimizing asphalt mixtures. A comprehensive database of 296 samples was created, incorporating 14 input variables related to asphalt properties, aggregate gradation, and fiber characteristics. Six machine learning models—TabPFN, ANN, SVR, RF, XGBoost, and LightGBM—were developed and compared, with hyperparameter optimization applied to five of them using NSGA-II. The results indicated that all models performed satisfactorily, with TabPFN achieving the best performance metrics, including the lowest RMSE (0.28) and highest R² (0.88). SHAP analysis was employed to identify the most influential variables affecting ST, revealing that nine variables accounted for 92% of the total contribution, with Ag9.5, FT, Ag4.75, AC, and Du being the most significant. The study also quantified favorable parameter ranges for enhancing ST. To enhance usability, a graphical user interface (GUI) was developed, integrating prediction capabilities with SHAP-based explanations, making the framework more accessible for practical applications.
Methodology
The study established a database of asphalt concrete samples and selected relevant input variables. Six machine learning models were developed and compared, with hyperparameter optimization performed for five models. SHAP analysis was utilized to interpret model predictions and identify influential variables. A GUI was created to enhance accessibility.
Results
All six models demonstrated satisfactory predictive capabilities, with TabPFN achieving the best performance metrics: RMSE of 0.28, MAE of 0.21, MAPE of 18.01%, MAD of 0.14, R² of 0.88, and a composite score of 0.91. SHAP analysis revealed that nine variables accounted for 92% of the total contribution to ST.
Implications
The proposed framework can significantly improve the efficiency and accuracy of asphalt concrete mixture design, providing engineers with a reliable tool for predicting material performance and optimizing mixtures based on data-driven insights.
Unleashing the Potential of Large Language Models: A Blueprint for Real-Time, Enterprise-Ready Deployments
Large Language Models
NLP
- Introduces a comprehensive LLMOps architecture for real-time deployments.
- Addresses critical issues like knowledge staleness and hallucinations in LLMs.
- Incorporates continuous learning and human feedback to enhance model reliability.
- Focuses on governance and compliance for enterprise applications.
Read more
Unleashing the Potential of Large Language Models: A Blueprint for Real-Time, Enterprise-Ready Deployments
Summary
This paper presents a novel architecture for deploying large language models (LLMs) in real-time enterprise environments, addressing critical limitations such as knowledge staleness, catastrophic forgetting, and hallucinations. The proposed framework integrates continuous data ingestion, continual learning, retrieval-augmented generation, and human-in-the-loop feedback to create a dynamic LLM operations (LLMOps) architecture. This architecture allows LLMs to adapt to changing information landscapes, ensuring timely and reliable outputs. The authors emphasize the importance of governance and compliance in enterprise settings, proposing a structured approach to manage trade-offs between latency, cost, and factuality. The framework is designed to enhance the operational capabilities of LLMs, making them more dependable and effective in high-stakes sectors like healthcare and finance. Through extensive experimentation, the authors validate their design patterns and demonstrate how their architecture can transform LLMs into robust systems that meet organizational and regulatory requirements.
Methodology
The authors propose a framework that combines real-time data ingestion, continual learning, and human feedback into a single operational pipeline. They utilize established software design patterns to manage trade-offs among latency, cost, and accuracy while ensuring compliance with regulatory requirements. The architecture includes components like an adaptive ingestion pattern orchestrator, continual learning mechanisms, and retrieval strategies to mitigate hallucinations.
Results
The proposed architecture was validated through extensive experiments, resulting in Pareto-optimal configurations that balance latency, cost, and accuracy. The framework demonstrated improved adaptability and reliability of LLMs in real-time operational contexts, particularly in high-risk sectors.
Implications
The findings suggest that the proposed LLMOps architecture can significantly enhance the deployment of LLMs in enterprise settings, making them more responsive to real-time data and regulatory demands. This has potential applications in various fields, including healthcare, finance, and customer support, where timely and accurate information is critical.
On the Identifiability of Masked Prediction: Mode Blindness and Mask Schedules
Theory
NLP
Generative Models
- Introduces an ε-identifiability modulus to unify rapid-mixing recovery and slow-mixing obstructions.
- Demonstrates mode blindness under large-context mode pinning, where shifts in mode weights minimally affect the objective.
- Establishes that schedule-averaged residual mode uncertainty governs sensitivity to mode weights.
- Validates theoretical findings through empirical tests on both computable laws and real corpora.
Read more
On the Identifiability of Masked Prediction: Mode Blindness and Mask Schedules
Summary
This paper investigates the identifiability of masked prediction methods, which learn representations by optimizing a schedule-weighted collection of conditional distributions. The authors focus on scenarios where data exhibits two well-separated global modes, revealing that the mask schedule significantly influences whether near-optimal conditional predictions can identify the underlying joint distribution. They introduce an ε-identifiability modulus to quantify the maximum distributional error permissible while maintaining a given excess risk. The findings indicate that under large-context mode pinning, reweighting the modes can lead to substantial shifts in the joint law while minimally affecting the masked objective. The authors demonstrate that mode blindness occurs when the mask schedule is dominated by large contexts, rendering the model insensitive to global mode weights. They also establish that low-visibility masks can recover sensitivity to these weights, providing explicit recovery guarantees. The theoretical insights are validated through empirical tests on computable laws, gradient training, and real-world data, illustrating the practical implications of their findings.
Methodology
The authors employ theoretical analysis to derive the ε-identifiability modulus and explore its implications for masked prediction. They analyze the effects of different mask schedules on mode sensitivity and conduct empirical experiments to validate their theoretical claims.
Results
The study reveals that mode blindness is a consequence of large-context mode pinning, where significant shifts in mode weights do not substantially alter the masked objective. The ε-identifiability modulus remains macroscopic even under small excess risk budgets, indicating a robust relationship between mask schedules and identifiability. The findings also show that low-visibility masks can restore sensitivity to mode weights, leading to explicit recovery guarantees.
Implications
The insights from this research could inform the design of more effective masked prediction models in various applications, particularly in natural language processing and generative modeling, by emphasizing the importance of mask schedules in achieving identifiability.
Structured Memory for Edge Language Models: Persistent Context and Corpus Retrieval via O(1) SSM State Injection
NLP
Large Language Models
Efficient ML
- Introduction of PRECOG, which allows for O(1) context retrieval and injection in SSMs.
- Demonstration of SMC for managing persistent context in a hierarchical memory structure.
- Achieved a significant reduction in prefill latency, making the system interactive for edge devices.
- Formal proof that PRECOG matches the quality of in-context RAG, ensuring reliability of results.
Read more
Structured Memory for Edge Language Models: Persistent Context and Corpus Retrieval via O(1) SSM State Injection
Summary
This paper addresses the limitations of Transformer architectures in language models, particularly in retrieval-augmented generation (RAG) scenarios. The authors introduce PRECOG (Pre-Computed Context Injection), a novel retrieval mechanism that reduces the context-ingestion cost from O(Lcontext) to O(1) by pre-encoding document corpora as fixed-size, position-agnostic recurrent hidden states in State-Space Models (SSMs). This allows for efficient retrieval and injection of context at query time, bypassing the need for in-context re-ingestion. The authors also present Structured Memory Consolidation (SMC), which organizes hidden states into a hierarchical persistent memory, enabling efficient management of accumulated context. The proposed methods are implemented on TENNs-LLM, a 1.2B-parameter gated-SSM language model, demonstrating a significant reduction in prefill latency from approximately 27 seconds to less than 6 milliseconds, achieving a speedup of around 4500 times. This advancement makes the system suitable for interactive applications on edge hardware, showcasing the architectural advantages of SSMs over traditional Transformers.
Methodology
The authors developed PRECOG by pre-encoding document corpora into SSM hidden states offline. At query time, the best-matching state is retrieved and injected into the model's recurrent state, allowing the model to process only the user query. SMC organizes past interaction states into a structured memory, facilitating efficient retrieval and integration of both short-term and long-term memory.
Results
The implementation of PRECOG on the TENNs-LLM model resulted in a dramatic reduction in prefill latency from approximately 27 seconds to under 6 milliseconds, achieving a speedup of about 4500 times. The quality of answers generated matched that of traditional in-context RAG methods, confirming the effectiveness of the proposed retrieval mechanism.
Implications
The findings suggest that SSMs can significantly enhance the efficiency of language models, particularly in edge computing environments where latency and memory constraints are critical. The ability to maintain persistent context and retrieve relevant information quickly opens up new possibilities for interactive applications in various domains, including personal assistants and real-time data processing.
EulerLoRA: Rank-Driven Jump Dynamics for Calibrated Parameter-Efficient Fine-Tuning
Efficient ML
Computer Vision
- EulerLoRA extends LoRA by introducing stochasticity, allowing for multiple predictive trajectories.
- It maintains the deterministic LoRA transformation in expectation while providing predictive uncertainty estimates.
- The method significantly reduces the number of trainable parameters compared to traditional LoRA-Ensemble approaches.
- EulerLoRA was tested on multiple vision datasets and showed competitive performance.
Read more
EulerLoRA: Rank-Driven Jump Dynamics for Calibrated Parameter-Efficient Fine-Tuning
Summary
EulerLoRA introduces a stochastic extension of Low-Rank Adaptation (LoRA) aimed at enhancing parameter-efficient fine-tuning while addressing the limitations of standard LoRA, which produces a single deterministic model and lacks predictive uncertainty estimation. By sampling structured variations along the rank-one components of shared low-rank adapters, EulerLoRA generates multiple predictive trajectories while maintaining the deterministic transformation of LoRA in expectation. The method was evaluated on vision transformers across various datasets, including CIFAR-10, CIFAR-100, and HAM10000, as well as out-of-distribution detection on SVHN. The results indicate that EulerLoRA achieves comparable or improved performance relative to strong LoRA-Ensemble baselines, while significantly reducing the number of trainable parameters—approximately 3 million for EulerLoRA compared to about 10 million for a rank-8, 16-adapter LoRA-Ensemble. This demonstrates that effective predictive diversity can be achieved with a smaller number of shared adapters, making it a promising approach for efficient model adaptation.
Methodology
EulerLoRA employs a stochastic formulation of LoRA based on finite-step rank-driven dynamics. It samples rank configurations over internal steps to create structured updates along learned rank-one directions, allowing for the generation of multiple predictive trajectories from shared low-rank adapter parameters.
Results
EulerLoRA demonstrated comparable or improved performance on vision tasks compared to LoRA-Ensemble baselines, achieving this with approximately 69% fewer trainable parameters. This indicates that useful predictive diversity can be derived from a limited number of shared adapters.
Implications
The findings suggest that EulerLoRA can be a valuable tool for efficient model fine-tuning in various applications, particularly in scenarios where computational resources are limited or where uncertainty estimation is critical for model reliability.
MMFGU: Multimodal Federated Graph Unlearning
Federated Learning
Graph Learning
Multimodal
- MMFGU addresses the unique challenges of multimodal federated graph unlearning.
- The framework allows for targeted removal of specific data while preserving related information.
- MMFGU achieves a 41.5× speedup over traditional full retraining methods.
- The proposed method outperforms 12 existing unlearning baselines in utility and effectiveness.
Read more
MMFGU: Multimodal Federated Graph Unlearning
Summary
The paper introduces MMFGU, a novel framework for multimodal federated graph unlearning, addressing the challenges of removing specific user-requested information from multimodal graph models without compromising retained data. In the context of multimodal federated graph learning (MM-FGL), users may request the deletion of accounts, interactions, or specific content while retaining other related data. Existing methods primarily focus on broader entity or client removal, failing to accommodate the nuanced needs of multimodal data. MMFGU tackles three main challenges: selectively removing requested information without damaging retained content, preventing recovery of deleted information through remaining modalities, and ensuring that related traces do not re-enter the global model post-aggregation. The framework employs target-specific representation decoupling, mapping heterogeneous requests into unified representations, and utilizing lightweight probes to repair residual traces. Experimental results demonstrate that MMFGU effectively fulfills unlearning requests, maintains the utility of retained graph data, and achieves a significant speedup in processing time compared to full retraining.
Methodology
MMFGU employs a target-specific representation decoupling approach, which maps heterogeneous unlearning requests into unified target carriers. It selectively decouples requested representations while preserving retained semantics, utilizes lightweight probes to expose and repair residual traces, and employs prototype-guided screening to identify affected clients for targeted unlearning responses.
Results
The experimental evaluation across 12 datasets and 2 downstream tasks shows that MMFGU achieves superior performance in unlearning effectiveness while maintaining downstream task performance. It approaches the performance of full retraining in terms of unlearning effectiveness and achieves a 41.5× speedup over traditional methods.
Implications
The proposed MMFGU framework has significant implications for privacy-preserving machine learning, particularly in applications involving sensitive multimodal data such as social networks, healthcare, and personalized recommendations. It enables organizations to comply with data privacy regulations while still benefiting from collaborative learning.
UniPolymer: A Unified Framework for Property Prediction, Structure Recommendation, and Evaluation in Polyimide Design
Generative Models
Graph Learning
Optimization
- UniPolymer integrates property prediction, structure generation, evaluation, and recommendation for polyimide design.
- The PITg-Curated dataset contains 10,066 deduplicated polyimide repeating units with Tg tags.
- The framework employs self-supervised learning and a continuous-discrete joint Tg representation for better structure generation.
- UniPolymer achieved a property prediction accuracy of R² = 0.93 and a candidate evaluation pass rate of 73.79%.
Read more
UniPolymer: A Unified Framework for Property Prediction, Structure Recommendation, and Evaluation in Polyimide Design
Summary
The paper presents UniPolymer, a novel framework aimed at improving the design of polyimides by addressing the challenges associated with predicting glass transition temperatures (Tg) and generating candidate structures that meet specific property requirements. Traditional methods often generate structures that do not align with target properties, leading to inefficiencies in material development. UniPolymer integrates property prediction, target-conditioned generation, candidate evaluation, and structure recommendation into a single framework. A key innovation is the construction of the PITg-Curated dataset, which includes 10,066 deduplicated polyimide repeating units tagged with Tg values. The framework employs self-supervised chemical semantic learning to establish a reliable structure-property relationship, enhancing structural consistency and fusing multi-scale information. It utilizes a continuous-discrete joint Tg representation to guide the autoregressive generation of molecular structures using SELFIES. The generated candidates are evaluated against a frozen property predictor and ranked based on their deviation from the target Tg, ensuring that only the most relevant candidates proceed to experimental validation. Experimental results demonstrate that UniPolymer achieves a property prediction accuracy of R² = 0.93 and a candidate evaluation pass rate of 73.79%, outperforming existing baselines. The predicted Tg values align closely with molecular dynamics simulation results, significantly reducing the number of candidates entering costly experimental stages.
Methodology
UniPolymer employs a unified approach that includes self-supervised chemical semantic learning for structure-property mapping, a continuous-discrete joint Tg representation for guiding structure generation, and a candidate evaluation process that ranks candidates based on their predicted Tg deviations and structural constraints.
Results
UniPolymer achieved a property prediction accuracy of R² = 0.93 and a candidate evaluation pass rate of 73.79%, which are improvements over existing methods. The predicted Tg values for the recommended candidates were in high agreement with molecular dynamics simulations, reducing the number of candidates that proceed to experimental validation.
Implications
The UniPolymer framework has the potential to significantly streamline the design process for polyimides, reducing costs and development time by ensuring that only high-quality candidates are selected for experimental validation. This could lead to more efficient material development in various applications, including aerospace and flexible electronics.
HindSearch: Trajectory-Level Hindsight Critique for Search-Augmented Reinforcement Learning
Reinforcement Learning
NLP
Large Language Models
- Introduction of HindSearch, a hindsight self-distillation procedure for search-augmented RL.
- Utilization of a frozen judge to provide directive critiques for failed trajectories.
- Achieved 39.4% average exact match on a standard benchmark suite, outperforming existing methods.
- Demonstrated that hindsight critiques significantly improve learning outcomes.
Read more
HindSearch: Trajectory-Level Hindsight Critique for Search-Augmented Reinforcement Learning
Summary
The paper introduces HindSearch, a novel approach to enhance search-augmented reinforcement learning (RL) agents by implementing a trajectory-level hindsight critique mechanism. Traditional search-augmented language model (LM) agents are trained using a binary exact-match reward system, which only indicates whether a trajectory succeeded or failed without providing insights into the reasons behind failures. HindSearch addresses this limitation by employing a frozen judge that critiques failed trajectories post-rollout, using the gold answer to generate directive critiques that guide the learning process. This critique serves as an auxiliary on-policy distillation signal, allowing the student model to learn from its mistakes more effectively. The authors demonstrate the effectiveness of HindSearch on a standard seven-benchmark suite with the Qwen2.5-3B-Instruct model, achieving an impressive average exact match (EM) score of 39.4%, significantly outperforming previous search-RL baselines. The study highlights the importance of hindsight in improving learning outcomes, as removing the judge's access to the gold answer diminishes the performance gains, underscoring the value of the critique mechanism in training search-augmented RL agents.
Methodology
HindSearch employs a trajectory-level hindsight critique (TLHC) approach, where after each rollout, failed trajectories are evaluated by a frozen judge that has access to the gold answer. The judge generates critiques that inform the student model's learning process through an auxiliary on-policy distillation loss. This method allows the model to receive corrective feedback rather than just evaluative signals, enhancing its ability to learn from mistakes.
Results
The implementation of HindSearch resulted in a 39.4% average exact match score on a standard seven-benchmark search-QA suite, marking a significant improvement over prior search-RL baselines. The study also found that removing the judge's access to the gold answer led to a substantial decrease in performance, indicating the critical role of hindsight in the learning process.
Implications
HindSearch has the potential to improve the training of search-augmented RL agents across various applications, particularly in scenarios where understanding the reasons behind failures is crucial. This approach could lead to more robust and effective models in natural language processing tasks, enhancing their ability to retrieve and reason over information.
LLM-Guided Retrieval for Prediction of Molecular Perturbation Responses
NLP
Large Language Models
- Introduces LLM-Guided Retrieval (LGR) for predicting molecular perturbation responses.
- Frames the prediction task as a retrieve-and-aggregate problem, simplifying the modeling process.
- Demonstrates significant improvements in prediction accuracy, especially for unseen cell lines.
- Highlights the importance of retrieval quality over model complexity in zero-shot prediction scenarios.
Read more
LLM-Guided Retrieval for Prediction of Molecular Perturbation Responses
Summary
This paper addresses the challenge of predicting transcriptomic responses to small-molecule perturbations across various cell lines, a crucial aspect of drug discovery. The authors propose a novel framework called LLM-Guided Retrieval (LGR), which reformulates the prediction task as a retrieve-and-aggregate problem. Instead of exhaustive profiling of drug-cell combinations, LGR utilizes a large language model (LLM) to rank candidate neighbor drugs that have been profiled in the target cell line. The model then aggregates their observed expression deltas to predict the response of an unmeasured drug. The evaluation is conducted on the Tahoe-100M single-cell perturbation atlas, focusing on unseen-drug, unseen-cell-line, and open-world scenarios. The results demonstrate that LGR consistently outperforms traditional baselines, particularly in generalizing to unseen cell lines, achieving higher correlation and lower error rates. The findings suggest that the quality of retrieval is more critical than the complexity of the predictive model, highlighting the utility of LLMs as effective biological prior modules in this context.
Methodology
The methodology involves using a large language model to select biologically relevant compounds from a restricted candidate pool based on their profiles in the target cell line. The selected compounds' expression deltas are then aggregated using a fixed mean aggregator to predict the response of an unmeasured drug.
Results
LGR outperformed traditional baselines such as drug mean and chemistry-based kNN models, particularly excelling in generalization to unseen cell lines. The model achieved higher correlation and lower error rates in predictions, indicating improved directional accuracy of gene regulation.
Implications
The findings suggest that LGR could enhance drug discovery processes by providing more accurate predictions of molecular responses, thereby facilitating the identification of effective drug candidates. The approach may also be applicable in other areas of functional genomics and personalized medicine.
Statistical Mechanics of Learning on Product Wasserstein Manifolds
Theory
Optimization
- Reframing distributional constraints as geometric priors enhances understanding of learning dynamics.
- Development of a hierarchical mean-field theory that captures layer-wise Wasserstein geometries in deep networks.
- Introduction of quantum Wasserstein distance as a metric for variational quantum circuits.
- Novel optimization algorithms (Hierarchical DisCo-SGD and Quantum DisCo) that improve training stability and generalization.
Read more
Statistical Mechanics of Learning on Product Wasserstein Manifolds
Summary
This paper challenges the conventional view of statistical mechanics in learning, which typically sees constraints on weight distributions as limitations that reduce model capacity. Instead, the authors propose that these constraints define the intrinsic geometry of the learning process. They formulate deep neural networks and variational quantum circuits as gradient flows on product Wasserstein manifolds, where each layer of a neural network corresponds to a classical Wasserstein space and the parameters of a quantum circuit correspond to a quantum Wasserstein space. This geometric perspective allows for a reinterpretation of capacity reduction as a feature of the constraint manifold's metric structure. The authors develop a hierarchical mean-field theory for deep networks, extend this framework to quantum settings using quantum Wasserstein distances, and introduce two algorithms—Hierarchical DisCo-SGD and Quantum DisCo—that optimize learning by following geodesics on the manifold. Empirical results demonstrate that this approach improves generalization, stabilizes training, and mitigates issues such as barren plateaus in quantum circuits compared to traditional unconstrained methods.
Methodology
The authors utilize a geometric framework based on product Wasserstein manifolds to model the learning process in deep neural networks and variational quantum circuits. They derive a hierarchical mean-field theory to extend the analysis to deep architectures and propose new algorithms for optimization that follow geodesics on the manifold.
Results
Experiments conducted on teacher-student problems, standard image classification tasks, and small variational quantum classifiers indicate that the proposed methods lead to better generalization, more stable training, and a reduction in the severity of barren plateaus compared to unconstrained and norm-based baselines.
Implications
This work suggests a new paradigm for understanding learning in both classical and quantum contexts, allowing for the incorporation of biologically inspired or hardware-specific distributional information into learning systems. It opens avenues for further research in optimizing neural networks and quantum circuits by leveraging geometric properties.
Hypergradient-based Bilevel Reinforcement Learning with Improved Sample Complexity
Reinforcement Learning
Optimization
Theory
- Introduction of a Hessian-free bilevel RL algorithm called Approximate Hypergradient Optimization (AHO).
- Improvement of sample complexity from ˜O(ϵ−3) to ˜O(ϵ−2) while maintaining iteration complexity of O(ϵ−1).
- Removal of the Polyak-Łojasiewicz condition assumption on the outer-level objective function.
- Utilization of Boltzmann policy optimality for unrealizable policy classes to enhance convergence.
Read more
Hypergradient-based Bilevel Reinforcement Learning with Improved Sample Complexity
Summary
This paper presents a novel hypergradient-based bilevel reinforcement learning (RL) algorithm that addresses the limitations of existing methods in terms of scalability and sample complexity. Bilevel RL is a framework that can model various RL problems, including meta-learning and hierarchical task decomposition. Traditional bilevel RL algorithms often rely on Hessian-based hypergradients, which can be computationally intensive, or penalty-based approximations that lead to high sample complexity. The authors propose a Hessian-free algorithm called Approximate Hypergradient Optimization (AHO) that leverages the optimality of the Boltzmann policy for entropy-regularized discounted RL objectives. This approach achieves an iteration complexity of O(ϵ−1) and a significantly improved sample complexity of ˜O(ϵ−2), surpassing the previous state-of-the-art sample complexity of ˜O(ϵ−3). The paper also removes the assumption of the Polyak-Łojasiewicz (PL) condition on the outer-level objective function, which was a limitation in prior works. The proposed algorithm is designed to be scalable and effective for non-convex bilevel RL problems, making it a significant advancement in the field.
Methodology
The authors develop a hypergradient-based bilevel RL algorithm that avoids Hessian computations by utilizing the optimality of the Boltzmann policy. They provide conditions under which this optimality can be applied to unrealizable policy classes. The algorithm employs gradient-shifted functions to approximate the hypergradient, allowing for the removal of the PL condition and enhancing sample complexity.
Results
The proposed AHO algorithm achieves an iteration complexity of O(ϵ−1) and a sample complexity of ˜O(ϵ−2). This represents a significant improvement over previous methods, which had a sample complexity of ˜O(ϵ−3). The algorithm's design allows it to operate effectively without the assumptions of unique minimizers or the PL condition, making it more broadly applicable.
Implications
The advancements presented in this paper could lead to more efficient and scalable applications of bilevel reinforcement learning in various domains, including meta-learning, hierarchical task decomposition, and reinforcement learning from human feedback. The improved sample complexity could facilitate faster convergence in practical RL scenarios.
PiDDM: Physics-Informed Differentiable Degradation Modeling for Lithium-Ion Battery State-of-Health Prediction
Time Series
- PiDDM integrates physics-based degradation mechanisms into machine learning for battery SOH prediction.
- The framework significantly reduces prediction errors compared to traditional models.
- PiDDM maintains physical consistency in capacity fade predictions across diverse operating conditions.
- The model effectively captures accelerated end-of-life degradation without unrealistic predictions.
Read more
PiDDM: Physics-Informed Differentiable Degradation Modeling for Lithium-Ion Battery State-of-Health Prediction
Summary
The paper presents PiDDM, a novel physics-informed differentiable degradation modeling framework aimed at improving the prediction of lithium-ion battery state-of-health (SOH). Traditional data-driven machine learning models often struggle with generalization across varying cycling protocols and can yield physically implausible predictions during long-term extrapolation. To overcome these challenges, PiDDM integrates empirical Arrhenius degradation kinetics related to solid electrolyte interphase (SEI) growth and loss of lithium inventory (LLI) into the neural network training objective. This incorporation of degradation physics into the loss function ensures that the model adheres to physically consistent monotonic capacity fade across diverse operating conditions. The model was evaluated using a public dataset of 55 batteries cycled under six different protocols, demonstrating superior performance compared to standard multilayer perceptron (MLP) and baseline physics-informed neural network (PINN) models. Notably, PiDDM achieved the lowest average prediction error and effectively captured accelerated end-of-life degradation without the non-physical capacity regeneration seen in other models. These findings highlight the potential of integrating physical principles into machine learning frameworks for enhanced predictive accuracy and reliability in battery health monitoring.
Methodology
The methodology involves developing a physics-informed neural network that incorporates degradation kinetics related to SEI growth and LLI loss into its training objective. This is achieved by embedding these physical constraints into the loss function, which regularizes the model's predictions to ensure they remain physically plausible over time.
Results
PiDDM outperformed both a standard multilayer perceptron and a baseline physics-informed neural network in terms of average prediction error. In extrapolation tests, where models were trained on the first 90% of battery cycle life, PiDDM accurately predicted the remaining 10% without exhibiting non-physical capacity regeneration, demonstrating improved predictive accuracy and physical consistency.
Implications
The findings suggest that incorporating physical degradation models into machine learning frameworks can enhance the reliability and accuracy of battery health predictions, which is crucial for the effective management of lithium-ion batteries in various applications, including electric vehicles and renewable energy systems.
Nonlinear Laplacians Improve Signed-Directed Graph Learning
Graph Learning
- Introduction of a non-linear Laplacian operator specific to signed-directed graphs (NLSD).
- NLSD allows for selective message passing based on the alignment of potential discrepancies with edge directions.
- Development of the NLSD-GNN framework that efficiently integrates signed and directional data.
- Extensive evaluation shows superior performance in node classification and link prediction tasks.
Read more
Nonlinear Laplacians Improve Signed-Directed Graph Learning
Summary
This paper addresses the limitations of existing linear Laplacians in the context of signed-directed graphs, which represent complex relationships with both positive and negative interactions. The authors introduce a novel non-linear Laplacian operator tailored for signed-directed networks (NLSD), which enhances the message-passing capabilities of Graph Neural Networks (GNNs). The NLSD operator calculates node-specific potentials based on features, allowing for selective message passing that aligns with edge directions. This approach leads to the development of the NLSD-GNN framework, which efficiently integrates signed and directional information. The authors conducted extensive experiments on node classification and link prediction tasks, demonstrating that NLSD-GNN outperforms existing methods across various datasets, thereby showcasing the effectiveness of non-linear Laplacians in graph learning.
Methodology
The authors extend the definition of existing Laplacians to create a non-linear operator that accounts for both edge direction and sign. The NLSD-GNN framework utilizes this operator to project node feature vectors into scalar potentials, which are then used to select a subgraph for message passing. The method incorporates edge dropouts and symmetrization to facilitate efficient computation.
Results
The NLSD-GNN framework demonstrated significant improvements over existing methods in benchmark datasets for transductive node classification and link prediction, validating the effectiveness of the proposed non-linear Laplacians.
Implications
The findings suggest that non-linear Laplacians can enhance the performance of GNNs in complex relational data, opening avenues for more effective modeling of social networks, recommendation systems, and other applications involving signed-directed graphs.
AdvPlan-Bench: Adversarial Evaluation of Structured Plan-Generation Agents
Theory
Reinforcement Learning
Generative Models
- AdvPlan-Bench provides a novel framework for evaluating structured plan-generation agents in adversarial settings.
- The benchmark includes metrics for assessing plan quality, response-budget sensitivity, and candidate diversity.
- Empirical results show that using a sampled best-response policy improves plan evaluation outcomes compared to single-sample responses.
- The study emphasizes the need for adversarial context in evaluating plan-generation agents, moving beyond isolated plan assessments.
Read more
AdvPlan-Bench: Adversarial Evaluation of Structured Plan-Generation Agents
Summary
The paper introduces AdvPlan-Bench, an offline benchmark designed for the adversarial evaluation of structured plan-generation agents. Traditional evaluations often assess plans in isolation, neglecting the influence of adversarial responses. AdvPlan-Bench addresses this gap by providing a framework that evaluates plans based on their performance against sampled adversarial responses. The benchmark includes a formal setup with typed action chains, synthetic quality scores, and various diagnostic metrics such as BLUE-vs-RED advantage and Nash-gap diagnostics. The authors conducted experiments across 150 synthetic scenarios, demonstrating that a sampled best-response policy significantly reduces the advantage of the original plan, highlighting the importance of considering adversarial contexts in plan evaluation. The benchmark is not intended for operational planning but serves as a reproducible tool for studying adversarial plan evaluation and multi-agent interactions.
Methodology
The authors developed AdvPlan-Bench as a reproducible offline benchmark that evaluates structured plans against adversarial responses. It utilizes typed action chains, synthetic quality scoring, and various diagnostic metrics to assess the performance of plan-generation agents. The evaluation is conducted through a series of synthetic scenarios designed to test the robustness of plans in adversarial contexts.
Results
In the experiments, a sampled best-response policy reduced the BLUE advantage from 0.518 to 0.486 and the BLUE win rate from 0.900 to 0.820 compared to a single-sample response. An offline LLM-policy contract baseline achieved a BLUE advantage of 0.496 and a win rate of 0.700, while a two-stage multi-agent council reached a BLUE advantage of 0.509 and a win rate of 0.813. The study also found a high inter-rater agreement of 0.978 in the rubric-sensitivity analysis.
Implications
AdvPlan-Bench has the potential to enhance the evaluation of AI agents in adversarial environments, providing insights into the robustness and adaptability of planning strategies. It can inform the development of more resilient AI systems capable of handling complex, competitive scenarios.
RamanPFN: learning from Raman spectral structure with a tabular foundation model
Theory
Efficient ML
Interpretability
- RamanPFN effectively encodes Raman spectral structure for improved predictive modeling.
- The framework combines Global Compositional Unmixing and Local Vibrational Subspace Encoding.
- RamanPFN achieved a 19.6% reduction in root-mean-square error and a 9.0% reduction in classification error.
- The approach allows for better handling of high-dimensional, small-sample Raman datasets.
Read more
RamanPFN: learning from Raman spectral structure with a tabular foundation model
Summary
RamanPFN introduces a novel framework for analyzing Raman spectroscopy data, which is crucial for non-destructive molecular characterization in various fields such as materials science and biomedicine. The challenge in Raman spectroscopy lies in the high-dimensional nature of the data, where informative variations occur across both local and distant spectral regions, often with limited labeled samples. Traditional methods like latent-variable chemometrics and deep spectral networks have limitations in capturing the intricate spectral structure. RamanPFN addresses these issues by employing a dual-scale spectral representation that encodes dependencies before inference using the TabPFN model. This framework consists of two main components: Global Compositional Unmixing, which organizes the entire spectrum into non-negative coordinates, and Local Vibrational Subspace Encoding, which captures independent variations in contiguous wavenumber regions. The representations are evaluated separately and combined for predictions. The effectiveness of RamanPFN was validated across 150 tasks from 74 public Raman datasets, demonstrating a significant reduction in root-mean-square error and classification error compared to direct TabPFN inference. This establishes RamanPFN as a robust interface between high-dimensional Raman measurements and tabular inference, enhancing predictive accuracy in Raman spectroscopy.
Methodology
RamanPFN employs a dual-scale spectral representation framework that includes Global Compositional Unmixing to create non-negative coordinates over the complete spectrum and Local Vibrational Subspace Encoding to represent contiguous wavenumber regions. These representations are processed separately and integrated at the prediction level using the TabPFN model, which performs in-context inference without task-specific parameter fitting.
Results
RamanPFN demonstrated an average reduction of 19.6% in root-mean-square error across 129 regression targets and a 9.0% reduction in classification error across 21 classification tasks when compared to direct TabPFN inference, showcasing its effectiveness in improving predictive accuracy in Raman spectroscopy.
Implications
The development of RamanPFN has significant implications for fields that rely on Raman spectroscopy for molecular characterization, including materials science, biomedicine, and industrial process monitoring. By enhancing predictive modeling capabilities, RamanPFN can facilitate more accurate analyses and interpretations of complex spectral data.
An analysis of machine learning approaches for enhancing decision-making in complex discrete choice tasks
Theory
- Machine learning methods can enhance discrete choice modeling by overcoming limitations of traditional parametric approaches.
- Semi-parametric and non-parametric models generally outperform parametric models in learning individual choice rules.
- Model performance improves significantly with increased training data and determinism of choice rules.
- The twinned neural network model showed the best performance in a case study using real energy policy preference data.
Read more
An analysis of machine learning approaches for enhancing decision-making in complex discrete choice tasks
Summary
This paper investigates the application of machine learning methods to improve discrete choice modeling, a common tool for preference elicitation in policy-making. Traditional parametric models often impose restrictive assumptions about decision-making processes, which can limit their effectiveness. The authors evaluate four machine learning models—multinomial logistic regression, generalized additive model, twinned neural network, and Gaussian process—assessing their ability to learn and predict five significant choice rules relevant to behavioral and social sciences. Through Monte Carlo experiments, the study examines model performance under varying conditions, such as the number of attributes in choice alternatives and the determinism of choice rules. Results indicate that semi-parametric and non-parametric models generally outperform parametric models, with performance improvements observed with increased training choice sets and choice rule determinism. A case study using real energy policy preference data further demonstrated the effectiveness of the twinned neural network model. This research highlights the potential of machine learning to enhance discrete choice modeling by allowing for greater flexibility in capturing individual preferences, thus informing better decision-making processes in complex environments.
Methodology
The study employs Monte Carlo simulations to evaluate the performance of four machine learning models in learning discrete choice rules. It compares parametric, semi-parametric, and non-parametric approaches across various experimental contexts, focusing on the ability to recover individual choice preferences from limited data.
Results
The findings reveal that semi-parametric and non-parametric models consistently outperform parametric models across all tested choice rules and contexts. Performance improvements range from 6% to 96% with increased training choice sets and from 0% to 55% with increased choice rule determinism. In a case study, the twinned neural network achieved the best Bayesian Information Criterion (BIC) score of 13.351.
Implications
This research has significant implications for policy-making, suggesting that machine learning can provide more accurate and flexible models for understanding individual preferences in complex decision-making scenarios. It encourages the adoption of non-parametric approaches in behavioral modeling, which could lead to better-informed policies and strategies.
Pseudorandom Streams within Diffusion Models Act as Learnable Inputs That Affect Generation Quality
Generative Models
Optimization
Theory
- Pseudorandom streams can be treated as learnable inputs in diffusion models.
- The study distinguishes between next-value predictability and exploitability of pseudorandom orbits.
- Different pseudorandom orbits lead to significant variations in diffusion loss and generation quality.
- The empirical analysis reveals a power-law relationship between probe loss and real-data diffusion loss.
Read more
Pseudorandom Streams within Diffusion Models Act as Learnable Inputs That Affect Generation Quality
Summary
This paper investigates the role of pseudorandom streams in diffusion models, proposing that these streams can act as learnable inputs that significantly influence both training and generation quality. The authors highlight that traditional diffusion models rely on stochastic inputs, which, on finite-precision hardware, are represented as deterministic numerical orbits generated by pseudorandom rules. The study introduces a small multilayer perceptron (MLP) to predict the next value of these orbits, assessing their predictability. A diffusion probe is employed to replace real images with random tensors, allowing the evaluation of how well the diffusion model can exploit the structure of these orbits. The experiments reveal that different pseudorandom orbits lead to varying diffusion losses and generation quality on datasets like MNIST and CIFAR-10. The findings suggest that pseudorandom sources are not merely implementation details but can be structured inputs that affect model performance. The paper also establishes a framework for analyzing the learnability of these orbits, distinguishing between next-value predictability and exploitability within the diffusion system, and demonstrating strong correlations between the learned measures and generation degradation.
Methodology
The authors developed an experimental framework that includes empirical-rank uniformization, Gaussianization, mean–variance calibration, and screening for strong dynamical structures. They utilized a multilayer perceptron to predict the next value of pseudorandom orbits and a diffusion probe to assess the impact of these orbits on the diffusion model's training objective.
Results
The experiments demonstrated that different pseudorandom orbits produced markedly different diffusion losses and generation quality on the MNIST and CIFAR-10 datasets. The MLP and diffusion probe losses showed strong rank correlations with generation degradation, and after normalization, both losses followed an empirical power law with different exponents for the two datasets.
Implications
This research implies that the choice of pseudorandom number generators can significantly impact the performance of diffusion models, suggesting that future work should consider the structure of random inputs as a critical factor in model design and training. It opens avenues for further exploration into how different types of randomness can be optimized for better generative performance.
Feature Interaction Modeling for Physics-Informed Neural Networks and Neural Operators
Theory
Efficient ML
Optimization
- Introduction of FM-PINN, enhancing the modeling of variable interactions in PINNs.
- Development of FM-Operator and FM-DeepONet for improved accuracy in nonlinear PDEs.
- Demonstrated substantial accuracy gains on shock-dominated equations.
- Effective modeling of cross-group interactions among various features.
Read more
Feature Interaction Modeling for Physics-Informed Neural Networks and Neural Operators
Summary
This paper introduces a novel approach to enhance the expressiveness of physics-informed neural networks (PINNs) and neural operator learning by integrating feature interaction modules derived from factorization machines (FMs). The authors propose FM-PINN, which captures spatio-temporal variable interactions to improve approximation accuracy for high-order parameterized partial differential equations (PDEs). The methodology involves grouping spatial coordinates, time, physical parameters, and boundary conditions into independent feature sets to model their interactions effectively. The paper further develops FM-Operator and FM-DeepONet, architectures designed to tackle nonlinear conservation laws and problems with sharp gradients or discontinuities. Numerical experiments demonstrate significant accuracy improvements on challenging shock-dominated equations, indicating the potential of this approach for physics-consistent modeling of PDEs with complex dependencies.
Methodology
The authors embed factorization machine-derived feature interaction modules into PINNs and neural operators. They propose three architectures: FM-PINN, which models pairwise relationships among spatial and temporal variables; FM-Operator, which captures relationships among sensor-level input functions and query coordinates; and FM-DeepONet, which refines features in the DeepONet framework. The approach leverages the bi-interaction mechanism of factorization machines to efficiently model interactions among continuous PDE inputs.
Results
Numerical tests reveal that the proposed FM-PINN outperforms traditional PINNs on high-dimensional smooth problems. FM-Operator achieves superior accuracy on sharp-gradient problems with fewer parameters, while FM-DeepONet significantly improves solution accuracy for problems with discontinuities. Overall, the proposed methods demonstrate substantial accuracy gains on challenging PDEs, particularly those dominated by shocks.
Implications
The findings suggest that incorporating feature interaction modeling can significantly enhance the performance of physics-informed neural networks and neural operators, particularly in applications involving complex PDEs. This approach could lead to more accurate and efficient solutions in various fields of science and engineering where PDEs are prevalent.
Flow Matching with Missing Data
Generative Models
Theory
Optimization
- MDFM treats missing data as latent variables and uses resampling to compute flow matching loss.
- Under MCAR, training with incomplete data yields the same objective as with complete data.
- One oracle completion per example is optimal under a fixed evaluation budget.
- The method preserves conditional variability, outperforming traditional point imputation.
Read more
Flow Matching with Missing Data
Summary
The paper introduces Missing-Data Flow Matching (MDFM), a novel approach to flow matching that addresses the challenge of incomplete training data, which is common in real-world applications. Traditional flow matching requires fully observed data, but many datasets have missing values due to various reasons such as sensor failures or incomplete surveys. MDFM treats the missing coordinates as latent variables and utilizes resampling from their conditional distribution to compute the flow matching loss. The authors prove that under the assumption of missing completely at random (MCAR), the objective of training with incomplete data is equivalent to that of complete data, thus ensuring that the learning process remains unaffected by missingness. The paper also provides a finite-sample analysis that reveals key insights regarding the optimal number of completions per example and the impact of the completion model's accuracy on bias. Experimental results validate the theoretical findings, demonstrating that MDFM preserves conditional variability better than traditional point imputation methods and performs competitively against established imputation techniques on real tabular datasets.
Methodology
The authors propose a framework where missing data points are treated as latent variables. Instead of using point imputation, they draw multiple completions from the conditional distribution of the missing data given the observed data. The flow matching loss is averaged over these completions. Theoretical proofs establish the equivalence of the incomplete and complete data objectives under MCAR, and a finite-sample analysis provides insights into the optimal design choices for the method.
Results
The experiments confirm that MDFM maintains the conditional variability of the generated data, which is often lost in point imputation approaches. The method shows competitive performance against strong classical and deep imputation baselines, particularly excelling in preserving covariance in real tabular datasets. The theoretical results regarding variance and bias are also supported by empirical evidence.
Implications
MDFM has significant implications for generative modeling in scenarios where data is often incomplete. It can be applied in various fields such as healthcare, finance, and social sciences, where missing data is prevalent. The method enhances the robustness of generative models, allowing them to learn from incomplete datasets without compromising the quality of the generated outputs.
Beckmann Transport Models: From Autonomous Flows to One-Step Maps
Generative Models
Optimization
Theory
- Establishes that time-independent flow matching can validly transport distributions when the target is singular.
- Introduces a conservation equation for the one-step transport map, facilitating direct learning from samples.
- Corrects the loss function for equilibrium matching, enhancing the performance of generative models.
- Demonstrates the applicability of the framework on image benchmarks, achieving superior results.
Read more
Beckmann Transport Models: From Autonomous Flows to One-Step Maps
Summary
This paper introduces a novel approach to flow matching using time-independent velocity fields, termed autonomous flows, to effectively map between two distributions, particularly when the target distribution is singular and supported on a lower-dimensional manifold. The authors demonstrate that the one-step generative map derived from these flows is the unique solution to a conservation equation, allowing for direct learning from samples. This work not only provides a theoretical foundation for the flux constraint in Beckmann's transportation problem but also corrects inconsistencies in existing methods. The proposed framework recovers established models, such as the closed-form Poisson-flow generative model, and offers a new loss function for equilibrium matching. The effectiveness of the autonomous flow and one-step map is validated through experiments on ImageNet 256x256, showing improved performance over previous methods.
Methodology
The authors utilize a theoretical framework based on divergence equations to establish the validity of autonomous flows in transporting distributions. They derive a conservation equation for the one-step map and propose a new loss function for training. The methodology includes experiments on 2D atomic targets and image datasets to validate the theoretical findings.
Results
The paper proves that the time-independent flow matching drift can transport the base distribution to the target when the latter is supported on a lower-dimensional manifold. It also establishes a conservation equation for the one-step map and shows that the corrected autonomous flow improves performance on image benchmarks compared to existing methods.
Implications
The findings have significant implications for generative modeling, particularly in simplifying the architecture and dynamics of generative models. The framework could lead to more efficient training and inference processes in various applications, including image generation and optimal transport problems.
A Neurosymbolic Approach for Explainable Early Diagnosis of Alzheimer's Disease
Audio & Speech
NLP
Interpretability
- Introduction of NeSyQuaKE, a neurosymbolic framework for automated extraction of qualitative knowledge from clinical audio.
- Utilization of pretrained foundation models for robust transcription and information extraction from verbal fluency tests.
- Construction of a Bayesian Network to reason about cognitive impairment markers and their relationships.
- Successful recovery of known clinical knowledge and identification of novel relationships between linguistic markers.
Read more
A Neurosymbolic Approach for Explainable Early Diagnosis of Alzheimer's Disease
Summary
This paper presents a novel automated pipeline, NeSyQuaKE, that leverages a neurosymbolic approach to extract qualitative knowledge from audio recordings of verbal fluency tests for the early diagnosis of Alzheimer's Disease (AD). Traditional methods for identifying AD markers are labor-intensive and require expert analysis, which limits scalability. The proposed method utilizes pretrained foundation models to transcribe and analyze raw audio data, extracting clinically relevant variables that are then structured into a Bayesian Network (BN). This BN facilitates reasoning about the relationships between cognitive impairment markers and their progression. The authors demonstrate that NeSyQuaKE not only recovers established clinical knowledge but also uncovers new qualitative relationships between linguistic markers and cognitive decline. The framework addresses the challenge of bridging the gap between unstructured audio data and structured clinical knowledge, ensuring that the derived metrics are mathematically consistent and verifiable. The evaluation on real-world datasets indicates the effectiveness of the approach in generating explainable insights into AD progression.
Methodology
The NeSyQuaKE framework integrates pretrained foundation models for audio transcription and information extraction, followed by the construction of a Bayesian Network to analyze the relationships between cognitive impairment markers. The QuaKE algorithm is employed for reasoning about these relationships, allowing for the derivation of qualitative influence statements that capture trends in the data.
Results
The evaluation of NeSyQuaKE on clinical verbal fluency test datasets showed that the system effectively recovers established clinical knowledge and generates new hypotheses regarding the qualitative relationships between cognitive impairment and its markers. The framework demonstrated a high degree of accuracy in identifying relevant linguistic features associated with Alzheimer's Disease.
Implications
The proposed framework has significant implications for the early diagnosis of Alzheimer's Disease, enabling scalable and automated analysis of verbal fluency tests. It can enhance clinical practices by providing explainable insights into cognitive decline, potentially leading to earlier interventions and improved patient outcomes.
Rethinking Pretraining for Specialized Design Data: Evidence from the JONES-19 Cultural Design Dataset
Computer Vision
- Domain-general pretraining improves performance but is not always necessary for specialized design data.
- Learning from scratch with multi-crop augmentation can effectively recover performance gains from pretraining.
- High-quality, curated datasets may be more beneficial than large-scale datasets in specialized design domains.
- The study emphasizes the unique challenges posed by design data compared to typical computer vision benchmarks.
Read more
Rethinking Pretraining for Specialized Design Data: Evidence from the JONES-19 Cultural Design Dataset
Summary
This paper investigates the effectiveness of pretraining strategies for Convolutional Neural Networks (CNNs) when applied to specialized design datasets, specifically the JONES-19 dataset, which consists of ornament designs based on The Grammar of Ornament. The authors compare two training strategies: (1) using ImageNet pretraining to leverage domain-general visual knowledge, and (2) training from scratch on the JONES-19 dataset. Through controlled experiments with ResNet18 and ResNet50, the study evaluates the impact of these strategies on classification accuracy. The findings reveal that while ImageNet pretraining enhances performance, training from scratch with multi-crop augmentation can recover and even surpass these gains. This suggests that for structured design data, localized design-driven representations are sufficient for effective learning, challenging the necessity of large-scale pretraining. The authors argue for the importance of curating smaller, high-quality datasets that encapsulate formal design principles over merely focusing on dataset scale, thereby advancing the understanding of visual knowledge transfer in specialized domains.
Methodology
The authors conducted controlled experiments using ResNet18 and ResNet50 architectures to compare the performance of CNNs trained with and without ImageNet pretraining, as well as with and without multi-crop augmentation, on the JONES-19 dataset.
Results
The study found that while ImageNet pretraining improved classification accuracy, training from scratch with multi-crop sampling effectively recovered these gains, indicating that local design-driven representations are sufficient for learning from specialized datasets.
Implications
The results suggest that in specialized design fields, focusing on the curation of smaller, high-quality datasets that reflect specific design principles may lead to better model performance than relying on large-scale datasets. This has potential applications in design scholarship and machine learning, promoting a more nuanced understanding of visual knowledge acquisition.
Technological Advances in Detecting and Managing Cognitive Impairment in Older Adults: Trends, Challenges, and Future Directions
Multimodal
- The paper synthesizes recent technological advancements in detecting cognitive impairment in older adults.
- A unified taxonomy and early-detection framework are proposed to enhance screening and intervention strategies.
- EEG and deep learning models show strong diagnostic accuracy, but many studies lack rigorous external validation.
- Clinical utility of blood-based biomarkers for Alzheimer's diagnosis is highlighted, along with approved anti-amyloid therapies.
Read more
Technological Advances in Detecting and Managing Cognitive Impairment in Older Adults: Trends, Challenges, and Future Directions
Summary
This paper addresses the growing challenge of cognitive decline in older adults, particularly focusing on Mild Cognitive Impairment (MCI) and dementia. It reviews recent technological advancements in detection and management, emphasizing the integration of neurophysiological signals (like EEG), neuroimaging techniques (MRI and PET), blood-based biomarkers, and digital markers. The authors propose a unified taxonomy and an integrative early-detection framework that connects screening to timely interventions. They highlight the diagnostic accuracy of various methods, including EEG markers and deep learning models, while cautioning that many results are based on small datasets lacking external validation. The paper also discusses the clinical utility of blood tests for Alzheimer's diagnosis and the approval of anti-amyloid therapies, alongside the maturation of multidomain lifestyle prevention strategies. Despite significant progress, challenges such as standardization, data privacy, and equitable deployment remain. The authors conclude that the future of cognitive impairment management lies in trustworthy, multimodal systems that facilitate early detection and personalized care.
Methodology
The authors conducted a comprehensive review of the literature on technological advancements in cognitive impairment detection and management, synthesizing findings from various disciplines and proposing a unified framework for early detection and intervention.
Results
The review indicates that EEG markers and advanced deep learning models can achieve high diagnostic accuracy for cognitive impairment. However, many reported results are based on limited datasets. The paper also notes the clinical approval of blood tests for Alzheimer's and the emergence of effective lifestyle interventions.
Implications
The findings suggest that integrating advanced technologies can lead to more accurate and timely detection of cognitive impairment, ultimately improving personalized care for older adults. Addressing the identified challenges could enhance the deployment of these technologies in clinical settings.
Isotonic Bradley-Terry Model for Paired Comparison Data
Theory
Optimization
Interpretability
- Proposes an isotonic Bradley-Terry model to improve prediction of win probabilities and player rankings.
- Addresses model misspecification by learning the inverse link function alongside rate parameters.
- Utilizes a (sub-)gradient method for parameter learning and isotonic regression for link function estimation.
- Demonstrates improved performance through experiments on synthetic and real-world sports data.
Read more
Isotonic Bradley-Terry Model for Paired Comparison Data
Summary
This paper addresses prediction problems in paired comparison data, such as predicting win probabilities between unmatched players and ranking players based on their strengths. Traditional models like the Bradley-Terry and Thurstone-Mosteller models utilize fixed inverse link functions to predict win probabilities based on learned rate parameters representing player strengths. However, these fixed functions can lead to model misspecification. To overcome this limitation, the author proposes an isotonic Bradley-Terry model that alternately learns both the rate parameters using a (sub-)gradient method and the inverse link function through isotonic regression. This approach ensures a monotonic improvement in training error and can handle situations where data is insufficient for strict ranking by allowing for exact ties. The effectiveness of the proposed model is validated through numerical experiments on synthetic data and real-world datasets from sports leagues, demonstrating improved prediction and ranking performance compared to traditional methods.
Methodology
The proposed model alternates between learning rate parameters using a (sub-)gradient method and estimating the inverse link function through isotonic regression. This iterative approach mitigates model misspecification and enhances prediction accuracy.
Results
The isotonic Bradley-Terry model showed significant improvements in predicting win probabilities and ranking players compared to traditional models, as evidenced by numerical experiments with both synthetic and real-world datasets from football, baseball, and tennis.
Implications
The proposed model can be applied in various domains involving paired comparisons, such as sports analytics, preference surveys, and any scenario where ranking based on comparative data is necessary. Its ability to adaptively learn the inverse link function enhances its applicability and robustness.
Disagree to Accelerate: Closing the Loop on Diffusion Feature Forecasts
Generative Models
Efficient ML
Computer Vision
- Introduces RACER, a closed-loop controller that adjusts trust in forecasts based on runtime reliability signals.
- Demonstrates that forecast reliability can be inferred from the disagreement between two cached forecasts.
- Achieves significant improvements in sampling speed and quality over existing open-loop methods across multiple models.
- Provides a deterministic error bound for the proposed shrinkage mechanism, ensuring reliability in predictions.
Read more
Disagree to Accelerate: Closing the Loop on Diffusion Feature Forecasts
Summary
This paper addresses the challenge of accelerating diffusion sampling in generative models by introducing a novel approach called RACER (Reliability-Aware Closed-loop controller with Exact Repay). Traditional methods have focused on improving the accuracy of feature forecasters, but they often fail to account for the variability in forecast reliability across different steps. The authors propose that the reliability of forecasts can be inferred from the disagreement between two cached forecasts, which serves as a runtime signal indicating when to trust a forecast. RACER utilizes this signal to implement a closed-loop control mechanism that adjusts the trust in forecasts dynamically. When uncertainty is detected, RACER reduces the reliance on the forecast and, if necessary, refreshes the feature by performing an additional denoiser evaluation, which is later compensated by skipping a subsequent evaluation. The authors derive a deterministic error bound for this shrinkage process and validate it empirically. The results demonstrate that RACER significantly outperforms existing open-loop baselines across multiple models and benchmarks, achieving faster sampling rates without sacrificing quality. This work highlights the importance of not only improving forecasting techniques but also effectively managing the trust in those forecasts during the sampling process.
Methodology
The authors developed RACER, a training-free closed-loop controller that utilizes the disagreement between two forecasts as a reliability signal. The controller adjusts the trust in forecasts dynamically, shrinking uncertain forecasts towards the last computed feature and refreshing features when necessary. The methodology includes deriving a deterministic error bound for the shrinkage and conducting empirical evaluations across various models and benchmarks.
Results
RACER outperformed the strongest open-loop baseline across multiple models (SD3.5-Large, FLUX.1-dev, Wan2.1-14B, and HunyuanVideo) and benchmarks (DrawBench, VBench, and COCO). It achieved a sampling speed improvement of 5.4 times on SD3.5 while maintaining quality, with a latency reduction to 0.76 times that of the baseline. The method also demonstrated robustness across different forecasting designs.
Implications
The findings suggest that managing the trust in forecasts is crucial for efficient diffusion sampling, which can lead to faster and more reliable generative models. This approach could be applied to other areas of machine learning where prediction reliability varies, potentially improving performance in real-time applications.
Fused Bayesian Flow Networks for Dual-Target Molecular Design
Generative Models
- FusedBFN integrates dual-target information through a unified distribution fusion framework.
- It is the first formulation of a fused Bayesian flow network for dual-target generation, utilizing pretrained models without additional training.
- The framework incorporates a chemically aware alignment strategy for better integration of structural features from both targets.
- Extensive experiments validate the effectiveness of FusedBFN in generating high-affinity dual-target molecules.
Read more
Fused Bayesian Flow Networks for Dual-Target Molecular Design
Summary
The paper introduces FusedBFN, a novel framework for dual-target molecular design that aims to generate 3D molecules capable of interacting with two target proteins simultaneously. This approach addresses the limitations of existing dual-target methods, which often focus on sequence generation or rely on predictive drift terms that do not fully integrate structural information from both targets. FusedBFN employs a product-of-experts formulation to fuse distributions conditioned on two target contexts within a unified continuous parameter space. To overcome the scarcity of dual-target structural data, the authors leverage a pretrained target-aware Bayesian flow network (BFN) model as a backbone, allowing for the extension of knowledge from single-target datasets. The framework also introduces a chemically aware prior-based alignment method and a prior-free pocket alignment strategy to construct aligned dual-target contexts. Extensive experiments demonstrate that FusedBFN generates molecules with strong binding affinity toward both targets while maintaining favorable molecular properties, marking a significant advancement in dual-target drug design.
Methodology
The methodology involves formulating dual-target generation as a distribution fusion problem using a product-of-experts approach. The authors utilize a pretrained target-aware Bayesian flow network as a backbone and introduce alignment strategies that consider chemical semantics to construct dual-target contexts. The Bayesian updating process propagates fused target-context information into the parameter space, facilitating the generation of molecules that can bind to both targets.
Results
The results show that FusedBFN can successfully generate molecules with strong binding affinities for both target proteins while preserving desirable molecular properties. The framework outperforms existing dual-target approaches, demonstrating its effectiveness in addressing the challenges of dual-target drug design.
Implications
The implications of this research extend to the field of drug discovery, particularly in developing polypharmacological compounds that can target multiple proteins simultaneously. This could lead to more effective treatments for complex diseases and enhance the therapeutic efficacy of new drugs.
Empowering Credit Risk Detection in Weixin Pay with Billion-Scale Deep Graph Learning
Graph Learning
- Introduces a novel framework (RAOS) for credit risk detection using GNNs.
- Addresses scalability issues in industrial GNN applications by utilizing overlapping subgraphs.
- Implements budget-constrained sampling to preserve critical risk patterns.
- Develops a cross-subgraph consistency alignment mechanism for improved representation.
Read more
Empowering Credit Risk Detection in Weixin Pay with Billion-Scale Deep Graph Learning
Summary
This paper addresses the critical challenge of credit risk detection in digital financial ecosystems, particularly within Weixin Pay, where billions of users are involved. The authors highlight the limitations of traditional risk detection methods that rely on expert rules and statistical feature engineering, which fail to capture the complex, non-linear relationships inherent in user behavior and risk events. To tackle this issue, they propose a novel Risk-Aware Overlapping Subgraph Learning (RAOS) framework that leverages Graph Neural Networks (GNNs) to effectively mine risk patterns from heterogeneous user-risk graphs. The RAOS framework introduces a method for constructing base partitions that ensure load balance while performing budget-constrained sampling to select informative long-tail nodes. This approach preserves critical risk diffusion patterns and mitigates noise. Additionally, a cross-subgraph consistency alignment mechanism is designed to harmonize local representations into a globally consistent latent space. The proposed method demonstrates significant improvements in risk detection accuracy and scalability when tested on Weixin Pay's production dataset, outperforming existing strategies.
Methodology
The authors developed the RAOS framework, which constructs base partitions for load balancing and employs budget-constrained sampling to select informative long-tail nodes. They also designed a cross-subgraph consistency alignment mechanism to ensure representation consistency across overlapping nodes, facilitating effective risk detection in large-scale graphs.
Results
The experiments conducted on Weixin Pay's production dataset revealed that the RAOS framework significantly outperformed existing credit risk detection strategies, achieving higher accuracy and scalability in identifying potential fraud risks among billions of users.
Implications
The findings suggest that the RAOS framework can be effectively applied in real-world financial systems to enhance credit risk detection capabilities, thereby minimizing financial losses and promoting the sustainability of digital financial services.
Stochastic Sequential Search in Very-High-Dimensional Feature Selection
Optimization
Efficient ML
Theory
- Introduces Stochastic Sequential Search (SSS) to improve feature selection in very-high-dimensional spaces.
- Utilizes a budgeted sampled step operator to evaluate a fixed number of candidates, reducing computational cost.
- Demonstrates that sSFFS retains high criterion values while requiring significantly fewer evaluations compared to traditional methods.
- Shows superior performance on benchmark datasets like madelon, gisette, and reuters corpus.
Read more
Stochastic Sequential Search in Very-High-Dimensional Feature Selection
Summary
This paper addresses the challenges of feature selection in very-high-dimensional spaces, where traditional sequential subset search methods become impractical due to their exhaustive nature. The authors introduce a novel approach called Stochastic Sequential Search (SSS), specifically focusing on the stochastic counterpart of floating search (sSFFS). By implementing a budgeted sampled step operator, the method evaluates a fixed number of candidate features per step, drawn from dependency-aware statistics learned online during the search process. This approach allows for efficient exploration of feature interactions while maintaining reachability of all features. The experimental results demonstrate that sSFFS retains high criterion values across various benchmark datasets, significantly outperforming traditional methods in both efficiency and effectiveness. The authors provide a verified implementation of their method in the Feature Selection Toolbox, making it accessible for further research and application.
Methodology
The authors replace the exhaustive candidate evaluation in traditional sequential search with a budgeted sampled step, where a fixed number of candidates are selected using temperature-controlled softmax sampling from online learned statistics. This method allows for efficient exploration of feature interactions while ensuring that all features remain reachable throughout the search process.
Results
The sSFFS method was tested on various datasets, including the 500-dimensional madelon benchmark, where it achieved at least 97% of the full-SFFS criterion value while using only a quarter of the evaluations. On the 5,000-dimensional gisette dataset, sSFFS maintained criterion values near 0.99 across subset sizes, outperforming traditional methods. In the 10,105-dimensional reuters corpus, sSFFS demonstrated superior accuracy and efficiency compared to Best Individual Features (BIF) and Dependency-Aware Features (DAF) methods.
Implications
The proposed stochastic feature selection method has significant implications for high-dimensional data analysis, particularly in fields such as text categorization and bioinformatics, where the number of features can be extremely large. The ability to efficiently select relevant features while considering their interactions can lead to improved model performance and reduced computational costs.
Upper-Expectile Multi-Step Q-Learning for Off-Policy Reinforcement Learning
Reinforcement Learning
Robotics
Theory
- ENQ reduces pessimistic bias in off-policy reinforcement learning by using asymmetric expectile loss.
- The method is a drop-in replacement for traditional n-step TD learning, requiring no additional action likelihoods or complex constraints.
- Theoretical guarantees show that ENQ maintains convergence properties and reduces bias under both deterministic and stochastic dynamics.
- Empirical results demonstrate that ENQ is competitive with LQL across multiple tasks while achieving higher training efficiency.
Read more
Upper-Expectile Multi-Step Q-Learning for Off-Policy Reinforcement Learning
Summary
This paper introduces Expectile n-step Q-learning (ENQ), a novel approach to off-policy reinforcement learning that addresses the pessimistic bias associated with multi-step returns. Traditional n-step temporal-difference (TD) learning suffers from this bias due to the reliance on suboptimal logged actions, which can degrade performance over long horizons. ENQ replaces the symmetric n-step TD loss with an asymmetric expectile loss, allowing for more effective reward propagation while minimizing the influence of poor action continuations. The authors prove that the ENQ operator is a γn-contraction and demonstrate that, under deterministic dynamics, the bias vanishes at the optimal action-value function Q*. They also show that ENQ is competitive with Long-Horizon Q-learning (LQL) across various tasks, achieving higher training throughput and benefiting from a ten-critic ensemble in controlled experiments. The method is simple to implement, requiring only a single expectile level as a hyperparameter, making it a flexible alternative to existing off-policy learning methods.
Methodology
The authors propose the Expectile n-step Q-learning (ENQ) method, which modifies the standard n-step TD update by employing an asymmetric expectile loss function. This approach emphasizes high-return logged actions and requires only evaluations at the initial and final states of the n-step update, avoiding the need for intermediate evaluations or trajectory-level constraints. Theoretical analysis is provided to establish the properties of the ENQ operator, and extensive empirical evaluations are conducted across various manipulation and navigation tasks.
Results
ENQ was tested across 27 task instances using a fixed expectile level of τ = 0.8 and a backup horizon of n = 4. The results indicate that ENQ is competitive with LQL in terms of overall performance and achieves higher training throughput. The method also shows robustness in its performance without requiring task-specific tuning.
Implications
The ENQ method has significant implications for improving the efficiency and effectiveness of off-policy reinforcement learning, particularly in scenarios where logged data is sparse or expensive to collect. Its simplicity and flexibility make it a promising candidate for various applications in robotics and continuous control tasks.
DFSC: Error-Controlled Differentiable Mittag-Leffler Propagation for Fractional Scientific Machine Learning
Theory
Efficient ML
Optimization
- DFSC separates known fractional propagation from unknown corrections, optimizing both simultaneously.
- The framework includes an adaptive algorithm for error control and certified diagnostics.
- DFSC significantly reduces computational time for repeated queries while maintaining accuracy.
- The system is validated across multiple physical domains and experimental conditions.
Read more
DFSC: Error-Controlled Differentiable Mittag-Leffler Propagation for Fractional Scientific Machine Learning
Summary
The paper introduces DFSC, a novel framework for fractional scientific machine learning that leverages differentiable Mittag-Leffler propagation. DFSC is designed to efficiently handle fractional differential equations by separating known fractional propagation from data-driven corrections, allowing for joint optimization of fractional orders and neural network parameters. The core component, the Mittag-Leffler Spectral Layer (MLSL), is implemented as a PyTorch layer that supports various operator paths and allows for adaptive error control. The framework includes a certified algorithm that provides error budgets and diagnostics, enhancing reliability in numerical computations. DFSC demonstrates significant performance improvements, reducing repeated-query times on CPU and GPU by factors of 4.61 to 16.22, while maintaining accuracy across a range of applications. The authors validate DFSC through extensive testing, including synthetic learning tasks and real experimental conditions, establishing it as a robust tool for scientific machine learning tasks involving fractional dynamics.
Methodology
The authors developed DFSC as a PyTorch-based framework that incorporates the Mittag-Leffler Spectral Layer (MLSL) for differentiable fractional propagation. The framework allows for adaptive error control by adjusting the truncation depth or Lanczos dimension based on user-defined tolerances. It supports various operator paths and integrates seamlessly with neural network architectures, enabling joint optimization of fractional orders and neural parameters.
Results
DFSC achieved a median bound/error effectivity of 1.246 across 59 reference cases, demonstrating its reliability in error estimation. The framework reduced repeated-query times by 4.61 to 7.11 times on CPU and 13.07 to 16.22 times on an RTX 5070, showcasing its efficiency. The validation included 27 inverse matrix cases, confirming the framework's capability to find full-rank local curvature.
Implications
DFSC provides a powerful tool for researchers and practitioners in scientific machine learning, particularly for applications involving fractional differential equations. Its ability to efficiently handle complex dynamics while maintaining accuracy opens new avenues for modeling and solving real-world problems in various fields such as physics, engineering, and finance.
Frugal Bayesian Optimization: Scalable Surrogates for Data- and Resource-Limited Discovery
Optimization
Efficient ML
- Gaussian Process-based BO has high computational costs without superior performance.
- Scalable surrogate models can achieve equal or better optimization results at lower costs.
- A surrogate-recommendation framework is introduced to aid in selecting appropriate models.
- FruBO serves as a reproducible baseline for Bayesian Optimization in resource-constrained environments.
Read more
Frugal Bayesian Optimization: Scalable Surrogates for Data- and Resource-Limited Discovery
Summary
This paper presents a systematic study of Bayesian Optimization (BO) with a focus on computational efficiency and optimization performance. The authors benchmark four surrogate models—Gaussian Processes (GPs), Random Forests, NGBoost, and Bayesian Adaptive Spline Surfaces—across various benchmark functions and real-world datasets in fields such as materials science and robotics. The findings reveal that GP-based BO incurs significant computational overhead without providing superior optimization results compared to more scalable alternatives. To address this issue, the authors introduce a surrogate-recommendation framework that helps select the most suitable surrogate model based on dataset characteristics. This work establishes Frugal Bayesian Optimization (FruBO) as a reproducible, compute-aware baseline for BO, offering practical guidance for surrogate selection in resource-limited scenarios.
Methodology
The authors conducted a comprehensive evaluation of four surrogate models by applying them to eight benchmark functions and nine real-world datasets. They analyzed the trade-offs between optimization quality and computational frugality, focusing on time and memory overhead associated with each surrogate model.
Results
The study found that Gaussian Process-based BO consistently resulted in the highest computational costs without delivering better optimization or sample efficiency. In contrast, the other surrogate models demonstrated comparable or superior performance with significantly reduced computational demands. The surrogate-recommendation framework successfully predicted suitable models based on dataset characteristics, enhancing the practicality of BO in limited-resource settings.
Implications
The findings suggest that researchers and practitioners can optimize their resource allocation by selecting more efficient surrogate models for Bayesian Optimization. This has significant implications for scientific discovery and engineering applications where data and computational resources are limited.
Assessing the Generalization of Graph Neural Networks for Fault Location Across Increasing Distributed Energy Resource Penetration Levels
Graph Learning
Time Series
- STGATv2 outperforms traditional and other neural network models in fault location tasks.
- Generalization performance varies significantly with DER penetration levels.
- Training at higher DER penetration levels leads to better retention of performance at lower levels.
- STGATv2 demonstrates robustness against measurement noise, maintaining high F1 scores.
Read more
Assessing the Generalization of Graph Neural Networks for Fault Location Across Increasing Distributed Energy Resource Penetration Levels
Summary
This paper addresses the challenges of accurately locating faults in power distribution networks, particularly as the penetration of distributed energy resources (DERs) increases. The authors investigate the performance of spatio-temporal graph neural networks (STGNNs), specifically the spatio-temporal graph attention network (STGATv2), in comparison to traditional machine learning methods and other neural network architectures. The study benchmarks STGATv2 against purely temporal (GRU) and purely spatial (GATv2) models, as well as traditional machine learning approaches, to evaluate their effectiveness in fault location across various DER penetration levels (10%, 25%, and 50%). The results indicate that STGATv2 consistently outperforms the other models, achieving macro F1 scores of 92-94% in in-distribution scenarios. Notably, the generalization of the models varies with DER penetration; training at higher penetration levels (50%) retains performance at lower levels, while training at lower levels (10%) leads to significant performance degradation at higher levels. Under realistic measurement noise, STGATv2 maintains over 85% F1 score, highlighting its robustness compared to GRU, which drops to as low as 33.5% F1. This study emphasizes the importance of topological awareness in fault location methods for active distribution networks.
Methodology
The authors developed a spatio-temporal graph neural network framework (STGNN) that combines temporal modeling through a gated recurrent unit (GRU) with spatial feature extraction using an improved graph attention layer. The framework processes node-level temporal representations and aggregates outputs for fault location predictions.
Results
STGATv2 achieved macro F1 scores of 92-94% in in-distribution tests. It retained 81-84% F1 score when trained at 50% DER penetration and tested at lower levels, while performance dropped significantly for models trained at 10% DER. Under noise conditions, STGATv2 maintained over 85% F1, contrasting with GRU's drop to 33.5% F1.
Implications
The findings suggest that incorporating spatio-temporal modeling and topological awareness in fault location methods can significantly enhance the reliability of power distribution networks, especially as DER penetration increases. This has implications for grid management and the integration of renewable energy sources.
Agentic Graph Token Reasoning
Large Language Models
Graph Learning
Reinforcement Learning
- Introduces Agentic Graph Token reasoning for dynamic graph analysis.
- Transforms graph tokenization from a static preprocessing step into an iterative reasoning process.
- Utilizes a three-stage training pipeline to enhance model performance and adaptability.
- Achieves significant improvements over existing methods across multiple graph domains.
Read more
Agentic Graph Token Reasoning
Summary
This paper introduces Agentic Graph Token (AGT) reasoning, a novel approach that enhances graph analysis by integrating graph tokenization into the reasoning process of large language models (LLMs). Traditional methods utilize static graph tokens, which encode a predefined graph view before the model encounters the target question, limiting the model's ability to adapt its reasoning based on the evolving context. In contrast, AGT allows the model to dynamically select which graph view to encode at each reasoning step, effectively treating graph tokenization as an iterative process. The authors propose a three-stage training pipeline: (1) self-supervised tasks for reading heterogeneous graph tokens, (2) a token-robust trajectory stage with a consistency regularizer, and (3) preference optimization that aligns graph-token evidence with node-text evidence. Extensive evaluations across seven graph domains demonstrate that AGT significantly outperforms existing baselines and exhibits strong zero-shot transfer capabilities to unseen domains. This approach not only improves performance on challenging samples but also represents a shift towards a more agentic paradigm in LLM-based graph analysis.
Methodology
The methodology involves a three-stage training process: (1) self-supervised tasks to familiarize the model with heterogeneous graph tokens, (2) a token-robust trajectory stage that employs a consistency regularizer to ensure coherence between graph-token and node-text evidence, and (3) preference optimization that rewards reasoning trajectories aligning both types of evidence.
Results
The AGT model outperforms a wide range of baseline methods across seven different graph domains, showing substantial improvements in accuracy and reasoning capabilities. The model also exhibits strong zero-shot transfer abilities to previously unseen domains, indicating its robustness and adaptability.
Implications
The findings suggest that integrating agentic reasoning into graph analysis can significantly enhance the performance of LLMs in understanding and processing relational data. This approach could be applied in various fields such as social network analysis, bioinformatics, and recommendation systems, where dynamic reasoning over complex relationships is crucial.
Multi-Source Dynamic Graph Learning for Compound-Flood Forecasting in Managed Coastal Systems
Graph Learning
Time Series
- Introduces an anchored dynamic-graph forecasting framework for compound flood prediction.
- Integrates multi-source hydrometeorological and operational data to enhance local forecasts.
- Evaluates forecasting performance using both global error metrics and event-scale characteristics.
- Demonstrates improved reliability for sustained high-water plateaus in managed coastal systems.
Read more
Multi-Source Dynamic Graph Learning for Compound-Flood Forecasting in Managed Coastal Systems
Summary
This paper addresses the challenge of forecasting compound flooding in managed coastal systems, where hydrological conditions and water management activities are monitored across multiple stations. Traditional forecasting models often fail to accurately predict prolonged high-water plateaus, which are critical for flood early warning. The authors propose an anchored forecasting framework that integrates cross-site information through state- and lead-dependent bounded residual corrections. This framework utilizes a multi-source regime representation to adaptively calibrate inter-site relationships, allowing for targeted adjustments while maintaining the stability of local forecasts. The study emphasizes the importance of evaluating event-scale high-water characteristics rather than relying solely on global error metrics. Experiments demonstrate that the proposed method enhances prediction reliability for sustained high-water conditions while preserving accuracy during routine hydrological scenarios. The findings support improved flood early warning and water management decision-making.
Methodology
The authors formulate high-water forecasting as a distributed observation problem, combining various environmental and operational factors. They develop a channel-wise PatchTST model to estimate target trajectories and utilize an input-dependent heterogeneous graph to summarize the monitoring network. The framework employs a multi-source regime representation to control corrections based on target and lead times, with lead-dependent bounds to limit adjustments.
Results
The proposed framework outperformed traditional forecasting methods in predicting sustained high-water episodes while maintaining accuracy during normal conditions. The evaluation metrics included both mean squared error (MSE) and mean absolute error (MAE), alongside event-specific metrics for detection, onset, peak, and duration of high-water episodes.
Implications
The findings suggest that integrating multi-source data can significantly enhance flood forecasting capabilities, which is crucial for early warning systems and effective water management strategies in coastal regions. The approach may also be applicable to other environmental forecasting challenges where distributed observations are relevant.
Interpretable Machine Learning for Traffic Congestion Prediction: Unveiling the Impact of Different COVID-19 Periods
Time Series
Interpretability
- Traffic congestion prediction is significantly affected by COVID-19 and associated restrictions.
- Bidirectional LSTM (Bi-LSTM) outperforms other machine learning models in predicting traffic congestion.
- Integrated Gradients and SHAP techniques enhance the interpretability of machine learning predictions.
- New COVID-19 cases have a strong negative influence on traffic congestion during lockdowns.
Read more
Interpretable Machine Learning for Traffic Congestion Prediction: Unveiling the Impact of Different COVID-19 Periods
Summary
This study investigates traffic congestion prediction in Alameda County, California, during pre-lockdown, lockdown, and post-lockdown periods of the COVID-19 pandemic. The authors highlight the significant changes in urban travel behavior due to the pandemic and government restrictions, which complicate congestion prediction. They incorporate external data such as weather conditions, seasonality, and COVID-19-related variables, employing Recursive Feature Elimination with Cross-Validation to identify key features while mitigating overfitting. The study evaluates multiple machine learning models, including Support Vector Regression (SVR), Multiple Linear Regression (MLR), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM) networks. The LSTM model is optimized using an adaptive parameter selection approach to enhance accuracy. Results indicate that the bidirectional LSTM (Bi-LSTM) consistently outperforms other models across all periods, attributed to its ability to capture temporal dependencies effectively. To improve interpretability, Integrated Gradients (IG) and SHapley Additive exPlanations (SHAP) are applied, revealing that new COVID-19 cases negatively impact traffic congestion during lockdowns, while post-pandemic behaviors show increased caution among travelers. The findings underscore the lasting effects of the pandemic on traffic patterns and the importance of incorporating COVID-19 factors in congestion prediction models.
Methodology
The study employs Recursive Feature Elimination with Cross-Validation to identify important features from external data sources. It trains and optimizes multiple machine learning models, including SVR, MLR, RNN, and LSTM, with a focus on adaptive parameter tuning for the LSTM model. The models are evaluated using Normalized Root Mean Square Error.
Results
The Bi-LSTM model consistently outperformed SVR, MLR, and RNN across all COVID-19 periods. The analysis revealed that new COVID-19 cases negatively influenced traffic congestion during lockdowns, while post-pandemic behaviors indicated increased caution among travelers, affecting traffic patterns.
Implications
The findings suggest that traffic congestion prediction models must incorporate COVID-19-related factors to enhance accuracy. The study provides insights into how urban mobility has changed due to the pandemic, which can inform future traffic management strategies and urban planning.
Question Begets Question: Self-Evolving Curriculum for Reinforcement Fine-Tuning on Competition Mathematics
Reinforcement Learning
Large Language Models
NLP
- Introduces a scalable method (QbQ) for generating diverse synthetic training data.
- Demonstrates that models can improve by training on problems they can mostly solve, rather than their hardest failures.
- Achieves a significant performance increase (pass@1 from 5.6% to 16.5%) through a self-evolving curriculum.
- Highlights the limitations of static training methods and the benefits of dynamic data tailoring.
Read more
Question Begets Question: Self-Evolving Curriculum for Reinforcement Fine-Tuning on Competition Mathematics
Summary
This paper addresses the challenges of fine-tuning language models for complex tasks, specifically competition mathematics, where traditional methods struggle due to data scarcity, lack of reasoning traces, and apparent learning ceilings. The authors introduce a novel approach called Question-begets-Question (QbQ), which generates diverse synthetic problem variants to enhance training data. They employ reinforcement learning exclusively on problem statements and final answers, avoiding reliance on reasoning traces. The study reveals that static training methods plateau in performance, while their self-evolving curriculum dynamically tailors training data based on the model's current capabilities. This iterative approach allows the model to break through performance ceilings, achieving a pass rate of 16.5% on the AIME benchmark after 20 training rounds, compared to lower rates with traditional methods. The findings suggest that training on problems the model can mostly solve leads to greater improvements, challenging conventional wisdom about learning from mistakes.
Methodology
The authors developed a self-evolving curriculum that involves three stages: self-evaluation of the model's performance on previous problems, generation of new problem variants based on those the model can mostly solve, and training on these new variants using reinforcement learning without exposing the model to reasoning traces.
Results
The self-evolving curriculum led to a pass rate of 16.5% on the AIME benchmark after 20 rounds of training, significantly higher than the initial 5.6% and the 14.5% achieved through static methods. This indicates that the model can effectively learn from tailored problem variants.
Implications
The findings suggest that dynamic and adaptive training strategies can significantly enhance the performance of language models on complex tasks. This approach could be applied to other domains where data is scarce or reasoning traces are unavailable, potentially improving AI capabilities in various fields.
Who Belongs in the Eval Set? A Capability-Taxonomy-Driven Pipeline for Curating Regression Eval Sets in Agent-Extensibility Platforms
NLP
Large Language Models
Optimization
- Introduces a novel pipeline for curating regression evaluation sets in agent-extensibility platforms.
- Utilizes a capability taxonomy to assess and manage incoming customer evaluation sets.
- Incorporates a hybrid classifier and an Invocation Quality rater to evaluate query capabilities and quality.
- Addresses the limitations of existing evaluation frameworks that focus on customer-side evaluations.
Read more
Who Belongs in the Eval Set? A Capability-Taxonomy-Driven Pipeline for Curating Regression Eval Sets in Agent-Extensibility Platforms
Summary
This paper addresses the challenge of curating regression evaluation sets for agent-extensibility platforms, particularly in the context of Microsoft 365 Copilot. The authors identify a regression-economics paradox where each customer onboarding brings a unique evaluation set, but the platform must manage a limited regression set due to release cadence constraints. Existing frameworks focus on customer-side evaluations, leaving a gap in platform-side curation. The proposed solution is a capability-taxonomy-driven pipeline that processes incoming evaluation sets by projecting queries into a predefined capability taxonomy. The pipeline consists of three main components: a classifier that determines the capabilities exercised by each query, an Invocation Quality (IQ) rater that scores the thoroughness of capability exercise, and a consolidator that makes decisions on query admission based on coverage and quality. This approach allows for a dynamic and efficient curation of regression sets that adapt to incoming customer evaluations while maximizing capability coverage.
Methodology
The methodology involves a three-component pipeline: a classifier that assesses each query against a capability taxonomy, an Invocation Quality rater that evaluates how well queries exercise capabilities, and a consolidator that makes admission decisions based on coverage and quality metrics. This approach allows for a systematic and efficient evaluation of incoming queries, ensuring that the regression set remains relevant and effective.
Results
The implementation of the proposed pipeline in Microsoft 365 Copilot demonstrates its effectiveness in managing regression evaluation sets. The pipeline successfully categorizes incoming queries, evaluates their quality, and maintains a regression set that captures a diverse range of capabilities while adhering to release constraints.
Implications
The findings suggest that the capability-taxonomy-driven approach can be applied to various regression evaluation set curation problems across different platforms. This could lead to improved efficiency in managing evaluation sets, better resource allocation, and enhanced performance of agent-extensibility platforms.
ReBRAC-v2: The Return of the King
Reinforcement Learning
Robotics
Optimization
- ReBRAC-v2 modernizes a conventional behavior-regularized actor-critic algorithm while maintaining simplicity.
- The algorithm integrates an exact-likelihood normalizing flow actor and a classification-based residual critic.
- A shared configuration was developed through Bayesian proposals, allowing for efficient adaptation across tasks.
- ReBRAC-v2 achieved state-of-the-art performance on OGBench and D4RL benchmarks, outperforming existing methods.
Read more
ReBRAC-v2: The Return of the King
Summary
The paper introduces ReBRAC-v2, a modernized version of the behavior-regularized actor-critic algorithm for offline reinforcement learning (RL). The authors explore whether significant improvements can be achieved by enhancing a conventional algorithm while maintaining its simplicity. ReBRAC-v2 employs an exact-likelihood normalizing flow as the RL actor and integrates various regularization techniques, including mixed behavior cloning, to optimize performance. The methodology includes a classification-based residual critic, staged optimization, and multi-sample action selection. The authors conducted extensive experiments across six challenging OGBench tasks, utilizing a shared configuration derived from approximately 600 Bayesian proposals. The results demonstrate that ReBRAC-v2 outperforms existing methods, achieving an average score of 74.8 across ten OGBench categories, significantly higher than the next-best result of 52.3. Additionally, it ranks first in eight categories and achieves the highest scores on D4RL AntMaze and Adroit benchmarks. The findings indicate that systematic engineering and integration of modern techniques can yield state-of-the-art performance without abandoning the foundational principles of offline RL.
Methodology
ReBRAC-v2 employs a behavior-regularized actor-critic framework, utilizing an exact-likelihood normalizing flow as the actor. It combines mixed behavior cloning with MSE and MAE regularization, and uses a classification-based residual critic. The training process involves staged optimization and multi-sample action selection to enhance performance.
Results
ReBRAC-v2 achieved an average score of 74.8 across ten OGBench categories, significantly surpassing the next-best result of 52.3. It ranked first in eight categories and achieved the highest scores of 90.2 on D4RL AntMaze and 33.6 on Adroit, demonstrating superior performance compared to other contemporary methods.
Implications
The findings suggest that conventional offline RL algorithms can be significantly enhanced through systematic modernization, potentially leading to broader applications in areas where online exploration is costly or unsafe. This approach may influence future research in offline RL and related fields.
Mitigating Backdoors via Decoy Shortcuts and Knowledge Decoupling
Computer Vision
- Introduces Trapping and Removing (TR) defense mechanism against backdoor attacks.
- Utilizes a lightweight shortcut branch to trap backdoor knowledge during training.
- Employs knowledge decoupling with entropy-based weight assignment for improved learning.
- Demonstrates effectiveness across multiple datasets and model architectures.
Read more
Mitigating Backdoors via Decoy Shortcuts and Knowledge Decoupling
Summary
This paper addresses the critical issue of backdoor attacks in deep neural networks, particularly when training with third-party data. The authors identify that backdoor behaviors are often absorbed by simpler parallel branches during joint training. To counteract this, they propose a novel defense mechanism called Trapping and Removing (TR), which introduces a lightweight shortcut branch as a 'honeypot' to capture backdoor knowledge during training. This approach allows for the removal of backdoors post-training by discarding the shortcut, without the need for additional data. Furthermore, the authors implement a knowledge decoupling strategy that uses entropy-based weight assignment to ensure that poisoned samples are directed through the honeypot while the main network focuses on benign data. An automatic shortcut generation strategy is also introduced to enhance generalization across different model architectures. The effectiveness of the TR method is demonstrated through extensive experiments on four benchmark datasets and five model architectures, showing that it successfully mitigates various backdoor attacks while maintaining performance on benign data.
Methodology
The authors propose a training-time defense mechanism that includes a decoy shortcut branch to trap backdoor knowledge. They implement a knowledge decoupling strategy that dynamically directs poisoned samples to the shortcut while allowing the main network to focus on benign samples. An adaptive strategy for automatic shortcut generation is also introduced to enhance the method's applicability across various architectures.
Results
The TR method was evaluated on four benchmark datasets and five model architectures, showing robust performance against ten different backdoor attack strategies. The results indicate that the proposed approach effectively mitigates backdoor threats while preserving the model's performance on benign data, outperforming existing state-of-the-art defenses.
Implications
The findings suggest that the TR method can be a practical solution for enhancing the security of deep learning models against backdoor attacks, particularly in scenarios where training data is sourced from third parties. This has significant implications for safety-critical applications, such as autonomous driving and healthcare, where model integrity is paramount.
Gated Q-learning: Add Off-Policy Bias to Taste
Reinforcement Learning
Theory
Efficient ML
- Introduces Gated Q-learning, a new framework for managing off-policy bias in Q-learning.
- Utilizes a state-action-dependent gating mechanism to balance eligibility trace attenuation.
- Provides theoretical proofs regarding the contraction properties of the expected operator.
- Empirical results show improved initial learning speed compared to traditional Q(λ) methods.
Read more
Gated Q-learning: Add Off-Policy Bias to Taste
Summary
This paper addresses the longstanding challenge of managing off-policy bias in Q-learning, particularly in the context of multistep credit assignment, which is essential for sample-efficient reinforcement learning. Traditionally, practitioners have faced a binary choice: either eliminate bias at the expense of truncated eligibility traces (as in Watkins’ Q(λ)) or ignore bias for faster learning, leading to detrimental errors (as in Peng’s Q(λ)). The author introduces Gated Q-learning, a novel algorithmic framework that interpolates between these extremes using a continuous, state-action-dependent gating mechanism. This mechanism selectively attenuates eligibility traces in an exploration-aware manner, allowing for longer credit-assignment horizons without the reliance on importance sampling. The paper provides a rigorous theoretical foundation for this approach, proving that the expected operator remains a contraction mapping and deriving its exact fixed point. Empirical evaluations demonstrate that Gated Q-learning enables faster initial learning compared to both extremes, offering a customizable alternative to importance sampling and enhancing the effective multistep horizon and off-policy bias management in Q-learning agents.
Methodology
The methodology involves the development of Gated Q-learning, which employs a continuous gating mechanism that adjusts eligibility traces based on the state-action pair. The paper includes a theoretical analysis of the value-function operator and conducts extensive empirical evaluations through a hyperparameter sweep in a random-walk environment to visualize the performance trade-offs.
Results
The results indicate that Gated Q-learning allows for longer credit-assignment horizons and faster initial learning compared to both Watkins’ Q(λ) and Peng’s Q(λ). The empirical evaluations provide detailed heatmaps illustrating the influence of various hyperparameters on performance, confirming the effectiveness of the proposed method.
Implications
The implications of this research extend to improving the efficiency and effectiveness of Q-learning algorithms in reinforcement learning, particularly in environments where off-policy learning is crucial. Gated Q-learning could enhance the performance of deep reinforcement learning architectures and offline reinforcement learning applications.
AOS: Adaptive Optimizer Switching via Training-State Signals for Faster Convergence and Better Generalization
Optimization
- AOS-R dynamically switches optimizers based on real-time training signals.
- The framework preserves momentum across optimizer transitions to avoid performance drops.
- AOS-R achieves faster convergence and better generalization compared to traditional single-optimizer methods.
- Empirical results show AOS-R outperforms individual optimizers in multiple benchmarks.
Read more
AOS: Adaptive Optimizer Switching via Training-State Signals for Faster Convergence and Better Generalization
Summary
The paper introduces AOS-R (Adaptive Optimizer Switching, Rule-Based), a novel framework designed to enhance the training of deep neural networks by dynamically switching between optimizers based on the current training state. Traditional single-optimizer approaches are inadequate as they do not adapt to the distinct phases of optimization. AOS-R addresses this by monitoring six online gradient-space signals: gradient noise scale, Hutchinson curvature trace, loss stagnation, update stability ratio, gradient stability index, and loss improvement ratio. This allows the framework to switch among AdamW, SGD with momentum, and Lion optimizers as the optimization landscape evolves. The method includes state-preserving momentum transfer to maintain momentum across transitions and a learning-rate bridge to prevent accuracy degradation. Empirical results demonstrate that AOS-R achieves superior performance on CIFAR-100 with a top-1 accuracy of 78% in 81 epochs, significantly faster than the other optimizers. Across eight model-dataset benchmarks, AOS-R outperformed individual optimizers in six cases, achieving a mean accuracy gain of +0.4 percentage points and a convergence speedup of 0.80x over AdamW under a unified hyperparameter setting.
Methodology
AOS-R employs a signal-driven approach that utilizes six interpretable online signals to determine the optimal optimizer at any given time during training. The framework includes mechanisms for state-preserving momentum transfer and a learning-rate continuity bridge to ensure smooth transitions between optimizers.
Results
AOS-R achieved a top-1 accuracy of 78% on CIFAR-100 in 81 epochs, which is 26% faster than AdamW, 43% faster than SGD-M, and 16% faster than Lion. It also demonstrated superior performance across eight model-dataset combinations, achieving the best accuracy in six cases and a mean accuracy improvement of +0.4 percentage points, along with a 0.80x speedup in convergence compared to AdamW.
Implications
The AOS framework has significant implications for optimizing deep learning training processes, particularly in scenarios where different phases of training require different optimization strategies. It can lead to more efficient training workflows and improved model performance across various tasks.
CRIP: Channel Level Representation Injection for Personalized One-Shot Federated Learning
Federated Learning
- Introduces CRIP, a personalized OSFL framework that focuses on channel-level feature alignment.
- Addresses the limitations of existing OSFL methods that rely on parameter-level aggregation under domain heterogeneity.
- Demonstrates that representation-space alignment is more effective than parameter-space methods for cross-domain knowledge transfer.
- Validates the proposed method through extensive experiments on multiple domain-heterogeneous benchmarks.
Read more
CRIP: Channel Level Representation Injection for Personalized One-Shot Federated Learning
Summary
The paper introduces CRIP, a novel framework for personalized one-shot federated learning (OSFL) that addresses the challenges posed by domain heterogeneity among clients. Traditional OSFL methods often struggle with knowledge transfer due to the lack of iterative communication and the reliance on parameter-level aggregation, which can lead to ineffective model performance in heterogeneous environments. CRIP operates in the representation space by aligning channel-level features from various clients, allowing for selective fusion of compatible features. This approach mitigates the noise introduced by incompatible feature distributions and enhances the model's adaptability to the target client's domain. The authors conducted extensive experiments on benchmark datasets such as DomainNet, PACS, and Office-Home, demonstrating that CRIP outperforms both local models and existing state-of-the-art methods, confirming the effectiveness of representation-space personalization in scenarios with extreme domain heterogeneity.
Methodology
CRIP employs a channel-level representation injection strategy where each client uploads its feature extractor to a central server. The server then broadcasts these extractors back to all clients. Each client measures the representational similarity of its features with those from other clients and selectively fuses compatible features, enhancing local model performance without iterative updates.
Results
The experimental results indicate that CRIP consistently outperforms local models and existing state-of-the-art OSFL methods across various domain-heterogeneous benchmarks, demonstrating its effectiveness in handling domain shifts and improving personalization in federated learning.
Implications
The findings suggest that CRIP can significantly enhance the performance of federated learning systems in real-world applications where data distribution varies across clients, such as in healthcare or finance. This approach could lead to more efficient and privacy-preserving collaborative learning frameworks.
HP-JEPA: Hierarchical Partitioning for Multi-Resolution Graph Joint-Embedding Predictive Learning
Graph Learning
- HP-JEPA introduces a hierarchical partitioning approach for graph representation learning, addressing the limitations of fixed-resolution methods.
- The framework allows for context-target latent prediction at multiple resolutions, capturing diverse structural patterns.
- Experiments show HP-JEPA outperforms the baseline Graph-JEPA in most tasks, particularly in accuracy across different graph sizes.
- Resolution-specific representations can be integrated to enhance downstream model performance by leveraging complementary information.
Read more
HP-JEPA: Hierarchical Partitioning for Multi-Resolution Graph Joint-Embedding Predictive Learning
Summary
The paper introduces HP-JEPA, a novel framework for graph self-supervised learning that enhances joint-embedding predictive architectures (JEPAs) by employing hierarchical partitioning for multi-resolution graph representation. Traditional JEPAs typically rely on a single predefined graph partition, which limits their ability to capture diverse structural patterns across different scales. HP-JEPA addresses this limitation by organizing graphs into an ordered set of coarse-to-fine partitions, allowing for context-target latent prediction at multiple resolutions. This method utilizes an online encoder, an exponential-moving-average target encoder, and a latent predictor to generate resolution-specific representations. These representations can then be integrated through concatenation or task-specific weighting, enabling downstream models to leverage complementary information from local, regional, and global graph structures. The authors validate HP-JEPA through experiments on seven graph classification benchmarks and one regression benchmark, demonstrating its superiority over the fixed-resolution Graph-JEPA baseline in six out of eight tasks, particularly excelling in accuracy across various graph sizes. The findings underscore the effectiveness of hierarchical multi-resolution partitioning in enhancing transferable graph representation learning.
Methodology
HP-JEPA employs a hierarchical partitioning strategy to create multiple graph resolutions, enabling context-target prediction at each level. It utilizes an online encoder for visible regions and an exponential-moving-average target encoder for held-out regions, with a latent predictor to estimate target representations. The resulting resolution-specific embeddings are integrated through concatenation or weighted selection based on task relevance.
Results
HP-JEPA demonstrated superior performance compared to the fixed-resolution Graph-JEPA baseline, achieving better results in six out of eight tasks across various benchmarks. Size-stratified analyses indicated that HP-JEPA consistently outperformed Graph-JEPA in most graph-size quartiles, showcasing its effectiveness in learning transferable representations.
Implications
The findings suggest that hierarchical multi-resolution approaches can significantly improve graph representation learning, making HP-JEPA applicable in various domains where graph-structured data is prevalent, such as social networks, biological networks, and recommendation systems. This method could enhance the performance of downstream tasks like classification and regression by providing richer and more nuanced representations.
Freeze, Then Select: Structured Field Adapters and Stability-Validated Weak Selection for PDE Discovery from Sparse Observations
Theory
Optimization
Time Series
- Introduction of a freeze-then-select method for PDE discovery that separates equation selection from neural optimization.
- Development of a structured field adapter that combines learned spatial features with temporal spline coefficients.
- Implementation of Stability-Validated Weak Selection (SVWS) to enhance the reliability of term selection across weak-form systems.
- Demonstration of high exact support recovery rates across multiple sparse observation regimes.
Read more
Freeze, Then Select: Structured Field Adapters and Stability-Validated Weak Selection for PDE Discovery from Sparse Observations
Summary
This paper presents a novel approach for discovering partial differential equations (PDEs) from sparse observations, addressing the challenges of reconstructing continuous fields and selecting the correct differential terms. The authors introduce a 'freeze-then-select' methodology that decouples equation selection from neural optimization. This method employs a structured field adapter that factors the field into learned spatial features and temporal coefficients represented by cubic splines. After training, the field is frozen, and Stability-Validated Weak Selection (SVWS) is used to identify recurrent terms across independent weak-form systems, refit candidate supports, and select the final equation based on held-out weak-form systems. The approach is evaluated across various sparse MDBench regimes, demonstrating superior performance in exact support recovery compared to classical and neural baselines, particularly in the challenging Kuramoto–Sivashinsky dynamics. Additionally, the method successfully recovers the power-law form of an unknown nonlinear diffusion function from sparse, noisy observations, showcasing its robustness and versatility.
Methodology
The proposed methodology involves a two-stage process: first, a structured field adapter is trained to reconstruct a continuous field from sparse observations without a PDE residual. This field is then frozen, and SVWS is applied to generate candidate supports, refit coefficients, and validate the final equation across independently constructed weak-form systems. The structured adapter utilizes cubic B-spline representations for temporal coefficients, allowing for smooth temporal evolution and analytic derivatives.
Results
The method achieves the highest exact support recovery rates in all six sparse MDBench regimes tested, with particularly notable improvements over classical and neural baselines in the Kuramoto–Sivashinsky dynamics. The SVWS extension also successfully recovers the power-law form of an unknown nonlinear diffusion function from sparse, noisy data.
Implications
This work has significant implications for data-driven discovery of governing equations in physical systems, particularly in scenarios with limited observations. The ability to accurately recover PDEs from sparse data can enhance modeling and simulation in various scientific and engineering applications.
Cross-Resolution Semantic Learning for Graph Domain Adaptation
Graph Learning
- Introduces the concept of semantic resolution shift in GDA.
- Proposes a novel method (CReSL) that learns soft source-to-target resolution correspondence.
- Constructs a multi-resolution representation bank with a shared GNN.
- Implements Cross-Resolution Prototype Transport and Cross-Resolution Target Grafting for improved adaptation.
Read more
Cross-Resolution Semantic Learning for Graph Domain Adaptation
Summary
This paper addresses the challenge of Graph Domain Adaptation (GDA), which involves transferring predictive knowledge from labeled source graphs to unlabeled target graphs under distribution shifts. The authors identify a limitation in existing methods that typically align representations or regularize graph structures without considering how class-discriminative knowledge learned at different source neighborhood ranges should be routed across target ranges. They introduce the concept of 'semantic resolution shift,' which refers to the cross-domain change in propagation resolutions where class-discriminative evidence is strongest. To tackle this issue, the authors propose Cross-Resolution Semantic Learning (CReSL), a novel GDA method that learns soft source-to-target resolution correspondence informed by cross-domain class structure. CReSL constructs a multi-resolution representation bank using a shared Graph Neural Network (GNN) and learnable resolution embeddings, and it introduces two key components: Cross-Resolution Prototype Transport (CRPT) and Cross-Resolution Target Grafting (CRTG). CRPT refines target posteriors and resolution correspondence using class-resolution prototypes, while CRTG enforces prediction consistency for instance-level adaptation under class uncertainty. The proposed method demonstrates significant improvements over strong baselines in extensive experiments across various graph benchmarks with diverse domain shifts.
Methodology
CReSL employs a multi-resolution representation bank created by a shared Graph Neural Network and learnable resolution embeddings. It utilizes Cross-Resolution Prototype Transport to refine target posteriors and establish resolution correspondence, and Cross-Resolution Target Grafting to ensure prediction consistency during instance-level adaptation.
Results
The experiments conducted on various graph benchmarks reveal that CReSL outperforms several strong representative baselines in most settings, indicating its effectiveness in addressing the challenges posed by semantic resolution shifts in GDA.
Implications
The findings suggest that CReSL can enhance the transferability of graph learning models across different domains, reducing the need for extensive target annotations and improving model deployment in diverse environments.
Learning Optimal Dynamic Matching via Graph Neural Networks
Reinforcement Learning
Graph Learning
Optimization
- Introduces a value-based reinforcement learning framework for dynamic matching on evolving graphs.
- Demonstrates an event-time reduction that simplifies decision-making in dynamic environments.
- Utilizes Graph Neural Networks to approximate continuation values, enhancing scalability and adaptability.
- Achieves superior performance in dynamic matching scenarios compared to traditional heuristics.
Read more
Learning Optimal Dynamic Matching via Graph Neural Networks
Summary
This paper addresses the challenge of dynamic matching in markets where the composition of participants evolves over time. The authors propose a value-based reinforcement learning framework that operates on finite, evolving weighted graphs. They develop an infinite-horizon continuous-time model that incorporates stochastic arrivals, node-type transitions, edge realizations, and exogenous exits. A key contribution is the event-time reduction, which allows the planner to act immediately after each exogenous event without losing optimality. The optimal edge-wise Q-function is characterized by a single continuation-value function on post-decision residual graphs, simplifying the learning process from state-action values to graph values. The authors utilize a Graph Neural Network (GNN) to approximate the value function, trained via temporal-difference learning, and implement a forward-greedy matching heuristic for action selection. The framework is evaluated in two scenarios: a binary-type benchmark and a kidney paired donation benchmark. In both cases, the learned policy demonstrates superior performance compared to traditional immediate and threshold-greedy rules, effectively balancing immediate match value with the preservation of future opportunities.
Methodology
The authors develop a reinforcement learning framework that employs Graph Neural Networks (GNNs) to approximate the value function for dynamic matching problems. They derive an event-time reduction to simplify decision-making and use temporal-difference learning for training the GNN. The approach is evaluated through simulations in two distinct environments, focusing on the timing tradeoff in dynamic matching.
Results
In the binary-type benchmark, the learned policy significantly outperformed both immediate greedy and threshold-greedy strategies by effectively preserving valuable nodes for future matches. In the kidney paired donation benchmark, the policy adapted to varying exit predictability, performing comparably to immediate greedy under unpredictable conditions and outperforming it under reliable exit signals.
Implications
The findings suggest that dynamic matching policies can be optimized using GNNs, leading to more effective market designs in various applications such as organ donation, ride-sharing, and online platforms. This approach can enhance decision-making in environments where participant dynamics are complex and evolving.
Evaluating Forecasting Techniques for Hardware Errors on a Large-scale HPC System
Time Series
- Forecasting effectiveness varies significantly based on the temporal structure of hardware errors.
- LSTM and Transformer models perform well for regularly occurring and stable error types.
- Sparse and burst-dominated errors are challenging to predict using existing methods.
- The study provides empirical guidance on the conditions under which forecasting can be effective.
Read more
Evaluating Forecasting Techniques for Hardware Errors on a Large-scale HPC System
Summary
This paper investigates the application of time series forecasting techniques to predict hardware errors in high-performance computing (HPC) systems, specifically using seven years of production logs from the Theta supercomputer. The authors analyze the temporal structure of hardware error logs to determine the effectiveness of classical statistical methods and modern deep learning models, including LSTM and Transformer architectures. The study finds that the predictability of hardware errors is highly dependent on their temporal characteristics; errors that occur regularly and exhibit structural stability can be accurately modeled, while sparse and burst-dominated errors pose significant challenges. The research does not propose a ready-to-deploy prediction framework but instead offers empirical insights into when forecasting is effective and suggests avenues for enhancing forecasting accuracy in HPC error analysis. This work contributes to the understanding of error dynamics in HPC systems and highlights the potential of predictive modeling in improving system reliability.
Methodology
The authors utilized a dataset of hardware error logs from the Theta supercomputer, transforming raw logs into structured time series representations. They benchmarked various forecasting models, including classical statistical methods and deep learning architectures, to evaluate their predictive performance across different error categories.
Results
The results indicate that only certain types of hardware errors exhibit sufficient regularity for meaningful prediction. LSTM and Transformer models were found to capture informative temporal structures effectively, while classical statistical methods remained competitive in specific scenarios. The study highlights the limitations of forecasting for sparse and burst-dominated error types.
Implications
The findings suggest that while some hardware errors can be predicted effectively, there is a need for further research to enhance forecasting methods for less predictable error types. This work lays the groundwork for future studies aimed at improving system reliability in HPC environments through better predictive modeling.
Reflection or Re-Generation? Why LLM Revision Fails Where Human Revision Succeeds
NLP
Large Language Models
Theory
- Introduction of the Human–LLM Reflection Framework (HRF) for controlled comparison of human and LLM revisions.
- LLM reflection is characterized as conditioned re-generation, failing to accumulate task-relevant information.
- Human revisions consistently yield positive information gain, while LLM revisions can result in negative information gain.
- The study identifies the revision step as the primary source of failure in LLMs, independent of input quality.
Read more
Reflection or Re-Generation? Why LLM Revision Fails Where Human Revision Succeeds
Summary
This paper investigates the differences between human and large language model (LLM) revision behaviors, particularly focusing on the concept of reflection in improving reasoning. The authors introduce the Human–LLM Reflection Framework (HRF), a controlled two-pass protocol that allows for a direct comparison of revision processes between humans and LLMs across various settings. The study employs an information-theoretic analysis to assess the effectiveness of revisions, revealing that LLMs often fail to improve their outputs during reflection, instead exhibiting behavior akin to conditioned re-generation. The findings indicate that LLMs show negligible or even negative information gain during revision tasks, contrasting with the positive information gain observed in human revisions. The paper also highlights that the failure of LLMs is localized to the revision step, rather than the quality of the initial input, and identifies variability in the mechanisms of failure across different tasks and models.
Methodology
The authors developed the HRF, a two-pass protocol where both human annotators and LLMs generate initial responses and then revise them based on prior outputs. The study utilized two discrete tasks for rigorous analysis: MalAlgoQA (mathematical reasoning) and IMDb-Rating (sentiment evaluation), alongside an open-ended temporal reasoning benchmark. Five LLMs were evaluated under identical conditions to assess their revision capabilities.
Results
The results demonstrate that LLMs exhibit two distinct failure modes during reflection: near-zero information gain on objective tasks and negative information gain on subjective tasks. In contrast, human revisions consistently improve responses, indicating a fundamental difference in the revision processes of humans and LLMs. The failure is attributed to the inability of LLMs to effectively detect and correct errors during the revision step.
Implications
These findings suggest that current LLMs may require enhancements in their reflective capabilities to match human-like error detection and correction. The insights could inform future research on improving LLM architectures and training methodologies to better facilitate genuine revision processes.
Smooth Reparameterizations of Functions on Simplicial Product Spaces: Applications to Probabilistic Tensor Decomposition and Functional Data Registration
Optimization
Theory
- Introduction of smooth reparameterizations for optimization on simplicial product spaces.
- Generalization of KKT point results from Hadamard parameterization to broader smooth transformations.
- Development of a Riemannian Gradient Descent algorithm that outperforms traditional methods.
- Application of the proposed methods to tensor factorization and functional data registration.
Read more
Smooth Reparameterizations of Functions on Simplicial Product Spaces: Applications to Probabilistic Tensor Decomposition and Functional Data Registration
Summary
This paper addresses optimization problems defined on product spaces of simplices, particularly focusing on applications in low-rank discrete multivariate probability distributions and functional data registration using the Square Root Velocity Function (SRVF) representation. The authors propose a novel approach of replacing the product simplex with a smooth, elementwise strictly convex reparameterization, transforming the problem into an unconstrained optimization problem on a manifold. They demonstrate that this reparameterization allows for the mapping of second-order Karush–Kuhn–Tucker (KKT) points on the smooth manifold to weak second-order KKT points on the product simplex. The paper introduces a Riemannian Gradient Descent (RGD) algorithm that outperforms traditional Projected Gradient Descent (PGD) methods, yielding a more accurate representation of function shapes during curve registration. The authors generalize previous results on Hadamard parameterizations to a broader class of smooth transformations, enabling the application of Riemannian optimization techniques to a wider range of problems, including matrix and tensor factorization.
Methodology
The authors utilize smooth reparameterizations of functions defined on simplicial product spaces to transform constrained optimization problems into unconstrained ones on manifolds. They derive a Riemannian Gradient Descent algorithm to solve these reparameterized problems, leveraging properties of second-order KKT conditions to ensure optimality.
Results
The proposed Riemannian Gradient Descent algorithm shows superior performance compared to Projected Gradient Descent in both simplex constrained tensor factorization and functional data registration tasks, achieving smoother and more accurate results.
Implications
The findings suggest that smooth reparameterizations can significantly enhance optimization techniques in various applications, particularly in machine learning tasks involving tensor decompositions and shape analysis. This approach may lead to more efficient algorithms in fields requiring complex data representations.
Caliber: Cross-Architecture Extraction-Cost Control for Score-Returning APIs
Theory
Efficient ML
- Caliber recasts model extraction defense as a controllable problem with predictable degradation.
- Introduces a margin-normalized calibration law applicable across various architectures without specific tuning.
- Demonstrates effectiveness across multiple model-dataset combinations with minimal error margins.
- Establishes a closed-form relationship between noise variance and utility targets.
Read more
Caliber: Cross-Architecture Extraction-Cost Control for Score-Returning APIs
Summary
The paper introduces Caliber, a novel defense mechanism against model extraction attacks targeting score-returning APIs. It formulates the problem of noise selection as a calibration issue, balancing the degradation of the supervision signal against the cost for attackers attempting to recover clean logits. Caliber employs independent and identically distributed Gaussian noise to the internal logits, establishing two key properties: monotone agreement degradation and per-input recovery cost. The first property ensures that as noise increases, the agreement with clean predictions decreases, while the second provides a closed-form minimax lower bound on the number of queries required to recover clean logits. The authors demonstrate that the noise required for a specified utility reduction varies across different architectures, and they normalize this noise by the squared median top-two logit margin. The effectiveness of Caliber is validated through experiments across over thirty model-dataset combinations, achieving mean absolute relative errors of 0.6–1.4%. The results indicate that surrogate models generally reflect the configured degradation, thereby limiting the effectiveness of extraction attempts.
Methodology
Caliber employs a Gaussian noise perturbation strategy on model logits, normalizing noise variance by the squared median top-two logit margin. It establishes a logistic relationship between noise and utility, allowing for calibration of noise levels to meet specified degradation targets. The methodology includes theoretical proofs of monotone degradation and a minimax lower bound on recovery costs.
Results
The experiments conducted across more than thirty model-dataset combinations showed that Caliber achieved mean absolute relative errors of 0.6–1.4% in utility targets. Additionally, end-to-end extraction studies confirmed that the performance of the surrogate models aligned with the calibrated degradation, effectively limiting the attacker's ability to extract clean logits.
Implications
Caliber has significant implications for the protection of intellectual property in deep learning models, particularly in commercial API environments. By providing a controllable defense mechanism, it enables service providers to mitigate the risks associated with model extraction while maintaining the usability of their APIs.
From field-scale to large-scale spectral libraries: Tabular foundation models in soil spectroscopy
Theory
Efficient ML
- TabPFN outperformed traditional regression models in predicting soil properties from spectral data.
- Explicit dimensionality reduction is not strictly required for strong performance in soil spectroscopy.
- Combining TabPFN with PLS latent variables yielded the best overall predictions.
- The study provides evidence-based guidance for model selection in soil spectroscopy.
Read more
From field-scale to large-scale spectral libraries: Tabular foundation models in soil spectroscopy
Summary
This paper explores the application of machine learning models, particularly tabular foundation models, in soil spectroscopy to predict soil properties from visible and near-infrared (vis-NIR) and mid-infrared (MIR) spectral data. The authors conducted a systematic investigation of various regression models and dimensionality reduction techniques across 85 regression tasks using open benchmark datasets in pedometrics. They compared the performance of an in-context learning tabular foundation model (TabPFN) against traditional methods such as convolutional neural networks (CNN), rule-based regression (Cubist), Random Forest, and partial least squares regression (PLSR). The study found that TabPFN consistently outperformed classical baselines, even when applied directly to full spectra, indicating that explicit dimensionality reduction is not always necessary for effective modeling. However, further improvements were noted when combining TabPFN with PLS latent variables, suggesting a complementary relationship between modern and traditional methods in spectroscopic calibration. The findings provide valuable insights for selecting appropriate calibration models in soil spectroscopy across different operational scales.
Methodology
The authors systematically investigated various regression models and dimensionality reduction techniques across 85 regression tasks. They compared TabPFN, CNN, Cubist, Random Forest, and PLSR using both full spectral data and features derived from PCA and PLS. Performance metrics were used to evaluate the effectiveness of each model in predicting soil properties.
Results
TabPFN consistently delivered the best performance across all scales, including tasks involving large spectral libraries with tens of thousands of soil samples. Notably, it surpassed classical baselines when applied to full spectra. The combination of PLS latent variables with TabPFN resulted in the highest prediction accuracy, demonstrating the effectiveness of integrating modern and traditional modeling approaches.
Implications
The findings suggest that machine learning, particularly tabular foundation models, can significantly enhance the accuracy of soil property predictions from spectral data. This has implications for digital soil mapping and precision agriculture, where rapid and cost-effective soil analysis is crucial. The study also encourages the integration of traditional methods with modern machine learning techniques to improve predictive modeling in soil science.
Analysing User Reviews to Identify User Concerns Around Permissions in AI Apps
NLP
- Developed a machine learning model to classify user reviews related to permissions in AI apps.
- Achieved an accuracy of 82% in classifying permission-related reviews.
- Identified that user concerns are more sentiment-driven rather than focused on specific permissions.
- Utilized AI-generated reviews to create a labeled dataset, reducing the need for manual annotation.
Read more
Analysing User Reviews to Identify User Concerns Around Permissions in AI Apps
Summary
This paper addresses the growing concern over user privacy and security in AI mobile applications by analyzing user reviews to classify concerns related to app permissions. The authors highlight that while AI is increasingly integrated into mobile apps, developers often lack knowledge of security best practices, leaving users to navigate their own privacy concerns. The study presents a machine learning model that classifies user reviews into permission-related categories, utilizing AI-generated reviews to create a labeled training set from a larger corpus of human-written reviews. This innovative approach eliminates the need for manual annotation, enabling the classification of permission-related reviews with an accuracy of 82%. The analysis reveals that users tend to express their concerns based on sentiment towards the app rather than specific permission types, which has implications for developers and platform administrators in understanding user attitudes and improving app transparency. The paper contributes to the field by providing a robust classifier and insights into user concerns regarding app permissions in AI applications.
Methodology
The methodology involved multiple stages: data curation, selection of training and test data, and machine learning model training. The authors collected both human-generated and AI-generated reviews, using the latter to identify relevant training examples. They trained binary neural network classifiers to classify security reviews and then isolated permission-related reviews, employing K-means clustering to analyze the topics discussed.
Results
The machine learning model successfully classified permission-related reviews with an accuracy of 82%. The analysis of the reviews indicated that users organize their concerns based on their sentiment towards the app rather than the specific types of permissions requested.
Implications
The findings suggest that developers should focus on enhancing transparency and user control over permissions in AI applications. Understanding user sentiment can guide improvements in app design and communication strategies, ultimately fostering greater user trust and satisfaction.
SILVA Networks as Structured Implicit Layers and Vector Attractors via Dynamic Interaction Fields
Graph Learning
Theory
Optimization
- SILVA Networks separate stimulus, local, and global interactions in a fixed-point architecture.
- The architecture is versatile, applicable to images, molecules, citation networks, and graphs.
- Local interactions are essential for graph tasks, while global interactions enhance long-range classification.
- SILVA allows for the training and visualization of internal dynamics, improving interpretability.
Read more
SILVA Networks as Structured Implicit Layers and Vector Attractors via Dynamic Interaction Fields
Summary
This paper introduces SILVA Networks, a novel architecture designed to address the complexities of learning representations that incorporate direct input, local structures, and broader contextual influences. Traditional implicit neural layers often merge these influences into a single update, obscuring the individual contributions of each component. SILVA Networks decompose this process into distinct roles: stimulus input, local interactions, global interactions, damping, and readout, all within a fixed-point architecture. This separation allows for a clearer understanding of how different factors contribute to the learning process. The architecture is adaptable across various domains, including images, molecular structures, citation networks, and long-range graph benchmarks, by defining domain-specific nodes and interactions. Experimental results indicate that local interactions are crucial for graph tasks, while global interactions provide significant benefits in long-range node classification. The SILVA framework enables the training, visualization, and diagnosis of internal dynamics, offering a more interpretable approach to implicit representations in machine learning.
Methodology
The SILVA Networks architecture employs a fixed-point layer structure that separates the roles of stimulus input, local and global interactions, and damping. It utilizes a damped update mechanism to propagate input through an interacting state until a vector attractor is reached. This approach is tested across various domains with specific adaptations for each context, allowing for a comprehensive analysis of the contributions from different interaction types.
Results
Experiments demonstrate that local interactions are critical for performance in graph-related tasks, while the benefits of global interactions are most pronounced in long-range node classification benchmarks. The architecture's ability to visualize and diagnose internal dynamics enhances understanding of how different components influence learning outcomes.
Implications
The SILVA Networks framework has the potential to improve the interpretability and performance of machine learning models across diverse applications, particularly in fields requiring complex relational data analysis, such as molecular chemistry and social network analysis. Its structured approach may lead to advancements in the design of neural architectures that better capture the intricacies of data interactions.
Convex Neural Energy Elements: Monolithic Finite-Element Assembly of Geometry-Parameterized Neural Operators with Stability and Error Guarantees
Theory
Optimization
- Introduces convex neural energy elements to address structural failures in neural operators for assembly.
- Establishes a regularization principle that ensures the nullspace of the regularizer contains the physics nullspace.
- Demonstrates significant improvements in accuracy and computational efficiency in finite-element assembly.
- Provides theoretical error bounds and experimental validation for the proposed method.
Read more
Convex Neural Energy Elements: Monolithic Finite-Element Assembly of Geometry-Parameterized Neural Operators with Stability and Error Guarantees
Summary
This paper addresses the limitations of the neural-operator element method (NOEM) in computational mechanics, particularly the failure of individually trained neural operators to serve as reusable assembly elements due to structural issues. The authors introduce convex neural energy elements, which are geometry-parameterized and architecturally convex, ensuring that the induced energy's Hessian is positive semidefinite. This is achieved through a hypernetwork-generated quadratic form and a regularization principle that aligns the regularizer's nullspace with the physics nullspace, thus eliminating biases that lead to spurious minima in optimization. The paper provides theoretical guarantees for error bounds related to energy-to-solution accuracy and geometry generalization, validated through experiments. The proposed method demonstrates significant improvements in accuracy and efficiency, achieving relative L2 errors of 0.6–1.0% in various geometries and a 175× speedup in boundary-quantity workloads. The findings suggest that making energy the learned object transforms neural operators from single-use surrogates into reusable elements, inheriting classical assembly guarantees.
Methodology
The authors developed convex neural energy elements by defining a scalar energy function that is architecturally convex in boundary degrees of freedom and smoothly parameterized by geometry. They employed a hypernetwork to generate a positive-semidefinite quadratic form and implemented a regularization principle to ensure the stability of the assembled elements. The methodology includes theoretical analysis and experimental validation on various geometrical configurations.
Results
The proposed convex neural energy elements achieved relative L2 errors of 0.6–1.0% across different geometries, including complex layouts. The method demonstrated a 175× speedup in setup times for boundary-quantity workloads and maintained positive definiteness in assembled systems, confirming the theoretical guarantees of the approach.
Implications
The findings have significant implications for computational mechanics, enabling the development of reusable neural operators that can efficiently handle a variety of geometrical configurations without the need for retraining. This could lead to advancements in simulations and modeling in engineering and physics applications.
Relative Parameter Importance in Task-Agnostic Replay-Free Continual Learning
NLP
Generative Models
Theory
- Introduction of a novel relative parameter-importance measure for informed regularization in continual learning.
- Focus on replay-free and task-id-free settings to mitigate catastrophic forgetting while enabling knowledge transfer.
- Demonstrated improvements over state-of-the-art continual learning methods in text classification tasks.
- Insights provided for extending the method to generative language models.
Read more
Relative Parameter Importance in Task-Agnostic Replay-Free Continual Learning
Summary
This paper addresses the challenges of continual learning (CL) in deep neural networks, specifically focusing on scenarios where there is no access to training data from prior tasks and no task identification at inference time. The authors introduce a novel measure called relative parameter importance, which assesses the significance of each parameter concerning both current and past tasks. This measure allows for a more informed regularization strategy, where parameters deemed crucial for maintaining past-task stability are heavily regularized, while those with lower importance can be updated more freely. The proposed method consists of two phases: a Look-Ahead phase, where an auxiliary model is trained on new task data to derive importance estimates, and a Main Continual Learning phase, where the main model is trained with a regularization objective based on these estimates. The authors demonstrate that their approach outperforms existing state-of-the-art CL methods in both class-incremental and domain-incremental learning scenarios, providing insights for extending the method to text generation tasks.
Methodology
The methodology involves two main phases: (1) Look-Ahead Phase, where a temporary auxiliary model is initialized with the frozen past-task model and trained on new task data to derive parameter importance estimates; (2) Main Continual Learning Phase, where the main model is trained using a regularization objective based on the derived relative parameter importance, allowing for selective updates of parameters.
Results
The proposed method showed significant improvements in performance on both class-incremental and domain-incremental learning tasks compared to existing continual learning methods. The results indicate that the relative parameter importance measure effectively balances stability and plasticity, facilitating backward knowledge transfer.
Implications
The findings suggest that the relative parameter importance measure can enhance continual learning frameworks, particularly in scenarios where data privacy and memory constraints limit traditional methods. Additionally, the insights gained may inform future research in applying continual learning techniques to generative models.
Understanding and Correcting Low-Frequency Bias in EEG Foundation Model
Time Series
- Identification of a persistent low-frequency bias in EEG foundation models.
- Explanation of the bias through the interaction of EEG's spectral structure and neural networks' learning preferences.
- Introduction of FAME, a frequency-balanced masked autoencoding framework.
- FAME achieves state-of-the-art performance on 24 out of 41 evaluated tasks.
Read more
Understanding and Correcting Low-Frequency Bias in EEG Foundation Model
Summary
This paper addresses a significant issue in EEG foundation models: a persistent low-frequency bias that affects the performance of these models across various scales and objectives. The authors identify that increasing the scale of pretraining data or model capacity does not consistently improve downstream performance due to this bias. They link the low-frequency bias to the interaction between the 1/f α-like spectral structure of EEG signals and the tendency of neural networks to learn low-frequency components preferentially. The authors propose a novel framework called FAME (Frequency-balanced Autoencoding framework for Masked EEG) that aims to correct this bias by reconstructing time-frequency activity in predefined EEG bands. FAME standardizes reconstruction targets within each band and assigns equal weight to all band-specific losses, promoting a more balanced learning signal across the EEG spectrum. The framework is evaluated on 41 downstream tasks in OmniEEG-Bench, achieving state-of-the-art performance on 24 tasks, thereby demonstrating the effectiveness of frequency-balanced supervision in enhancing the transferability of EEG representations.
Methodology
The authors conducted frequency-resolved analyses of existing EEG foundation models to identify the low-frequency bias. They proposed the FAME framework, which reconstructs masked EEG inputs by standardizing reconstruction targets within predefined frequency bands and equalizing the loss weights across these bands. This approach mitigates the dominance of low-frequency components in the learning process.
Results
FAME was evaluated on 41 downstream tasks in the OmniEEG-Bench, achieving state-of-the-art performance on 24 tasks. The results indicate that frequency-balanced supervision leads to more spectrally balanced representations and enhances the effectiveness of EEG pretraining.
Implications
The findings suggest that addressing low-frequency bias can significantly improve the performance of EEG models, making them more effective for various applications in biomedical engineering and neuroscience. This work may influence future research directions in EEG representation learning and pretraining paradigms.
Sensitivity Analysis of GRU, LSTM and Transformer Encoder in Classification of Automated Driving Systems
Time Series
- Evaluation of GRU, LSTM, and Transformer models for classifying ADSs using telematics data.
- Strong clean-data performance with macro F1-scores above 0.90 for all models.
- Introduction of a modular robustness evaluation framework to assess model performance under telematics degradation.
- Significant vulnerability of models to temporal jitter, revealing limitations in current sequence models.
Read more
Sensitivity Analysis of GRU, LSTM and Transformer Encoder in Classification of Automated Driving Systems
Summary
This paper investigates the effectiveness of three sequence-based classification models—Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM) networks, and Transformer encoder models—in identifying Level 2 automated driving systems (ADSs) using vehicle telematics data. The authors evaluate the models on their ability to classify different ADSs, including Comma Openpilot, Tesla Autopilot, and Cadillac Super Cruise, achieving strong performance with macro F1-scores of 0.92 (GRU), 0.90 (LSTM), and 0.93 (Transformer) on clean data. The study introduces a modular robustness evaluation framework that simulates realistic telematics degradation through various corruption types and severity levels. Results indicate that while all models maintain high accuracy under clean conditions, they exhibit significant vulnerability to temporal jitter, with macro F1-scores dropping to 0.44–0.50, highlighting a critical limitation in current sequence models for ADS identification. This research contributes to the understanding of model robustness in the context of ADS classification and emphasizes the need for independent verification of ADS behavior.
Methodology
The study employs three sequence-based models (GRU, LSTM, Transformer) to classify ADSs based on vehicle telematics data. A modular robustness evaluation framework is introduced, simulating realistic telematics degradation through five corruption families at varying severity levels. Performance is measured using macro F1-scores to assess both clean and corrupted data.
Results
All models achieved strong clean-data performance with macro F1-scores of 0.92 (GRU), 0.90 (LSTM), and 0.93 (Transformer). However, under conditions of temporal jitter, model performance dropped significantly to macro F1-scores of 0.44–0.50, indicating a critical failure mode.
Implications
The findings underscore the importance of robust model evaluation in real-world applications of ADS classification. The research highlights the need for independent monitoring systems that can reliably verify ADS behavior, which is crucial for safety, regulatory compliance, and insurance assessments.
Adaptivity via a Parallel Architecture for Stochastic Gradient Methods
Optimization
Theory
Efficient ML
- Introduces a parallel framework for static gradient methods to enhance adaptivity.
- Utilizes multiple processors to concurrently search for optimal iteration budgets.
- Achieves a (p, αp)-approximation guarantee for convergence efficiency.
- Simplifies convergence analysis for nonconvex objective functions.
Read more
Adaptivity via a Parallel Architecture for Stochastic Gradient Methods
Summary
This paper presents a novel parallel framework designed to enhance the adaptivity of static gradient methods in stochastic optimization. The proposed method, denoted as GD(x0, T), takes an initial point x0 and a specified number of iterations T, with a step size determined by a function S(T). The framework utilizes p processors that concurrently search for an optimal value of T based on a geometric sequence, ensuring that the gradient descent process meets desired convergence criteria. Each processor operates through an infinite sequence of stages, executing GD(x0, Tj,i) at each stage, where Tj,i is determined by a prescribed function h(j, i). The efficiency of this parallel architecture is quantified by its (p, αp)-approximation guarantee, which ensures that for any integer T, there exists a processor and stage such that T ≤ Tj,i ≤ T* < αpT. The construction achieves a specific approximation bound, facilitating a simpler convergence analysis for gradient descent on nonconvex functions while maintaining adaptivity to parameters like Lipschitz smoothness and stochastic gradient characteristics. The paper also discusses the limitations of static gradient methods and the advantages of adaptive methods, positioning the proposed framework as a significant step towards improving optimization efficiency in large-scale machine learning tasks.
Methodology
The methodology involves a parallel architecture where multiple processors execute static gradient methods concurrently, each searching for an appropriate number of iterations T based on a geometric sequence. The framework is designed to ensure that the convergence conditions are satisfied while allowing for a simpler analysis of the gradient descent process.
Results
The results demonstrate that the proposed parallel framework achieves a (p, αp)-approximation, allowing for efficient convergence to stationary points in nonconvex optimization problems. The framework's design ensures that it can adaptively adjust to various problem characteristics, improving the overall performance of static gradient methods.
Implications
The findings suggest that this parallel framework can significantly enhance the efficiency and robustness of optimization in large-scale machine learning applications, particularly in scenarios where problem characteristics are unknown or variable. This could lead to improved training times and convergence rates in deep learning and other stochastic optimization tasks.
How Much Does a Reasoning Summary Reveal? An Observability Ladder for Large Language Models
Large Language Models
NLP
Interpretability
- Introduces an observability ladder to assess the effectiveness of different information access levels in predicting model correctness.
- Demonstrates that reasoning summaries are less effective than full reasoning traces, especially when the prompt is visible.
- Finds that the last words of the reasoning trace can predict correctness as well as or better than summaries.
- Highlights the importance of considering both the display of information and the reader's perspective in evaluating model outputs.
Read more
How Much Does a Reasoning Summary Reveal? An Observability Ladder for Large Language Models
Summary
This paper introduces an observability ladder for large language models (LLMs) that evaluates how different levels of information access affect the ability to predict the correctness of model responses. The authors investigate the utility of reasoning summaries, full reasoning traces, and internal signals in assessing model outputs across various benchmarks using five open-weight models. The study reveals that reasoning summaries provide significant predictive power when the prompt is not visible, but their effectiveness diminishes when the prompt is available. The full reasoning trace consistently outperforms summaries, particularly in terms of providing dense and discriminative cues for uncertainty and self-correction. The findings highlight the importance of considering both the display of information and the reader's perspective when evaluating model correctness, suggesting that monitorability is a joint property that should be clearly defined in future research.
Methodology
The authors conducted experiments using five open-weight models across three benchmarks, training matched linear correctness predictors based on varying access levels to model outputs: final responses, self-summaries, full reasoning traces, and internal signals. They compared the predictive power of these different access levels in terms of AUROC scores.
Results
The study found that reasoning summaries achieved a mean AUROC of 0.774 when the prompt was withheld, compared to 0.813 for the full reasoning trace. When the prompt was visible, the summary's predictive gain dropped significantly to +0.019. The last words of the trace provided comparable or better predictive power than summaries, indicating that the trace contains richer information for assessing correctness.
Implications
The findings suggest that developers and researchers should prioritize providing full reasoning traces over summaries for better model evaluation and monitoring. Additionally, the results emphasize the need for clear definitions of monitorability in LLMs, considering both the information displayed and the context of the reader.
AutoCause: A Python framework that automates expert decisions in environmental time-series causal discovery
Time Series
- AutoCause automates the decision-making process in environmental time-series causal discovery.
- The framework integrates multiple causal-discovery methods and provides a consensus approach for link evaluation.
- It records all analytical decisions, enhancing reproducibility and auditability.
- Evaluation on 145 datasets shows improved precision for majority-supported links compared to single-method links.
Read more
AutoCause: A Python framework that automates expert decisions in environmental time-series causal discovery
Summary
The paper presents AutoCause, an open-source Python framework designed to automate expert decisions in the causal discovery of environmental time-series data. Traditional causal discovery methods require significant expert input regarding method selection, conditional independence tests, and other analytical decisions, which can lead to inconsistencies and hinder reproducibility. AutoCause addresses these challenges by formalizing the configuration and evaluation steps surrounding established causal-discovery algorithms. It integrates four causal-discovery methods from three families and includes non-causal reference models, allowing for a multi-method consensus approach. The framework records each decision made during the analysis, derives defaults from an extended causal-audit module, and permits domain-informed overrides. The authors evaluated AutoCause on 145 datasets from various sources, demonstrating that majority-supported links are more precise than those derived from single methods. This framework aims to convert inconsistent expert practices into a more auditable and repeatable analysis process, while still allowing for causal interpretation by the analyst.
Methodology
AutoCause utilizes a workflow that wraps four established causal-discovery methods, including constraint-based and regression approaches. It incorporates an extended causal-audit module for pre-discovery diagnostics, allowing the framework to inform decisions on method selection, conditional independence tests, lag horizons, and sample size adequacy. The framework processes data through a single function call and generates outputs that include raw and FDR-corrected results, consensus tiers, and graph-recovery metrics.
Results
The evaluation of AutoCause on 145 datasets revealed that majority-supported links were more precise than those identified by individual methods, particularly in synthetic benchmarks. However, performance varied against real-world river topology data. The framework effectively transformed expert decision-making into a structured, repeatable process, with recorded configurations enabling reproducibility.
Implications
AutoCause has the potential to standardize causal discovery practices in environmental science, making analyses more reproducible and transparent. It can facilitate better decision-making in environmental management and policy by providing clearer insights into causal relationships within complex time-series data.
GraphIR: Architecture-Level Search States for LLM-Guided Neural Architecture Evolution
Large Language Models
Optimization
Graph Learning
- GraphIR provides a structured architecture state for LLM-guided neural architecture search.
- It organizes candidate architectures into three views: computation skeleton, mutation surface, and validity envelope.
- GraphIR effectively addresses representation mismatches in traditional code-based NAS methods.
- The method shows superior performance across multiple benchmarks while ensuring efficient NAS processes.
Read more
GraphIR: Architecture-Level Search States for LLM-Guided Neural Architecture Evolution
Summary
The paper introduces GraphIR, an architecture-aware intermediate representation designed to enhance neural architecture search (NAS) by providing a structured candidate state for large language models (LLMs) to guide architecture evolution. Traditional code-level NAS methods face a representation mismatch, where the flexibility of executable code does not align with the architecture state necessary for effective mutation. GraphIR addresses this by organizing candidate architectures into three complementary views: a computation skeleton that outlines tensor flow, a mutation surface that highlights editable components, and a validity envelope that captures interface contracts and dependencies. The authors evaluate GraphIR using NAS-Dependency, a benchmark comprising 120 questions across six dependency-reasoning dimensions, demonstrating its effectiveness in identifying dependencies and diagnosing risks. The results show that GraphIR outperforms existing methods across six downstream benchmarks while maintaining comparable model sizes and efficient NAS processes when integrated into OpenEvolve. This work emphasizes the importance of a mutation-oriented architecture state in bridging the gap between executable neural programs and LLM-guided architecture evolution.
Methodology
GraphIR employs static analysis and execution tracing to extract relevant architecture information from neural network programs. It organizes this information into a compact representation that includes a computation skeleton, mutation surface, and validity envelope, which are then used to guide LLMs in architecture mutation tasks.
Results
GraphIR was evaluated against NAS-Dependency and demonstrated superior performance in identifying dependencies and diagnosing risks. It achieved the best overall search performance across six downstream benchmarks, including CLRS, while maintaining comparable model sizes and efficient end-to-end NAS times when integrated into OpenEvolve.
Implications
The development of GraphIR has significant implications for the field of neural architecture search, particularly in enhancing the capabilities of LLMs to effectively guide architecture evolution. This could lead to more efficient and effective neural network designs, ultimately improving performance in various machine learning tasks.
The Fourth Quadrant: A Stylized View of Benign Misfitting
Theory
- Introduces the concept of 'benign misfitting' where good test performance is associated with high training error.
- Establishes a stylized model for supervised linear regression to illustrate the phenomenon.
- Identifies specific training set size regimes where span predictors generalize well despite high training error.
- Demonstrates that one-pass SGD can achieve low test error while exhibiting large training error.
Read more
The Fourth Quadrant: A Stylized View of Benign Misfitting
Summary
This paper explores a novel concept termed 'benign misfitting' in the context of supervised linear regression. The authors present a stylized model where training error can be misleading, particularly in a deterministic (d + 1)-dimensional single-spike model. In this model, all training vectors share a common informative spike coordinate, while the remaining components are considered nuisance. The training labels are uniform, and fresh test points are drawn from a Gaussian distribution. The study reveals that in certain regimes, predictors that generalize well may exhibit worse training performance than a trivial predictor that outputs zero. This phenomenon, termed 'benign misfitting,' occurs in a specific range of training set sizes, where the test performance is good despite high training error. The authors compute the train-test tradeoff curve for span predictors and demonstrate that one-pass stochastic gradient descent (SGD) can achieve small test error while maintaining large training error, thus aligning with the concept of benign misfitting. The findings challenge traditional notions of overfitting and underfitting, suggesting that good test performance can coexist with poor training performance under certain conditions.
Methodology
The authors utilize a deterministic (d + 1)-dimensional single-spike model for supervised linear regression, analyzing the relationship between training and test errors. They compute the exact train-test tradeoff curve for span predictors and investigate the performance of one-pass stochastic gradient descent (SGD) in this context.
Results
The study reveals that in the benign misfitting regime, effective predictors can have diverging training errors while maintaining low test errors. The best span predictor begins to generalize when the number of training points exceeds a certain threshold, and SGD can match the performance of the best span predictor up to a logarithmic factor.
Implications
The findings suggest that traditional metrics for assessing model performance may need reevaluation, particularly in scenarios where high training error does not preclude good generalization. This could influence future research on model selection and training strategies in machine learning.