AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
48
Papers today
8h
Update frequency
7
Days of history
F2STNet: Fair and Federated Spectral-Temporal Modeling for Graph Forecasting
Federated Learning
Graph Learning
Time Series
- F2STNet combines spectral encoding and lightweight temporal modeling for efficient spatiotemporal forecasting.
- The Fairness-aware Federated Aggregation (FFA) scheme dynamically adjusts client weights to improve equity among heterogeneous participants.
- Experimental results show that F2STNet consistently outperforms existing methods in forecasting accuracy across multiple datasets.
- The framework is particularly effective in enhancing fairness metrics in federated learning settings.
Read more
F2STNet: Fair and Federated Spectral-Temporal Modeling for Graph Forecasting
Summary
The paper introduces F2STNet, a novel federated learning framework designed for spatiotemporal forecasting on graph-structured data. This framework addresses the challenges posed by decentralized and heterogeneous data, which complicate sequence modeling and collaborative training. F2STNet integrates several key components: truncated graph-Fourier features for spectral encoding, a lightweight diagonal state-space temporal encoder for efficient long-range dependency modeling, and a Fairness-aware Federated Aggregation (FFA) mechanism that adjusts client contributions based on validation losses and a fairness schedule. The spectral branch captures the graph's frequency structure, while the state-space layer ensures linear complexity in handling temporal dependencies. The authors conducted experiments on three real-world datasets—PeMS04, HZMetro, and KnowAir—demonstrating that F2STNet outperforms state-of-the-art baselines in forecasting accuracy and enhances fairness in federated settings, particularly on the PeMS04 dataset. This work highlights the potential for federated learning to improve spatiotemporal forecasting while addressing issues of fairness and efficiency.
Methodology
F2STNet employs a federated learning approach that integrates truncated graph-Fourier features for spectral analysis, a diagonal state-space encoder for temporal modeling, and a fairness-aware aggregation method that reweights client updates based on predictive losses. This combination allows for effective modeling of both spatial and temporal dependencies while ensuring fairness across clients.
Results
F2STNet demonstrated superior forecasting accuracy compared to state-of-the-art methods on the PeMS04, HZMetro, and KnowAir datasets. Additionally, in federated experiments on PeMS04, it improved metrics related to worst-client performance and client dispersion, indicating enhanced fairness in model performance across diverse clients.
Implications
The findings suggest that F2STNet can be effectively applied in privacy-sensitive environments where decentralized data is prevalent, such as traffic management and environmental monitoring. The framework's focus on fairness also opens avenues for equitable model performance across diverse data sources, which is critical in real-world applications.
Detecting an Effect Is Not Learning to Act on It: A Reward-SNR Floor for LLM Acquisition Agents
Large Language Models
Theory
Interpretability
- Detecting an effect does not equate to learning to act on it; a reward-SNR floor limits the learnability of acquisition policies.
- Structured Hypothesis Embeddings (SHE) provide a method for generating user intent hypotheses but show conditional downstream value.
- No learned acquisition policy outperformed random selection across multiple datasets due to insufficient SNR.
- The paper establishes a necessary condition for policy learnability based on the SNR floor, which is validated through positive controls.
Read more
Detecting an Effect Is Not Learning to Act on It: A Reward-SNR Floor for LLM Acquisition Agents
Summary
This paper addresses the critical distinction between detecting that an auxiliary signal helps on average and learning to act on it for individual instances. The author argues that a reward-SNR (Signal-to-Noise Ratio) floor governs the feasibility of learning a per-instance acquisition policy from offline data. The study reveals that even when an auxiliary signal appears beneficial, it may not lead to a deployable policy that outperforms random selection. The author introduces Structured Hypothesis Embeddings (SHE), where a frozen LLM generates ranked, confidence-scored hypotheses about user intents, which are then used as input for a recommender system. However, the results show that the downstream value of SHE is conditional on the backbone architecture and the regime of data, with significant improvements observed only in specific contexts. The paper concludes that learned acquisition strategies fail across various granularities due to the datasets falling below the SNR detectability floor, suggesting that a design-time regime gate is a more effective approach than a learned policy.
Methodology
The paper employs a theoretical framework to establish the reward-SNR detectability floor and introduces Structured Hypothesis Embeddings (SHE) as a method for generating user intent hypotheses. The effectiveness of SHE is evaluated across three public datasets (MIND, REES46, Amazon-Beauty) to assess its faithfulness and downstream value, while also testing the learnability of acquisition policies.
Results
The results indicate that SHE is a faithful signal but its downstream value is backbone- and regime-conditional. The learned acquisition strategies consistently failed to outperform random selection across all tested granularities, with all datasets falling below the SNR detectability floor, thus confirming the theoretical predictions of the paper.
Implications
The findings suggest that machine learning systems relying on auxiliary signals should consider the SNR of the reward effect when designing acquisition policies. The proposed design-time regime gate could enhance the effectiveness of such systems in practice, particularly in scenarios where learning per-instance policies is unfeasible.
Efficient Hypergradient Descent for Inverse Reinforcement Learning
Reinforcement Learning
Optimization
Efficient ML
- Introduces a structured Fisher-based hypergradient for ML-IRL, enhancing computational efficiency.
- Derives sample-based estimators for the implicit hypergradient, linking it to the trajectory Fisher information matrix.
- Proposes a streaming SCFD solver that reduces storage complexity from O(d²) to O(md), improving runtime.
- Demonstrates competitive performance in policy quality and reward-ranking on standard benchmarks.
Read more
Efficient Hypergradient Descent for Inverse Reinforcement Learning
Summary
This paper addresses the challenges of Inverse Reinforcement Learning (IRL), particularly in the context of maximum-likelihood IRL (ML-IRL), which is formulated as a bilevel optimization problem. The authors propose a novel approach to compute hypergradients efficiently by leveraging the relationship between the inner Hessian and the Fisher information matrix of the policy. They derive a structured Fisher-based hypergradient, which is closely related to Natural Hypergradient Descent (NHGD). To overcome scalability issues associated with large Fisher matrices, the authors introduce a streaming spectral sketch method to approximate the required inverse-Fisher-vector product, significantly reducing memory requirements. The proposed method is evaluated against a first-order stochastic bilevel baseline in both discrete and continuous control environments, demonstrating competitive policy performance and improved computational efficiency. This work represents a significant advancement in the efficient computation of hypergradients for ML-IRL, enabling better reward function recovery from expert demonstrations.
Methodology
The authors formulate ML-IRL as a bilevel optimization problem, derive a hypergradient based on the Fisher information matrix, and utilize a streaming spectral sketch to approximate the inverse-Fisher-vector product, avoiding the need for dense Fisher matrix construction.
Results
The proposed method shows competitive policy performance and strong reward-ranking quality in experiments on CartPole and LQR environments, while significantly reducing memory usage and improving computational efficiency compared to explicit Fisher solvers.
Implications
This research has implications for improving the efficiency of reward function learning in IRL, potentially enabling broader applications in robotics and adaptive control systems where expert demonstrations are available.
TideRL: Boosting Agentic RL Goodput with Readiness-Aware Scheduling
Reinforcement Learning
Large Language Models
Efficient ML
- TideRL improves RL training goodput significantly, achieving up to 5.6× improvement over synchronous baselines.
- The system employs Continuous Task Batching, Resource-Aware Ref-Actor Pipelining, and Elastic Resource Scaling to optimize resource usage.
- It enhances KV cache hit rate by 1.58× and reduces per-step training time by up to 44.3%.
- TideRL addresses inefficiencies in multi-turn agentic workloads, particularly in managing GPU resources and task states.
Read more
TideRL: Boosting Agentic RL Goodput with Readiness-Aware Scheduling
Summary
The paper introduces TideRL, a novel reinforcement learning (RL) system designed to enhance the training throughput (goodput) for large language models (LLMs) engaged in multi-turn agentic workloads. Unlike traditional RL approaches that focus on raw GPU occupancy, TideRL emphasizes the importance of minimizing GPU waiting times and redundant computations during training. The system incorporates three key components: Continuous Task Batching (CTB), which preserves useful rollout states; Resource-Aware Ref-Actor Pipelining (RA2P), which optimally selects between streaming and aggregation based on task readiness; and Elastic Resource Scaling (ERS), which dynamically reallocates resources based on workload demands. The authors demonstrate that TideRL significantly improves RL training goodput by up to 5.6 times compared to synchronous baselines and over 33% compared to asynchronous ones, while maintaining similar task performance. Additionally, it enhances the key-value (KV) cache hit rate, reduces per-step training time, and minimizes total waiting time, addressing critical inefficiencies in existing RL systems.
Methodology
TideRL employs a readiness-aware architecture that integrates Continuous Task Batching to retain useful task states, Resource-Aware Ref-Actor Pipelining to optimize execution modes based on task readiness, and Elastic Resource Scaling to dynamically adjust GPU allocations. This co-design of memory management, execution pipelines, and resource allocation allows for improved throughput in multi-turn RL scenarios.
Results
The implementation of TideRL resulted in a maximum improvement of 5.6× in RL training goodput over synchronous systems and over 33% over asynchronous systems. The KV cache hit rate improved by 1.58×, per-step training time decreased by up to 44.3%, and total waiting time was reduced by up to 77.6%. These results indicate a significant enhancement in the efficiency of RL training for agentic workloads.
Implications
TideRL's advancements in RL training methodologies could lead to more efficient training of large language models and other AI systems that require multi-turn interactions with external environments. This could facilitate the development of more capable autonomous agents in various applications, including robotics, virtual assistants, and interactive AI systems.
Adaptive Symmetry Discovery for Dynamical System Identification
Theory
Time Series
Efficient ML
- Identifying dynamical systems can be significantly improved by leveraging symmetries, allowing for shorter trajectory data.
- The proposed method enables automatic discovery of symmetry groups from single trajectories.
- Theoretical guarantees for symmetry discovery are provided, contrasting with existing heuristic methods.
- Tools from group representation theory and Cayley graphs are utilized, offering new techniques for studying symmetries.
Read more
Adaptive Symmetry Discovery for Dynamical System Identification
Summary
This paper addresses the challenge of dynamical system identification, particularly focusing on systems that exhibit symmetries due to underlying physical laws. The authors propose a novel approach to adaptive symmetry discovery, enabling the identification of system parameters from a single trajectory, even when the symmetry group is unknown. They demonstrate that known symmetries allow for significantly shorter trajectories for identification compared to generic cases. The paper introduces a method for automatic symmetry discovery that learns the symmetry group directly from the trajectory data, achieving optimal trajectory lengths similar to those in known-symmetry scenarios. The analysis employs group representation theory and the properties of Cayley graphs, providing foundational insights into the role of symmetries in dynamical systems. The findings suggest that symmetry discovery can be performed with minimal overhead, enhancing sample efficiency in system identification.
Methodology
The authors utilize group representation theory and Cayley graph properties to analyze the identifiability of dynamical systems. They propose a method for automatic symmetry discovery that incorporates learned symmetries into the identification process, allowing for effective parameter recovery from single trajectories.
Results
The study shows that when symmetries are known, the required trajectory length for system identification is drastically reduced. Furthermore, the automatic symmetry discovery method achieves similar efficiency as the known-symmetry case, demonstrating that symmetry discovery incurs negligible overhead under mild conditions.
Implications
The findings have significant implications for scientific discovery, enabling more efficient identification of dynamical systems across various fields, including biology and physics. The techniques introduced may also inspire further research into the role of symmetries in machine learning and dynamical systems.
Transformer Geometry Observatory TGO-IV: Developmental Topology Observatory
Theory
Interpretability
NLP
- Introduction of the first topological observatory for Transformer representations, extending previous frameworks to include topological data analysis.
- Development of a topology-centric analysis framework that constructs Vietoris–Rips simplicial complexes to study the evolution of representation topology.
- Proposal of a suite of complementary topological observatories to quantify the dynamics of topological features across Transformer layers.
- Investigation of the systematic evolution of representation topology and its implications for understanding Transformer learning dynamics.
Read more
Transformer Geometry Observatory TGO-IV: Developmental Topology Observatory
Summary
The paper introduces TGO-IV, a novel framework for analyzing the evolution of representations in Transformers using topological data analysis (TDA). While previous studies have focused on local geometric properties or isolated layers, TGO-IV aims to provide a comprehensive understanding of how representation point clouds transform across layers. By constructing Vietoris–Rips simplicial complexes from token-level representations, the authors investigate the persistent topological signatures that emerge throughout the forward pass of the network. This framework includes various topological observatories such as Persistence Diagrams, Barcode Diagrams, and Betti Curves, allowing for a systematic characterization of the evolution of representation topology. The study addresses critical research questions regarding the systematic evolution of representation topology across Transformer depth and whether topological transitions occur in specific layers. The findings suggest that the representation manifold undergoes significant topological changes, providing new insights into the learning dynamics of Transformers beyond traditional geometric and spectral analyses.
Methodology
The authors construct Vietoris–Rips simplicial complexes from token-level representation point clouds and analyze their persistent homology across Transformer layers. This involves using various topological tools such as Persistence Diagrams, Barcode Diagrams, and Betti Curves to quantify the evolution of topological features.
Results
The study reveals that the representation topology evolves systematically across Transformer layers, with significant topological transitions occurring. The findings indicate that the representation manifold does not merely preserve its structure but undergoes fundamental changes during the forward pass.
Implications
The insights gained from this study could enhance the interpretability of Transformers, providing a deeper understanding of how these models learn and evolve representations. This could inform future research on model design and optimization in both NLP and computer vision tasks.
Toward Human Rights Benchmarking for LLMs: A Pilot Methodology
Large Language Models
NLP
- Introduction of HumRightsBench, the first benchmark for evaluating LLM reasoning in human rights law.
- Adaptation of the IRAC framework to the IRAP framework for structured human rights legal reasoning.
- Development of a scenario corpus validated by human rights professionals.
- Pilot results show significant variance in LLM performance, indicating the benchmark's utility and the urgency of improving AI reasoning in human rights contexts.
Read more
Toward Human Rights Benchmarking for LLMs: A Pilot Methodology
Summary
This paper addresses the lack of evaluation benchmarks for large language models (LLMs) in the context of human rights law. The authors introduce HumRightsBench, the first expert-validated benchmark designed to assess LLMs' reasoning capabilities regarding human rights legal analysis. They adapt the IRAC framework to create the IRAP framework, which includes Issue Identification, Rule Recall, Rule Application, and Proposed Remedies. The benchmark is grounded in authentic, scenario-based prompts validated by human rights lawyers and professionals. Pilot results demonstrate that LLMs exhibit varying accuracy scores (0.339-0.577) across different reasoning tasks, indicating that while some models perform reasonably well, they are not yet adequate for complex human rights reasoning tasks. The findings highlight the need for improved AI evaluations in this critical area, as current models struggle particularly with detecting obligation violations, a fundamental aspect of human rights law.
Methodology
The authors developed the HumRightsBench benchmark by adapting the IRAC framework to create the IRAP framework, which breaks down human rights reasoning into four subtasks. They created scenario-based prompts grounded in international human rights law, validated by experts in the field. The benchmark was tested on three leading LLMs, and performance was assessed based on accuracy across various reasoning tasks.
Results
The pilot study revealed that the overall accuracy of the tested LLMs ranged from 0.339 to 0.577, with significant variance across different tasks. The models performed particularly poorly in detecting obligation violations, underscoring the limitations of current LLMs in handling complex human rights legal reasoning.
Implications
The development of HumRightsBench has the potential to advance the evaluation of AI systems in legal contexts, particularly in human rights. It emphasizes the necessity for rigorous benchmarks to ensure that LLMs can adequately support legal determinations and uphold human rights standards.
SoftmaxGRPO: Learning to Reason using Softmax Advantage Group Estimation
Reinforcement Learning
Theory
Optimization
- SoftmaxGRPO is a drop-in replacement for GRPO that uses temperature-scaled softmax advantages.
- It maintains bounded weights regardless of prompt difficulty, addressing issues of gradient concentration on easy prompts.
- The paper derives exact objectives for binary and bounded scalar rewards, highlighting limitations of existing methods.
- Empirical results demonstrate significant performance improvements over GRPO in various tasks.
Read more
SoftmaxGRPO: Learning to Reason using Softmax Advantage Group Estimation
Summary
The paper introduces Softmax Advantage Group Estimation (SoftmaxGRPO), a novel approach to group-based reinforcement learning that addresses the limitations of existing methods like GRPO, particularly in handling prompt difficulty and reward normalization. Traditional group-based objectives often lead to poor allocation of learning signals, especially under binary rewards, where easy prompts receive disproportionate attention. SoftmaxGRPO replaces z-score-normalized group advantages with temperature-scaled softmax advantages, ensuring bounded weights across varying prompt difficulties. The authors derive the exact finite-group population objective for binary rewards and establish its MaxRL limit, while also demonstrating that a universal finite-group scalar objective cannot exist without additional assumptions. Empirical results show that SoftmaxGRPO reallocates gradient budgets effectively, improving performance on tasks with weak rewards, such as DeepMath and Poetry, outperforming GRPO under identical conditions. This method enhances learning efficiency and effectiveness in both verifiable and non-verifiable tasks.
Methodology
The authors propose SoftmaxGRPO as a one-line modification to GRPO, utilizing temperature-scaled softmax advantages instead of z-score normalization. They derive the finite-group population objective for binary rewards and analyze its behavior in the low-temperature limit. The methodology includes empirical evaluations across different tasks, comparing performance against GRPO and other variants.
Results
SoftmaxGRPO achieved a pass rate of 51.8% on the DeepMath task with verifiable rewards and improved a 1.5B instruction-tuned model's performance from 35.0% to 68.0% on the Poetry task using lightweight text-similarity rewards. These results indicate a consistent improvement over GRPO under identical reward conditions.
Implications
The introduction of SoftmaxGRPO has significant implications for reinforcement learning, particularly in scenarios where reward signals are weak or noisy. By improving the allocation of learning signals, this method can enhance model performance in complex tasks such as summarization, open-ended question answering, and creative generation, making it a valuable tool for post-training optimization.
From token probabilities to calibrated confidence: An empirical study of mathematical question answering
NLP
Large Language Models
- Token probabilities from LLMs are often overconfident and poorly calibrated.
- Aggregating probabilities over the full reasoning trajectory leads to better confidence estimates.
- Multi-pass methods, including self-verification and Monte Carlo Dropout, can yield calibrated confidence.
- Post-hoc calibration methods like Platt scaling and isotonic regression significantly reduce calibration error.
Read more
From token probabilities to calibrated confidence: An empirical study of mathematical question answering
Summary
This paper investigates the confidence estimation of large language models (LLMs) in the context of mathematical question answering. It addresses the challenge of overconfidence in token probabilities, which are often poorly calibrated. The authors explore both single-pass and multi-pass confidence estimation methods, comparing how different aggregation strategies of token probabilities can yield more reliable confidence scores. They introduce multi-pass methods such as self-verification through re-prompting and Monte Carlo Dropout, which provide additional confidence signals. The study also evaluates post-hoc calibration techniques, specifically Platt scaling and isotonic regression, to improve the alignment of confidence estimates with empirical accuracy. The findings indicate that aggregating token probabilities over the entire reasoning trajectory enhances the discriminative power of confidence estimates, while multi-pass methods can effectively yield calibrated confidence. The paper concludes that calibration efficiency varies with dataset difficulty and that transferability of calibration mappings is model and dataset dependent.
Methodology
The authors systematically compare single-pass confidence estimators based on different subsets and aggregations of token probabilities. They also evaluate multi-pass methods, including self-verification and Monte Carlo Dropout, to derive additional confidence signals. Finally, they assess the effectiveness of post-hoc calibration techniques, Platt scaling and isotonic regression, in aligning confidence estimates with empirical accuracy.
Results
The study finds that aggregating token probabilities over the entire sequence improves the discriminative power of confidence estimates. Multi-pass methods provide calibrated confidence estimates, with the in-situ self-verification variant achieving significant efficiency gains. Both Platt scaling and isotonic regression reduce in-domain calibration error, with isotonic regression being effective with as few as 50 examples. Calibration efficiency varies with dataset difficulty, and the transferability of calibration mappings is asymmetric across different datasets and models.
Implications
The findings have significant implications for deploying LLMs in high-stakes domains where understanding the reliability of model outputs is crucial. Improved confidence estimation and calibration can enhance trust in automated systems used in fields like finance, law, and healthcare.
Uncertainty-Aware Ensemble Deep Randomized Neural Networks for Classification
Theory
Efficient ML
- Introduction of IF-dRVFL and IF-edRVFL models to enhance robustness against noise and outliers.
- Adaptive sample weighting based on intuitionistic fuzzy scores improves discrimination among data points.
- Stacked robust hidden layers reduce the impact of noisy features during training.
- Extensive experiments validate the superiority of the proposed models over existing approaches.
Read more
Uncertainty-Aware Ensemble Deep Randomized Neural Networks for Classification
Summary
This paper addresses the limitations of existing deep randomized neural networks (dRVFL and edRVFL) in handling noisy and outlier-contaminated datasets. The authors propose two novel frameworks: intuitionistic fuzzy dRVFL (IF-dRVFL) and intuitionistic fuzzy edRVFL (IF-edRVFL). These models leverage intuitionistic fuzzy theory to assign adaptive weights to training samples based on their membership and non-membership degrees. This approach allows for effective discrimination between clean, noisy, and outlier data points, enhancing the robustness of the models. The proposed frameworks incorporate stacked robust hidden layers that mitigate the propagation of noise through the network. Extensive experiments on UCI and KEEL benchmark datasets demonstrate that IF-dRVFL and IF-edRVFL outperform existing state-of-the-art methods, both fuzzy and non-fuzzy, in terms of classification accuracy and robustness against noise. The source code for the models is made available for further research and application.
Methodology
The authors developed IF-dRVFL and IF-edRVFL models that utilize intuitionistic fuzzy theory to compute membership and non-membership degrees for training samples. These degrees are used to assign adaptive weights, allowing the models to effectively handle noise and outliers. The models are evaluated through extensive experiments on benchmark datasets, comparing their performance against existing state-of-the-art methods.
Results
The experimental results indicate that both IF-dRVFL and IF-edRVFL significantly outperform traditional dRVFL and edRVFL models, as well as other fuzzy and non-fuzzy approaches, particularly in the presence of Gaussian noise. The proposed models demonstrate improved classification accuracy and robustness, confirming their effectiveness in real-world applications.
Implications
The findings suggest that incorporating intuitionistic fuzzy theory into deep randomized neural networks can greatly enhance their applicability in noisy environments, making them suitable for various real-world classification tasks, including those in fields like healthcare, finance, and image recognition.
Exact Rank and Convex Calibration Dimension Lower Bounds for the Multi-Label F1 Loss
Theory
- Determined the exact rank of the F1 loss matrix as s² - s + 2.
- Established the column-affine dimension of the F1 loss as s² - s + 1.
- Proved a quadratic lower bound for the convex calibration dimension of the F1 loss.
- Constructed a specific distribution to analyze Bayes optimal predictions related to the F1 loss.
Read more
Exact Rank and Convex Calibration Dimension Lower Bounds for the Multi-Label F1 Loss
Summary
This paper investigates the multi-label F1 loss, a critical performance metric in multi-label classification, by determining its exact rank and establishing lower bounds for its convex calibration dimension. The author identifies that the F1 score matrix, along with its shifted and unshifted loss matrices, has a rank of s² - s + 2, while the column-affine dimension is s² - s + 1. The study further explores the Bayes geometry of the F1 loss, constructing a distribution that reveals all supersets of a fixed core label set as Bayes optimal. This leads to a lower bound for the convex calibration dimension, showing that CCdim(LF1) is at least (2/3)√3s², thus establishing that CCdim(LF1) = Θ(s²). The findings provide a deeper understanding of the structure of the F1 loss and its implications for developing efficient algorithms in multi-label classification.
Methodology
The author employs mathematical proofs involving matrix rank and affine dimension analysis, alongside the construction of specific distributions to analyze the Bayes geometry of the F1 loss. The study utilizes concepts from decision theory and convex analysis to derive lower bounds for the convex calibration dimension.
Results
The paper concludes that the exact rank of the F1 loss matrix is s² - s + 2, and the column-affine dimension is s² - s + 1. Additionally, it establishes that the convex calibration dimension CCdim(LF1) is at least (2/3)√3s², leading to the result that CCdim(LF1) = Θ(s²).
Implications
The results have significant implications for the design of algorithms in multi-label classification, particularly in understanding the complexity of loss functions and the development of efficient calibrated surrogates. This can enhance the performance of machine learning models in applications requiring multi-label predictions.
Loss-Resilient Wireless Video Token Communication over Block Fading Channels
Computer Vision
Generative Models
Efficient ML
- Development of a video-token importance metric distinguishing I- and P-tokens for effective reconstruction.
- Introduction of an importance-aware packetization scheme that reduces information loss during transmission.
- Design of an online CSI-aware packet allocation algorithm that optimizes packet distribution based on channel conditions.
- Demonstration of improved video quality and resilience against packet loss in wireless transmission scenarios.
Read more
Loss-Resilient Wireless Video Token Communication over Block Fading Channels
Summary
This paper presents a novel framework for wireless video token communication (WVTC) designed to enhance video transmission over block fading channels. The authors address the challenge of token loss during transmission, which can severely impact video reconstruction due to the temporal dependencies among video tokens. The proposed WVTC framework evaluates the importance of video tokens based on their structural and temporal characteristics, assigning higher priority to structural I-tokens while measuring the importance of P-tokens through local representation changes. A shuffling mixed I/P-token packetization scheme is introduced to disperse critical tokens across packets, thereby minimizing the risk of losing multiple important tokens in a single packet. Additionally, an online channel state information (CSI)-aware packet allocation algorithm is developed, which optimally allocates packets based on their importance and current channel conditions. The receiver employs a detokenizer that reconstructs missing video content from the received tokens without requiring retransmission. The numerical results demonstrate that the proposed framework significantly improves perceptual quality and allows for more graceful degradation of video quality under increasing packet error rates.
Methodology
The methodology involves evaluating token importance based on structural and temporal characteristics, implementing a mixed I/P-token packetization scheme to disperse critical tokens, and designing an online scheduler that allocates packets based on current block channel state information. The receiver utilizes a detokenizer to reconstruct video content from the successfully received tokens.
Results
The proposed WVTC framework showed improved perceptual quality of video transmission and demonstrated more graceful degradation under conditions of increasing packet error rates compared to conventional methods.
Implications
The findings suggest that the WVTC framework can be effectively applied in real-time video communication systems, particularly in environments with unreliable wireless channels, enhancing the robustness and quality of video streaming applications.
Test-Time Augmentation for LLMs: When Input Diversity Beats Output Diversity at Matched Compute
NLP
Large Language Models
Efficient ML
- Test-Time Augmentation (TTA) introduces input diversity to improve LLM accuracy during inference.
- Semantic rephrasing significantly outperforms self-consistency in accuracy and cost-effectiveness.
- TTA is particularly beneficial for mid-tier models, maximizing the efficiency of compute usage.
- The study provides practical insights into the number of augmentations and multi-modal strategies.
Read more
Test-Time Augmentation for LLMs: When Input Diversity Beats Output Diversity at Matched Compute
Summary
This paper investigates the effectiveness of Test-Time Augmentation (TTA) in improving the accuracy of Large Language Models (LLMs) during inference while managing compute costs. Unlike traditional methods like self-consistency that focus solely on output diversity, TTA introduces input diversity by perturbing the input data and aggregating predictions from multiple transformed versions. The authors conduct a systematic comparison of TTA against established methods, including chain-of-thought prompting and self-consistency, across six diverse datasets. Their findings reveal that semantic rephrasing, one of the input-side strategies, consistently outperforms the baselines in terms of accuracy and cost-effectiveness, achieving approximately 1.8 times more accuracy per dollar spent compared to self-consistency. The study also highlights the cost-effectiveness of TTA for mid-tier models, suggesting that varying input can convert compute into accuracy more efficiently than varying reasoning paths alone. The implementation of TTA is made publicly available, providing a practical tool for enhancing LLM performance without the need for retraining.
Methodology
The authors conducted a matched-compute comparison of three input-side strategies (semantic rephrasing, lexical perturbations, and visual transformations) against output-side methods like chain-of-thought prompting and self-consistency. They evaluated these methods across six datasets, analyzing accuracy gains and cost-effectiveness.
Results
The results indicated that semantic rephrasing yielded statistically significant accuracy improvements over the baselines, outperforming self-consistency on five out of six tasks. TTA provided approximately 1.8 times more accuracy per dollar compared to self-consistency, demonstrating its efficiency in converting compute into accuracy.
Implications
The findings suggest that practitioners can achieve better performance from LLMs by employing input-side diversity techniques like TTA, particularly in scenarios where compute resources are limited. This approach can enhance the deployment of LLMs in various applications, including multilingual tasks and sentiment analysis.
Observational Policy Ranking for SMB Financial Guidance from Multi-Action Accounting Logs
Reinforcement Learning
Optimization
Theory
- Introduces Observational Policy Ranking for SMB financial guidance from accounting logs.
- Develops CAR-PL, a covariate-adjusted action-wise R-learner for multi-hot logs.
- Demonstrates CAR-PL's effectiveness in achieving the highest Gross Profit estimates.
- Highlights the significance of recommendation concentration in policy comparison.
Read more
Observational Policy Ranking for SMB Financial Guidance from Multi-Action Accounting Logs
Summary
This paper addresses the challenge of providing timely financial guidance to small and medium-sized businesses (SMBs) using historical accounting logs, which often reflect self-selected and co-occurring business changes rather than randomized recommendations. The authors introduce a novel approach termed Observational Policy Ranking, which formulates the problem of selecting business-change categories based on pre-decision financial information. They present Covariate-Adjusted Residual Policy Learning (CAR-PL), an action-wise R-learner that operates on multi-hot logs and incorporates observational support to regularize selection. The methodology is validated using a dataset of 85,078 company-month observations from 7,505 firms, comparing CAR-PL against several benchmarks including an uplift T-Learner and a contextual value model. The results indicate that CAR-PL achieves the highest Gross Profit point estimate, while the T-Learner excels in Revenue estimates. The study emphasizes the importance of recommendation concentration and the stability of policy selection across different financial KPIs, ultimately supporting the objective-specific ranking of financial guidance for SMBs.
Methodology
The authors formulated the problem as observational policy ranking, utilizing a dataset of accounting logs to select from 34 business-change categories. They introduced CAR-PL, which combines action-wise R-learning with support shrinkage to stabilize category selection. The performance of CAR-PL was compared against an uplift T-Learner, a contextual value model, and other non-personalized policies using a shared model-assisted scoring rule.
Results
CAR-PL achieved the highest Gross Profit point estimate (0.084), while the T-Learner had the highest Revenue point estimate (0.085). The contextual value model led in Quick Ratio estimates (0.062). Notably, CAR-PL and the T-Learner were not statistically separated on growth KPIs, indicating comparable effectiveness despite differing category selections.
Implications
The findings suggest that SMBs can benefit from data-driven financial guidance derived from historical accounting logs, potentially improving decision-making processes. The methodology can be applied in various financial advisory contexts, enhancing the scalability of expert guidance through automated systems.
Correlation flow governs learning at criticality
Theory
- Establishes a direct link between correlation propagation and the Neural Tangent Kernel (NTK).
- Identifies a critical point in the weight-bias variance plane for effective correlation propagation.
- Demonstrates that orthogonal initialization suppresses finite-size corrections compared to Gaussian initialization.
- Proves that the NTK is proportional to output correlation at infinite depth.
Read more
Correlation flow governs learning at criticality
Summary
This paper investigates the relationship between correlation propagation and learning dynamics in deep neural networks, particularly focusing on the role of initialization. The authors utilize mean-field theory and random matrix theory to establish a connection between correlation flow and the Neural Tangent Kernel (NTK) in infinitely wide and deep networks. They identify a critical point in the weight-bias variance plane where correlation propagation to infinite depth occurs, leading to a vanishing end-to-end Jacobian with depth. This finding reveals that the NTK becomes directly proportional to output correlation at infinite depth, highlighting a previously unnoticed equivalence between information propagation and learning dynamics. The study also demonstrates that orthogonal initialization can suppress finite-size corrections that arise under Gaussian initialization, thereby clarifying the impact of different initialization methods. The theoretical predictions are validated through quantitative analysis on finite-width, finite-depth networks, emphasizing the importance of orthogonal initialization at criticality in controlling deep learning dynamics.
Methodology
The authors combine mean-field theory and random matrix theory to analyze the dynamics of deep neural networks. They focus on the correlation propagation and the behavior of the NTK in the sequential limit where width approaches infinity before depth. The study employs theoretical analysis and quantitative validation on networks of finite width and depth.
Results
The paper finds that correlation propagation to infinite depth is achievable only at a specific critical point in the weight-bias variance plane. At this critical point, the end-to-end Jacobian vanishes algebraically with depth, leading to the conclusion that the NTK is exactly proportional to output correlation at infinite depth. Additionally, orthogonal initialization is shown to effectively control the dynamics of deep learning by suppressing harmful finite-size corrections.
Implications
The findings suggest that careful initialization strategies, particularly orthogonal initialization, can significantly enhance the training dynamics of deep neural networks, potentially leading to more efficient learning in large-scale models. This work contributes to the theoretical understanding of deep learning and may inform future architectural designs and initialization techniques.
Real-Time Climate Risk Assessment for Supply Chain Resilience: A Data-Driven Nowcasting Framework for Colombian Agriculture
Time Series
- Development of a real-time climate risk assessment framework for Colombian agriculture.
- Integration of short-term climate forecasting with supply chain risk modeling.
- Prototype implementation demonstrates feasibility using historical data without satellite imagery.
- Short-term precipitation nowcasts can inform actionable risk indicators for supply chains.
Read more
Real-Time Climate Risk Assessment for Supply Chain Resilience: A Data-Driven Nowcasting Framework for Colombian Agriculture
Summary
This paper introduces a methodological framework for real-time climate risk assessment aimed at enhancing supply chain resilience in Colombian agriculture. The framework utilizes data-driven nowcasting techniques to address the challenges posed by climate variability, which significantly affects agricultural production and logistics. The proposed approach integrates short-term climate forecasting based on historical meteorological data with supply chain risk modeling, establishing a conceptual early warning system. A prototype implementation demonstrates the feasibility of this framework using historical meteorological and agricultural time series data, without the need for satellite imagery or complex computer vision systems. The methodology incorporates explicit risk mapping, threshold-based categorization, and stakeholder-oriented risk signals to facilitate decision-making in supply chains. Experimental results indicate that short-term precipitation nowcasts can be effectively translated into actionable risk indicators, aiding anticipatory decisions related to inventory management, sourcing, and transportation. This work addresses critical gaps in climate-adaptive supply chain management, particularly in developing countries where advanced remote sensing infrastructure is limited.
Methodology
The methodology involves integrating historical ground-based meteorological observations with agricultural statistics to establish climate-production relationships. It employs nowcasting techniques to generate short-term climate forecasts and translates these into threshold-based risk categories and early warning signals. A prototype was validated in a controlled environment using synthetic scenarios based on Colombian climate variability.
Results
The results from synthetic and historical data experiments show that the framework can produce actionable risk indicators based on short-term precipitation forecasts, which support anticipatory supply chain decisions regarding inventory, sourcing, and transport.
Implications
The implications of this work extend to improving supply chain resilience in agriculture by providing timely climate information that can enhance decision-making processes. It lays the groundwork for future field deployments and collaborations with national institutions to refine and operationalize the framework.
Optimistic Rates for Multiclass PAC Learning
Theory
- Introduction of optimistic rates that scale with oracle risk in multiclass PAC learning.
- Establishment of upper and lower bounds for optimal excess risk based on Natarajan and Daniely–Shalev-Shwartz dimensions.
- Development of a new relative compression theorem that enhances existing learning architectures.
- Extension of results to list learning, addressing the complexity of multiple hypotheses.
Read more
Optimistic Rates for Multiclass PAC Learning
Summary
This paper addresses the limitations of worst-case bounds in multiclass PAC learning, particularly when the best classifier is nearly correct. The authors introduce optimistic rates, which provide guarantees that scale with the oracle risk. They establish that for classes characterized by Natarajan dimension (dN) and Daniely–Shalev-Shwartz dimension (dDS), the optimal excess risk can be expressed as a function of the oracle risk (L★) and the dimensions. The paper closes the gap in existing bounds by presenting an upper bound that combines the cover–menu–compression architecture with a new relative compression theorem. This upper bound is shown to hold uniformly across different alphabet sizes, even when the learner is unaware of L★ or the confidence level. The authors also provide a matching lower bound using a pair-Assouad scheme, demonstrating that both terms in the upper bound are necessary. Additionally, the results extend to list learning, yielding optimistic rates and lower bounds that reflect the complexity of the task. The main theorems are verified using machine-checking in Lean 4, ensuring robustness in the findings.
Methodology
The authors employ a combination of theoretical analysis and machine-checking to derive optimistic rates for multiclass PAC learning. They utilize the cover–menu–compression architecture and introduce a new relative compression theorem to establish upper bounds. Lower bounds are derived using a pair-Assouad scheme and fiber arguments related to pseudo-cubes. The methodology also includes a detailed examination of the learning process in the context of both realizable and agnostic settings.
Results
The paper presents an upper bound for the optimal excess risk that is eΘ(L★dN/n + dDS/n), which holds uniformly across alphabet sizes. The lower bound is shown to be necessary and matches the upper bound, confirming the tightness of the results. The findings extend to list learning, providing optimistic rates and lower bounds that reflect the complexity of learning from multiple hypotheses.
Implications
The results have significant implications for the design of learning algorithms in multiclass settings, particularly in scenarios where the best hypothesis is nearly correct. The optimistic rates can lead to more efficient learning strategies that adapt to the level of realizability, potentially improving performance in practical applications such as image classification and natural language processing.
Power law graph attention: exact generalization of scaled dot-product attention, empirical collapse at inference
Large Language Models
NLP
Theory
- PLGA generalizes SDPA by introducing a learned bilinear operator derived from input data.
- The architecture allows for exact containment of SDPA, demonstrating a significant mathematical relationship.
- An inference-collapse phenomenon is observed, where outputs remain invariant under input perturbations.
- The model exhibits a learned singularity condition, indicating a unique structure in its attention dynamics.
Read more
Power law graph attention: exact generalization of scaled dot-product attention, empirical collapse at inference
Summary
This paper introduces the Power Law Graph Attention (PLGA) mechanism, which generalizes the conventional scaled dot-product attention (SDPA) used in large language models (LLMs) by replacing its fixed bilinear form with a learned, input-generated bilinear operator. The architecture, termed the Large Language Model from Power Law Decoder Representations (PLDR-LLM), is rigorously defined and verified against reference releases. The author presents several key findings, including the exact containment of SDPA within the PLGA framework when the learned operator equals the identity matrix, and the emergence of an inference-collapse phenomenon where the model's outputs become invariant to input changes, allowing for a simplified inference process. The paper also discusses the implications of self-organized criticality in the attention dynamics and presents empirical results demonstrating the model's stability and performance. The findings suggest that PLGA offers a more flexible and mathematically robust alternative to traditional attention mechanisms, with potential applications in enhancing the efficiency and effectiveness of LLMs.
Methodology
The paper employs a mathematical framework to define the PLGA operator and its properties, utilizing concepts from Perron–Frobenius theory and self-organized criticality. The architecture is validated through rigorous proofs and empirical measurements, including blockwise training and scoring methods, and machine-checked proofs in Lean 4.
Results
The PLGA mechanism was shown to maintain operator invariance with relative fluctuations of 10^-6 or lower, indicating high stability. The learned singularity condition was confirmed, with the interaction tensor remaining strictly positive and exhibiting a Perron–Frobenius structure. The model's performance was validated against benchmarks, demonstrating consistent results across different scoring protocols.
Implications
The findings suggest that PLGA could enhance the design of future LLMs by providing a more adaptable and mathematically grounded attention mechanism. This could lead to improvements in inference efficiency and model robustness, potentially impacting various applications in natural language processing and beyond.
Coordinating the Unknown Lipschitz Constant in Multiplayer Bandits
Theory
Optimization
- Introduces a meta-algorithm, mECAB, for estimating the Lipschitz constant in multiplayer bandits.
- Analyzes three information structures affecting coordination among players.
- Proves regret bounds that highlight the importance of information structure in achieving coordination.
- Demonstrates that agreement can be achieved without communication through dithered quantization.
Read more
Coordinating the Unknown Lipschitz Constant in Multiplayer Bandits
Summary
This paper investigates cooperative multi-agent bandit problems in continuous action spaces where the Lipschitz constant is unknown. The authors consider three distinct information structures: (A) unobserved actions with common rewards, (B) observed actions with independent rewards, and (C) unobserved actions with independent rewards. They propose a meta-algorithm, mECAB, which estimates the Lipschitz constant through uniform exploration of a coarse grid, selects a discretization of the joint action space, and applies a cooperative bandit method to the resulting discrete problem. The key challenge lies in the players' inability to communicate once learning begins, necessitating a shared discretization based on their independent observations. The authors demonstrate that common rewards and observable actions facilitate agreement on discretization without additional cost, while in their absence, a dithered quantization method can achieve agreement at no leading order regret cost. The paper provides regret guarantees for all three scenarios and includes simulations comparing Lipschitz-adaptive discretization with non-adaptive methods.
Methodology
The authors develop a meta-algorithm that estimates an upper confidence bound on the Lipschitz constant through uniform exploration of a coarse grid. They then fix a discretization of the joint action space and apply a cooperative multiplayer multi-armed bandit (MAB) algorithm to the induced discrete problem. The analysis includes proving regret bounds for different information structures and conducting simulations to compare adaptive and non-adaptive discretization strategies.
Results
The paper establishes regret guarantees for all three information structures, demonstrating that common rewards and observable actions facilitate coordination without additional costs. In scenarios lacking these features, the authors show that a dithered quantization approach can still achieve agreement among players, maintaining a competitive regret performance. The simulations indicate that Lipschitz-adaptive discretization outperforms non-adaptive methods.
Implications
The findings have significant implications for decentralized applications in areas such as wireless communication, distributed sensing, and multi-agent systems, where agents must coordinate their actions under information constraints. The proposed methods can enhance decision-making efficiency in environments where communication is limited or unavailable.
Efficient Reinforcement Learning for Long-Horizon Tool-Use Agentic Tasks
Reinforcement Learning
Efficient ML
Robotics
- Introduction of SINKFLEX-RL, a modular RL training system for dual-control environments.
- Utilization of group-relative policy optimization to enhance training efficiency without a separate value model.
- Implementation of sink-aware FlexAttention to optimize memory usage during long-context training.
- Demonstrated improvements in validation rewards and memory efficiency in preliminary experiments.
Read more
Efficient Reinforcement Learning for Long-Horizon Tool-Use Agentic Tasks
Summary
This paper presents SINKFLEX-RL, a modular training system designed for reinforcement learning (RL) in dual-control tool-use environments, addressing the challenges of long-horizon tasks that require reasoning over user goals, domain policies, and delayed rewards. The authors highlight the inefficiencies of traditional on-policy rollouts in RL, which can lead to excessive memory usage due to long contexts. SINKFLEX-RL integrates several components: a Gymnasium-compatible environment wrapper, a VERL-style rollout dataflow, and group-relative policy optimization (GRPO) that eliminates the need for a separate value model. Additionally, it introduces a sink-aware FlexAttention mechanism that optimizes memory usage while maintaining model-specific scaling under various masking conditions. The paper evaluates the system through preliminary experiments on the Ï„2-Bench retail benchmark, demonstrating significant improvements in validation rewards and memory efficiency. The findings underscore the importance of integrating environment interfaces, RL dataflow, and attention mechanisms to facilitate memory-efficient training for long-horizon agentic tasks.
Methodology
The authors developed SINKFLEX-RL by wrapping dual-control environments in a Gymnasium-compatible interface, employing group-relative policy optimization (GRPO) for actor updates, and implementing a sink-aware FlexAttention path that combines causal and sliding-window masks with differentiable sink scaling. This approach aims to reduce memory overhead while maintaining effective training for long-horizon tasks.
Results
In preliminary experiments on the Ï„2-Bench retail benchmark, the validation reward increased from 0.25 to 0.44 during training, indicating improved agent performance. Additionally, the optimized attention path reduced peak VRAM usage by 19.7%, demonstrating significant memory efficiency gains while handling longer sequences.
Implications
The findings suggest that integrating efficient attention mechanisms and RL dataflows can lead to more scalable and effective training systems for complex, long-horizon tasks in interactive environments. This has potential applications in areas such as customer service automation, robotics, and any domain requiring sophisticated tool use and user interaction.
Full-Feature versus Limited-Input Machine Learning for Residential Energy Estimation: A Comparative Analysis of RECS and ResStock Under Realistic Input Constraints
Efficient ML
- Machine learning models for residential energy estimation often rely on extensive feature sets that are impractical for real-world applications.
- The study demonstrates a significant drop in predictive accuracy when models are limited to easily obtainable inputs.
- Targeted modeling for homogeneous populations can enhance accuracy, suggesting a need for cohort-specific approaches.
- Tree-based ensemble models can serve as effective emulators for national-scale residential energy datasets.
Read more
Full-Feature versus Limited-Input Machine Learning for Residential Energy Estimation: A Comparative Analysis of RECS and ResStock Under Realistic Input Constraints
Summary
This paper investigates the trade-off between predictive accuracy and input accessibility in residential energy estimation using machine learning models. It compares two datasets: the Residential Energy Consumption Survey (RECS) and the simulation-based ResStock dataset. A structured machine-learning pipeline was developed to estimate total energy, space heating, and space cooling. Full-feature models established performance benchmarks, achieving high predictive accuracy (R² ≈ 0.90 for ResStock and R² ≈ 0.73 for RECS). However, when restricted to ten low-burden variables, model performance dropped to R² ≈ 0.61 for RECS and R² ≈ 0.62 for ResStock. Notably, a targeted model for a homogeneous cohort improved accuracy to R² ≈ 0.85. The findings suggest that while tree-based ensemble models can effectively emulate national-scale energy datasets, the availability of features and the dataset's origin are crucial for model performance. The study advocates for a tiered deployment strategy, utilizing low-burden models for preliminary assessments and detailed audits for critical decisions.
Methodology
The authors developed a structured machine-learning pipeline to estimate residential energy consumption using two datasets (RECS and ResStock). They first established benchmarks with full-feature models and then restricted the input to ten low-burden variables to simulate realistic deployment conditions. Various algorithms, including CatBoost, XGBoost, LightGBM, Random Forest, and Neural Networks, were evaluated for their predictive performance.
Results
Full-feature models achieved R² values of approximately 0.90 for ResStock and 0.73 for RECS. When limited to ten accessible inputs, performance decreased to R² ≈ 0.61 for RECS and R² ≈ 0.62 for ResStock. However, a targeted model for a specific cohort improved accuracy to R² ≈ 0.85, indicating the effectiveness of focused modeling.
Implications
The findings highlight the importance of considering input accessibility in residential energy modeling. They suggest that low-burden models can be useful for preliminary assessments, while more detailed audits are necessary for high-stakes decisions. This approach can enhance energy efficiency planning and support better decision-making in residential energy management.
Multi-Relational Knowledge Graph Enhanced Embedding for Trajectory-User Linking
Graph Learning
Time Series
- MakeTUL is the first method to integrate knowledge graph representation learning into Trajectory-User Linking.
- The approach organizes trajectory data into a multi-relational knowledge graph, enhancing the representation of mobility semantics.
- MakeTUL significantly outperforms existing methods in trajectory-user linking tasks across multiple datasets.
- The dual-branch classification layer effectively combines global and sequential evidence for improved classification accuracy.
Read more
Multi-Relational Knowledge Graph Enhanced Embedding for Trajectory-User Linking
Summary
The paper introduces MakeTUL, a novel approach for Trajectory-User Linking (TUL) that leverages multi-relational knowledge graphs to enhance trajectory representation learning. Traditional methods often treat features such as Point of Interest (POI), temporal, and semantic information independently, which limits their effectiveness in capturing the structural knowledge inherent in trajectories. MakeTUL organizes visit-time, POI-category, and transfer-speed data into a multi-relational mobility knowledge graph, allowing for a more integrated representation of mobility semantics. The method enriches POI representations with high-order co-occurrence patterns and combines these with temporal and categorical information to improve the learning of ordered mobility patterns. A dual-branch classification layer is employed to merge global structural and sequential evidence for better decision-making. The effectiveness of MakeTUL is validated through experiments on multiple datasets, where it outperforms existing methods across various metrics, demonstrating the advantages of incorporating multi-relational knowledge and structural priors in trajectory classification tasks.
Methodology
MakeTUL utilizes a multi-relational mobility knowledge graph to represent various aspects of trajectories, including visit-time, POI-category, and transfer-speed. It enhances POI representations with high-order co-occurrence patterns and employs a trajectory sequence learning module to capture ordered mobility patterns. A dual-branch classification layer is used to integrate global structural evidence and sequential evidence for decision-making.
Results
MakeTUL achieved the best performance on all five evaluation metrics (ACC@1, ACC@5, Macro-Precision, Macro-Recall, and Macro-F1) across six user-scale settings. Notably, on the Foursquare-JKT dataset with 1,400 users, it exceeded the strongest baselines by 10.67% in ACC@1 and 12.70% in Macro-F1, demonstrating the effectiveness of its multi-relational knowledge and structural priors.
Implications
The findings suggest that incorporating multi-relational knowledge graphs can significantly enhance the accuracy of trajectory-user linking, which has implications for user mobility analysis and personalized location-based services. This approach could be applied in various fields, including urban planning, transportation, and social network analysis.
SAGE: SLO-Aware Adaptive Retrieval for Production RAG Systems
NLP
Large Language Models
Efficient ML
- SAGE dynamically adjusts retrieval budgets based on query difficulty, improving efficiency.
- The method achieves 95% SLO compliance under a 5-second latency constraint.
- SAGE reduces retrieval costs by 51% while maintaining high answer quality.
- The policy generalizes across multiple datasets and LLM architectures without retraining.
Read more
SAGE: SLO-Aware Adaptive Retrieval for Production RAG Systems
Summary
The paper introduces SAGE, a novel SLO-aware adaptive retrieval policy designed for Retrieval-Augmented Generation (RAG) systems operating under strict service level objectives (SLOs) regarding tail latency and infrastructure costs. Traditional retrieval pipelines utilize a fixed retrieval budget that fails to account for the varying difficulty of queries, leading to inefficiencies such as over-retrieving for simple queries and under-serving complex ones. SAGE addresses this issue by dynamically predicting the optimal number of passages to retrieve for each query based on lightweight features derived from initial retrieval scores and rank statistics. The policy is trained offline using imitation learning from an oracle that approximates optimal latency-quality trade-offs. Empirical results demonstrate that SAGE significantly improves SLO compliance, achieving 95% compliance under a 5-second P95 latency SLO, compared to only 30% for the best static baseline. Additionally, SAGE reduces P95 latency by 36% and retrieval costs by 51%, with only a minor decrease in Exact Match (EM) performance. The approach generalizes well across different datasets and large language model (LLM) families, making it a practical solution for production RAG systems.
Methodology
SAGE employs a learned adaptive retrieval policy that utilizes lightweight features from initial retrieval processes, such as score distributions and rank gaps, to predict the optimal number of passages to retrieve for each query. It is trained offline using imitation learning from an oracle that simulates optimal latency-quality trade-offs, ensuring minimal runtime overhead during inference.
Results
SAGE demonstrated a significant improvement in SLO compliance, achieving 95% compliance under a 5-second P95 latency SLO, compared to 30% for the best static baseline (k=20). It also reduced P95 latency by 36% and retrieval costs by 51%, with only a 2 percentage point decrease in Exact Match (EM) performance. The policy effectively generalized across various datasets and LLM families, yielding consistent SLO improvements.
Implications
The findings suggest that SAGE can be a valuable tool for enhancing the efficiency and responsiveness of RAG systems in production environments, particularly in applications requiring strict adherence to latency and cost constraints. Its ability to generalize across different datasets and models indicates potential for widespread adoption in various user-facing applications.
IADD-TR: Intervention-Aware Dynamics Decoupling with Targeted Regularization for Model-Based Reinforcement Learning
Reinforcement Learning
Robotics
Optimization
- IADD-TR effectively decouples dynamics modeling into distinct stages to mitigate confounding bias.
- The framework incorporates targeted regularization to enhance policy learning and robustness.
- Extensive experiments show that IADD-TR achieves competitive returns with improved sample efficiency.
- The approach addresses the limitations of traditional MBRL methods by focusing on the interplay between actions and environmental evolution.
Read more
IADD-TR: Intervention-Aware Dynamics Decoupling with Targeted Regularization for Model-Based Reinforcement Learning
Summary
This paper presents IADD-TR, a novel framework for Model-Based Reinforcement Learning (MBRL) that addresses the challenges of confounding bias in both dynamics modeling and policy learning. Traditional MBRL methods often treat the transition model and critic as monolithic predictors, which can lead to biased data and poor generalization. IADD-TR introduces Intervention-Aware Dynamics Decoupling (IADD) to factorize transitions into an action-intervention stage and an action-free natural evolution stage, utilizing a zero-action anchor for robust generalization. Additionally, it employs Targeted Regularization (TR) to enhance policy learning by incorporating a residual correction based on action density, yielding doubly robust policy-gradient estimation. The framework was evaluated through extensive experiments on five MuJoCo tasks, demonstrating competitive performance and improved sample efficiency compared to existing methods.
Methodology
The methodology involves two main components: Intervention-Aware Dynamics Decoupling (IADD) and Targeted Regularization (TR). IADD separates the dynamics into an action-intervention stage and a natural evolution stage, utilizing a zero-action anchor for better generalization. TR enhances the critic's learning process by applying a correction based on action density, optimizing a targeted loss to improve policy-gradient estimation.
Results
The experiments conducted on five MuJoCo tasks revealed that IADD-TR not only achieved competitive returns but also demonstrated improved sample efficiency compared to existing MBRL methods. This indicates that the proposed framework effectively mitigates the issues related to confounding bias in both dynamics modeling and policy learning.
Implications
The findings suggest that IADD-TR could be applied in various real-world scenarios where sample efficiency is critical, such as robotics, autonomous driving, and other decision-making environments. By addressing confounding bias, the framework may lead to more reliable and generalizable reinforcement learning agents.
TACTICL: Task-Aware Compression of Tabular ICL Models
Efficient ML
- TACTICL enables significant layer pruning while retaining model performance.
- The framework integrates lightweight adapters to maintain in-context learning capabilities.
- Automated optimization of compression configurations is a key feature of TACTICL.
- The method demonstrates robustness against data shifts, enhancing its applicability.
Read more
TACTICL: Task-Aware Compression of Tabular ICL Models
Summary
The paper presents TACTICL, an automated framework designed for the task-aware compression of tabular in-context learning (ICL) models. As foundation models for tabular tasks have shown impressive performance, they also incur high inference costs. TACTICL addresses this by simultaneously pruning transformer layers and integrating lightweight adapters trained on specific downstream tasks, thus maintaining the model's in-context adaptability. The authors evaluate TACTICL on 47 benchmark datasets, demonstrating that it can effectively remove up to 85% of the layers without significantly impacting performance. The framework also shows resilience to data shifts, preserving in-context learning capabilities while reducing computational demands. This approach allows for a robust exploitation of the depth-wise redundancy in tabular foundation models, combining task-specific adaptation with structured compression.
Methodology
TACTICL employs a structured pruning strategy combined with the training of lightweight adapter modules. It automates the selection of layers to prune based on their contribution to task performance, optimizing a compression configuration that balances model size and predictive accuracy. The framework is designed to maintain the flexibility of the original model, allowing it to adapt to new tasks without retraining from scratch.
Results
The evaluation on 47 TabArena datasets indicates that TACTICL can maintain the performance of TabPFNv2.5 while achieving nearly linear reductions in computational costs corresponding to the number of layers removed. The framework successfully preserves the in-context learning ability and demonstrates robustness to data shifts.
Implications
TACTICL's approach to model compression has significant implications for deploying tabular foundation models in resource-constrained environments. By reducing inference costs while maintaining adaptability, it opens avenues for more efficient machine learning applications in various domains, particularly where tabular data is prevalent.
Cross-View Feature Matching: Survey, Benchmarking, and Foundation-Model Perspectives
Computer Vision
- Introduction of a hierarchical taxonomy for cross-view feature matching methods.
- Comprehensive benchmarking of state-of-the-art techniques under consistent evaluation protocols.
- Distillation of key design principles from recent advancements in matching strategies.
- Identification of open challenges and future research directions in the field.
Read more
Cross-View Feature Matching: Survey, Benchmarking, and Foundation-Model Perspectives
Summary
This paper presents a comprehensive survey of cross-view feature matching, a critical problem in computer vision that focuses on establishing reliable correspondences between images captured from different viewpoints. The authors highlight the evolution of the field from task-specific models to more unified and generalizable correspondence models, particularly influenced by the rise of vision foundation models (VFMs). The survey introduces a structured taxonomy that categorizes existing methods based on various dimensions, including feature extraction, matcher types, and training strategies. This framework facilitates a coherent analysis and comparison of diverse approaches. The authors also benchmark state-of-the-art methods under consistent protocols, allowing for fair performance evaluations. Key design principles are distilled from recent advancements, emphasizing the shift towards robust and efficient matching strategies. Furthermore, the paper discusses open challenges in the field, such as improving efficiency, robustness in extreme conditions, and enhancing cross-domain generalization. Overall, this survey serves as a structured reference for understanding the current landscape and future directions of cross-view feature matching.
Methodology
The authors conducted a systematic review of existing literature on cross-view feature matching, categorizing methods into a structured taxonomy. They analyzed recent advancements in various matching techniques, including GNN, Transformer, Mamba, and diffusion-based methods. A unified benchmarking approach was employed to evaluate the performance of representative methods under consistent protocols.
Results
The survey reveals a significant evolution in cross-view feature matching techniques, highlighting the transition from traditional handcrafted methods to modern deep learning frameworks. The benchmarking results provide insights into the strengths and limitations of various approaches, showcasing the impact of VFMs on improving matching accuracy and robustness.
Implications
The findings of this survey have implications for advancing robust perception systems in computer vision, particularly in applications such as structure-from-motion, visual localization, and simultaneous localization and mapping (SLAM). The structured taxonomy and benchmarking can guide future research and development in cross-view feature matching.
ProbGuard: Calibrated Safety Risk Estimation from LLM Output Distributions
Large Language Models
NLP
- ProbGuard is the first architecture-agnostic guardrail that uses probabilistic risk estimation for LLM safety.
- It formulates safety risk based on early output distributions, allowing for better calibration of safety probabilities.
- The methodology employs Monte Carlo sampling to estimate safety risks without accessing hidden states of LLMs.
- ProbGuard shows superior performance in calibration and early intervention compared to existing guardrails.
Read more
ProbGuard: Calibrated Safety Risk Estimation from LLM Output Distributions
Summary
The paper introduces ProbGuard, a novel guardrail for Large Language Models (LLMs) that addresses the limitations of existing safety assessment methods, which typically rely on deterministic classification of outputs. Recognizing that safety assessment is inherently uncertain, especially during the early stages of output generation, ProbGuard leverages the probabilistic information contained in LLM output distributions. By formulating safety risk as the probability of generating unsafe content, the authors utilize Monte Carlo sampling to estimate this risk based on early output distributions. This approach allows for the calibration of safety probabilities and enables early intervention to halt unsafe outputs. The authors demonstrate that ProbGuard significantly improves calibration performance across various model-dataset combinations, achieving a reduction in average Brier score and Expected Calibration Error (ECE) by 79.6% and 71.9%, respectively, compared to the best baseline. Additionally, it effectively limits the success rate of jailbreak attacks to 1% after only ten decoding steps, showcasing its potential for real-time safety monitoring in LLM applications.
Methodology
ProbGuard employs a probabilistic framework to assess safety risks by analyzing the output distributions of LLMs. It formulates the safety risk as the probability of unsafe content generation and utilizes Monte Carlo sampling to estimate this probability from early output distributions. The guardrail encodes these distributions as probability-weighted representations, ensuring architecture-agnostic applicability.
Results
ProbGuard achieves the best calibration performance across nine model-dataset combinations, reducing the average Brier score by 79.6% and ECE by 71.9% compared to the best baseline. It also limits the attack success rate to 1% across six jailbreak attack scenarios after observing only the first ten decoding steps.
Implications
The development of ProbGuard has significant implications for enhancing the safety of LLMs in real-world applications. By enabling early detection and intervention against unsafe outputs, it can improve the reliability of LLMs in sensitive domains such as healthcare, finance, and content moderation. This approach can also inform future research on probabilistic safety assessments in AI systems.
Adaptive Supervised Anchoring for On-Policy Self-Distillation
NLP
Large Language Models
- Identification of rollout-conditioned signal degradation as a critical failure mode in OPSD.
- Introduction of context-valid supervision as a design principle for effective self-distillation.
- Development of Supervised Distillation Steering (SDS) that combines context-separated supervision with adaptive anchoring.
- Demonstration of improved task acquisition and general capability retention across multiple model scales and tasks.
Read more
Adaptive Supervised Anchoring for On-Policy Self-Distillation
Summary
This paper addresses the limitations of on-policy self-distillation (OPSD) in language models, particularly the issue of rollout-conditioned signal degradation, where the quality of trajectories sampled from the student affects the teacher's ability to provide effective supervision. The authors propose a novel framework called Supervised Distillation Steering (SDS), which separates two supervision pathways: one that retains guidance based on the student’s actual rollouts and another that applies supervised cross-entropy on canonical ground-truth contexts. This separation helps avoid the incompatibility of imposing target tokens on erroneous prefixes. The strength of the canonical-context anchor is adaptively adjusted based on token-level alignment between rollouts and targets, enhancing guidance during early training and relaxing it as the model improves. The proposed method is evaluated across various model scales and tasks, demonstrating improved task acquisition and general capability retention compared to traditional OPSD and supervised fine-tuning (SFT).
Methodology
The authors conducted controlled experiments to diagnose the impact of trajectory quality on the effectiveness of OPSD. They proposed the SDS framework, which separates supervision pathways and adaptively adjusts the strength of canonical-context anchoring based on token-level alignment. The framework was evaluated on various model scales and tasks, including domain adaptation and general reasoning benchmarks.
Results
The SDS framework showed significant improvements in task acquisition over traditional OPSD methods while maintaining better general capabilities compared to supervised fine-tuning. The experiments revealed that the adaptive anchoring mechanism effectively enhanced learning dynamics and performance across different model sizes and tasks.
Implications
The findings suggest that improving context quality is essential for effective on-policy self-distillation in language models. The proposed SDS framework could be applied to enhance the training of language models in various applications, particularly in scenarios where task-specific knowledge is critical.
ReOrder-OPD: Reliability-Aware Prompt Ordering for On-Policy Distillation
NLP
Large Language Models
Generative Models
- Introduces prompt-level teacher continuation reliability (R) as a new metric for assessing supervision quality in OPD.
- Demonstrates that ordering prompts by reliability significantly improves OPD performance compared to random or ascending orders.
- Establishes a practical method for estimating reliability using ROUGE-5 scores, enabling effective prompt scheduling.
- Shows consistent performance gains across different model families and tasks, highlighting the versatility of the proposed method.
Read more
ReOrder-OPD: Reliability-Aware Prompt Ordering for On-Policy Distillation
Summary
This paper introduces ReOrder-OPD, a novel approach to on-policy distillation (OPD) that focuses on the reliability of teacher supervision in guiding student-generated trajectories. The authors identify that traditional methods rely on local confidence metrics, which do not adequately assess whether a teacher can successfully continue a student's prefix to a correct answer. To address this, they define a new metric called prompt-level teacher continuation reliability (R), which quantifies the probability of a teacher reaching a correct answer from a given student prefix. The paper demonstrates that prompts with higher reliability lead to greater gains in OPD performance. The authors propose a method to estimate R using the maximum ROUGE-5 F1 score between independent student rollouts and correct teacher trajectories, allowing for effective prompt ordering. ReOrder-OPD sorts prompts based on this reliability proxy and utilizes fresh on-policy rollouts for training. The experimental results show consistent improvements across various model families and tasks, indicating that prompt ordering enhances the effectiveness of within-trajectory supervision. Overall, this work contributes to the understanding of supervision quality in OPD and presents a practical solution for improving model training.
Methodology
The authors define prompt-level teacher continuation reliability (R) as the probability of a teacher reaching a correct answer from a student prefix. They estimate R using the maximum ROUGE-5 F1 score between independent student rollouts and correct teacher trajectories. ReOrder-OPD sorts prompts based on this reliability proxy and employs fresh on-policy rollouts for training.
Results
The experiments reveal that prompts with higher reliability lead to larger OPD gains. ReOrder-OPD consistently outperforms random and ascending prompt orders across various settings, including mathematics and code tasks, demonstrating the effectiveness of prompt ordering in enhancing OPD performance.
Implications
The findings suggest that improving the reliability of teacher supervision can significantly enhance the training of student models in OPD scenarios. This approach may be applicable to various tasks involving language models and could inform future research on effective supervision strategies in machine learning.
How Simple Can It Get? From Interpretable Equations to Readable Rules for Financial Decision Making
Interpretability
- Controlled simplification of interpretable classifiers can yield more readable rules.
- Pruning features can be done with minimal impact on predictive performance.
- Human assessments indicate varying preferences for model representations based on professional background.
- A predictive model for fidelity loss during simplification is introduced.
Read more
How Simple Can It Get? From Interpretable Equations to Readable Rules for Financial Decision Making
Summary
This paper addresses the challenge of creating interpretable machine learning models in finance, where regulatory requirements necessitate explainable decision-making. The authors propose a novel approach that simplifies an interpretable classifier, specifically a monomial equation, into more readable forms such as pruned monomials, directional if-then rules, and scorecards. The study evaluates the trade-offs between model simplicity, predictive performance, and fidelity to the original model. The authors find that simplification can often be achieved with minimal loss in predictive accuracy, while enhancing readability. A human assessment reveals that different professional backgrounds influence preferences for various representations. Additionally, the paper introduces a method to predict the fidelity of simplified models based on the original model's characteristics, providing insights into how simplifications affect performance. Overall, the research contributes to the understanding of how to balance interpretability and usability in financial decision-making models.
Methodology
The authors start with a fitted monomial classifier and systematically simplify it into various forms, including pruned monomials and directional rules. They evaluate each representation's predictive performance, fidelity to the original model, and perceived readability through human assessments. Additionally, they derive a model-based prediction of fidelity loss associated with simplifications.
Results
The study demonstrates that simplifications, such as pruning, can be performed with little cost to predictive performance. The results show that while fidelity may decrease, simpler representations can still function effectively as classifiers. Human assessments reveal that simplification enhances perceived readability, with preferences varying among finance professionals and AI/ML researchers.
Implications
The findings suggest that financial institutions can adopt simpler, more interpretable models without significantly sacrificing performance, potentially improving compliance with regulatory requirements. The approach may also inform the design of machine learning models in other regulated domains where explainability is crucial.
LegoLM: Structured Weight Sharing for Large Language Models
Large Language Models
Efficient ML
NLP
- Identification of two failure modes in weight sharing for LLMs: distributional mismatch and outlier dominance.
- Introduction of three data-free adaptations to resolve these issues: scalar-block encoding, percentile-selective replacement, and boundary-layer protection.
- Demonstrated superior performance in compression and quality preservation compared to existing methods like PTQ-8bit.
- Outlier dominance increases with model scale, highlighting the critical need to preserve outlier weights.
Read more
LegoLM: Structured Weight Sharing for Large Language Models
Summary
LegoLM introduces a structured weight-sharing compression framework specifically designed for large language models (LLMs). The paper identifies two primary failure modes in traditional global weight sharing: distributional mismatch and outlier dominance. Distributional mismatch occurs when transformer layers with varying weight scales lead to a scale-mismatch penalty that increases with the dimensionality of weight blocks. Outlier dominance refers to the issue where a small fraction of weights, which lie beyond the decision threshold, cannot be accurately represented by centroids, resulting in significant quality degradation. LegoLM addresses these issues through three innovative adaptations: scalar-block encoding to mitigate scale mismatch, percentile-selective replacement to preserve outlier weights, and boundary-layer protection for the first and last transformer blocks. The framework demonstrates impressive performance, achieving minimal perplexity degradation while maintaining high compression ratios across different model sizes, outperforming existing methods such as PTQ-8bit. The findings also reveal that the impact of outlier dominance escalates with model size, emphasizing the importance of preserving outlier weights for maintaining model quality. Overall, LegoLM presents a competitive and efficient approach to compressing LLMs without the need for extensive data or retraining.
Methodology
The LegoLM framework employs a structured weight-sharing approach that includes scalar-block encoding to eliminate scale mismatch, percentile-selective replacement to retain outlier weights, and boundary-layer protection for critical transformer layers. The framework is evaluated on models like GPT-2 small and Mistral-7B, using perplexity as a measure of quality degradation.
Results
LegoLM achieves a perplexity degradation of only +0.03% at 4.41× compression on Mistral-7B, outperforming PTQ-8bit in both quality and compression ratio. It also maintains accuracy within noise at 5.12× compression on downstream tasks, exceeding PTQ-8bit’s compression ratio while matching its accuracy.
Implications
The findings suggest that LegoLM can significantly enhance the deployment of large language models by enabling efficient compression without sacrificing performance, making it feasible to run larger models on consumer hardware. This has potential applications in various NLP tasks where model size and efficiency are critical.
A Mechanistic Diagnostic of Rank Collapse in Post-Norm Decoder Transformers
NLP
Large Language Models
Theory
- Post-Norm rank collapse is a two-stage process involving attention amplification and gradient shrinkage.
- Causal attention increases token similarity at initialization, leading to high-similarity states that are difficult to repair.
- Collapsed networks exhibit a high loss floor and vanishing gradients, limiting their optimization potential.
- Experimental results validate the theoretical framework, showing alignment with predicted token similarity growth and gradient behavior.
Read more
A Mechanistic Diagnostic of Rank Collapse in Post-Norm Decoder Transformers
Summary
This paper investigates the phenomenon of rank collapse in Post-Norm decoder-only Transformers, which has led to a preference for Pre-Norm architectures due to the sensitivity of Post-Norm training to initialization and learning rates. The authors present a two-stage analysis of rank collapse, identifying causal attention as a mechanism that amplifies token similarity during initialization, while the RMSNorm backward factor contributes to gradient decay during training. They characterize the properties of collapsed networks, noting that they can only achieve suboptimal solutions with vanishing gradients in collapsed layers. Experimental results corroborate the theoretical predictions, demonstrating the mechanisms behind token similarity growth and gradient contraction. The findings provide insights into the limitations of Post-Norm architectures and suggest that understanding these dynamics is crucial for improving training stability in deep Transformers.
Methodology
The authors employ a two-stage analysis framework to examine the dynamics of token similarity and gradient behavior in Post-Norm Transformers. They utilize controlled initialization measurements and training diagnostics on 48-layer decoder-only Transformers to validate their theoretical predictions regarding token similarity growth and gradient contraction.
Results
The experiments confirm the predicted growth of token similarity during initialization and the contraction of gradients during training in collapsed networks. The results show that collapsed training runs remain near a predicted frequency loss, consistent with the characterized properties of collapsed networks.
Implications
The insights gained from this study could lead to improved training methodologies for deep Transformers, particularly in understanding how architectural choices affect optimization dynamics. This could enhance the performance and stability of large-scale models in NLP and other applications.
Accelerated Learning of High Dimensional Functions with a Tensor-Featured Training Network
Optimization
Efficient ML
Theory
- Introduction of tensor features in DNNs enhances optimization efficiency.
- Two-step optimization process improves convergence of DNN parameters.
- Randomized tensor decomposition significantly reduces storage costs.
- Method effectively trains models in high dimensions (5 to 40).
Read more
Accelerated Learning of High Dimensional Functions with a Tensor-Featured Training Network
Summary
This paper presents a novel method to accelerate the optimization of high-dimensional functions using deep neural networks (DNNs) by introducing contextual features into the first layer of the DNN. The authors propose a two-step optimization process where DNN parameters are optimized using standard gradient descent while keeping the input-feature basis fixed. After optimizing the DNN parameters, the feature layer is updated before resuming DNN optimization. The feature layer consists of two types of functions: rank-1 features that can be evaluated quickly in a matrix-free manner and more complex tensor features that require tensor network (TN) decomposition strategies. The authors leverage a randomized tensor decomposition strategy to efficiently decompose high-dimensional functions, achieving a significant reduction in storage costs by at least 8 orders of magnitude. The proposed method allows for efficient training of models in dimensions ranging from 5 to 40, addressing the challenges of the curse of dimensionality in scientific machine learning.
Methodology
The authors utilize a two-step optimization process for DNNs, where DNN parameters are optimized using gradient descent while keeping the input features fixed. The feature layer is updated after parameter optimization, and tensor decomposition strategies are employed to manage high-dimensional functions efficiently.
Results
The proposed method demonstrates a substantial reduction in storage costs for high-dimensional function representation and allows for efficient training of DNNs across multiple dimensions, showcasing improved convergence and optimization performance.
Implications
This work has significant implications for scientific computing and machine learning, particularly in applications requiring the representation and optimization of high-dimensional functions, such as solving partial differential equations and other complex computational problems.
Share First, Route What Remains: A Unified Framework for Token-Adaptive MoE Computation
Efficient ML
NLP
Large Language Models
- Introduces a unified framework for token-adaptive MoE computation, UniF-MoE.
- Reveals a dependency between reusable and token-specific computation in MoE models.
- Implements a sequential decision-making process for shared and residual expert allocation.
- Demonstrates improved accuracy and efficiency trade-offs over traditional MoE approaches.
Read more
Share First, Route What Remains: A Unified Framework for Token-Adaptive MoE Computation
Summary
This paper presents a unified framework for token-adaptive mixture-of-experts (MoE) computation, termed UniF-MoE. The authors identify a critical dependency between reusable computation and token-specific demands within MoE models, arguing that the processes of shared modeling, fine-grained computation, and dynamic routing should not be treated independently. Instead, they propose a sequential approach: first sharing reusable computation, then routing what remains. The UniF-MoE framework implements this principle by partitioning experts into aligned blocks, where a shared-demand score determines the extent of shared computation and the number of residual experts needed. The framework employs a Gram regularizer to promote diverse routing directions and reduce expert overlap. Experimental results on benchmark datasets such as DomainBed and GLUE demonstrate that UniF-MoE achieves superior predictive performance while also reducing computational costs, inference latency, and memory usage compared to existing MoE models.
Methodology
The authors decompose the computation of MoE models into key-value channels and analyze the interactions between co-activated experts. They develop the UniF-MoE framework, which integrates shared computation and residual expert routing into a single decision-making process. The framework uses a shared-demand score to determine the allocation of computation and employs a Gram regularizer to ensure diverse routing paths.
Results
Experiments on DomainBed and GLUE benchmarks show that UniF-MoE outperforms both static and dynamic MoE models in terms of predictive performance while simultaneously reducing activated computation, inference latency, and memory usage. The findings indicate that the coordinated approach to shared and residual computation leads to significant efficiency gains.
Implications
The proposed framework has potential applications in various domains requiring efficient computation, such as natural language processing and computer vision, where large models need to balance performance with resource constraints. The insights into the dependency between shared and residual computation could inform future designs of MoE architectures.
Generator-Guided Inverse Sampling for Lévy-Driven Generative Models
Generative Models
Theory
Efficient ML
- Introduces a generator-based characterization of Lévy-driven Markov processes for inverse sampling.
- Develops a practical inverse sampler that decomposes reverse dynamics into distinct components.
- Utilizes neural networks for large jump rate management while maintaining interpretability through analytical distributions.
- Demonstrates application in OFDM-SISO channel estimation, showcasing robustness against heavy-tailed noise.
Read more
Generator-Guided Inverse Sampling for Lévy-Driven Generative Models
Summary
This paper explores inverse sampling for Lévy-driven generative models through the lens of Markov generators. Unlike traditional diffusion models, Lévy-driven dynamics incorporate infinite jump activities, complicating the characterization of their reverse processes using score information alone. The authors analyze both forward and reversed generators, revealing that the reversed jump component behaves as a state-dependent Markov jump process governed by a nonlocal density ratio. This insight leads to the development of a structured reverse sampler that separates the dynamics into diffusion, small jump, and large jump components. The proposed sampler is computationally efficient and specifically designed for isotropic linear Lévy stochastic differential equations (SDEs) with symmetric α-stable jump components. The neural network is utilized to manage the rate of large jump activities, while jump amplitudes are derived from analytically determined conditional distributions, enhancing interpretability and controllability. The authors also adapt their sampler for observation-guided sampling, demonstrating its effectiveness in estimating OFDM-SISO channels under mixed Gaussian and impulsive noise. Simulations indicate robust estimation performance, balancing complexity and effectiveness.
Methodology
The authors analyze the forward and reversed generators of Lévy-driven processes, deriving a structured approach to inverse sampling. They decompose the reverse dynamics into diffusion, small jump, and large jump components, employing neural networks for large jump rates and analytical methods for jump amplitudes. Efficient implementation techniques are introduced to optimize performance.
Results
The proposed sampler shows robust performance in simulations, particularly in the context of OFDM-SISO channel estimation under mixed Gaussian and impulsive noise. The results indicate a favorable tradeoff between computational complexity and estimation accuracy.
Implications
The findings suggest that Lévy-driven generative models can effectively handle complex data distributions with heavy tails and impulsive components, making them suitable for various applications in signal processing and beyond, particularly in scenarios involving nonlocal dynamics.
SoftMCC: An MCC-Brier Calibration Bridge for Threshold-Free Model Selection under Class Imbalance
Theory
- SoftMCC provides a threshold-free approach to model selection that reduces dependency on hard predictions.
- The framework achieves better stability and reproducibility in model rankings compared to traditional metrics.
- SoftMCC's core score is a covariance-normalized association that reduces to MCC for hard predictions.
- Empirical results indicate that SoftMCC does not always improve the utility of selected models despite better ranking stability.
Read more
SoftMCC: An MCC-Brier Calibration Bridge for Threshold-Free Model Selection under Class Imbalance
Summary
This paper introduces SoftMCC, a novel framework for model selection in imbalanced binary classification that addresses the limitations of traditional Matthews correlation coefficient (MCC) metrics, which are heavily dependent on thresholding. SoftMCC operates on probability-valued confusion counts and provides a calibrated identity specific to MCC, allowing for a more stable and interpretable model selection process. The framework is designed to maintain a connection to hard MCC while assessing the quality of probability predictions. The empirical evaluation of SoftMCC across 18 different settings demonstrates its superior stability and reproducibility in ranking models compared to established metrics like AUPRC and MCC@0.5. The study also highlights the calibration sensitivity of SoftMCC, revealing that while it performs well in ranking models, it does not necessarily guarantee improved utility of the selected models. The findings suggest that SoftMCC can serve as a valuable tool for practitioners dealing with class imbalance in binary classification tasks.
Methodology
The methodology involves a formal analysis of SoftMCC as a covariance-normalized probability-label association, alongside an empirical evaluation across 18 imbalanced tabular datasets with 12 repeated splits per setting. The study compares SoftMCC against established metrics like MCC@best, F1@best, AUROC, and AUPRC, focusing on ranking reproducibility and selected-model utility.
Results
SoftMCC achieved the best stability mean rank of 2.31 and the highest mean tie-corrected Kendall’s W of 0.659 in the empirical evaluation. The Friedman test indicated significant differences (p = 0.007) between SoftMCC and other metrics. However, selected-model utility showed no significant advantage, with only F1@best surviving Holm correction (p = 0.014). Label permutation and temperature scaling affected rankings, indicating sensitivity to calibration.
Implications
The findings suggest that SoftMCC can be a useful tool for model selection in scenarios with class imbalance, particularly in fields like fraud detection and medical diagnosis. However, practitioners should be cautious about the potential lack of utility improvement when using SoftMCC for model selection.
DEFT: Data-Efficient Frequency-domain Top-k Sampling via Inverse Discrete Fourier Transform for Spatiotemporal Dynamical Systems Modeling
Time Series
Efficient ML
Theory
- DEFT efficiently samples dominant Fourier modes to generate training data for spatiotemporal dynamical systems.
- The method reduces data requirements by 40% while maintaining high predictive accuracy.
- DEFT achieves R² values exceeding 0.99 in battery degradation modeling, demonstrating its robustness.
- The framework includes a theoretical generalization bound and a criterion for optimal mode selection.
Read more
DEFT: Data-Efficient Frequency-domain Top-k Sampling via Inverse Discrete Fourier Transform for Spatiotemporal Dynamical Systems Modeling
Summary
The paper introduces DEFT, a novel frequency-domain data sampling method aimed at efficiently modeling spatiotemporal dynamical systems governed by partial differential equations (PDEs). Traditional approaches either rely on costly physics-based simulations or require extensive training data, which can lead to poor generalization in dynamic conditions. DEFT addresses these challenges by identifying dominant Fourier modes of a system and generating training data through the inverse discrete Fourier transform (IDFT). The authors derive a generalization bound for the method and provide a principled criterion for selecting the number of modes (K). The framework is validated through experiments on canonical PDEs, demonstrating superior performance in scenarios dominated by few frequency components. Additionally, DEFT reduces data requirements by 40% with minimal loss in predictive accuracy and achieves high predictive accuracy (R² > 0.99) in battery degradation modeling, showcasing its applicability across different battery chemistries with limited fine-tuning data. Overall, DEFT presents an effective approach for data-efficient operator learning in complex dynamical systems.
Methodology
DEFT employs a three-tier framework that integrates physics-based modeling, frequency-domain analysis, and operator learning. It utilizes spectral analysis to identify dominant frequencies and synthesizes diverse training waveforms through IDFT, which are then labeled by a physics model. An adaptive sampling algorithm ranks frequency components based on energy contribution and generates a large training set from a small observation set.
Results
The experiments validate DEFT's effectiveness in three domains: canonical PDEs, data-value filtering in diffusion-sorption and Burgers equations, and battery degradation modeling. The method outperformed traditional sampling techniques, achieving significant reductions in data requirements and maintaining high predictive accuracy across various test datasets.
Implications
DEFT has the potential to revolutionize data sampling in scientific machine learning, particularly for applications in engineering and physics where data acquisition is expensive and time-consuming. Its ability to generalize across different systems with minimal fine-tuning could facilitate advancements in battery technology and other spatiotemporal modeling tasks.
FlowScout: From Execution Feedback to Reliable Tool-Using Agent Workflows
Large Language Models
Graph Learning
Optimization
- FLOWSCOUT generates tool-integrated agentic workflows from historical task-solving records.
- The framework utilizes a directed graph representation for workflows, incorporating LLM and tool-calling nodes.
- Monte Carlo tree search is employed to refine workflows based on execution feedback.
- Experimental results show substantial improvements in tool invocation correctness and execution stability.
Read more
FlowScout: From Execution Feedback to Reliable Tool-Using Agent Workflows
Summary
The paper introduces FLOWSCOUT, an innovative framework designed to enhance the reliability and usability of agentic workflows in automation systems that utilize large language models (LLMs). Traditional methods of constructing agentic workflows are often manual and require significant domain expertise, leading to challenges in generating high-quality workflows. Previous approaches primarily focused on LLM-centric workflows, which abstract tool executions and limit their practical applicability. FLOWSCOUT addresses these issues by generating tool-integrated workflows from historical task-solving records. It models workflows as directed graphs comprising LLM nodes and tool-calling nodes, and employs a Monte Carlo tree search guided by execution feedback to refine these workflows. The framework begins by mining common tool coordination patterns from historical data to create an initial workflow, which is then iteratively improved based on execution results. The authors evaluate FLOWSCOUT across four task domains—finance, sports, travel, and weather—demonstrating significant improvements in tool invocation correctness and execution stability compared to existing baselines.
Methodology
FLOWSCOUT represents agentic workflows as directed graphs and formulates their generation as a graph search problem. It first constructs an initial workflow by mining common tool coordination patterns from historical records, then refines this workflow using Monte Carlo tree search based on execution feedback from historical user queries.
Results
The evaluation of FLOWSCOUT revealed that it improved tool invocation correctness by at least 92.69% and execution scores by at least 17.66% compared to baseline methods. Additionally, it reduced performance variation across repeated runs significantly, while incurring a higher runtime cost of at least 24.12% compared to some baselines.
Implications
The development of FLOWSCOUT has significant implications for automating complex tasks using LLMs, as it enhances the reliability and interpretability of workflows. This can lead to more effective automation systems in various domains, including finance, sports, travel, and weather forecasting.
Real Data Closes Synthetic-to-Real Gap in Optical Chemical Structure Recognition
Computer Vision
Multimodal
- Real labeled training images significantly improve OSCR performance, closing the synthetic-to-real gap.
- The effectiveness of vision LoRA adaptation varies depending on the base model used.
- Base model selection is crucial, especially when real data is limited, as performance disparities decrease with more real data.
- Controlled experiments confirm the importance of real data in enhancing model accuracy across different benchmarks.
Read more
Real Data Closes Synthetic-to-Real Gap in Optical Chemical Structure Recognition
Summary
This paper addresses the challenge of Optical Chemical Structure Recognition (OSCR), which is crucial for converting chemical drawings into machine-readable formats. While existing models perform well on synthetic images, their accuracy drops significantly on real-world documents. The authors start with the Qwen2.5-VL-7B model, achieving over 91% accuracy on synthetic images but only below 16% on real benchmarks. To improve performance, they fine-tune 21 recognizers using a combination of synthetic and labeled real images from patents and journals. The study finds that incorporating real training data is the most effective strategy, with accuracy on the ACS benchmark improving from 0.15 to 0.46 as the proportion of real data increases. The paper also explores the impact of different model architectures and adaptation strategies, revealing that the effectiveness of vision LoRA varies by base model. The results indicate that the choice of base model and the mixture of training data are critical for optimizing performance in OSCR tasks.
Methodology
The authors fine-tuned 21 different recognizers on a mixture of synthetic and labeled real images. They varied the Vision Language Model (VLM) base, the fraction of real training data, and the adaptation strategy using LoRA. Controlled experiments were conducted to assess the impact of these factors on model performance across multiple benchmarks.
Results
The study found that adding labeled real training images led to the most significant improvements in accuracy. For example, the Qwen2.5-VL model's exact match on ACS increased from 0.15 without real data to 0.46 with 50.2% real data. The strongest model, GLM-4.1V-9B, achieved an exact match of 0.96 on clean renders and varying scores on real datasets (0.49 on ACS, 0.65 on CLEF-IP, etc.). The performance gap between different base models decreased as the proportion of real data increased.
Implications
This research has significant implications for the field of computational chemistry, particularly in enhancing the usability of chemical literature by improving the accuracy of structure recognition systems. It suggests that incorporating real-world data into training processes is essential for developing robust AI models capable of handling diverse depiction styles in chemical drawings.
Approximation Rates for Metaplectic Neural Networks
Theory
Efficient ML
- Introduction of metaplectic Barron spaces as an extension of classical Barron spaces.
- Establishment of embeddings between metaplectic Barron spaces and Sobolev spaces.
- Development of a neural metaplectic dictionary for improved function approximation.
- Demonstration of superior performance in approximating Schrödinger equation solutions compared to classical architectures.
Read more
Approximation Rates for Metaplectic Neural Networks
Summary
This paper presents quantitative approximation results for shallow neural networks utilizing a metaplectic operator-based dictionary. The authors extend the concept of Barron spaces through the introduction of metaplectic Barron spaces, which are defined using a symplectically motivated extension of the Fourier transform called the metaplectic transform. They establish embeddings between metaplectic Barron spaces and Sobolev spaces, demonstrating that the metaplectic Barron norm of derivatives can be controlled by suitable linear combinations of polynomials of the function. The paper further introduces a neural metaplectic dictionary and proves Monte-Carlo approximation bounds for metaplectic Barron functions using finite linear combinations of dictionary atoms. A deep neural network architecture based on this dictionary is proposed and tested for approximating solutions to time-dependent Schrödinger equations, showing improved performance over traditional physics-informed neural network architectures. This work not only enhances the theoretical understanding of neural network approximation capabilities but also provides a foundation for designing problem-adapted neural network architectures for scientific machine learning applications.
Methodology
The authors utilize a theoretical framework involving metaplectic transforms to define metaplectic Barron spaces. They establish mathematical embeddings to connect these spaces with Sobolev spaces and derive approximation bounds using Monte-Carlo methods. A deep neural network architecture is constructed based on the metaplectic dictionary to validate the theoretical findings through empirical testing.
Results
The paper successfully establishes that metaplectic Barron functions can be approximated with finite linear combinations of dictionary atoms, leading to explicit decay rates for approximation errors. The proposed deep neural network architecture outperforms classical physics-informed neural networks in approximating solutions to time-dependent Schrödinger equations.
Implications
The findings suggest that metaplectic neural networks could be effectively applied in scientific computing, particularly in solving high-dimensional partial differential equations. The theoretical advancements may also guide the development of more efficient neural network architectures tailored for specific applications in machine learning.
Boundary-Seeking Policy Gradient for Safe Reinforcement Learning
Reinforcement Learning
Optimization
Robotics
- BSPG explicitly separates reward improvement and boundary regulation in policy updates.
- The method ensures convergence of the constraint residual to zero while maintaining reward ascent.
- BSPG achieves higher rewards and tighter boundary adherence compared to existing methods.
- The paper provides a theoretical foundation for the boundary structure in constrained policy optimization.
Read more
Boundary-Seeking Policy Gradient for Safe Reinforcement Learning
Summary
This paper introduces the Boundary-Seeking Policy Gradient (BSPG) method for safe reinforcement learning, which aims to maximize rewards while adhering to safety constraints defined in Constrained Markov Decision Processes (CMDPs). The authors highlight that traditional gradient-based methods often do not exploit the structural property that optimal policies lie on the constraint boundary when constraints are active. BSPG addresses this by separating the update into two components: a tangential component that improves rewards while maintaining cost, and a normal component that adjusts the policy towards the active boundary based on the constraint residual. The authors prove that under ideal conditions, the constraint residual converges to zero, and the policy satisfies the Karush-Kuhn-Tucker (KKT) conditions at convergence. The effectiveness of BSPG is demonstrated through experiments on a Safety-Gymnasium navigation task, where it outperforms existing baselines in terms of reward and boundary adherence.
Methodology
The BSPG method combines a tangential update that preserves cost while maximizing reward with a normal update that adjusts the policy towards the constraint boundary based on the signed residual. The authors derive an algebraic Lagrangian representation for the update and prove convergence properties under exact gradients and regularity conditions.
Results
In experiments conducted on a standard Safety-Gymnasium navigation task, BSPG demonstrated superior performance by achieving higher rewards and maintaining closer adherence to the safety boundary compared to baseline methods. The theoretical results confirm that the constraint residual converges to zero, and the policy satisfies the KKT conditions at convergence.
Implications
The proposed BSPG method has significant implications for safe reinforcement learning applications in safety-critical domains such as autonomous driving and robotics, where adherence to safety constraints is essential while maximizing performance.
Online Learning of Scale Parameters in Score-Driven Filters
Time Series
Optimization
Theory
- The paper presents a novel approach to online learning of scale parameters in score-driven filters.
- Gain selection is framed as a conditional predictive decision problem with a Kullback-Leibler objective.
- Dynamic-regret bounds for mirror updates are established under convexity and regularity conditions.
- Simulations and empirical results indicate that bounded mirror gains outperform constant gains in volatile markets.
Read more
Online Learning of Scale Parameters in Score-Driven Filters
Summary
This paper investigates the online learning of scale parameters, referred to as gains, in score-driven filters, which are models that utilize scaled log-likelihood scores to update parameters based on observations. The authors treat the gain as a decision variable and explore its impact on predictive density and state transitions. They establish that each admissible gain leads to a reachable next state and a one-step-ahead predictive density, framing gain selection as a conditional predictive decision problem with a Kullback-Leibler objective. The paper introduces a stochastic gradient derived from the negative raw product of consecutive scores for scalar unscaled gains and discusses the implications of positive scaling. The authors also present dynamic-regret bounds for projected and discounted mirror updates under specific conditions. Through simulations, they demonstrate the effects of scaling, link geometry, and persistence on performance. An empirical analysis of equity-index volatilities shows that the bounded mirror gain approach generally outperforms constant gain methods, particularly in volatile market conditions.
Methodology
The authors utilize a theoretical framework that combines stochastic gradient methods with mirror descent geometries to optimize gain selection in score-driven filters. They derive dynamic-regret bounds and conduct simulations to analyze the effects of different gain strategies on predictive performance.
Results
The results indicate that the bounded mirror gain approach generally matches or exceeds the performance of constant gain methods, particularly in multi-crisis market scenarios. The simulations illustrate the influence of scaling and link geometry on the effectiveness of the learning process.
Implications
The findings suggest that incorporating online learning of scale parameters can enhance the performance of score-driven models in financial applications, particularly in dynamic and volatile environments. This approach may lead to more robust modeling of time-varying parameters in various economic and financial contexts.
Generalized Convexity and Smoothness via Conjugate Duality: Optimization Theory for Deep Neural Networks
Optimization
Theory
- Introduction of H(ψ)-convexity and H(Ψ)-smoothness to unify optimization frameworks for DNNs.
- Generalized GD and SGD are proposed with a proven optimal learning rate of 1.
- DNN training is reformulated as a composite optimization problem focusing on gradient energy and Jacobian norm.
- The framework provides insights into the impact of architectural designs and training configurations on convergence.
Read more
Generalized Convexity and Smoothness via Conjugate Duality: Optimization Theory for Deep Neural Networks
Summary
This paper addresses the gap between the empirical success of deep neural networks (DNNs) trained with stochastic gradient descent (SGD) and the limitations of classical optimization theory, which often relies on assumptions of convexity and smoothness that are not applicable to DNN objectives. The author introduces a unified optimization framework by generalizing classical concepts of convexity and smoothness through Legendre functions and convex conjugation, leading to the definitions of H(ψ)-convexity and H(Ψ)-smoothness. These concepts unify both convex and non-convex, as well as smooth and non-smooth functions, revealing a duality between smoothness and convexity. The paper proposes generalized gradient descent (GD) and generalized SGD, proving that the optimal learning rate for generalized GD is exactly 1, thus simplifying the learning rate tuning process. Additionally, the author reformulates DNN training as a composite optimization problem, demonstrating that convergence is achieved by reducing gradient energy and controlling the induced norm of the network Jacobian. The paper also introduces the gradient correlation factor and model capacity risk to analyze the effects of network architecture and training configurations on convergence. Extensive experiments validate the theoretical findings, showing alignment between theoretical predictions and empirical training dynamics.
Methodology
The paper employs a theoretical approach to generalize classical optimization concepts using Legendre functions and convex conjugation. It defines new forms of convexity and smoothness, derives generalized optimization algorithms, and reformulates DNN training as a composite optimization problem. The analysis includes mathematical proofs of convergence rates and empirical validation through experiments across various architectures and datasets.
Results
The proposed generalized GD and SGD demonstrate convergence properties under the new definitions of convexity and smoothness. The optimal learning rate for generalized GD is shown to be 1, and the framework successfully captures the dynamics of DNN training, aligning theoretical predictions with empirical results across different configurations.
Implications
This work provides a new theoretical foundation for understanding DNN training dynamics, which could lead to more effective optimization strategies and improved training methodologies. The insights into architectural influences and training configurations may inform future designs of neural networks.
Hallucinations and Constraints : Regulating surgical workflow recognition beyond accuracy
Robotics
Theory
Time Series
- Hallucinations in AI can be defined as topological errors, which are critical in medical imaging.
- Linear temporal logic predicates can effectively regulate these hallucinations in surgical workflow recognition.
- The proposed methodology improves accuracy in surgical phase recognition by about 10% while minimizing errors.
- The study highlights the importance of mathematical guarantees in the regulation of AI in medical applications.
Read more
Hallucinations and Constraints : Regulating surgical workflow recognition beyond accuracy
Summary
This paper addresses the issue of hallucinations in artificial intelligence, particularly in the context of medical imaging and surgical workflow recognition. The authors argue that hallucinations, defined as topological errors in predictions, can be systematically measured and regulated. They propose that linear temporal logic predicates can be used to define and enforce constraints on these errors, particularly in the application of automatic surgical phase recognition during robot-assisted hysterectomy. By applying probabilistic graphical models to enforce these constraints, the authors demonstrate an improvement in accuracy by approximately 10% while significantly reducing topological errors. This approach suggests that mathematical guarantees of correctness can complement empirical methods in regulating machine learning applications in medical contexts, paving the way for safer AI integration in surgical procedures.
Methodology
The authors utilize linear temporal logic predicates to define topological errors in surgical workflow recognition. They apply probabilistic graphical models to enforce these predicates in the context of robot-assisted hysterectomy, allowing for the systematic regulation of hallucinations in AI predictions.
Results
The implementation of constrained AI using linear temporal logic predicates resulted in a 10% increase in accuracy for automatic surgical phase recognition, while also eliminating a significant number of topological errors, demonstrating the effectiveness of this regulatory approach.
Implications
The findings suggest that integrating mathematical constraints into AI systems can enhance their reliability and safety in medical applications, particularly in surgical settings. This approach could lead to more robust AI systems that are better suited for real-world medical challenges.
ProTAGAD: A Foundation Model for TAG Anomaly Detection with Decoupled Topological and Textual Prototypes
Graph Learning
- ProTAGAD addresses the Blurred-Anomaly-Boundary (BAB) issue in TAG anomaly detection.
- The model utilizes decoupled topological and textual prototypes to isolate anomaly evidence.
- Extensive experiments show ProTAGAD achieves state-of-the-art performance on 14 benchmark datasets.
- The decoupled design effectively preserves complementary information from both modalities.
Read more
ProTAGAD: A Foundation Model for TAG Anomaly Detection with Decoupled Topological and Textual Prototypes
Summary
The paper introduces ProTAGAD, a novel foundation model designed for anomaly detection in Text-Attributed Graphs (TAGs). Unlike traditional Graph Anomaly Detection (GAD) methods that primarily focus on structural irregularities, ProTAGAD addresses the need to jointly consider both topological patterns and textual semantics to effectively identify nuanced anomalies. The authors identify a significant challenge in existing GAD approaches, termed the Blurred-Anomaly-Boundary (BAB) issue, which arises from the deep cross-modality coupling in holistic message-passing schemes. This coupling obscures subtle anomaly signals, making it difficult to distinguish between normal and anomalous nodes. To overcome this, ProTAGAD employs a decoupled approach, utilizing separate prototype banks for topological normality and textual anomaly detection. This design allows for independent modeling of each modality, thus preserving the integrity of anomaly cues. The model was evaluated across 14 diverse benchmark datasets, demonstrating state-of-the-art performance in zero-shot cross-domain settings. The results indicate that ProTAGAD not only mitigates the BAB issue but also enhances the model's generalization capabilities across different domains.
Methodology
ProTAGAD employs a prototype-based framework that decouples the learning of textual and topological representations. It constructs dual prototype banks to independently model structural normality and semantic consistency, combining their anomaly scores only at the decision level to avoid cross-modal interference.
Results
ProTAGAD achieved the best AUROC on seven out of eight unseen target graphs, with an average rank of 1.12 across 14 TAG datasets. The results confirm the effectiveness of the decoupled design in alleviating the BAB issue and enhancing cross-domain generalization.
Implications
The findings suggest that decoupled representation learning can significantly improve anomaly detection in complex graph structures, with potential applications in fraud detection, social network moderation, and cybersecurity.
Do Judges Behave Like Algorithms?
Interpretability
- Judges often behave algorithmically, with decisions resembling interpretable formulas.
- Inconsistencies in judicial decisions can lead to unequal treatment of defendants.
- Machine learning models can identify key variables influencing judicial decision-making.
- The study emphasizes the importance of understanding when rules versus standards are applied in judicial contexts.
Read more
Do Judges Behave Like Algorithms?
Summary
This paper investigates whether judges in the judicial system behave like algorithms by following predictable, formula-like rules in their decision-making processes. The authors explore this question in the context of misdemeanor bail hearings in Harris County, Texas, where they analyze judicial decisions to determine if they are consistent and based on identifiable variables such as criminal history, age, and charge type. The study employs machine learning models to assess the decision-making patterns of magistrate judges, measuring variable importance to identify key factors influencing their rulings. The findings indicate that while judges often exhibit algorithmic behavior, there are notable inconsistencies in their decisions, leading to unequal treatment of similar defendants. The authors argue that understanding these discrepancies can enhance the justice system by highlighting areas where individualized standards may be more appropriate than rigid rules.
Methodology
The authors analyzed over 22,000 unique bail hearing cases decided by 21 magistrate judges in Harris County, Texas. They trained machine learning models for each judge to assess decision-making patterns and variable importance metrics to identify key factors influencing their rulings.
Results
The analysis revealed that judges generally exhibit algorithmic behavior, with their decisions being captured by small, interpretable formulas. However, significant inconsistencies were found among judges, leading to unequal treatment of similar defendants in certain cases.
Implications
The findings suggest that understanding judicial decision-making can inform improvements in the justice system, particularly in identifying when individualized standards should be applied instead of rigid rules. This could lead to more equitable outcomes in judicial processes.
Beyond the Capability Boundary: Zeroth-Order Optimization for Self-Evolving LLM Agents
NLP
Large Language Models
Optimization
- Introduces a zeroth-order self-evolution framework for LLM agents to learn beyond capability boundaries.
- Utilizes perturbation of LoRA parameters to adapt to difficult examples without trajectory annotations.
- Implements a parallel perturbation inference mechanism to reduce time consumption in optimization.
- Demonstrates improved performance on deep research benchmarks, especially on challenging tasks.
Read more
Beyond the Capability Boundary: Zeroth-Order Optimization for Self-Evolving LLM Agents
Summary
This paper addresses the limitations of self-evolving large language model (LLM) agents, which struggle to learn beyond their inherent capability boundaries due to their inability to sample correct trajectories on difficult examples. The authors propose a novel zeroth-order self-evolution framework that allows LLM agents to adapt to challenging instances by perturbing the parameters of low-rank adaptation (LoRA) modules without requiring trajectory annotations. The method involves perturbing the LoRA parameters, running the agent, and computing loss differences to estimate gradients for updating the parameters. This process forms a closed self-evolution loop, where successful trajectories generated by the optimized agent are used for supervised fine-tuning. To enhance efficiency, the authors introduce a parallel perturbation inference mechanism and an adaptive lookup mechanism, along with an answer perplexity loss that stabilizes optimization. Experimental results demonstrate that this approach significantly improves the performance of LLM agents on various deep research benchmarks, particularly on difficult examples, showcasing its effectiveness in breaking through capability boundaries.
Methodology
The proposed methodology involves perturbing LoRA parameters of LLMs to compute loss differences, which are then used to estimate gradients for parameter updates. A parallel perturbation inference mechanism is employed to efficiently evaluate multiple perturbations, while an adaptive lookup mechanism reduces repeated tool calls. The answer perplexity loss is introduced as a smoother feedback signal for optimization.
Results
The experiments show that the proposed method consistently outperforms strong baselines, achieving better performance on deep research benchmarks, particularly on difficult examples where traditional methods fail. The approach successfully discovers more effective trajectories, validating its capability to push LLM agents beyond their inherent limitations.
Implications
This research has significant implications for the development of more capable LLM agents that can autonomously improve their performance on complex tasks without the need for extensive annotated data. It opens avenues for further exploration in self-evolving systems and optimization techniques in machine learning.
Beyond Routing: Decoupling Expert Dispatch and Aggregation in Sparse Mixture-of-Experts
NLP
Large Language Models
Efficient ML
- Decoupling expert dispatch from aggregation can lead to improved performance in Sparse MoE models.
- The highest-scored expert selected by the router is often not the optimal choice for aggregation.
- FDAA effectively learns token-adaptive aggregation weights without rerouting experts or additional computations.
- Experiments show significant performance gains across multiple datasets, indicating the robustness of the proposed method.
Read more
Beyond Routing: Decoupling Expert Dispatch and Aggregation in Sparse Mixture-of-Experts
Summary
This paper investigates the roles of dispatch and aggregation in Sparse Mixture-of-Experts (MoE) models, which traditionally couple these functions using the same scores for both selecting experts and weighting their outputs. The author proposes a decoupling approach, focusing on the aggregation of outputs from a fixed set of selected experts while keeping their identities and computations constant. The study reveals that the highest-scored expert selected by the router is often not the best choice for aggregation, indicating a mismatch between selection and commitment. To address this, the author introduces Fixed-Dispatch Adaptive Aggregation (FDAA), a lightweight post-compute head that optimizes aggregation weights directly under the language-modeling objective, while keeping the router and experts frozen. The results demonstrate significant improvements in performance across various datasets, highlighting the importance of separating expert selection from aggregation in MoE architectures.
Methodology
The study employs a causal decomposition approach to analyze the roles of dispatch and aggregation in MoE models. It maintains fixed expert selections and computations while varying aggregation weights. The FDAA method is introduced as a trainable aggregation head optimized under the language-modeling objective, with experiments conducted on pretrained models like OLMoE-1B-7B and DeepSeek-V2-Lite.
Results
The introduction of FDAA resulted in a statistically significant improvement in full-horizon cross-entropy by 0.0160 ± 0.0039 on OLMoE-1B-7B and improved performance on WikiText-103, C4, and Penn Treebank datasets. The router's top-scored expert was identified as the best choice only 17.2% of the time, emphasizing the selection-commitment mismatch. The FDAA method outperformed static calibration methods and demonstrated strong transferability across different architectures.
Implications
The findings suggest that separating expert selection from aggregation can enhance the efficiency and effectiveness of MoE models, potentially leading to better performance in various NLP tasks. This decoupling approach may influence future designs of MoE architectures and their applications in large-scale language models.