AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
66
Papers today
8h
Update frequency
7
Days of history
From Tokens to Watt-hours: Analytical Energy Estimation for LLM Inference on Modern GPUs
Large Language Models
Efficient ML
- Introduces a novel methodology for estimating LLM inference energy on modern GPUs.
- Separates energy estimates into components for detailed analysis of energy consumption.
- Provides a framework for comparing models and optimizing for energy efficiency.
- Focuses on transparency and reproducibility in energy estimation.
Read more
From Tokens to Watt-hours: Analytical Energy Estimation for LLM Inference on Modern GPUs
Summary
This paper addresses the growing concern of energy consumption in large language model (LLM) inference, particularly in the context of environmental sustainability. The authors propose a novel analytical methodology for estimating the energy required for LLM inference on NVIDIA H100-class GPUs without the need for direct runtime measurements. The methodology combines transformer FLOP accounting, calibrated memory-traffic factors, and hardware-specific energy coefficients for FP16/BF16 tensor-core computations. It distinguishes between prompt prefill and autoregressive decoding, enabling detailed energy estimates for input tokens, output tokens, and complete inference requests. The energy consumption is further decomposed into various components, including compute, parameter-access, key-value cache write, and attention read, allowing for a comprehensive analysis of how model size, context length, and generated-token count affect energy usage. The proposed estimator is designed to be transparent and reproducible, providing a useful tool for model comparison, green-coding analysis, and design-time evaluations of LLM workloads, without replacing physical power measurements.
Methodology
The authors developed a semi-analytical estimator that combines parameter-scaled transformer FLOP modeling, calibrated memory-traffic estimation, and hardware-specific energy coefficients. This estimator is designed for scenarios where direct power measurement is impractical, providing assumption-explicit energy estimates for GPU-side operations.
Results
The methodology allows for detailed energy estimates that can be normalized per token and broken down into specific components. It highlights how different factors, such as model size and context length, influence energy consumption, and suggests actionable strategies for reducing energy usage in LLM inference.
Implications
The findings have significant implications for AI practitioners and researchers, particularly in the context of sustainability. The methodology can guide decisions on model selection and deployment strategies, promoting energy-efficient practices in AI development and deployment.
EvoCause: LLM-Guided Evolution of Causal Graphs for Root Cause Analysis
Graph Learning
Large Language Models
Interpretability
- EvoCause integrates LLM-guided graph editing with deterministic validation to refine causal graphs for RCA.
- The framework allows for the incorporation of expert feedback without modifying the underlying causal discovery algorithms.
- TeleRCA, a large-scale expert-annotated benchmark, is introduced to evaluate RCA methods.
- EvoCause outperforms traditional methods in both synthetic and real-world scenarios, emphasizing the role of alarm title information.
Read more
EvoCause: LLM-Guided Evolution of Causal Graphs for Root Cause Analysis
Summary
EvoCause addresses the challenge of root cause analysis (RCA) in modern telecommunication and cloud systems, where correlated alarm cascades occur during component failures. Traditional methods rely on causal graphs derived from observational data, but these graphs are static and do not incorporate expert feedback from historical incidents. EvoCause introduces a novel framework that utilizes large language models (LLMs) to propose semantically plausible edits to causal graphs based on expert constraints, while deterministic code ensures the structural integrity of the graphs. The paper also presents TeleRCA, a benchmark dataset containing expert-annotated RCA data from a real-world telecommunication network. Experimental results demonstrate that EvoCause significantly improves RCA performance and graph reconstruction metrics compared to baseline methods, highlighting the importance of expert feedback and alarm title information in refining causal graphs.
Methodology
EvoCause employs a two-step process: first, it uses an LLM to generate potential edits to the causal graph based on expert constraints; second, it applies deterministic procedures to validate these edits, ensuring the graph remains acyclic and structurally valid. The best-performing graph is selected based on alignment with expert annotations.
Results
EvoCause shows significant improvements over unrefined causal discovery methods, with increases in Node F1, Case EM, and Graph F1 by 11.59, 9.40, and 4.59 percentage points, respectively, while reducing nSHD by 0.2379. The analysis of the TeleRCA dataset reveals that anonymizing alarm titles leads to a decrease in RCA performance metrics, underscoring the value of human-readable information.
Implications
The findings suggest that integrating expert knowledge into causal graph refinement can enhance the accuracy of RCA in complex systems. The release of TeleRCA provides a valuable resource for future research in event-sequence-based RCA methods, potentially leading to more robust diagnostic tools in telecommunication and cloud services.
When Does Explicit View Routing Work? A Controlled Study of Multi-View Graph-Text Alignment
Graph Learning
Multimodal
NLP
- Introduces a controlled framework to differentiate between semantic routing and architectural channelization in graph-text retrieval.
- Demonstrates that correct routing improves retrieval performance for labels and properties, but not consistently for topology.
- Shows that a joint model can outperform separately trained specialists in aggregate retrieval scores.
- Evaluates the effects of paraphrase augmentation on retrieval quality, revealing trade-offs between robustness and canonical retrieval.
Read more
When Does Explicit View Routing Work? A Controlled Study of Multi-View Graph-Text Alignment
Summary
This paper investigates the effectiveness of explicit view routing in multi-view graph-text alignment, particularly focusing on the conditions under which semantic routing can be distinguished from architectural channelization. The authors present a controlled version of the Multi-View Graph-Text Alignment (MV-GTA) model that utilizes deterministic text segments, isolated text encoders, and view-specific graph heads. By employing a causal test comparing correct routing against deranged samples, the study reveals that correct routing significantly enhances label and property retrieval performance on molecular datasets BBBP and BACE. The findings indicate that while routing is effective for label and property retrieval, topology specialization is inconsistent across datasets. The authors also explore the impact of paraphrase augmentation on retrieval quality, noting that while it improves unseen-template performance, it can also lead to a decline in canonical retrieval under certain conditions. Overall, the study provides insights into the diagnostic capabilities of routing in graph-text retrieval systems and highlights the complexities involved in achieving consistent semantic specialization across multiple views.
Methodology
The authors implemented a controlled version of MV-GTA using deterministic text segments and isolated encoders for each view. They conducted experiments comparing correct routing with deranged samples, where segments were misaligned with their intended channels. Relevance was defined using external labels and RDKit descriptors, allowing for a causal analysis of retrieval performance.
Results
The study found that correct routing improved label and property nDCG scores by 0.305 to 0.685 compared to deranged training. The expected graph head outperformed the best wrong head by 0.303 to 0.453. In a matched three-seed comparison, a joint model achieved higher mean retrieval scores than three separately trained specialists. Paraphrase augmentation improved unseen-template nDCG by 0.140 to 0.147, although consistency training sometimes reduced canonical retrieval.
Implications
The findings suggest that explicit routing can enhance retrieval performance in graph-text systems, particularly for specific semantic aspects. This has potential applications in fields such as cheminformatics and knowledge graph retrieval, where accurate alignment of structured and unstructured data is crucial. The study also highlights the importance of controlled experiments in understanding model behavior and improving multi-view learning strategies.
Beyond KV Reconstruction: Functional Reconstruction for MLA Draft Models in Speculative Decoding
Large Language Models
Efficient ML
Optimization
- Conversion from MHA/GQA to MLA can lead to a loss of proposal agreement necessary for effective speculative decoding.
- The proposed functional reconstruction method optimizes draft quality without requiring verifier supervision or changes to the inference graph.
- The method is converter-agnostic and can be applied to various model architectures and configurations.
- Functional reconstruction improves acceptance rates in 37 out of 64 task configurations evaluated.
Read more
Beyond KV Reconstruction: Functional Reconstruction for MLA Draft Models in Speculative Decoding
Summary
This paper addresses the challenges associated with converting multi-head attention (MHA) and grouped-query attention (GQA) models to multi-head latent attention (MLA) for efficient long-context language model inference. While MLA reduces memory traffic during decoding by using compact latent states instead of a key-value (KV) cache, the conversion process can degrade the draft quality necessary for speculative decoding. The authors propose a novel functional reconstruction method that optimizes the MLA attention modules to better align with the outputs of the original MHA/GQA models, thereby improving draft acceptance rates during speculative decoding. This method operates as a converter-agnostic, training-time-only post-conversion stage, requiring no additional supervision or changes to the inference structure. The evaluation of various model configurations demonstrates that the proposed approach significantly enhances acceptance rates in many cases, while maintaining or improving throughput.
Methodology
The authors introduce a functional reconstruction method that optimizes MLA attention modules to reproduce the outputs of original MHA/GQA blocks on calibration hidden states. This end-to-end reconstruction is performed during training and does not require any changes to the converted cache structure or inference graph. The evaluation involves testing multiple model configurations across different tasks to assess the impact of the proposed method on draft acceptance and throughput.
Results
The functional reconstruction method resulted in material acceptance improvements in 37 out of 64 matched task configurations, with 26 remaining unchanged and one showing a material decrease. Additionally, 12 of the improved configurations also exhibited increased throughput, indicating that the method effectively enhances both draft acceptance and generation speed.
Implications
The findings suggest that optimizing draft quality during the conversion process can significantly enhance the performance of speculative decoding in language models. This has potential applications in improving the efficiency and effectiveness of large language models in real-time applications, where rapid and accurate token generation is critical.
RAGuard: A Layered Defense Framework for Retrieval-Augmented Generation Systems Against Data Poisoning
NLP
Large Language Models
- RAGuard introduces a two-layer defense framework against data poisoning in RAG systems.
- The first layer uses adversarial fine-tuning to improve the retriever's ability to downrank malicious passages.
- ZKIP operates without requiring poison labels or ground-truth answers, making it adaptable to unseen attack types.
- The combined defense effectively reduces attack success rates to 0.000 while maintaining retrieval quality.
Read more
RAGuard: A Layered Defense Framework for Retrieval-Augmented Generation Systems Against Data Poisoning
Summary
The paper introduces RAGuard, a novel defense framework designed to protect Retrieval-Augmented Generation (RAG) systems from data poisoning attacks. RAG systems, which combine large language models (LLMs) with external corpora, are vulnerable to corpus poisoning where malicious passages can mislead the model's outputs. RAGuard employs a two-layer defense mechanism: the first layer involves adversarial fine-tuning of a dense retriever on synthetic poisoned documents to downrank harmful passages. The second layer, known as the Zero-Knowledge Inference Patch (ZKIP), is a label-free, black-box filter that evaluates the impact of each retrieved document on the model's output by analyzing semantic shifts and output entropy changes when a document is removed. The evaluation of RAGuard on poisoned datasets, including Natural Questions and BEIR, demonstrates that while adversarial retriever training reduces attack success rates, ZKIP effectively eliminates them, achieving an attack success rate of 0.000 across various configurations. The authors also discuss the computational costs associated with ZKIP and propose methods to mitigate these costs, while identifying the limitations of existing defenses against keyword-preserving poisons.
Methodology
RAGuard employs a two-layer defense strategy: the first layer involves adversarial training of a dense retriever on synthetic poisoned documents, while the second layer, ZKIP, assesses the impact of each retrieved document on the model's output by performing counterfactual evaluations without requiring external labels or ground-truth answers.
Results
The evaluation of RAGuard on poisoned Natural Questions and BEIR datasets shows that adversarial retriever training alone reduces attack success rates but does not eliminate them. In contrast, ZKIP successfully drives the attack success rate to 0.000 across all configurations, while keeping Recall@5 within 0.03 of the clean-corpus baseline.
Implications
RAGuard's layered defense framework has significant implications for enhancing the robustness of RAG systems against data poisoning attacks, potentially improving the reliability of LLM outputs in real-world applications. The findings also suggest the need for ongoing research into defense mechanisms that can adapt to evolving attack strategies.
DAS-PMVC: A Framework for Partial Multi-View Clustering via Dual Alignment and Structure Enhancement
Graph Learning
Multimodal
Optimization
- DAS-PMVC addresses the partial view alignment problem in multi-view clustering.
- The framework employs a dual alignment strategy to enhance feature alignment and clustering accuracy.
- Utilizes multi-view graph convolutional networks for improved feature extraction.
- Experimental results show superior clustering performance compared to existing methods.
Read more
DAS-PMVC: A Framework for Partial Multi-View Clustering via Dual Alignment and Structure Enhancement
Summary
The paper introduces DAS-PMVC, a novel framework designed to address the challenges of partial multi-view clustering, particularly the partial view alignment problem (PVAP) caused by misalignment in multi-view data. The proposed framework consists of three main components: (1) anchor graph structure alignment, which establishes initial view alignment through joint embedding representations derived from anchor point relationships; (2) structure-enhanced feature learning, where the model utilizes multi-view graph convolutional networks to extract deep latent features and improve representation discriminative power; and (3) a dual alignment strategy that optimizes the alignment of latent features using contrastive learning loss and the Hungarian algorithm during training. The experimental results demonstrate that DAS-PMVC significantly outperforms existing state-of-the-art methods in clustering performance across various datasets, highlighting its effectiveness in handling misaligned multi-view data.
Methodology
DAS-PMVC integrates three core methodologies: (1) anchor graph structure alignment for initial view alignment, (2) structure-enhanced feature learning through multi-view graph convolutional networks, and (3) a dual alignment strategy that combines contrastive learning and the Hungarian algorithm to refine latent feature alignment during training.
Results
The experimental evaluations indicate that DAS-PMVC consistently outperforms current state-of-the-art clustering methods, demonstrating its robustness and effectiveness in managing partially aligned multi-view data.
Implications
The proposed framework has significant implications for applications in fields where multi-view data is prevalent, such as computer vision, social network analysis, and bioinformatics, providing a more effective approach to clustering in the presence of misalignment.
NMINE: Normalized Mutual Information Neural Estimation
Theory
- Introduction of a fully neural estimator for normalized mutual information.
- Combines MINE-based mutual information estimation with neural entropy estimation.
- Derives entropy estimates from neural divergence relative to uniform distributions.
- Demonstrates improved accuracy over traditional k-nearest-neighbor-based estimators.
Read more
NMINE: Normalized Mutual Information Neural Estimation
Summary
This paper introduces NMINE, a novel neural estimator for normalized mutual information (NMI) tailored for continuous multidimensional variables. Traditional methods for estimating mutual information face challenges in high-dimensional settings, particularly regarding numerical stability and accuracy. NMINE addresses these issues by integrating a MINE-based mutual information estimator with neural marginal entropy estimators inspired by MI-NEE. The proposed method utilizes the Donsker–Varadhan variational representation for mutual information estimation and learns the divergence between marginal distributions and a uniform reference distribution to derive entropy estimates. This approach replaces the conventional k-nearest-neighbor methods, which are prone to errors in high dimensions. The authors validate NMINE through experiments on Gaussian data across one to eight dimensions, demonstrating that it significantly outperforms a KSG-based NMI baseline in terms of accuracy. The findings suggest that neural estimation techniques can enhance the measurement of normalized dependencies in complex continuous settings, paving the way for broader applications in fields such as molecular dynamics and interpretable machine learning.
Methodology
The NMINE framework employs a neural network to estimate mutual information using the Donsker–Varadhan variational representation. It incorporates neural divergence estimation for marginal entropies, allowing for a more stable and accurate computation of normalized mutual information compared to existing nearest-neighbor methods.
Results
Experiments reveal that NMINE achieves higher accuracy in estimating normalized mutual information than the KSG-based baseline across various dimensions, indicating its effectiveness in handling continuous multidimensional data.
Implications
The NMINE estimator has potential applications in fields requiring accurate dependency measurement, such as molecular dynamics, interpretable machine learning, and representation learning, particularly in high-dimensional and complex data scenarios.
Real-Time Hard Peak Age-of-Information Safety with No-Regret Learning
Optimization
Theory
Reinforcement Learning
- Introduces OCO-PAoI-Hard, a novel scheduling framework for peak AoI safety in IoT systems.
- Achieves zero per-slot deadline violations under adversarial conditions while maintaining O(√T) regret.
- Utilizes a causal proposal-shield-update loop for real-time scheduling and feasibility enforcement.
- Provides comprehensive theoretical guarantees and empirical validation against multiple baselines.
Read more
Real-Time Hard Peak Age-of-Information Safety with No-Regret Learning
Summary
This paper addresses the critical challenge of ensuring that the peak Age of Information (AoI) in safety-critical IoT systems remains below a hard deadline in real-time scenarios. Traditional approaches have struggled to meet this requirement under adversarial conditions, often relying on restrictive assumptions or failing to guarantee per-slot deadline adherence. The authors propose a novel framework called OCO-PAoI-Hard, which transforms the scheduling problem into a time-varying constrained online convex optimization problem. This transformation allows for the enforcement of strict deadlines while maintaining no-regret learning behavior. The proposed algorithm utilizes a strictly causal proposal-shield-update loop that ensures feasibility through Euclidean projection, while also achieving O(√T) regret against any static comparator. The theoretical analysis includes closed-form regret bounds and a minimax lower bound, demonstrating the robustness of the approach. Empirical results show that OCO-PAoI-Hard achieves zero modeled-state deadline violations across multiple trials, outperforming several baseline methods significantly in terms of deadline adherence and regret.
Methodology
The authors develop the OCO-PAoI-Hard framework by recognizing that the peak AoI deadlines can be expressed as affine half-space constraints on resource allocation. They implement a strictly causal proposal-shield-update loop that projects proposed actions onto a feasible set, ensuring that every action adheres to the modeled deadlines. The algorithm is designed to operate under adversarial conditions, allowing for robust online learning.
Results
The OCO-PAoI-Hard framework demonstrated zero modeled-state deadline violations across all trials in a four-sensor adversarial fluid-model trap channel. In contrast, baseline methods exhibited significant per-slot miss rates, ranging from 1.65% to 64.0%. The empirical normalized regret remained below the theoretical envelope across a wide range of time horizons, confirming the effectiveness of the proposed method.
Implications
The findings have significant implications for the design of safety-critical IoT systems, such as those used in industrial control, autonomous vehicles, and remote operations, where maintaining timely information is crucial for system stability and safety. The proposed framework can enhance the reliability of real-time communication in these applications.
From Expert Reduction to Behavioral Divergence: Tracing Numerical State through Sparse MoE Inference
Large Language Models
Theory
Efficient ML
- Different expert-reduction orders can produce observable differences in sparse-MoE executions.
- The study identifies the significance of operand representation and accumulator precision in model compatibility.
- Two state boundaries are validated for reproducing controlled divergent trajectories in autoregressive decoding.
- Identical emitted tokens may not imply identical execution states, leading to potential output divergence.
Read more
From Expert Reduction to Behavioral Divergence: Tracing Numerical State through Sparse MoE Inference
Summary
This paper investigates the impact of expert-reduction orders on the execution of sparse mixture of experts (MoE) models, specifically focusing on the DeepSeek-V4-Flash framework. The author demonstrates that mathematically equivalent reduction orders can lead to significantly different execution states due to floating-point precision issues. By freezing the local MoE state and varying aggregation semantics, the study identifies how different operand representations and accumulator precisions affect the model's behavior. The research addresses three key questions: the relationship between reduction semantics and behavioral basins, the effects of operand representation and precision on compatibility, and the sufficiency of internal state boundaries for reproducing divergent trajectories. The findings reveal that identical emitted tokens do not guarantee identical execution states, highlighting the importance of understanding internal state boundaries in sparse-MoE models. The results suggest that backend implementations must ensure compatibility not only at the output level but also at various internal state boundaries to maintain consistent model behavior.
Methodology
The research employs a trace-freeze-fork intervention to analyze the effects of different expert aggregation semantics on the post-mHC state space. It utilizes a series of ablation studies to separate operand representation from accumulator precision, and it conducts exhaustive enumeration of cross-expert reduction orders to explore behavioral structures. The experiments are conducted using the native DeepSeek inference path, allowing for detailed control and instrumentation of expert aggregation and state transitions.
Results
The study finds that 720 A-mode orders yield 10 continuation basins, while 720 B-mode orders form 360 exact structural classes and 11 basins. Under specific prompts, B classes split into various continuation categories. The research confirms that exact reconstruction of post-mHC endpoints can reproduce downstream trajectories, and that full persistent state reconstruction can maintain continuity across token boundaries. These results establish the importance of internal state boundaries in reproducing model behavior.
Implications
The findings suggest that backend implementations of sparse-MoE models must account for the effects of reduction orders and floating-point precision to ensure consistent behavior across different hardware. This has implications for the design and qualification of machine learning systems, particularly in applications requiring high reliability and reproducibility.
LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger
Multimodal
- Introduces a Structured Evidence Ledger for multimodal agentic reasoning.
- Identifies four recurring failure patterns that affect trajectory faithfulness.
- Implements a Three-Layer Grounding Protocol to verify evidence integrity.
- Enhances reasoning efficiency with an Adaptive Dual-Path Dispatcher.
Read more
LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger
Summary
The paper introduces LedgerMind, a framework designed to enhance the fidelity of multimodal agents in visual question answering by addressing the limitations of traditional evaluation methods that focus solely on final-answer accuracy. LedgerMind treats the agent's trajectory as a provenance-constrained state machine, utilizing a Structured Evidence Ledger to normalize tool outputs and ensure that reasoning and decision claims are grounded in verifiable evidence. This structured approach allows for the identification and mitigation of four common failure patterns: unsupported intermediate reasoning, Phantom Grounding (where claims cite valid evidence but introduce unsupported entities), over-reasoning on simple queries, and repair-time amplification (where attempts to correct errors introduce new unsupported claims). The framework incorporates a Three-Layer Grounding Protocol, an Adaptive Dual-Path Dispatcher, and an Event-Triggered Verification-and-Repair engine to ensure that all claims are backed by tool-produced provenance. Experiments demonstrate that LedgerMind significantly improves both answer accuracy and trajectory-level faithfulness across multiple multimodal reasoning benchmarks.
Methodology
The authors developed LedgerMind as a structured framework that normalizes tool outputs into a Structured Evidence Ledger. This ledger contains detailed information about each piece of evidence, including its source, type, confidence, and dependencies. The framework employs a Three-Layer Grounding Protocol to ensure that claims are grounded in actual evidence, an Adaptive Dual-Path Dispatcher to optimize reasoning depth based on query complexity, and an Event-Triggered Verification-and-Repair engine to manage corrections without introducing unsupported claims.
Results
Experiments conducted on various multimodal reasoning benchmarks indicate that LedgerMind outperforms existing models in terms of both answer accuracy and trajectory-level faithfulness, effectively addressing the identified failure patterns and demonstrating the benefits of a provenance-constrained approach.
Implications
The findings suggest that adopting a structured evidence-based approach can significantly enhance the reliability and transparency of multimodal reasoning systems. This has potential applications in areas requiring high-stakes decision-making, such as healthcare, legal reasoning, and automated customer support, where understanding the basis of conclusions is critical.
THGFM: Dual-Branch Temporal Heterogeneous Graph Fusion Model
Graph Learning
- THGFM integrates cross-type transfer and relation-aware specialization in a unified framework.
- The model employs a dual-path architecture with SSTA and RTTA branches.
- Type-Conditioned Non-Competitive Gated Sum Fusion allows independent feature weighting.
- Rotary Temporal Attention directly incorporates relative time into attention mechanisms.
Read more
THGFM: Dual-Branch Temporal Heterogeneous Graph Fusion Model
Summary
The paper introduces THGFM, a novel model designed for learning on temporal heterogeneous graphs, which are characterized by diverse node and relation types that evolve over time. Traditional methods struggle to balance cross-type transfer and relation-aware specialization, often treating temporal information as an additive feature rather than integrating it into the core attention mechanism. THGFM addresses these challenges through a dual-path architecture comprising a Shared-Space Temporal Attention (SSTA) branch for efficient cross-type transfer and a Relational Type-Partitioned Temporal Attention (RTTA) branch for specialized relation modeling. The model employs a Type-Conditioned Non-Competitive Gated Sum Fusion (TC-NGSF) mechanism, allowing independent amplification or suppression of features from both branches based on the target node type. Additionally, THGFM incorporates Rotary Temporal Attention to directly integrate relative time into the attention scores, enhancing the model's ability to prioritize relevant neighbors based on temporal proximity. The model was evaluated on various academic graph benchmarks, demonstrating significant performance improvements over existing graph transformer models.
Methodology
THGFM utilizes a dual-branch architecture with a Shared-Space Temporal Attention branch for cross-type transfer and a Relational Type-Partitioned Temporal Attention branch for specialization. It employs Type-Conditioned Non-Competitive Gated Sum Fusion to adaptively combine features from both branches and incorporates Rotary Temporal Attention to enhance the influence of temporal dynamics on the attention scores.
Results
THGFM consistently outperformed baseline graph transformer models, achieving a +3.25% mean gain across six tasks. Notable peak relative gains included +12.37% on OAG-CS PV, +4.87% on PF-L2, and +1.18% on PF-L1, along with +4.24%, +3.73%, and +4.61% on OGBN-MAG, HTAG-ArXiv, and HTAG-DBLP, respectively.
Implications
The advancements presented by THGFM could enhance various applications involving dynamic relational systems, such as academic citation networks, e-commerce platforms, and scientific knowledge graphs, by providing more accurate and context-aware representations of complex relationships over time.
From Unsupervised Subgroups to Hypothetical State-Intervention Policies: An Evaluation of Selected Subgrouping Methods in Observational Health Data
Theory
Interpretability
Optimization
- Proposes a framework for subgroup analysis that avoids reliance on outcome data.
- Evaluates various unsupervised clustering methods for policy prioritization in health interventions.
- Findings suggest no single subgrouping method consistently dominates across different policy scenarios.
- Results should be interpreted as assumption-dependent evidence rather than definitive proof of intervention effectiveness.
Read more
From Unsupervised Subgroups to Hypothetical State-Intervention Policies: An Evaluation of Selected Subgrouping Methods in Observational Health Data
Summary
This paper addresses the challenges of subgroup analysis in observational health data, where conventional methods often yield unstable conclusions due to issues like multiplicity and small sample sizes. The authors propose a novel framework that constructs subgroups based solely on pre-treatment patient characteristics, without relying on outcome or treatment-effect information. This approach aims to create interpretable units for policy prioritization in healthcare interventions. The framework integrates ensemble causal discovery, inductive unsupervised clustering, and uncertainty-aware subgroup selection, followed by a held-out doubly robust policy evaluation. Various clustering methods, including K-means, Fuzzy C-means, and Bayesian Gaussian mixture models, are compared against a supervised causal-forest-derived comparator. The framework is applied to evaluate hypothetical interventions for obesity and glucose levels in diabetes risk, as well as smoking history related to sleep disturbance. The results indicate that while certain subgrouping methods yield favorable point estimates for policy utility, no single method consistently outperforms others across different scenarios. The findings emphasize the importance of interpreting results as assumption-dependent evidence rather than definitive proof of intervention benefits, highlighting the need for further validation in longitudinal studies.
Methodology
The authors developed an end-to-end framework combining ensemble causal discovery, inductive unsupervised clustering, and uncertainty-aware subgroup selection. They compared multiple clustering techniques, including K-means, Fuzzy C-means, and Bayesian Gaussian mixture models, against a supervised causal-forest-derived comparator. The evaluation involved hypothetical state-shift policies for health interventions, using a fixed budget for subgroup prioritization.
Results
The highest estimated ungated utilities were 0.799 for the BMI policy using Bayesian GMM, 0.735 for the glucose policy using hard or membership-weighted FCM, and 0.775 for the smoking-history policy using K-means. However, paired bootstrap comparisons showed no statistically significant differences among the evaluated subgroup policies after Holm adjustment.
Implications
The findings suggest that phenotype-first subgroups can provide transparent units for observational policy prioritization in healthcare. However, the observational nature of the data and the uncertainty in causal graphs necessitate cautious interpretation of the results. Future research should focus on validating these policies in external cohorts and exploring clinically implementable interventions.
Existence-Field Diffusion Model for Spatial Point Processes with Variable Cardinality
Generative Models
- Introduces a novel EFDM that models spatial point processes with variable cardinality.
- Overcomes limitations of existing methods by integrating cardinality and spatial structure in a unified diffusion process.
- Utilizes existence variables to represent the presence of points, allowing for continuous modeling.
- Demonstrates improved performance on synthetic and real-world datasets compared to previous approaches.
Read more
Existence-Field Diffusion Model for Spatial Point Processes with Variable Cardinality
Summary
This paper introduces the Existence-Field Diffusion Model (EFDM) for generative modeling of spatial point processes (SPP) characterized by variable cardinality. Traditional methods struggle to jointly model the number of points and their spatial configurations, often decoupling these aspects or relying on discrete operations that limit flexibility. The EFDM addresses these challenges by associating each potential point with an existence variable that indicates its degree of presence, allowing for a unified diffusion process that captures both spatial locations and cardinality without explicit transitions. The authors demonstrate that EFDM effectively models the joint distribution of cardinality and spatial structure, outperforming existing methods on datasets with varying cardinality, thus providing a more flexible framework for generative modeling in real-world applications.
Methodology
The EFDM employs a continuous diffusion process where both the positions of points and their existence variables are diffused to noise in the forward process and denoised in the reverse process. This allows for probabilistic sampling of existence at the end of generation, enabling a flexible representation of cardinality and spatial configurations without discrete transitions.
Results
The EFDM was evaluated on both synthetic and real-world datasets, showing superior modeling capabilities in capturing the dependency between cardinality and spatial structure. The results indicate that EFDM achieves stronger performance in conditional evaluations across varying cardinality levels compared to existing methods.
Implications
The EFDM has potential applications in various fields where spatial point processes are relevant, such as urban planning, molecular modeling, and environmental monitoring. Its ability to flexibly model variable cardinality and spatial configurations can enhance the understanding and generation of complex spatial phenomena.
Contrastive Concept Importance: Explaining Pairwise Class Decisions Through Automatically Extracted Concept Representations
Interpretability
Computer Vision
- Introduces Contrastive Concept Importance (CCI) for explaining pairwise class decisions.
- Provides signed attributions indicating support for target vs. foil classes.
- Evaluates method on ImageNet class pairs, revealing class-pair specific behaviors.
- Distinguishes between globally important concepts and those affecting specific class distinctions.
Read more
Contrastive Concept Importance: Explaining Pairwise Class Decisions Through Automatically Extracted Concept Representations
Summary
This paper introduces a novel method called Contrastive Concept Importance (CCI) to enhance the interpretability of machine learning models, particularly in the context of pairwise class decisions. Traditional feature attribution methods typically answer the question 'why P?' for a single class, but CCI addresses the more informative question 'why P rather than Q?' by attributing the logit margin between a target class (P) and a foil class (Q) to automatically extracted visual concepts. The method provides signed scores that indicate whether a concept supports the target class over the foil or vice versa, allowing for a nuanced understanding of concept contributions. The authors evaluate CCI on ImageNet class pairs using various techniques, including insertion and deletion curves and logit-wise decomposition analysis. The results demonstrate that CCI reveals class-pair specific behaviors that ordinary concept importance fails to capture, highlighting the importance of contrastive explanations in understanding model decisions, especially in cases of misclassification or class confusion.
Methodology
The methodology involves attributing the logit margin between a target class and a foil class to concepts from an automatically extracted visual concept basis. The resulting scores are signed and can be decomposed into target-logit and foil-logit effects, allowing for detailed analysis of concept contributions.
Results
The evaluation shows that CCI effectively captures class-pair specific model behavior, revealing insights that ordinary concept importance methods overlook. The method demonstrates the ability to assess whether concepts influence fine-grained distinctions or broader category evidence.
Implications
The findings suggest that CCI can significantly improve the interpretability of machine learning models, particularly in applications where understanding the rationale behind class distinctions is crucial, such as in medical imaging or autonomous systems.
From Interface to Inference: Eliciting Any-Order Inference from Any-Order Models
NLP
Large Language Models
Generative Models
- Identifies the limitations of autoregressive models in non-causal reasoning tasks.
- Proposes two novel approaches to achieve genuine any-order inference: insertion-based and latent-space masked diffusion.
- Demonstrates that positional uncertainty is a fundamental bottleneck in achieving any-order inference.
- Empirical results show significant improvements in performance for code generation and mathematical reasoning tasks.
Read more
From Interface to Inference: Eliciting Any-Order Inference from Any-Order Models
Summary
This paper addresses the limitations of autoregressive language models in performing any-order inference, which is crucial for tasks like code generation that require flexibility in structuring information. The authors identify a gap between the any-order prediction interface provided by masked diffusion models (MDMs) and the actual inference capabilities, attributing this to 'positional uncertainty.' They propose two innovative approaches to overcome this issue: (1) Insertion-based masked diffusion, which allows for token insertions and reduces fixed-position commitments, enabling generation across non-contiguous regions; and (2) Latent-space masked diffusion, which shifts the focus from token-level predictions to coarser semantic segments in latent space. Both methods are empirically validated, showing distinct any-order inference behaviors and improved performance in downstream tasks such as Python coding and mathematical problem-solving.
Methodology
The authors develop two complementary methods: Insertion-based masked diffusion, which allows for flexible token insertions, and Latent-space masked diffusion, which operates in a continuous latent space rather than at the token level. These methods are empirically tested on specific tasks, including Python code generation and mathematical problem-solving, using models fine-tuned for these purposes.
Results
The FlexMDM model, a 7B parameter model for Python coding, outperforms existing models on various benchmarks, while the 125M LatentMDM model shows superior performance compared to standard MDMs and autoregressive models under the same inference budget. Both approaches successfully demonstrate distinct any-order inference behaviors.
Implications
The findings suggest that masked diffusion models can be effectively adapted for tasks requiring non-linear reasoning and flexibility, such as code generation and complex problem-solving. This could lead to advancements in generative modeling techniques and applications in various fields requiring discrete reasoning.
S-CEReBrO: Breaking the Memory Barrier in Continuous EEG Monitoring
Time Series
Efficient ML
- S-CEReBrO introduces a Windowed Alternating Attention mechanism that maintains constant memory usage.
- The architecture can process EEG signals 100 times longer than traditional full self-attention models.
- It achieves state-of-the-art performance on multiple EEG analysis tasks while using fewer parameters.
- The model demonstrates a significant increase in inference throughput and reduced memory requirements.
Read more
S-CEReBrO: Breaking the Memory Barrier in Continuous EEG Monitoring
Summary
The paper introduces S-CEReBrO (Streaming CEReBrO), a novel architecture designed to enhance continuous EEG monitoring by overcoming memory limitations associated with Transformer-based models. Traditional global attention mechanisms in Transformers lead to memory overflow during prolonged signal processing, which is critical for applications like seizure detection. S-CEReBrO employs a Windowed Alternating Attention mechanism that divides attention computation into fixed-size spatiotemporal windows, ensuring constant memory usage regardless of signal duration. This innovation allows S-CEReBrO to process signals 100 times longer than full self-attention models and 3 times longer than low-rank linear attention models while requiring 55% less memory and achieving 2.1 times higher inference throughput. The model is pre-trained on over 25,000 hours of EEG recordings from more than 12,000 subjects and demonstrates state-of-the-art performance on 7 out of 11 downstream tasks, using up to 60% fewer parameters than existing models. This work significantly advances the efficiency and generalizability of EEG monitoring systems.
Methodology
The S-CEReBrO architecture utilizes a Windowed Alternating Attention mechanism that interleaves spatial and temporal attention blocks, allowing for local interactions and maintaining a constant memory footprint. The model processes EEG signals by tokenizing them into patches and employing masked autoencoding for representation learning. This approach ensures linear computational complexity and addresses the challenges posed by long-duration EEG monitoring.
Results
S-CEReBrO outperforms global attention baselines on 7 out of 11 public benchmarks related to EEG analysis, achieving state-of-the-art results while requiring 55% less memory than low-rank linear attention models and 60% fewer parameters than existing architectures. The model also demonstrates a 2.1 times increase in inference throughput.
Implications
The advancements presented in S-CEReBrO could lead to more efficient and effective continuous EEG monitoring systems, facilitating real-time applications such as seizure detection and brain-computer interfaces. The model's ability to handle long-duration signals with reduced memory requirements makes it suitable for deployment in resource-constrained environments.
Benchmarking ConvLSTM for One-Day-Ahead IMDAA Rainfall-Field Prediction across Four Indian Cities
Time Series
- ConvLSTM does not consistently outperform simpler forecasting models for rainfall prediction.
- FC-LSTM achieved the lowest domain-mean rainfall error in three out of four cities.
- Persistence model outperformed all others in high-rainfall day detection across all cities.
- The effectiveness of models varies significantly based on input types and evaluation metrics.
Read more
Benchmarking ConvLSTM for One-Day-Ahead IMDAA Rainfall-Field Prediction across Four Indian Cities
Summary
This study investigates the effectiveness of Convolutional Long Short-Term Memory networks (ConvLSTMs) for one-day-ahead rainfall-field prediction using Indian Monsoon Data Assimilation and Analysis (IMDAA) fields from four major cities: Bengaluru, Delhi, Kolkata, and Mumbai, spanning June to September from 1998 to 2020. The research compares ten different forecasting approaches, including naïve, statistical, tree-based, and neural models, utilizing both atmospheric-only and rainfall-history-plus-atmospheric inputs. The performance of these models is evaluated based on complete fields, domain-mean rainfall, spatial anomalies, and high-rainfall days. The findings reveal that ConvLSTM does not consistently outperform simpler models, with FC-LSTM achieving the lowest domain-mean rainfall error in three cities, while persistence was most effective in Delhi. ConvLSTM showed the best performance in spatial-anomaly error only in Mumbai, where rainfall fields exhibited greater short-term spatial continuity. Overall, neural models tended to underestimate rainfall magnitude and struggled with high-rainfall day predictions, while persistence consistently excelled in detection performance across all cities. The study concludes that the choice of architecture should be guided by comprehensive benchmarking across various performance metrics.
Methodology
The study employed a comparative analysis of ten forecasting models, including ConvLSTM, FC-LSTM, and persistence, using IMDAA rainfall data. Performance was assessed using metrics such as root-mean-square error (RMSE) for complete fields, domain-mean rainfall, and spatial anomalies, as well as detection rates for high-rainfall days. Two input formulations were tested: atmospheric-only and rainfall-history-plus-atmospheric.
Results
The results indicated that ConvLSTM did not consistently outperform simpler models, with FC-LSTM yielding the lowest domain-mean rainfall error in Bengaluru, Kolkata, and Mumbai, while persistence was best in Delhi. ConvLSTM only excelled in spatial-anomaly error in Mumbai. All neural models underestimated rainfall magnitude and had low detection rates for high-rainfall days, whereas persistence achieved the highest detection performance across all cities.
Implications
The findings suggest that while ConvLSTM has potential, simpler models may be more effective for short-term rainfall prediction in certain contexts. This research highlights the importance of model benchmarking and the need for tailored approaches in meteorological forecasting, particularly in regions with variable rainfall patterns.
Contrastive Reinforced Policy Optimization via Privileged Self-Distillation
Reinforcement Learning
Large Language Models
Optimization
- CRPO enhances OPSD by applying contrastive learning principles to mitigate exposure bias.
- The method uses predictive entropy to classify positions for effective distillation.
- CRPO shows significant improvements in training stability and generalization across diverse benchmarks.
- The theoretical foundation of CRPO aligns with logit-wise credit assignment, ensuring robust performance.
Read more
Contrastive Reinforced Policy Optimization via Privileged Self-Distillation
Summary
This paper introduces Contrastive Reinforced Policy Optimization (CRPO), a novel approach to On-Policy Self-Distillation (OPSD) that addresses the challenges of exposure bias in multi-turn agentic settings. Traditional OPSD methods suffer from unreliable position-level supervision due to the privileged information of the self-teacher, leading to reasoning route convergence and unclear optimization directions. CRPO reformulates the OPSD framework using contrastive learning principles, distinguishing between positive and negative positions based on predictive entropy. By conducting group-wise contrast, CRPO enhances training stability and generalization in long-horizon interactions. The authors validate their approach through extensive evaluations across 13 challenging benchmarks, demonstrating that CRPO consistently outperforms existing reinforcement learning and self-distillation methods, thereby providing a more reliable optimization signal for agentic tasks.
Methodology
CRPO reformulates the OPSD framework by treating the original context and the context with privileged information as two views. It employs predictive entropy to classify positions into positive and negative pairs, conducting group-wise contrast to preserve clear optimization directions. The method integrates contrastive learning principles to enhance the reliability of supervision signals during training.
Results
The evaluations across 13 benchmarks indicate that CRPO consistently outperforms traditional reinforcement learning and self-distillation methods, showing enhanced training stability and better generalization in long-horizon interactions. The results highlight the effectiveness of the contrastive approach in addressing the limitations of existing OPSD techniques.
Implications
The findings suggest that CRPO could be applied to improve the performance of large language models and other agentic systems that require reliable optimization signals in multi-turn interactions. This approach may lead to advancements in various applications, including dialogue systems, decision-making agents, and complex reasoning tasks.
AgentGFM: A Graph Foundation Model with Node-Agent Information-Flow Control
Graph Learning
- AgentGFM allows for node-specific information propagation decisions, enhancing adaptability across different graph structures.
- The model employs a shared end-to-end trainable policy for all nodes, promoting transferability of learned behaviors.
- The predict-act-observe-correct process enables nodes to refine their decisions based on feedback, improving information flow control.
- Extensive experiments show that AgentGFM outperforms traditional GFMs in various transfer scenarios.
Read more
AgentGFM: A Graph Foundation Model with Node-Agent Information-Flow Control
Summary
The paper introduces AgentGFM, a novel Graph Foundation Model (GFM) that addresses the limitations of existing models in transferring information across diverse graph structures. Traditional GFMs often rely on fixed propagation schemes that do not adapt well to the varying local structural patterns found in different graphs and nodes. This paper proposes a new approach where each node in the graph is treated as an autonomous agent capable of making decisions about information propagation based on its current state and feedback from the graph. The AgentGFM employs a predict-act-observe-correct process, allowing nodes to make informed decisions regarding source reception, signal-channel selection, and gain-aware halting. This method enhances the adaptability of information flow control, enabling nodes to follow distinct trajectories while still adhering to a shared, end-to-end trainable policy. The authors validate the effectiveness of AgentGFM through extensive experiments across various node-level, graph-level, and large-scale transfer scenarios, demonstrating its superior performance in handling diverse graph topologies.
Methodology
AgentGFM is built on an agent-based decision-making framework where each node acts as an agent. The model uses a predict-act-observe-correct cycle to facilitate adaptive information flow control. During the act phase, nodes make decisions about how to receive information and propagate it through the graph, which are then corrected based on the feedback received from the graph's responses.
Results
The experiments conducted demonstrate that AgentGFM significantly improves performance in node-level, graph-level, and large-scale transfer scenarios compared to existing GFMs. The model's ability to adapt to diverse topological structures is validated, showcasing its effectiveness in real-world applications.
Implications
AgentGFM has potential applications in various domains that utilize graph-structured data, such as social network analysis, recommendation systems, and knowledge graph reasoning. Its adaptive information flow control can enhance the performance of graph-based learning tasks across different contexts.
Compliance2LoRA: On-Demand Safety Alignment on Arbitrary Policy Subsets via Hypernetwork-Generated LoRA Adapters
Large Language Models
NLP
Reinforcement Learning
- Compliance2LoRA provides a unified framework for dynamic safety policy compliance in large reasoning models.
- The hypernetwork generates LoRA weights based on customizable policy inputs, enabling efficient model adaptation.
- The framework allows for on-demand adjustments to policy compliance without sacrificing task performance.
- Training involves both supervised and reinforcement learning techniques to enhance model generalization.
Read more
Compliance2LoRA: On-Demand Safety Alignment on Arbitrary Policy Subsets via Hypernetwork-Generated LoRA Adapters
Summary
The paper presents Compliance2LoRA, a novel framework designed to address the challenges of safety compliance in large reasoning models (LRMs) by enabling on-demand policy alignment. As LRMs become increasingly personalized for diverse user needs, the requirement for adherence to various safety policies grows. Traditional methods of training separate models for each policy subset are impractical due to combinatorial overhead and data scarcity for rare policies. Compliance2LoRA leverages a hypernetwork architecture to generate LoRA (Low-Rank Adaptation) weights dynamically based on user-specified policy subsets. This approach allows for efficient policy compliance without the need to maintain multiple model versions. The framework employs a combination of supervised fine-tuning and reinforcement learning to optimize the hypernetwork, ensuring that the resulting model maintains high performance across different reasoning tasks and datasets. The authors demonstrate the effectiveness of Compliance2LoRA in producing compliant responses while preserving the reasoning capabilities of the LRM, showcasing its potential for practical applications in safety-critical environments.
Methodology
The methodology involves training a hypernetwork that generates LoRA weights conditioned on multiple safety policy embeddings. The process includes converting policy descriptions into embeddings, applying self-attention mechanisms for aggregation, and generating LoRA weights through a multi-layer perceptron. The model is trained using both supervised fine-tuning and reinforcement learning techniques to optimize policy compliance.
Results
The results indicate that Compliance2LoRA achieves comparable safety performance to traditional methods while significantly reducing computational overhead. The framework demonstrates the ability to generalize across unseen policy combinations and maintains task performance across different reasoning models and datasets.
Implications
The implications of this work are significant for deploying large reasoning models in safety-sensitive applications, such as healthcare, finance, and content moderation, where compliance with varying safety policies is crucial. Compliance2LoRA offers a scalable solution for adapting models to diverse user requirements without the need for extensive retraining or storage of multiple model versions.
Cybersecurity Detection Classification with Reasoning-enabled Language Models
NLP
Large Language Models
Reinforcement Learning
- First detection-triage classifier trained to produce CoT reasoning for cybersecurity detections.
- Introduced a four-stage training process that enhances classifier performance and confidence estimation.
- Demonstrated that an untrained confidence calibrator leads to a collapse in high-confidence recall.
- A specialized 30B model outperforms general-purpose models in cybersecurity detection tasks.
Read more
Cybersecurity Detection Classification with Reasoning-enabled Language Models
Summary
This paper addresses the challenge of alert fatigue in Security Operations Centers (SOCs) caused by the overwhelming number of detections that analysts must triage. The authors propose a novel approach that utilizes chain-of-thought (CoT) reasoning in large language models (LLMs) to improve the classification of cybersecurity detections as either malicious or benign. Unlike previous methods that only trained LLMs to produce triage labels, this work emphasizes reasoning about the detections before labeling. The authors developed a four-stage training process that includes prompt optimization, self-training, reinforcement learning with verifiable rewards, and confidence calibration. This methodology not only enhances the accuracy of the triage classifier but also addresses the issue of confidence estimation, which is crucial for managing false positives and negatives. The results show that their system achieves an overall accuracy of 82.6%, with significant improvements in benign and malicious recall compared to traditional direct-label classifiers. The findings suggest that specialized training of LLMs for cybersecurity tasks can lead to better performance than general-purpose models, highlighting the importance of targeted training in this domain.
Methodology
The authors employed a four-stage training process: prompt optimization to enhance input quality, self-training to leverage unlabeled data, reinforcement learning with verifiable rewards to refine decision-making, and confidence calibration to accurately estimate the correctness of the model's outputs.
Results
The proposed system achieved 82.6% test accuracy, with a 43.0% improvement in benign recall and an 18.3% improvement in malicious recall at high-confidence operating points compared to a direct-label LLM classifier. The necessity of the confidence calibrator was confirmed, as an untrained model resulted in zero high-confidence recall.
Implications
This research has significant implications for improving the efficiency of cybersecurity operations by reducing alert fatigue through automated triage systems that leverage reasoning capabilities of LLMs. It suggests that specialized training can enhance the performance of AI in cybersecurity applications, potentially leading to more effective threat detection and response.
Search Strategies for Optimal Classification and Regression Trees
Optimization
Interpretability
Efficient ML
- Introduces a unified algorithmic framework for Optimal Decision Trees (ODTs) that allows for the comparison of various search strategies.
- Empirically evaluates 18 search strategies, marking one of the largest evaluations in the field.
- Demonstrates that the best search strategy significantly improves anytime performance for classification and runtime for regression.
- Clarifies the trade-offs between different search strategies, enhancing understanding of their strengths and weaknesses.
Read more
Search Strategies for Optimal Classification and Regression Trees
Summary
This paper addresses the challenges of scalability in learning Optimal Decision Trees (ODTs), which are compact and interpretable models that globally optimize objectives such as accuracy and constraints. The authors introduce a general algorithmic framework that allows for the instantiation and comparison of various search strategies for ODTs. By empirically evaluating 18 different search strategies, the study provides insights into their performance, revealing that the best strategy significantly enhances anytime performance for classification tasks and improves runtime for regression tasks by over an order of magnitude. The framework not only clarifies the strengths and weaknesses of individual strategies but also facilitates a unified understanding of the differences between existing ODT algorithms, which have previously been studied in isolation. This work contributes to the field by providing a comprehensive evaluation of search strategies and establishing a foundation for future research in ODT optimization.
Methodology
The authors developed a general algorithmic framework for ODTs that incorporates and generalizes existing search strategies. They conducted empirical evaluations of 18 search strategies, comparing their performance in terms of runtime and solution quality for both classification and regression tasks.
Results
The evaluation revealed that the best-performing search strategy outperformed state-of-the-art methods in terms of anytime performance for classification and achieved a runtime improvement of over an order of magnitude for regression tasks.
Implications
The findings suggest that the proposed framework can guide future research in ODT optimization, enabling the development of more efficient algorithms that can be applied in high-stakes domains such as healthcare, where interpretability and accuracy are crucial.
Kairos: Numerically Robust News Recommendation under Item Cold-Start via Cholesky-based LinUCB
Optimization
Efficient ML
Theory
- Addresses item cold-start challenges in news recommendation systems.
- Introduces a Cholesky-based approach for numerical stability in LinUCB.
- Integrates Matryoshka Representation Learning for efficient inference.
- Demonstrates significant efficiency gains in empirical evaluations.
Read more
Kairos: Numerically Robust News Recommendation under Item Cold-Start via Cholesky-based LinUCB
Summary
The paper presents Project Kairos, a novel framework for news recommendation that addresses the challenges of item cold-start in dynamic news environments. Traditional recommendation systems often rely on extensive interaction data, which is scarce in the fast-paced news domain where articles have a short Time-to-Live (TTL). Kairos employs a contextual online learning approach using the Linear Upper Confidence Bound (LinUCB) algorithm, which estimates recommendation payoffs while managing uncertainty. A key innovation is the replacement of the error-prone Sherman-Morrison matrix inversion with direct rank-1 updates of Cholesky factors, ensuring numerical stability and preserving the positive definiteness of the covariance matrix. Additionally, the integration of Matryoshka Representation Learning (MRL) allows for efficient candidate generation with minimal precision loss. Empirical evaluations using the Tagesschau API demonstrate that Kairos achieves a 4.85-fold efficiency gain without significantly compromising ranking precision, making it a promising solution for high-performance recommendation systems in resource-constrained environments.
Methodology
The methodology involves using the LinUCB algorithm for contextual bandit learning, which estimates user preferences based on contextual embeddings. The paper replaces traditional matrix inversion techniques with Cholesky factor updates to maintain numerical stability, and employs Matryoshka Representation Learning to optimize feature representation and inference efficiency.
Results
Empirical evaluations show that Kairos achieves a 4.85-fold increase in efficiency while maintaining ranking precision, indicating its effectiveness in overcoming the limitations of traditional recommendation systems in dynamic news environments.
Implications
The findings suggest that Kairos can significantly enhance news recommendation systems, particularly in regions with limited data availability. Its approach can be adapted for other domains facing similar cold-start issues, potentially improving user engagement and satisfaction.
TAPO: Transition-Aware Policy Optimization for LLM Agents
Reinforcement Learning
Large Language Models
Optimization
- TAPO leverages action-conditioned environmental feedback to enhance policy optimization in LLM agents.
- The framework alternates between standard policy optimization and transition supervision, improving sensitivity to environmental dynamics.
- TAPO requires no additional expert data or sampling costs, making it a lightweight enhancement for existing RL algorithms.
- Empirical results show consistent performance improvements across different environments and model scales.
Read more
TAPO: Transition-Aware Policy Optimization for LLM Agents
Summary
The paper introduces TAPO (Transition-Aware Policy Optimization), a novel framework aimed at enhancing the post-training of Large Language Model (LLM) agents through Reinforcement Learning (RL). Traditional RL methods primarily focus on sparse task rewards for policy optimization, neglecting the rich, action-conditioned environmental feedback available during online interactions. TAPO addresses this gap by alternating between policy optimization and transition supervision, leveraging rollout data to predict the next observation based on the current state and executed action. This dual approach not only improves the agent's policy but also enhances its understanding of environmental dynamics, leading to better decision-making in long-horizon tasks. The authors validate TAPO through systematic experiments on benchmarks like WebShop and ALFWorld, demonstrating consistent performance improvements across various model scales and RL algorithms. The findings underscore the importance of utilizing inherent transition feedback from rollouts as a valuable learning signal, which can significantly enhance the effectiveness of LLM agents in complex environments.
Methodology
TAPO employs a unified training framework that alternates between policy optimization and transition supervision. It utilizes existing rollout data to apply action-conditioned next-observation prediction, enhancing the model's understanding of environmental transitions while optimizing the policy simultaneously.
Results
The experiments conducted on WebShop and ALFWorld benchmarks reveal that TAPO consistently outperforms traditional policy optimization baselines, demonstrating improved task performance and decision-making capabilities across various model scales and RL algorithms.
Implications
The findings suggest that incorporating transition supervision into the training of LLM agents can significantly enhance their performance in complex, long-horizon tasks. This approach could be applied to various domains where LLM agents are utilized, potentially leading to more robust and capable autonomous systems.
Top-$k$ Pareto Bandits: Hypervolume Regret for Multi-Objective Slate Selection
Optimization
Reinforcement Learning
Theory
- Introduction of the Top-k Pareto bandit setting and α-approximate hypervolume regret.
- Development of the THV-UCB algorithm for greedy selection based on optimistic marginal hypervolume gains.
- Establishment of theoretical regret bounds that are gap-free and gap-dependent.
- Empirical evaluation demonstrating THV-UCB's superior performance in hypervolume maximization.
Read more
Top-$k$ Pareto Bandits: Hypervolume Regret for Multi-Objective Slate Selection
Summary
This paper addresses a stochastic multi-objective bandit problem where an agent selects a slate of k arms and observes their d-dimensional reward vectors under semi-bandit feedback. The focus is on approximating the Pareto frontier rather than identifying a single optimal arm. The authors introduce the concept of α-approximate hypervolume regret, which measures the performance of selected arms against the best achievable subset in hindsight. They propose an algorithm called THV-UCB, which uses optimistic estimates of marginal hypervolume contributions to select arms greedily. Theoretical analysis provides gap-free and gap-dependent regret bounds, demonstrating the effectiveness of small subsets in approximating Pareto fronts across various applications. The paper also empirically evaluates the algorithm against existing methods, showing superior performance in terms of cumulative regret and hypervolume maximization across different geometries and dimensions of the Pareto front.
Methodology
The authors formalize a stochastic multi-objective bandit problem where the agent selects k arms from n candidates and observes their d-dimensional reward vectors. The THV-UCB algorithm employs coordinate-wise confidence boxes and greedy selection based on optimistic marginal hypervolume contributions, allowing for effective pruning of dominated arms and initial forced exploration.
Results
The THV-UCB algorithm achieves the lowest cumulative α-regret and the highest hypervolume across four different geometries of Pareto fronts and varying dimensions. The theoretical analysis confirms a gap-free regret bound of ˜O(d√nkT) and a gap-dependent bound of ˜O(nk².5/∆min), which becomes polylogarithmic in T under certain conditions.
Implications
The findings suggest that small subsets can effectively approximate Pareto fronts in multi-objective optimization scenarios, which is valuable for applications in recommender systems, market making, and clinical trials. The theoretical and empirical results support the use of THV-UCB in real-world decision-making processes involving multiple conflicting criteria.
Context-Informed Ship Trajectory Prediction via Conditional Attention
Multimodal
Time Series
Robotics
- Introduces the Conditional Informer architecture for ship trajectory prediction.
- Utilizes a Conditional Attention mechanism to model the influence of environmental factors on vessel dynamics.
- Implements Modality Masking to address data intermittency and prevent shortcut learning.
- Achieves a 15.4% improvement in prediction accuracy over traditional baselines.
Read more
Context-Informed Ship Trajectory Prediction via Conditional Attention
Summary
This paper addresses the challenge of long-term ship trajectory prediction, which is crucial for maritime safety and autonomous navigation. Traditional Transformer-based models have improved prediction accuracy but often treat vessel motion as an isolated process, neglecting the influence of environmental factors such as weather. The authors propose a novel architecture called the Conditional Informer, which reformulates trajectory prediction as a conditional generation task. This architecture incorporates a Conditional Attention mechanism that allows vessel states to query environmental contexts, effectively capturing the physical dependencies between vessel dynamics and external conditions. Additionally, the paper introduces a Modality Masking training strategy to mitigate the effects of data intermittency, which is common in real-world scenarios. Experimental results demonstrate that the Conditional Informer outperforms existing baselines by 15.4% in prediction accuracy when context is available and significantly reduces fallback errors during sensor failures.
Methodology
The authors developed the Conditional Informer, an encoder-decoder architecture that employs Conditional Attention to explicitly model the relationship between vessel states and environmental contexts. The architecture uses separate embedding pathways for kinematic and environmental data, allowing for a directional influence of environmental factors on vessel dynamics. The Modality Masking strategy is introduced to enhance training stability and model performance during data dropouts.
Results
The Conditional Informer outperformed kinematic and concatenation-based baselines by 15.4% in prediction accuracy when environmental context was utilized. The implementation of Modality Masking significantly reduced fallback errors, achieving nearly an order of magnitude improvement compared to unconstrained models, thereby enhancing the model's reliability in real-world applications.
Implications
The proposed approach has significant implications for improving maritime navigation safety and efficiency, particularly in autonomous vessel technologies. By accurately predicting ship trajectories in the presence of environmental influences, the model can aid in proactive shipping lane management and enhance maritime domain awareness, especially during data transmission failures.
$β$-OPSD: Deriving with Policy Optimization, Training with Self-Distillation
NLP
Large Language Models
Reinforcement Learning
- β-OPSD generalizes OPSD by introducing a controllable parameter β for better optimization.
- The optimal policy is derived as a geometric interpolation between a reference policy and a privileged teacher.
- The method efficiently approximates complex RL objectives using token-level logit interpolation.
- Experiments show significant improvements in reasoning performance and optimization stability over vanilla OPSD.
Read more
$β$-OPSD: Deriving with Policy Optimization, Training with Self-Distillation
Summary
The paper introduces β-OPSD, an advanced framework for on-policy self-distillation (OPSD) aimed at enhancing reasoning capabilities in language models. Traditional OPSD, while promising, often suffers from brittleness and requires extensive engineering to function effectively. The authors identify that vanilla OPSD operates at a fixed parameter β = 1, which limits its flexibility. By generalizing this approach, they propose a family of OPSD objectives parameterized by β, allowing for controlled adjustments in the balance between proximity to a reference policy and guidance from a privileged teacher. The optimal policy for β-OPSD is derived as a geometric interpolation between these two policies. Instead of directly optimizing the complex reinforcement learning (RL) objective, the authors convert it into a distillation target, enabling efficient training through token-level logit interpolation. This method retains the simplicity of OPSD while improving optimization stability and performance. The paper demonstrates that β-OPSD significantly outperforms vanilla OPSD across various mathematical reasoning benchmarks, showcasing its potential to enhance reasoning performance in large language models without sacrificing efficiency.
Methodology
The authors derive the optimal policy for β-OPSD as a geometric interpolation between a reference policy and a privileged teacher. They implement this through a distillation target that is efficiently realized using token-level logit interpolation, allowing for a smooth curriculum of training targets. Additionally, they incorporate return-to-go credit assignment to enhance the alignment of updates with the overall sequence-level objective.
Results
Experiments conducted on mathematical reasoning benchmarks demonstrate that β-OPSD consistently improves optimization stability and reasoning performance compared to vanilla OPSD. For instance, on the Qwen3-1.7B model, β-OPSD achieves an improvement of up to 9.16 percentage points in average performance across multiple benchmarks.
Implications
The findings suggest that β-OPSD could be a valuable approach for enhancing the reasoning capabilities of large language models, making them more robust and efficient in learning from their own outputs. This has potential applications in various domains where reasoning and decision-making are critical.
Recursive transformers for semiconductor thermo-mechanical reliability
Efficient ML
- Conventional transformers are over-parameterized for small engineering datasets, leading to overfitting.
- Three recursive transformer architectures are proposed to optimize parameter efficiency and computational complexity.
- The models are validated on thermo-mechanical reliability analysis and Laplace PDE numerical solving tasks.
- Recursive weight-sharing transformers provide a practical solution for resource-constrained engineering applications.
Read more
Recursive transformers for semiconductor thermo-mechanical reliability
Summary
This paper addresses the challenges of using conventional transformer architectures in engineering design, particularly in semiconductor thermo-mechanical reliability analysis. Traditional transformers are often over-parameterized for the small datasets typical in this field, leading to overfitting and high computational costs. The author proposes three recursive transformer paradigms: Tiny Recursive Model (TRM), Depth Recursive Model (DEPTH), and a simple recursive transformer, focusing on reducing parameter count and computational complexity while maintaining predictive performance. The study systematically evaluates these models on two engineering tasks: thermo-mechanical reliability analysis of semiconductor packages and a Laplace PDE iterative numerical solver for capacitance fields. The findings indicate that recursive weight-sharing transformers effectively balance prediction accuracy, parameter efficiency, and computational cost, making them suitable for small-data engineering surrogate modeling, which is a limitation of conventional large-parameter transformers.
Methodology
The paper employs a hardware-aware evaluation of three recursive transformer paradigms, comparing their predictive performance, parameter count, and computational complexity (FLOPs). The models are tested on two engineering prediction tasks, emphasizing the need for architectures that are efficient in both memory and computation.
Results
The recursive weight-sharing transformers demonstrated improved predictive performance while significantly reducing the number of parameters and computational complexity compared to conventional transformer models. The results validate the effectiveness of these architectures in small-data engineering applications, particularly in semiconductor reliability prediction and capacitance field modeling.
Implications
The findings suggest that recursive transformers can be effectively utilized in engineering design workflows, particularly in scenarios where computational resources are limited. This could lead to faster and more efficient design processes in semiconductor manufacturing and other engineering fields.
Same Graph Cross-Task Transfer in GNNs: Protocols and Predictors
Graph Learning
- Formalizes same graph cross-task transfer between node classification and link prediction.
- Introduces a leakage-free evaluation protocol to ensure fair comparisons.
- Demonstrates that NC to LP transfer is generally beneficial, while LP to NC transfer is more fragile.
- Introduces CoTask Score (CTS) for summarizing joint task performance.
Read more
Same Graph Cross-Task Transfer in GNNs: Protocols and Predictors
Summary
This paper addresses the challenge of cross-task transfer learning in Graph Neural Networks (GNNs) by formalizing the reuse of supervision signals from node classification (NC) and link prediction (LP) tasks on the same graph. The authors identify issues with existing evaluation protocols that can lead to unreliable conclusions about transfer learning effectiveness. To overcome these issues, they propose a leakage-free evaluation protocol that standardizes node and edge splits, constructs a consistent observed graph, and employs fixed negatives for LP. The study reveals that cross-task transfer is asymmetric: NC to LP transfer is consistently beneficial on homophilic graphs, while LP to NC transfer is less reliable and can degrade performance unless specific conditions are met. The authors introduce the CoTask Score (CTS) to evaluate joint task performance and show that simple dataset statistics, particularly homophily, can guide the choice of transfer mechanisms. Overall, the paper provides a structured approach to understanding and leveraging cross-task transfer in GNNs, offering insights into when and how to effectively reuse representations across tasks.
Methodology
The authors developed a standardized evaluation protocol that fixes node and edge splits, constructs a consistent observed graph, and uses fixed negatives for link prediction. They evaluated five cross-task transfer strategies across three GNN architectures (GCN, GraphSAGE, GPS) under both transductive and inductive settings, analyzing the directional transfer effects.
Results
The experiments showed that NC to LP transfer consistently improves performance on homophilic graphs, while LP to NC transfer is less reliable and can degrade accuracy unless the graph structure favors it. The findings suggest that LP can serve as structural pretraining for NC under certain conditions, particularly when LP is easy and NC is unsaturated.
Implications
The results have significant implications for designing GNNs that can effectively leverage multiple tasks on the same graph, guiding practitioners on when to reuse representations and how to optimize model performance across different tasks.
Do You Really Need to Pretrain Q-Functions for Online RL Fine-Tuning?
Reinforcement Learning
Robotics
- Naive pretraining of Q-functions does not significantly enhance performance in online RL fine-tuning.
- The Q-function learned during pretraining is fundamentally different from the one learned during online fine-tuning.
- Initialization via Policy Ensemble (IPE) improves Q-function learning by leveraging diverse policy rollouts.
- IPE results in an average performance improvement of 26% over naive pretraining methods.
Read more
Do You Really Need to Pretrain Q-Functions for Online RL Fine-Tuning?
Summary
This paper investigates the necessity of pretraining Q-functions in the context of online reinforcement learning (RL) fine-tuning, particularly when starting from a pretrained policy. The authors challenge the conventional wisdom that suggests pretraining the Q-function on offline data is beneficial. Through systematic empirical analysis, they find that naive pretraining often yields minimal performance improvements compared to random initialization. The authors attribute this to a fundamental mismatch between the Q-function learned during pretraining and the one that online fine-tuning converges to. To address this issue, they propose a novel method called Initialization via Policy Ensemble (IPE), which trains multiple diverse policies and uses their pooled rollouts to enhance Q-function learning during online RL. The results demonstrate that IPE significantly outperforms naive Q-function pretraining, achieving an average improvement of 26% across various continuous control benchmarks.
Methodology
The authors conducted a systematic empirical analysis comparing the performance of pretrained Q-functions against randomly initialized Q-functions in online RL fine-tuning. They introduced IPE, which involves training multiple diverse policies and using their rollouts to bootstrap Q-function learning, thereby increasing the diversity of the data used for training.
Results
The study found that naive Q-function pretraining often provides little to no benefit compared to random initialization. In contrast, the proposed IPE method resulted in an average performance improvement of 26% across a suite of challenging continuous control tasks.
Implications
The findings suggest that practitioners in reinforcement learning should reconsider the reliance on Q-function pretraining, especially when fine-tuning pretrained policies. The proposed IPE method offers a promising alternative for improving online RL performance, which could be beneficial in various applications, including robotics and control systems.
Cost-Sensitive Conformal Prediction and Human-in-the-Loop Abstention for Imbalanced High-Stakes Decision Support: A Multi-Domain Benchmark
Theory
Efficient ML
Interpretability
- Marginal conformal prediction fails to adequately cover minority classes, with coverage dropping below 1%.
- Mondrian conformal prediction restores minority coverage, achieving an average improvement of 61.7 percentage points.
- Cost-controlled abstention mechanisms can lower expected decision costs by deferring ambiguous predictions to human experts.
- The study quantifies break-even thresholds for human review costs, establishing economic boundaries for automated decision-making.
Read more
Cost-Sensitive Conformal Prediction and Human-in-the-Loop Abstention for Imbalanced High-Stakes Decision Support: A Multi-Domain Benchmark
Summary
This paper addresses the challenges of uncertainty quantification in high-stakes decision-making systems, particularly in the context of imbalanced datasets where minority classes are often underrepresented. The authors benchmark various conformal prediction methods, including marginal conformal prediction (CP) and class-conditional (Mondrian) CP, across 15 real-world datasets. They demonstrate that marginal CP can severely under-cover minority classes, achieving less than 1% coverage in extreme cases. In contrast, Mondrian CP significantly improves minority coverage by an average of 61.7 percentage points. The study also introduces a cost-controlled abstention mechanism that allows for deferring ambiguous predictions to human experts, thereby reducing expected decision costs. The authors quantify dataset-specific break-even thresholds for human review costs, providing actionable insights for deploying cost-sensitive conformal prediction in decision support systems. Overall, the findings highlight the importance of tailored uncertainty quantification methods in high-stakes applications such as healthcare and fraud detection.
Methodology
The authors conducted a comprehensive benchmark comparing marginal CP, Mondrian CP, and cost-controlled abstention across 15 imbalanced datasets. They utilized 7 classification models, 3 probability calibration techniques, and performed 3,150 experimental runs to evaluate the performance of these methods in terms of minority class coverage and decision costs.
Results
The empirical results showed that marginal CP had severe undercoverage for minority classes, while Mondrian CP significantly improved coverage. The integration of Mondrian CP with cost-controlled abstention mechanisms led to reduced expected decision costs, demonstrating the effectiveness of this approach in high-stakes scenarios. The study also established dataset-specific break-even thresholds for human review costs, indicating where deferring decisions to experts is economically beneficial.
Implications
The findings have significant implications for high-stakes decision-making systems in fields such as healthcare, credit scoring, and fraud detection, where accurate uncertainty quantification is critical. The proposed methods can enhance the reliability of automated systems while managing costs associated with human review.
Policy Gradient Steering: Interventions from Behavioral Objectives
Reinforcement Learning
- Introduction of Policy Gradient Steering (PGS) as a method for behavioral adaptation in reinforcement learning.
- Demonstration of PGS's effectiveness in a two-route gridworld environment, chess, and competitive football.
- PGS allows for the creation of removable and composable behavioral interventions without retraining the base policy.
- The method shows that compatible tactical objectives can constructively accumulate.
Read more
Policy Gradient Steering: Interventions from Behavioral Objectives
Summary
This paper introduces Policy Gradient Steering (PGS), a novel method for dynamically altering the behavior of trained policies in reinforcement learning without the need for retraining. Existing steering methods often fail to effectively modify policies in certain environments, such as a two-route gridworld. PGS addresses this by framing steering as a reinforcement learning problem, where it accumulates gradients from a small set of rollouts to create a removable task vector that can influence policy decisions at inference time. The authors demonstrate the effectiveness of PGS through controlled experiments in a gridworld, chess puzzles, and competitive football, showcasing its ability to construct temporary and composable behavioral adaptations. The results indicate that PGS can successfully alter specific behaviors in multi-agent settings and that these adaptations can transfer across different contexts, providing a flexible interface for behavioral modifications in decision-making tasks.
Methodology
The methodology involves accumulating gradients from a small number of rollouts or demonstrations based on a temporary behavioral objective. PGS constructs an additive steering vector by assigning credit to actions that lead to specific outcomes, rather than relying on post-outcome representations. This approach allows for the scaling, composition, and removal of the steering vector at inference time, maintaining the integrity of the original policy.
Results
The empirical evaluations revealed that PGS effectively calibrated and reversed interventions in the gridworld environment. In chess, independently fitted PGS vectors demonstrated compositional properties, retaining their effects across different skill levels. In competitive football, PGS successfully altered team behaviors and showed transferability of effects across opponents, indicating its robustness in multi-agent scenarios.
Implications
The findings suggest that PGS can be utilized in various applications requiring dynamic behavioral adaptations, such as robotics, game AI, and other decision-making systems. Its ability to provide lightweight, inference-time control without modifying the base policy opens new avenues for real-time adjustments in complex environments.
Understanding Context Sampling in TabPFN on Small Tabular Datasets
Theory
Efficient ML
- Larger contexts in TabPFN yield more stable and accurate predictions.
- Context diversity, rather than representativeness, is crucial for accuracy.
- Expensive selection methods do not provide significant advantages over uniform random selection.
- Controlled experiments demonstrate the importance of diversity in context sampling.
Read more
Understanding Context Sampling in TabPFN on Small Tabular Datasets
Summary
This paper investigates the impact of context sampling on the performance of TabPFN, a model that utilizes in-context learning for classification tasks on small tabular datasets. The study focuses on three main questions: the variability of predictions across different random contexts, the relationship between context representativeness and accuracy, and the effectiveness of various context selection methods. Through experiments on 15 small OpenML datasets, the author finds that larger contexts lead to more stable and accurate predictions. The analysis reveals that while accuracy correlates with how well a context preserves the training distribution, the key factor influencing accuracy is actually the diversity of the context rather than its representativeness. Furthermore, the study concludes that more complex selection methods, such as K-Means and farthest-point sampling, do not significantly outperform uniform random selection, which is simpler and more cost-effective. The findings emphasize the importance of context diversity in achieving reliable performance in small-data scenarios.
Methodology
The methodology involves repeated random sub-sampling across 15 small OpenML datasets to assess the variability of predictions with different context sizes. Controlled experiments are conducted to evaluate the effects of context representativeness versus diversity on accuracy. Additionally, a mixed-effects analysis is utilized to analyze the relationship between context characteristics and model performance.
Results
The results indicate that increasing the context size reduces the variability of predictions, with the coefficient of variation of AUC decreasing significantly as context size increases. While accuracy initially appears to correlate with representativeness, controlled experiments reveal that diversity is the actual predictor of accuracy. More complex selection methods do not yield better performance than uniform random sampling, which effectively covers the feature space.
Implications
The findings suggest that practitioners using TabPFN on small datasets should prioritize context diversity over representativeness when selecting training examples. This insight can inform better practices in prototype selection, potentially leading to improved model performance in various applications involving small tabular datasets.
Lottery Tickets Are Not Deployment Tickets
Computer Vision
Theory
Efficient ML
- Lottery tickets can achieve accuracy comparable to dense models but may behave differently in deployment.
- Replacing a dense model with a lottery ticket can lead to significant changes in downstream decision-making.
- Clean-accuracy recovery does not guarantee compatibility with existing decision logic.
- A behavioral-compatibility distance metric is introduced to evaluate model behavior in practical deployment scenarios.
Read more
Lottery Tickets Are Not Deployment Tickets
Summary
This paper investigates the practical implications of using lottery tickets (LTs) as replacements for dense models in deployment scenarios. Previous literature has shown mixed results regarding the effects of sparsification and compression on model behavior, with some studies reporting benefits while others indicate adverse effects. The author focuses on the production-replacement question, assessing whether an accuracy-matched lottery ticket can replace an incumbent dense model without necessitating changes to downstream decision logic. The study audits various deployment-relevant behaviors, including calibration, out-of-distribution (OOD) response, and class-level reliability, and introduces a behavioral-compatibility distance metric to summarize deviations from expected performance. Experiments conducted on multiple datasets (CIFAR-10, CIFAR-100, Imagenette, Flowers-102, and FGVC-Aircraft) using different model architectures (ResNet, WideResNet, ConvNeXt, and Vision Transformer) reveal that while sparse models can achieve similar accuracy to dense counterparts, they often exhibit behavioral differences that can lead to significant changes in decision outcomes. Specifically, the findings indicate that replacing a dense model with a lottery ticket can alter 7% to 10% of accept-review decisions, thereby creating the operational burden that drop-in replacements aim to avoid. The paper concludes that clean-accuracy certification alone is insufficient for ensuring compatibility with existing deployment logic, emphasizing the need for additional audits to assess operational-point compatibility.
Methodology
The author conducts a series of experiments across multiple datasets and model architectures to evaluate the deployment-relevant behaviors of lottery tickets compared to dense models. The study utilizes a behavioral-compatibility distance metric to quantify deviations in model behavior and assesses the impact of these deviations on downstream decision-making processes.
Results
The experiments demonstrate that while lottery tickets can match the clean accuracy of dense models, they often lead to behavioral differences that result in 7% to 10% changes in accept-review decisions. This indicates that deploying lottery tickets as replacements for dense models can create operational challenges that require revalidation of downstream decision logic.
Implications
The findings suggest that organizations considering the deployment of sparse models must conduct thorough compatibility audits to ensure that operational decision-making remains consistent. This has implications for the practical application of sparsification techniques in real-world scenarios, where maintaining decision logic is critical.
Q-Steer: Action-Value Guidance for Molecular Policy Optimization
Reinforcement Learning
Generative Models
Optimization
- Q-Steer provides action-value guidance during molecular generation, addressing the myopic nature of traditional methods.
- The method utilizes a frozen PAVS-Q model to estimate the potential rewards of candidate tokens based on partial SMILES prefixes.
- Extensive experiments show consistent improvements in mean valid-unique scores across various molecular language models and optimizers.
- The findings highlight the importance of action identity in optimization performance.
Read more
Q-Steer: Action-Value Guidance for Molecular Policy Optimization
Summary
The paper introduces Q-Steer, a novel approach to molecular policy optimization that addresses the challenge of delayed feedback in oracle-limited molecular generation. Traditional methods receive rewards only after generating complete molecules, leading to myopic decision-making during the token generation process. Q-Steer employs an offline-trained, frozen prefix-action value scorer (PAVS-Q) to provide action-value guidance during rollouts. This guidance estimates the potential downstream reward of candidate next tokens based on the current partial SMILES prefix, allowing the optimizer to make more informed decisions without altering the optimizer update rule or the online oracle budget. The authors conducted extensive factorial studies across two molecular language model backbones and four optimizer families, demonstrating that Q-Steer consistently improves the mean valid-unique score across all tested configurations. The results indicate that action identity plays a crucial role in performance, with specific mechanisms showing that prefix-broadcast values are nearly neutral while shuffled action values negatively impact results. Overall, Q-Steer serves as a reusable tool that enhances molecular optimization by integrating action-level future value guidance.
Methodology
The authors developed Q-Steer as a rollout-time action-value steering primitive, using an offline-trained PAVS-Q model to score candidate next tokens during the sampling process. This model provides a normalized value bonus to the sampling logits based on the estimated downstream rewards, allowing the optimizer to maintain its original update rules while benefiting from enhanced decision-making during token generation.
Results
Q-Steer improved the mean valid-unique score in all tested configurations, achieving positive macro mean-score gains between +0.033 and +0.049 across eight backbone-optimizer combinations. The method demonstrated 18-20 task wins per cell in the factorial studies, indicating its effectiveness in enhancing molecular optimization performance.
Implications
The introduction of Q-Steer has significant implications for molecular design and optimization, particularly in drug discovery and materials science, where generating high-quality molecular structures efficiently is crucial. The method's modular nature allows for easy integration into existing molecular optimization frameworks, potentially leading to more effective and efficient exploration of chemical space.
Latent States in Neural Networks: Recovering the Temporal Structure of Drifting Data from Model Weights
Time Series
NLP
Multimodal
- Latent states can be recovered from the weights of neural networks trained on temporally drifting data.
- The study employs a hidden Markov model (HMM) to analyze the weight trajectories of classifiers.
- Classifiers show better generalization within the same latent state compared to across state boundaries.
- The identified latent states correlate more with class distribution shifts than with weight-space geometry.
Read more
Latent States in Neural Networks: Recovering the Temporal Structure of Drifting Data from Model Weights
Summary
This paper investigates whether the discrete regimes of temporally drifting data streams can be recovered from the weights of neural network models trained on such data. The author employs a hidden Markov model (HMM) to analyze the trajectory of model weights over time, focusing on two domains: multimodal misinformation detection using the Fakeddit dataset and sentiment analysis using the Yelp dataset. By training classifiers on consecutive temporal windows and aligning their weights, the study finds that the latent states identified by the HMM correspond to coherent phases in the data. The results indicate that classifiers generalize better to data from windows that share the same latent state as their training data, demonstrating a significant within-state transfer advantage. This advantage persists even after controlling for temporal proximity and is more pronounced than a naive partitioning of the data. The study concludes that the latent states, derived solely from model weights, correlate more strongly with shifts in class distribution than with the weight-space geometry, suggesting that these states capture relevant structural information for transfer learning beyond mere data distribution.
Methodology
The author trains classifiers on consecutive temporal windows of data and aligns their weights for comparison. A hidden Markov model (HMM) is then fitted to the chronologically ordered sequence of these weights to recover latent states that partition the timeline into coherent phases.
Results
The study finds that classifiers trained on data from windows sharing the same latent state generalize better than those trained on data across state boundaries. This within-state transfer advantage is statistically significant and exceeds that of a naive equal-size partition of the data. The effects are consistent across both the misinformation detection and sentiment analysis tasks, although they are somewhat attenuated in the more stable Yelp dataset.
Implications
The findings suggest that neural networks can capture and leverage temporal structures in data streams, which could enhance model robustness and transfer learning capabilities in dynamic environments. This approach may be applicable in various domains where data distributions change over time, such as finance, healthcare, and social media analysis.
Collaborative Weighting with Pessimistic Critic for Mitigating Overestimation in Off-Policy Reinforcement Learning
Reinforcement Learning
Robotics
Optimization
- Introduction of CWAC framework that combines collaborative weighting and pessimistic value estimation.
- Adaptive reweighting of TD-errors based on predictive uncertainty to improve learning stability.
- Stochastic sampling from return distributions to reduce overestimation and error propagation.
- Demonstrated superior performance and sample efficiency in continuous control benchmarks.
Read more
Collaborative Weighting with Pessimistic Critic for Mitigating Overestimation in Off-Policy Reinforcement Learning
Summary
This paper addresses the persistent overestimation bias and degraded training stability in off-policy reinforcement learning (RL) algorithms, particularly in continuous control tasks. The authors introduce a novel framework called Collaborative Weighting Actor-Critic (CWAC), which integrates a distributional critic to model return uncertainty and a collaborative weighting mechanism that reweights temporal-difference (TD) errors based on predictive uncertainty. This approach allows for robust learning from reliable samples while suppressing noisy updates. Additionally, the authors implement a stochastic pessimistic value estimation scheme that mitigates error propagation during policy improvement. CWAC can be easily integrated into existing off-policy algorithms like Soft Actor-Critic (SAC), Twin Delayed Deep Deterministic Policy Gradient (TD3), and Deep Deterministic Policy Gradient (DDPG) with minimal computational overhead. Empirical evaluations demonstrate that CWAC significantly enhances performance across various simulated tasks, achieving improved sample efficiency and training stability compared to state-of-the-art methods.
Methodology
The CWAC framework employs a distributional critic to estimate predictive uncertainty and incorporates a collaborative weighting mechanism that adaptively adjusts the influence of TD-errors based on this uncertainty. A stochastic pessimistic value estimation approach is utilized, where policy improvements are made through sampling from the return distribution rather than directly maximizing expected Q-values.
Results
The experimental results show that CWAC outperforms existing state-of-the-art off-policy reinforcement learning methods across eight challenging continuous control tasks, leading to significant improvements in both sample efficiency and training stability.
Implications
The proposed CWAC framework has the potential to enhance the reliability and efficiency of off-policy reinforcement learning applications in various domains, including robotics, autonomous driving, and industrial scheduling, by effectively mitigating overestimation biases.
Tight Sample Complexity for Low-Rank Adaptation: Matching Bounds and Rank Selection
Theory
Efficient ML
Large Language Models
- Establishes tight upper and lower bounds on the sample complexity of Low-Rank Adaptation.
- Introduces a rank-selection dichotomy, highlighting the effects of under-ranking and over-ranking.
- Validates theoretical predictions through experiments on synthetic and real datasets.
- Demonstrates that over-parameterization can harm performance in constrained empirical risk minimization.
Read more
Tight Sample Complexity for Low-Rank Adaptation: Matching Bounds and Rank Selection
Summary
This paper addresses the statistical properties of Low-Rank Adaptation (LoRA), a technique for fine-tuning large pretrained models. While LoRA has gained popularity due to its efficiency, its sample complexity and optimal rank selection have not been fully understood. The author establishes both upper and lower bounds on the excess risk associated with LoRA, closing the gap in existing literature. An upper bound of O(rd/n) is derived using local Rademacher complexity, while a matching minimax lower bound of Ω(rd/n) is proven through a Fano-type argument. The findings reveal a rank-selection dichotomy: for constrained empirical risk minimizers, the optimal rank matches the intrinsic rank, while for adaptive estimators, over-ranking does not incur a penalty. The theoretical predictions are validated through experiments on synthetic and real datasets, demonstrating a U-shaped relationship in validation loss with respect to rank, confirming the detrimental effects of over-ranking. This work provides crucial insights into the statistical complexity of LoRA fine-tuning and offers practical guidance for rank selection.
Methodology
The paper employs local Rademacher complexity to derive an upper bound on the excess risk for LoRA. A Fano-type argument is used to establish a matching lower bound. The rank-selection theorem is derived through a combination of bias analysis and the established bounds, with empirical validation conducted on synthetic and real datasets.
Results
The results include an upper bound of O(rd/n) on the excess risk for LoRA, a matching lower bound of Ω(rd/n), and a rank-selection theorem indicating that the optimal rank for constrained empirical risk minimization is the intrinsic rank. Experiments confirm the theoretical predictions, showing a U-shaped validation loss curve with significant loss inflation at larger ranks.
Implications
The findings provide essential insights for practitioners using LoRA, emphasizing the importance of selecting an appropriate rank to avoid performance degradation. The strategies proposed for rank selection can enhance the efficiency and effectiveness of fine-tuning large models.
LM-GRASP: Instance-Specific Language Models for Combinatorial Construction via Online Imitation Learning
Optimization
- LM-GRASP reformulates the constructive phase of GRASP as an online imitation learning task.
- The framework learns from high-quality solutions without requiring external data or offline pretraining.
- A decoder-only Transformer network serves as the constructive policy, capturing non-myopic dependencies.
- LM-GRASP outperforms traditional methods, achieving a notable reduction in makespan for PFSP.
Read more
LM-GRASP: Instance-Specific Language Models for Combinatorial Construction via Online Imitation Learning
Summary
The paper introduces LM-GRASP, a novel framework that integrates instance-specific language models into the Greedy Randomized Adaptive Search Procedure (GRASP) for combinatorial optimization. Unlike traditional methods that rely on pre-trained neural constructors, LM-GRASP employs online imitation learning to adaptively learn a constructive policy from high-quality solutions generated during the optimization process. The framework utilizes a local search procedure as an expert oracle and a decoder-only Transformer network to model the construction policy. This approach eliminates the need for static, hand-crafted heuristics, allowing the model to exploit structural regularities in the solution space. The authors evaluate LM-GRASP on the Permutation Flow-Shop Scheduling Problem (PFSP) using the challenging Taillard benchmark suite, demonstrating that it significantly outperforms existing methods, including GPU-accelerated GRASP variants. The results indicate that instance-specific, online-trained language models can effectively enhance metaheuristic search, particularly for complex combinatorial problems.
Methodology
LM-GRASP employs a hybrid metaheuristic approach that combines online imitation learning with a decoder-only Transformer network. The local search procedure acts as an expert oracle, refining candidate solutions into local optima. The model is trained iteratively through behavioral cloning on an evolving archive of elite solutions, requiring no explicit problem-specific feature encoding.
Results
LM-GRASP demonstrated a significant performance improvement, outperforming GPU-GRASP by an average of 28.4 makespan units on the PFSP benchmark. This improvement is comparable to the performance gains achieved through GPU acceleration, indicating that the online-trained language model effectively enhances the search process.
Implications
The findings suggest that integrating instance-specific, online-trained language models into metaheuristic frameworks can provide a practical alternative to traditional hand-engineered constructors, particularly for complex combinatorial optimization problems. This approach may lead to more efficient and adaptable optimization strategies across various domains.
Learning-Augmented and Randomized Algorithms for Line Aggregation with Delays
Theory
Optimization
- Introduction of a deterministic learning-augmented algorithm for line aggregation with delays.
- Development of a randomized algorithm with a competitive ratio better than existing deterministic benchmarks.
- Establishment of a new lower bound for randomized algorithms in this context.
- Combination of learning-augmented and randomized approaches to enhance algorithm performance.
Read more
Learning-Augmented and Randomized Algorithms for Line Aggregation with Delays
Summary
This paper investigates learning-augmented and randomized algorithms for online aggregation with delays on a line metric. The authors propose a deterministic learning-augmented Balance algorithm that achieves (4/λ + 1/λ²)-robustness and (4 + λ)-consistency, where λ is a parameter indicating the trust in advice. Additionally, they introduce a randomized algorithm that is (e + 1)-competitive against an oblivious adversary, surpassing the deterministic 5-competitive benchmark. The paper also establishes a new lower bound of e for randomized algorithms, improving upon previous bounds. Furthermore, a combined approach yields a randomized learning-augmented algorithm with (e/λ + 1/λ²)-robustness and (e + λ)-consistency. Numerical experiments validate the theoretical findings, demonstrating that the proposed algorithms outperform existing methods and maintain stability as the advice replacement rate increases.
Methodology
The authors utilize a combination of deterministic and randomized algorithmic strategies, incorporating learning-augmented techniques that adjust request urgency based on advice. They analyze the performance of these algorithms using worst-case competitive analysis and conduct numerical experiments to assess empirical performance.
Results
The deterministic learning-augmented algorithm achieves (4/λ + 1/λ²)-robustness and (4 + λ)-consistency. The randomized algorithm achieves an (e + 1)-competitive ratio, while the combined randomized learning-augmented algorithm reaches (e/λ + 1/λ²)-robustness and (e + λ)-consistency. The empirical results demonstrate that these algorithms outperform existing online algorithms.
Implications
The findings suggest that integrating learning-augmented methods with randomized algorithms can significantly improve performance in online decision-making scenarios, particularly in applications involving aggregation and service requests. This approach could be beneficial in fields such as logistics, data transmission, and inventory management.
Budget-Aware LLM Discovery via Cost-Calibrated Frontier Utility
Large Language Models
Optimization
Efficient ML
- Cost-blind credit assignment can lead to significant quality loss in LLM discovery under budget constraints.
- CostAda introduces a cost-calibrated frontier utility that improves resource allocation during the search process.
- The methodology allows for dynamic adjustments in exploration intensity and frontier allocation based on remaining budget.
- CostAda achieves superior performance with reduced budget usage compared to existing methods.
Read more
Budget-Aware LLM Discovery via Cost-Calibrated Frontier Utility
Summary
This paper addresses the challenges of large language model (LLM) discovery under explicit token budgets, where traditional adaptive discovery controllers fail to account for the varying costs of search actions. The authors introduce CostAda, a cost-calibrated adaptive controller that optimizes the search process by integrating cost considerations into the credit assignment for search actions. They demonstrate that existing methods, which rely solely on score progress, can lead to suboptimal outcomes as costs diverge and frontiers multiply. CostAda employs a novel utility function that values frontier progress relative to the realized action cost and remaining budget, allowing for more effective allocation of resources during the search process. The methodology includes controlling local exploration intensity, frontier allocation, and budgeted tactic intervention based on the cost-calibrated utility. The results show that CostAda achieves the same quality as the strongest baseline while using at most half the budget in twelve out of sixteen benchmark-backbone pairs, and it records the highest mean final quality across eight benchmarks under GLM-5 and GPT-5.4.
Methodology
The authors developed CostAda, an adaptive controller that utilizes a cost-calibrated frontier utility to guide the search process. This utility combines local frontier progress, global best improvement, realized action cost, and remaining budget to inform decision-making. The controller adjusts local exploration intensity, allocates resources to frontiers based on their cost-effectiveness, and implements budgeted interventions when necessary.
Results
CostAda demonstrated the ability to reach the quality of the strongest baseline while using at most half the budget in twelve out of sixteen benchmark-backbone pairs. Additionally, it achieved the highest mean final quality across all eight benchmarks tested under GLM-5 and GPT-5.4.
Implications
The findings suggest that incorporating cost considerations into adaptive search strategies can significantly enhance the efficiency and effectiveness of LLM discovery processes. This approach may be beneficial in various applications, including algorithm design, optimization tasks, and other domains where resource constraints are critical.
First-order Constrained Trilevel Optimization Over Distributed Networks for Robust Coreset Selection
Optimization
Federated Learning
Theory
- Introduces a novel framework for distributed robust coreset selection, addressing a significant gap in existing literature.
- Formulates the problem as a constrained trilevel optimization, which is more complex than traditional approaches.
- Proposes a distributed algorithm with proven non-asymptotic convergence guarantees.
- Demonstrates the effectiveness of the proposed method through extensive empirical evaluations.
Read more
First-order Constrained Trilevel Optimization Over Distributed Networks for Robust Coreset Selection
Summary
This paper addresses the challenge of robust coreset selection in distributed networks, particularly in the context of the Internet of Things (IoT), where large volumes of sensitive data are generated. The authors propose a novel framework called Federated First-order Constrained Trilevel Optimization (F2CTO) to tackle the problem of coreset selection while ensuring model robustness and privacy. The work formulates the coreset selection as a constrained trilevel optimization problem, which is more complex than traditional bilevel formulations. The F2CTO framework integrates a hierarchical composite value-function reformulation with a distributed alternating projected gradient algorithm, allowing for efficient optimization without the need to centralize sensitive data. The authors prove that their method achieves a non-asymptotic convergence rate of O(ϵ−3/2) for finding an ϵ-stationary point, demonstrating both theoretical and practical advancements in the field. Extensive empirical evaluations confirm the effectiveness and efficiency of F2CTO in real-world continual learning scenarios, highlighting its potential to enhance model performance while addressing privacy concerns.
Methodology
The authors propose the F2CTO framework, which reformulates the coreset selection problem into a trilevel optimization problem with hierarchical constraints. They utilize a hierarchical composite value-function approach and implement a distributed alternating projected gradient algorithm to solve the optimization problem efficiently.
Results
The proposed F2CTO method achieves a non-asymptotic convergence rate of O(ϵ−3/2) for finding an ϵ-stationary point. Empirical evaluations demonstrate that F2CTO effectively enhances model robustness and efficiency in continual learning tasks, outperforming existing methods.
Implications
The findings suggest that F2CTO can significantly improve the performance of machine learning models deployed in distributed environments, particularly in privacy-sensitive applications such as IoT. This work paves the way for future research in distributed optimization and robust learning frameworks.
Harnessing the Potential of Optimizing Data Mixtures via Bayesian Domain Reweighting
NLP
Large Language Models
Optimization
- Introduces a Bayesian domain weighting method to optimize data mixtures for LLMs.
- Addresses limitations of existing methods that rely on strong structural assumptions.
- Demonstrates stable and efficient learning of domain weights with reduced computational costs.
- Empirical results show improved performance over traditional function-fitting approaches.
Read more
Harnessing the Potential of Optimizing Data Mixtures via Bayesian Domain Reweighting
Summary
This paper addresses the challenge of optimizing data mixtures for Large Language Models (LLMs) by proposing a Bayesian domain weighting method. Traditional approaches rely on manual heuristics or function-fitting techniques that often fail to capture the complex relationships between different data domains, leading to suboptimal model performance. The authors highlight the limitations of existing methods, which assume structural properties like rank invariance that do not hold in practice, resulting in significant estimation biases when applied to larger datasets. To overcome these issues, the proposed method infers domain weights from a Dirichlet distribution, incorporating prior information learned from observations through a Gamma prior. This Bayesian approach stabilizes the optimization process and reduces computational overhead, allowing for efficient learning of domain weights. Experimental results demonstrate that the proposed method achieves stable convergence and identifies optimal data mixtures while requiring less data than traditional search-based methods. Overall, this work revitalizes the optimization-based domain weighting approach for large-scale applications, enhancing the performance of LLMs.
Methodology
The authors propose a Bayesian domain weighting method that utilizes a Dirichlet distribution to infer domain weights. This method incorporates prior information through a Gamma distribution, allowing for a more stable optimization process compared to traditional function-fitting techniques. The approach is designed to minimize validation losses while efficiently learning domain weights from the data.
Results
The experimental results indicate that the proposed Bayesian domain weighting method achieves stable convergence in learning domain weights and identifies optimal mixtures more effectively than existing methods. It demonstrates significant improvements in performance while consuming substantially less data, validating its efficacy for large-scale LLM training.
Implications
The findings suggest that the Bayesian domain weighting method can enhance the training of LLMs by optimizing data mixtures more effectively, potentially leading to better model performance in diverse applications. This approach could be particularly beneficial in scenarios with complex data distributions and varying domain importance.
Universality and Approximation Rates of Graph Neural Networks with Random Features
Graph Learning
Theory
- PENNs with random features can approximate any permutation-invariant or permutation-equivariant function on directed graphs.
- The paper provides a concrete architecture for universality, improving upon existing results in the literature.
- Upper bounds on approximation rates are derived, linking network complexity to approximation accuracy.
- The study addresses potential overfitting issues associated with random features and suggests averaging techniques to mitigate these effects.
Read more
Universality and Approximation Rates of Graph Neural Networks with Random Features
Summary
This paper investigates the expressiveness of message-passing graph neural networks (GNNs) enhanced by random node features. The authors establish a universality result for permutation-equivariant neural networks (PENNs), demonstrating that these networks can approximate any measurable permutation-invariant or permutation-equivariant function on directed graphs with multidimensional features. The study extends existing theoretical frameworks by specifying a concrete architecture (PENNs) that achieves universality without imposing strict assumptions on the features. Additionally, the authors derive upper bounds on approximation rates for k-times continuously differentiable functions, linking the complexity of PENN components to the desired approximation accuracy. The findings suggest that random features can improve the expressiveness of GNNs while also addressing concerns about generalization capabilities. The paper contributes to the understanding of GNNs' learning capabilities and their application in various domains where graph structures are prevalent.
Methodology
The authors utilize theoretical analysis grounded in classical approximation theorems for feedforward neural networks, focusing on the architecture of PENNs. They analyze the approximation capabilities of these networks with random features and derive upper bounds for approximation rates based on the complexity of the network components.
Results
The main results include a universality theorem showing that PENNs with random features can approximate any measurable permutation-invariant or permutation-equivariant function on directed graphs. Additionally, the authors provide upper bounds on the approximation rates for k-times continuously differentiable functions, establishing a relationship between the complexity of the PENN architecture and the desired approximation accuracy.
Implications
The findings have significant implications for the design and application of GNNs in various fields, including bioinformatics, social sciences, and finance. The ability to approximate complex functions with high accuracy enhances the potential of GNNs in real-world applications where graph data is prevalent.
Evaluation Protocols and Cross-Subject Generalization in EEG Emotion Recognition
Time Series
- Introduces a three-part EEG protocol record to clarify evaluation processes.
- Demonstrates significant performance discrepancies between subject-dependent and subject-disjoint evaluations.
- Establishes the importance of distinguishing between different evaluation settings in EEG studies.
- Proposes an auditable checklist for EEG reporting to enhance reproducibility.
Read more
Evaluation Protocols and Cross-Subject Generalization in EEG Emotion Recognition
Summary
This paper addresses the critical issue of evaluation protocols in EEG emotion recognition, emphasizing that reported accuracy is influenced by the entire evaluation process rather than solely the classifier used. The authors propose a structured methodology that separates the target quantity, development procedure, and reporting rules. They utilize an archived dynamical graph convolutional neural network (DGCNN) pathway on the SEED and SEED-IV datasets to illustrate their approach. The study reveals that subject-dependent evaluations yield high accuracies, but significant gaps exist when evaluating entirely held-out participants, indicating that current methodologies may not adequately account for subject identity and other factors. The authors introduce a three-part EEG protocol record to clarify the evaluation process and propose an auditable checklist for EEG reporting. The findings suggest that different evaluation settings should be treated as answers to distinct questions, highlighting the need for standardized practices in EEG emotion recognition research.
Methodology
The authors employed a dynamical graph convolutional neural network (DGCNN) on the SEED and SEED-IV datasets, implementing a structured evaluation protocol that includes subject-dependent, subject-disjoint, and cross-session evaluations. They analyzed checkpoint selection based on repeated test-set evaluations and compared different evaluation settings to assess performance consistency.
Results
The study found that the SEED dataset's accuracy was within 1.47 percentage points of the public reference value, while the SEED-IV dataset showed a 3.40-point unresolved difference. Checkpoint selection improved mean window accuracy significantly. However, the accuracy for entirely held-out participants was notably lower, with 0.5348 for SEED and 0.3954 for SEED-IV, indicating challenges in generalization across subjects.
Implications
The findings underscore the necessity for standardized evaluation protocols in EEG emotion recognition to ensure reproducibility and reliability of results. This work can inform future research practices and improve the robustness of models developed for emotion recognition using EEG data.
High-Order Markov Blanket Discovery via a k-Order Relaxation of the Faithfulness Assumption
Graph Learning
Theory
- Proposes a k-order relaxation of the faithfulness assumption for Markov blanket discovery.
- Introduces the k-order Markov blanket (kOMB) algorithm for identifying higher-order dependencies.
- Demonstrates that kOMB can recover Markov blankets under both true and empirical violations of faithfulness.
- Outperforms existing methods on benchmark datasets, emphasizing the need for higher-order relationship exploration.
Read more
High-Order Markov Blanket Discovery via a k-Order Relaxation of the Faithfulness Assumption
Summary
This paper addresses the challenge of discovering the Markov blanket (MB) of a variable from data, which is crucial for applications in Bayesian networks, causal discovery, and feature selection. Traditional methods often rely on the faithfulness assumption, which posits that conditional independencies in the distribution correspond to separations in the graphical structure. However, this assumption can fail in the presence of higher-order dependencies, such as XOR and parity relations, leading to spurious dependencies in finite samples. To overcome this limitation, the authors propose a k-order relaxation of the faithfulness assumption, which allows for the identification of parity-type relationships among k+2 variables. They introduce a proof-of-concept algorithm called k-order Markov blanket (kOMB) that leverages this relaxation for MB discovery. The empirical results demonstrate that kOMB effectively recovers the MB under both true and empirical violations of faithfulness, outperforming existing constraint-based methods on benchmark datasets. This work highlights the importance of considering higher-order relationships in MB discovery, paving the way for more robust algorithms in graphical model learning.
Methodology
The authors developed a k-order relaxation of the faithfulness assumption to capture relationships among k+2 variables. They implemented a proof-of-concept algorithm, kOMB, which utilizes this relaxation for Markov blanket discovery. The methodology includes theoretical proofs and empirical evaluations on benchmark datasets to validate the effectiveness of kOMB.
Results
The kOMB algorithm successfully recovered the Markov blanket of a variable even in scenarios where traditional methods failed due to violations of the faithfulness assumption. Empirical evaluations showed that kOMB outperformed existing constraint-based methods, particularly in cases involving higher-order dependencies among variables.
Implications
The findings suggest that incorporating higher-order dependencies into Markov blanket discovery can lead to more accurate models in Bayesian networks and causal inference. This approach could enhance feature selection processes and improve the robustness of machine learning algorithms in various applications.
Semi-Supervised Learning for Molecular Graphs via Ensemble Consensus
Graph Learning
Theory
Efficient ML
- Introduces a novel SSL method based on ensemble consensus for molecular graphs.
- Demonstrates significant improvements in predictive accuracy across diverse datasets.
- Individual models trained with ensemble consensus outperform traditional supervised ensembles.
- Reduces calibration error and enhances model robustness.
Read more
Semi-Supervised Learning for Molecular Graphs via Ensemble Consensus
Summary
This paper addresses the challenges of semi-supervised learning (SSL) in molecular sciences, where acquiring labeled data is costly and time-consuming, while large amounts of unlabeled data are available. The authors propose a novel SSL method that utilizes ensemble consensus instead of traditional label-preserving augmentations, which are difficult to design for molecular graphs due to the sensitivity of molecular properties to structural changes. The proposed method trains an ensemble of models where each member learns from labeled data using a standard supervised loss and from unlabeled data through a consensus loss that promotes agreement among the ensemble members. This approach not only enhances predictive accuracy across various molecular datasets and tasks but also improves model robustness and reduces calibration error. The authors provide theoretical motivation for their method, demonstrating that it effectively eliminates the performance gap between individual ensemble members and the full ensemble, with individual models outperforming traditional supervised ensembles in most cases.
Methodology
The authors propose an SSL framework that trains an ensemble of models, where each model learns from labeled data using a supervised loss and from unlabeled data using a consensus loss. This consensus loss encourages agreement among the predictions of the ensemble members, thus leveraging the collective knowledge of the ensemble without requiring explicit data augmentations.
Results
The proposed ensemble consensus method consistently improves predictive accuracy across various molecular datasets and tasks. It effectively eliminates the performance gap between individual ensemble members and the full ensemble, with individual models often outperforming traditional supervised ensembles. Additionally, the method shows a reduction in calibration error compared to other SSL approaches.
Implications
The findings suggest that ensemble consensus can be a powerful approach for SSL in molecular sciences, potentially accelerating the discovery of new molecules and materials by making better use of available unlabeled data. This method could be applied to other domains facing similar challenges with labeled data scarcity.
Information Bottleneck Learning for Faithful Time Series Forecasting Explanations
Time Series
Interpretability
- IB-Forecast provides a self-interpretable framework for multivariate time-series forecasting.
- The model decomposes forecasts into a learned periodic profile and a gated deviation readout.
- It employs an information bottleneck to ensure that only relevant historical data influences predictions.
- IB-Forecast achieves competitive accuracy while requiring only 14-20% of observations for low-error predictions.
Read more
Information Bottleneck Learning for Faithful Time Series Forecasting Explanations
Summary
The paper introduces IB-Forecast, a novel multivariate time-series forecasting framework that emphasizes both predictive accuracy and interpretability. As decision-making increasingly relies on forecasts in critical sectors like energy and healthcare, understanding the historical data behind predictions is essential. Existing interpretable forecasting models often fail to guarantee that their internal structures accurately reflect the evidence driving predictions. In contrast, IB-Forecast integrates an information bottleneck approach to ensure that explanations are faithful to the underlying data. The framework decomposes forecasts into a learned periodic component and a residual component, utilizing explainable masks over input tokens. This allows users to control explanation sparsity while maintaining competitive forecasting accuracy. The authors demonstrate that IB-Forecast achieves comparable forecasting performance to leading black-box models, while providing high-fidelity explanations with significantly fewer observations. The framework's design ensures that only the most relevant historical data influences predictions, thus enhancing both stability and generalization.
Methodology
IB-Forecast utilizes a self-interpretable architecture that decomposes each prediction into a learned periodic profile and a gated deviation readout. The model employs an information bottleneck to balance the retention of forecasting-relevant information while limiting unnecessary historical details. This is achieved through a binary mask that selects a small subset of historical inputs, allowing for end-to-end optimization and direct integration of explanations within the forecasting computation.
Results
Extensive experiments show that IB-Forecast matches the forecasting error of leading black-box models while providing faithful explanations at no additional inference cost. Under matched sparsity budgets, the model consistently outperforms gradient-based, occlusion-based, and optimization-based baselines across various datasets, demonstrating high explanation fidelity with only a fraction of the historical data.
Implications
The development of IB-Forecast has significant implications for fields where time-series forecasting is critical, such as energy management, transportation, and healthcare. By providing interpretable and faithful explanations, the framework can enhance trust in automated decision-making systems and facilitate better understanding of the factors influencing predictions.
Weak-to-Strong On-Policy Distillation
NLP
Large Language Models
Reinforcement Learning
- W2S-OPD allows strong models to improve using weaker models as teachers.
- The framework constructs a proxy teacher from a contrast pair of weak models.
- W2S-OPD consistently outperforms traditional on-policy distillation methods.
- The method enables students to surpass their domain teachers in performance.
Read more
Weak-to-Strong On-Policy Distillation
Summary
The paper introduces Weak-to-Strong On-Policy Distillation (W2S-OPD), a novel framework designed to enhance the performance of a strong student model by distilling knowledge from multiple weaker models. Traditional on-policy distillation methods assume that the teacher model is at least as capable as the student, which limits their effectiveness when no larger teacher exists or when training multiple domain experts is computationally expensive. W2S-OPD addresses these challenges by constructing a proxy teacher in logit space from a contrast pair of weak models, allowing the student to learn from the isolated capability direction without being constrained by the weaker teacher's distribution. The authors propose three methods for creating the proxy teacher: using a post-RL expert against its pre-RL initialization, contrasting larger and smaller base models, and conditioning a single base model on correct and incorrect hints. The framework is evaluated across four math reasoning and three code generation benchmarks, demonstrating that W2S-OPD consistently outperforms traditional OPD methods and enables the student to exceed the performance of its domain teacher.
Methodology
W2S-OPD constructs a proxy teacher from two weak models (a positive and a negative model) to isolate capability directions. The student model minimizes the per-token reverse KL divergence on its own rollouts to distill knowledge from this proxy teacher. Three contrast methods are employed: pre-RL vs. post-RL models, larger vs. smaller base models, and a single base model with correct and incorrect hints.
Results
W2S-OPD outperformed traditional OPD methods by significant margins, achieving improvements of 11.4% and 12.0% in math reasoning tasks. The student model was able to exceed the performance of the domain expert it learned from, demonstrating the effectiveness of the weak-to-strong distillation approach.
Implications
The findings suggest that W2S-OPD can be a cost-effective strategy for enhancing the capabilities of large language models, particularly when stronger teacher models are not available. This approach could lead to more efficient training processes in various applications of NLP and machine learning.
PlatformBid: An Auto-Bidding Benchmark from a Unified Advertising Platform's Perspective
Reinforcement Learning
Generative Models
Optimization
- Introduction of PlatformBid as a benchmark for auto-bidding from a unified platform perspective.
- Definition of three competitive settings: homogeneous, heterogeneous, and promotional competition.
- Evaluation of existing auto-bidding methods alongside the introduction of a novel method, BidFlow.
- Demonstration of BidFlow's effectiveness in dynamic environments and its practical deployment.
Read more
PlatformBid: An Auto-Bidding Benchmark from a Unified Advertising Platform's Perspective
Summary
The paper introduces PlatformBid, the first comprehensive benchmark for auto-bidding algorithms from a unified advertising platform's perspective, addressing the integration of Supply Side Platforms (SSP), Demand Side Platforms (DSP), and Ad Exchanges. Traditional auto-bidding methods primarily focus on maximizing advertiser conversions through DSP-centric approaches. However, with the rise of large ad platforms that consolidate these functions, there is a need for algorithms that also optimize platform revenue. The authors define three competitive settings within PlatformBid: homogeneous competition, heterogeneous competition, and promotional competition, which reflect real-world bidding scenarios. They evaluate various existing auto-bidding methods, including classical control methods, reinforcement learning (RL)-based methods, and generative approaches. Additionally, they propose a novel method called BidFlow, which utilizes flow-matching techniques to adapt to dynamic competitive environments. Experimental results indicate that BidFlow outperforms existing methods in the defined settings, and online experiments on Kuaishou show a 0.68% improvement in target cost, demonstrating the practical applicability of PlatformBid.
Methodology
The authors developed the PlatformBid benchmark by defining three competitive settings that simulate real-world bidding scenarios. They systematically evaluated existing auto-bidding methods, including classical control, reinforcement learning, and generative models, and proposed a new method called BidFlow based on flow-matching techniques to handle dynamic competition.
Results
The experimental results showed that BidFlow significantly outperformed existing auto-bidding methods across the defined settings. Additionally, online experiments conducted on Kuaishou demonstrated a 0.68% improvement in target cost, validating the effectiveness of the proposed benchmark and method.
Implications
PlatformBid provides a new framework for evaluating auto-bidding algorithms, encouraging further research in platform-centric strategies. The findings suggest that integrating SSP, DSP, and Ad Exchange functionalities can lead to improved revenue optimization for advertising platforms, influencing future algorithm development and deployment in the industry.
Early Verdicts, Better Budgets: Sequential Adaptive Rollout Allocation for Compute-Efficient RLVR
Reinforcement Learning
Large Language Models
Efficient ML
- SARA optimizes rollout allocation by leveraging early verdicts on prompt effectiveness.
- The method reduces the number of rollouts needed by 22% compared to dynamic sampling.
- SARA can be integrated into existing GRPO-style pipelines without additional prediction rollouts.
- The approach is validated through empirical results on mathematical reasoning and planning tasks using large language models.
Read more
Early Verdicts, Better Budgets: Sequential Adaptive Rollout Allocation for Compute-Efficient RLVR
Summary
This paper addresses the inefficiencies in Reinforcement Learning with Verifiable Rewards (RLVR) caused by the generation of saturated groups during rollout, which yield no policy-gradient signal. The authors introduce SARA (Sequential Adaptive Rollout Allocation), a novel method that optimizes the allocation of rollouts based on early verdicts of group effectiveness. SARA employs a Beta posterior to estimate the success rate of prompts and utilizes a two-threshold stopping rule to commit to effective groups and abandon saturated ones early. This approach allows for the reallocation of freed budget to new prompts, significantly reducing the number of rollouts needed. The authors demonstrate that SARA achieves comparable performance to existing methods while using 22% fewer rollouts, and when combined with dynamic sampling, it surpasses existing accuracy benchmarks with a substantial reduction in computational cost.
Methodology
The authors recast the rollout allocation problem as a budget-constrained sequential allocation problem. SARA maintains a Beta posterior over the success rates of prompts, evaluates their effectiveness using a closed-form predictor, and applies a two-threshold rule for decision-making. This allows for early abandonment of ineffective groups and reallocation of resources to new prompts.
Results
SARA matches the performance of dynamic sampling while utilizing 22% fewer rollouts. When combined with dynamic sampling, SARA achieves the highest accuracy with a reduction of 67% in rollouts, demonstrating its efficiency in resource allocation.
Implications
The findings suggest that SARA can significantly improve the efficiency of RLVR processes, making it a valuable tool for applications requiring large-scale rollout generation, such as mathematical problem solving and multi-step planning in large language models.
The Role of Causality in Algorithmic Recourse
Theory
Optimization
Interpretability
- Existing algorithmic recourse methods often ignore the causal impact of suggested changes, leading to strategic manipulation.
- The proposed causal performative recourse framework models the interaction between agents and the learning algorithm using structural causal models.
- The authors identify conditions for stable and optimal recourse solutions that can be computed efficiently.
- Empirical results show that causal recourse reduces incentives for gaming and enhances predictive accuracy.
Read more
The Role of Causality in Algorithmic Recourse
Summary
This paper addresses the challenges of algorithmic recourse in high-stakes classification settings, where individuals seek actionable changes to improve their predicted outcomes. The authors highlight that existing recourse methods often fail to consider whether the suggested feature changes lead to actual improvements in an individual's true qualifications, potentially incentivizing strategic behavior that undermines predictive accuracy. To tackle this issue, the authors propose a causal performative recourse framework that models the interactions between agents and the learning algorithm through a structural causal model. This framework allows for the analysis of how recourse-induced actions affect both feature modifications and true labels, leading to a non-convex optimization problem. The authors characterize conditions for the existence of performatively stable solutions and provide algorithms for efficient computation. Empirical evaluations on semi-synthetic and real credit datasets demonstrate that incorporating causal responses into recourse policies significantly reduces incentives for gaming the system and improves predictive performance.
Methodology
The authors develop a causal performative recourse framework that models the interactions between agents and the learning algorithm through a structural causal model. They formulate the resulting learning problem as a performative prediction problem and provide algorithms with formal convergence guarantees for obtaining causal recourse.
Results
The empirical evaluation on semi-synthetic and real-world datasets demonstrates that the proposed causal recourse framework converges effectively, showing significant benefits over naive recourse policies. The results indicate that accounting for causal responses leads to reduced incentives for gaming and improved predictive performance.
Implications
This work has significant implications for high-stakes decision-making systems, such as lending and hiring, where algorithmic recourse can be designed to promote genuine improvements in qualifications rather than mere compliance with model predictions. By incorporating causality, practitioners can develop more effective and robust recourse policies that maintain predictive accuracy and reduce strategic manipulation.
Prox: Training-Free FFN Activation Sparsity via Approximate Intermediate-Channel Salience in LLMs
Large Language Models
Efficient ML
- Prox leverages the SwiGLU intermediate state for effective channel selection without incurring high computational costs.
- The framework achieves up to a 1.99× speedup in end-to-end decoding at 70% FFN sparsity.
- Prox is compatible with existing quantization techniques and sparse attention mechanisms.
- The method shows superior performance compared to existing training-free baselines across various models and sparsity levels.
Read more
Prox: Training-Free FFN Activation Sparsity via Approximate Intermediate-Channel Salience in LLMs
Summary
The paper introduces Prox, a novel two-stage training-free framework aimed at enhancing activation sparsity in feed-forward networks (FFNs) within large language models (LLMs). FFNs are identified as significant contributors to memory traffic and computation during inference, making them prime candidates for optimization. Existing training-free methods often lead to model quality degradation at high sparsity levels due to ineffective channel-selection strategies. The authors propose that the SwiGLU intermediate state can serve as an effective channel-selection signal, although its direct computation is costly. Prox circumvents this by using a two-stage approach: the first stage constructs a shared mask based on input sparsity and quantized proxy weights, while the second stage computes the selected channels accurately, allowing for efficient sparse execution across all three projections of the FFN. The framework is evaluated across ten LLMs from six model families, demonstrating significant improvements in decoding speed and maintaining model accuracy.
Methodology
Prox employs a two-stage framework for sparse SwiGLU FFNs. In Stage 1, it constructs a shared channel mask using input sparsity and quantized proxy weights. In Stage 2, it computes the selected channels exactly, applying the mask to all three projections of the FFN. Custom CUDA and Triton kernels are utilized to optimize computation and reduce weight accesses.
Results
Prox consistently outperforms state-of-the-art training-free methods, particularly at 60-70% FFN sparsity. The framework achieves up to a 1.99× speedup in end-to-end decoding while maintaining minimal accuracy loss across ten evaluated LLMs.
Implications
The findings suggest that Prox can significantly enhance the efficiency of LLM inference, making it particularly beneficial for deployment in resource-constrained environments. Its compatibility with quantization and sparse attention opens avenues for further optimization in large-scale models.
Expanding Data-Agnostic Pivotal Instances Selection Models with Proximity Trees and Ensemble Learning
Interpretability
- Introduction of PivotTree, a hierarchical model for selecting pivotal instances.
- The model enhances interpretability by mimicking human cognitive processes in decision-making.
- Incorporation of proximity and oblique trees, as well as ensemble methods, to improve performance.
- Demonstrated effectiveness across diverse data modalities, outperforming existing strategies.
Read more
Expanding Data-Agnostic Pivotal Instances Selection Models with Proximity Trees and Ensemble Learning
Summary
This paper presents a novel approach to enhance interpretability in machine learning through a hierarchical model called PivotTree, which selects pivotal instances based on their similarity to input data. The authors argue that human decision-making often relies on comparing new cases to representative examples, and thus, their model aims to mimic this cognitive process. PivotTree not only serves as a pivot selection technique but also functions as a standalone predictive model. It incorporates pairs of pivots and utilizes proximity and oblique trees, along with ensemble methods, to improve versatility and effectiveness. The model is designed to be data modality-agnostic, leveraging pre-trained networks for data transformation. Extensive experiments across various datasets, including tabular data, text, images, and time series, demonstrate that PivotTree outperforms existing instance selection strategies and achieves competitive results against state-of-the-art interpretable models while maintaining a minimal number of pivots. This work addresses the critical need for transparency in machine learning models, aligning with the objectives of Explainable AI (XAI).
Methodology
The authors developed PivotTree, which uses a similarity-based decision tree structure to classify new instances. It employs both single and pairs of pivots for decision-making, integrating oblique and proximity splits to refine predictions. The model is further enhanced through ensemble learning techniques, specifically using Random Forests and splitting stump forests to balance accuracy and efficiency.
Results
The experiments showed that PivotTree outperformed alternative instance selection strategies and achieved competitive results against state-of-the-art interpretable models. The model maintained a minimal number of pivots while effectively handling various data modalities, demonstrating its robustness and versatility.
Implications
The findings suggest that PivotTree can be applied in various decision-making scenarios across multiple domains, including healthcare, finance, and text analysis, where interpretability and transparency are crucial. This approach could significantly improve user trust and understanding of machine learning models.
The Convergence Behavior of Adam under Heavy-Tailed Noise
Optimization
Theory
- First convergence guarantees for Adam under heavy-tailed noise.
- Generalization of the online-to-nonconvex conversion framework to accommodate heavy-tailed noise.
- Establishment of a complete discounted regret analysis for the vector-form Adam update.
- Demonstration of suboptimal iteration complexity that persists even in bounded-variance cases.
Read more
The Convergence Behavior of Adam under Heavy-Tailed Noise
Summary
This paper presents the first convergence guarantees for the Adam optimizer in the presence of heavy-tailed stochastic noise, a scenario increasingly relevant in modern deep learning. While existing Adam variants have demonstrated optimal iteration complexity under bounded-variance conditions, their performance under heavy-tailed noise—where stochastic gradients have a bounded p-th central moment for p in (1, 2]—has not been thoroughly investigated. The author generalizes the online-to-nonconvex conversion framework to include heavy-tailed martingale-difference noise, allowing for a comprehensive discounted regret analysis of the plain vector-form Adam update without imposing restrictive parameter couplings. The findings indicate that Adam can converge to (ρ, ϵ)-stationary points under heavy-tailed noise, albeit with suboptimal iteration complexity that is p-dependent. However, when the domain radius is known and utilized to control the output of the online learner, the convergence rate can be improved to match optimal complexity. This work enhances the theoretical understanding of Adam's robustness and limitations in heavy-tailed environments, providing insights that could influence future optimization strategies in deep learning.
Methodology
The author employs a generalized online-to-nonconvex conversion framework to analyze the Adam optimizer under heavy-tailed noise. This involves developing a discounted regret analysis that does not rely on restrictive parameter couplings and treats the numerical stabilizer as a fixed constant.
Results
The analysis shows that Adam converges to (ρ, ϵ)-stationary points under heavy-tailed noise, with a p-dependent and suboptimal iteration complexity. However, when the domain radius is known, the convergence rate can be optimized to match the best-known complexity.
Implications
The findings provide new theoretical insights into the behavior of Adam in challenging noise conditions, which could inform the design of more robust optimization algorithms for deep learning applications, particularly in scenarios where heavy-tailed noise is prevalent.
DHRCL: Training Code LLMs with Dense Hierarchical Rewards and Curriculum Learning
Large Language Models
Reinforcement Learning
Optimization
- Introduction of a three-stage hierarchical reward curriculum that coordinates multiple feedback types.
- Automatic stage progression based on validation trends, eliminating the need for fixed thresholds.
- Implementation of stage-aware token credit redistribution to optimize learning efficiency.
- Demonstrated consistent performance improvements across different model sizes.
Read more
DHRCL: Training Code LLMs with Dense Hierarchical Rewards and Curriculum Learning
Summary
The paper presents DHRCL, a novel reinforcement learning framework designed for training code-oriented large language models (LLMs). Traditional methods often rely on sparse rewards or static combinations of dense signals, which fail to capture the progressive dependencies among various programming capabilities such as syntax validity, execution success, and functional correctness. DHRCL addresses these limitations by introducing a Dense Hierarchical Rewards system that decomposes feedback into distinct categories: syntax validation, execution success, unit-test pass rate, and AST-based structural similarity. This feedback is organized into a three-stage curriculum—Syntax, Execution, Pass & Structural—where the duration of each stage is determined automatically based on recent validation trends. Additionally, DHRCL incorporates a stage-aware probability-based token credit redistribution mechanism that emphasizes established token patterns during early stages and allocates credit or blame according to the uncertainty of token decisions in later stages. Experimental results demonstrate that DHRCL outperforms various baseline methods across different model capacities, indicating its robustness and effectiveness in enhancing code generation capabilities.
Methodology
DHRCL employs a reinforcement learning framework that integrates a Dense Hierarchical Reward system with a Curriculum Learning approach. The rewards are decomposed into distinct categories, and a three-stage curriculum is established to guide the training process. The duration of each stage is dynamically adjusted based on validation trends, and a novel token credit redistribution mechanism is applied to optimize the learning process across different stages.
Results
The experiments conducted under a unified Qwen3-8B and KodCode protocol show that DHRCL consistently outperforms baseline methods, including binary, pass-rate, reward-model-based, and verifiable dense-reward approaches. The advantages of DHRCL are maintained across various model capacities (Qwen3-4B, Qwen3-8B, and Qwen3-14B), highlighting its effectiveness in improving code generation performance.
Implications
DHRCL has significant implications for the development of more effective code generation models, potentially enhancing the capabilities of LLMs in programming tasks. Its structured approach to reward design and curriculum learning could lead to advancements in automated code generation, debugging, and software development tools.
ClawTrack: Towards Trace-Level Evaluation and Improvement of Real-World Autonomous Agents
NLP
Large Language Models
Robotics
- ClawTrack introduces a dual-assessment benchmark that evaluates both task outcomes and reasoning processes.
- The Process Grader assesses reasoning quality across four dimensions, enabling detailed attribution of successes and failures.
- Result verification is identified as a critical bottleneck in the reasoning process.
- The framework is robust to evaluator choice, demonstrating consistent improvements in model performance.
Read more
ClawTrack: Towards Trace-Level Evaluation and Improvement of Real-World Autonomous Agents
Summary
The paper introduces ClawTrack, a novel dual-assessment benchmark designed to evaluate and improve the performance of autonomous agents, particularly those based on large language models (LLMs). Traditional evaluation methods primarily focus on the final outcomes of tasks, failing to provide insights into the reasoning processes that lead to those outcomes. ClawTrack addresses this gap by measuring both the Task Score (the outcome) and the Process Score (the reasoning process) across 320 tasks in 8 domains, utilizing 25+ deterministic mock services. The Process Grader evaluates each reasoning turn based on four dimensions: goal alignment, efficiency, information utilization, and result verification, supported by a comprehensive set of 12,541 task-specific rubric items. The evaluation of 21 models over more than 16,000 trials reveals that process scores can effectively attribute successes and failures to specific reasoning dimensions, highlighting the importance of result verification as a systematic bottleneck. The framework demonstrates robustness across different evaluator choices and shows that process-based trajectory filtering can lead to consistent improvements in model performance post-training. ClawTrack thus provides a structured approach to understanding and enhancing the capabilities of autonomous agents in complex workflows.
Methodology
The methodology involves the development of ClawTrack, which includes a comprehensive set of tasks and mock services for evaluating autonomous agents. The Process Grader uses multi-dimensional rubrics to score each reasoning turn, while an Outcome Grader assesses final results. The evaluation process encompasses over 16,000 trials across various models, allowing for a detailed analysis of reasoning processes.
Results
The results indicate that process scores can effectively distinguish between reliable reasoning and lucky successes, with the four evaluation dimensions being complementary. The study finds that result verification is a key area needing improvement, and that the framework can yield consistent post-training enhancements across different model scales.
Implications
The implications of this work are significant for the development and deployment of autonomous agents in real-world applications. By providing a structured evaluation framework, ClawTrack can help identify specific areas for improvement in agent reasoning, ultimately leading to more reliable and effective autonomous systems.
Actions Have Consequences: Detecting Outcome Performativity using Intervention Testing
Theory
- Introduction of Outcome Performativity A/B Detection (OPAB) for detecting Outcome Performativity.
- Derivation of sample complexity bounds for OPAB under different assumption classes.
- Empirical validation of OPAB's effectiveness in detecting Outcome Performativity.
- Identification of regions of indistinguishability where detection may be challenging.
Read more
Actions Have Consequences: Detecting Outcome Performativity using Intervention Testing
Summary
This paper addresses the phenomenon of Outcome Performativity, where predictions made by machine learning models causally influence the outcomes they predict. The authors propose a novel method called Outcome Performativity A/B Detection (OPAB) to detect this phenomenon through intervention testing. OPAB assesses the dissimilarity in outcome distributions produced by different prediction groups to determine if Outcome Performativity is present. The paper derives sample complexity bounds for OPAB under various assumption classes of Outcome Performativity and empirically validates these bounds. The findings indicate that OPAB can effectively detect Outcome Performativity in numerous scenarios, although there are regions of indistinguishability where the number of interventions may be insufficient for detection. This work has significant implications for practitioners in fields where sample acquisition is limited or ethically challenging, allowing for the identification of Outcome Performativity before deploying predictive models. A case study on the Open Bandits dataset demonstrates the efficacy of OPAB, and the authors suggest future research directions to further explore this area.
Methodology
The authors utilize intervention testing inspired by Randomized Controlled Trials (RCTs) to assess the dissimilarity in outcome distributions from different prediction groups. They derive sample complexity bounds for OPAB under various assumptions of Outcome Performativity and conduct empirical evaluations to validate their theoretical findings.
Results
The results show that OPAB can successfully detect Outcome Performativity in many cases, with established sample complexity bounds. However, the study also identifies regions where the required number of interventions is too high for practical detection, highlighting limitations in certain scenarios.
Implications
The findings suggest that OPAB can be a valuable tool for practitioners to detect Outcome Performativity before deploying machine learning models, particularly in sensitive domains like healthcare and finance where sample acquisition can be costly or unethical. This proactive approach can help mitigate biases and performance decay in predictive models.
Hierarchical Spatio-Temporal Transformer for Coherent Emergency Department Forecasting
Time Series
- HierSTT is a novel hierarchical Transformer framework for multi-level ED demand forecasting.
- The model incorporates a coherence-aware loss function to ensure consistency across hierarchical levels.
- A new nationwide Portuguese ED dataset is introduced, reflecting real-world data heterogeneity.
- HierSTT outperforms traditional statistical methods and non-hierarchical deep learning approaches.
Read more
Hierarchical Spatio-Temporal Transformer for Coherent Emergency Department Forecasting
Summary
The paper addresses the challenge of forecasting patient demand in Emergency Departments (EDs), which are critical components of healthcare systems facing unpredictable patient flows. Traditional forecasting methods often treat ED demand as an isolated problem at a single level, leading to incoherent predictions across different hierarchical levels (hospital, regional, national). To overcome this limitation, the authors propose HierSTT, a hierarchical Transformer-based framework that simultaneously predicts demand at multiple levels in a coherent manner. The model integrates a Temporal Fusion Transformer for national dynamics and spatio-temporal Transformer modules for regional and hospital-level forecasts, conditioned on higher-level predictions. A coherence-aware loss function is introduced to penalize inconsistencies between forecasts at different levels. The authors also present a new nationwide dataset from Portugal, encompassing 81 hospitals across 5 regions, which includes heterogeneous covariates. Experimental results demonstrate that HierSTT significantly outperforms existing non-hierarchical deep learning models and classical reconciliation methods, achieving a 32% reduction in average WAPE (Weighted Absolute Percentage Error) and producing coherent predictions across levels.
Methodology
The HierSTT framework employs a hierarchical structure with a Temporal Fusion Transformer for national-level demand and spatio-temporal Transformer encoder-decoder modules for regional and hospital-level predictions. The model is trained with a coherence-aware loss function that penalizes inconsistencies between forecasts at different levels, allowing for joint modeling of demand across the hierarchical structure.
Results
HierSTT achieved a 32% reduction in average WAPE compared to the best non-hierarchical deep learning baseline and outperformed all classical hierarchical reconciliation methods. The model produced near-coherent predictions across hospital, regional, and national levels, demonstrating its effectiveness in addressing the complexities of ED demand forecasting.
Implications
The proposed framework can significantly enhance ED planning and resource allocation by providing coherent and accurate demand forecasts. This can lead to improved patient care, optimized staffing, and better management of healthcare resources at multiple administrative levels.
LoRA Scaffolded Policy Optimization (LSPO): A Sampling-Time Low-Rank Scaffold for Recovering Reinforcement-Learning Gradient on Zero-Reward Cliff Prompts
Reinforcement Learning
Large Language Models
Optimization
- LSPO recovers lost gradients on cliff prompts in RLVR by using a low-rank adapter during training.
- The method integrates supervised learning with reinforcement learning without modifying the loss function.
- Empirical results show LSPO outperforms DAPO on all tested benchmarks, with significant improvements in performance metrics.
- The approach effectively converts a substantial portion of cliff prompts into usable learning signals.
Read more
LoRA Scaffolded Policy Optimization (LSPO): A Sampling-Time Low-Rank Scaffold for Recovering Reinforcement-Learning Gradient on Zero-Reward Cliff Prompts
Summary
The paper introduces LoRA Scaffolded Policy Optimization (LSPO), a novel approach designed to address a critical limitation in reinforcement learning from verifiable rewards (RLVR) when dealing with 'cliff' prompts—situations where all sampled rollouts fail, resulting in zero gradient updates. LSPO employs a sampling-time mechanism that detects these cliff prompts and utilizes a low-rank adapter to fit ground-truth solutions briefly. This adapter is used to re-roll the cliff prompts, generating successful completions that are then integrated back into the RL training batch with an importance-sampling correction. The method ensures that the base model receives only the RL gradient while the adapter is discarded after training, leading to a base-only model. The empirical evaluation on the DeepMath-103K dataset demonstrates that LSPO outperforms the DAPO baseline across multiple benchmarks, effectively converting approximately 43% of cliff prompts into usable policy-gradient signals. This approach not only enhances the learning efficiency of RL models but also addresses the structural blind spot inherent in traditional group-relative policy optimization methods.
Methodology
LSPO employs a two-optimizer framework where a low-rank adapter is trained via supervised learning on cliff prompts while the base model is updated through reinforcement learning. The adapter is used temporarily to generate successful rollouts, which are then spliced back into the training batch, allowing the base model to learn from previously unlearnable prompts.
Results
On the DeepMath-103K dataset, LSPO achieved superior performance compared to the DAPO baseline across all 16 benchmark cells, with notable gains of up to +10.7 points on AIME24/pass@4 and +3.8 points on average across all metrics. The method also demonstrated a conversion rate of approximately 43% for cliff prompts into usable gradient signals.
Implications
The LSPO framework has the potential to significantly improve the training efficiency of reinforcement learning models, particularly in scenarios where traditional methods struggle to provide learning signals. This could lead to advancements in mathematical reasoning tasks and other applications relying on RLVR.
Regularizing modality contribution drift in multimodal continual learning
Multimodal
- Identification and formalization of modality contribution drift (MCD) in MMCL.
- Introduction of Continual Modality Contribution Drift Regularization (CMCDR) to preserve modality contributions.
- CMCDR supports both replay-based and replay-free learning settings.
- Extensive experiments validate the effectiveness of CMCDR across various benchmarks.
Read more
Regularizing modality contribution drift in multimodal continual learning
Summary
This paper addresses the challenge of modality contribution drift (MCD) in multimodal continual learning (MMCL), where the relative contributions of different modalities can change as new tasks are learned. The authors introduce a novel metric, the MCD score, which quantifies the stability of modality contributions across incremental tasks. They propose Continual Modality Contribution Drift Regularization (CMCDR), a method designed to preserve the contribution structures of previously learned tasks while adapting to new modality dependencies. CMCDR is versatile, offering both replay-based and replay-free versions to accommodate different MMCL scenarios. The paper presents extensive experiments demonstrating that CMCDR effectively mitigates MCD and enhances performance in multimodal class-incremental learning and continual visual question answering tasks, highlighting the importance of maintaining stable modality contributions in dynamic learning environments.
Methodology
The authors developed an intervention-based diagnostic estimator to assess modality contributions by manipulating modality subsets and measuring the resulting predictive benefits. They defined the MCD score to quantify changes in contribution strength and reliance across tasks. CMCDR was then proposed to regularize these contributions, with implementations for both scenarios where old exemplars are available (replay-based) and where they are not (replay-free).
Results
The experiments conducted on multimodal class-incremental learning and continual visual question answering showed that CMCDR consistently reduced modality contribution drift and improved model performance across different MMCL settings and benchmarks, demonstrating its effectiveness in maintaining stable modality contributions.
Implications
The findings suggest that addressing modality contribution drift is crucial for improving the robustness of multimodal learning systems. CMCDR can be applied in various real-world applications where models need to learn from dynamic multimodal data without forgetting previously acquired knowledge, such as in robotics, autonomous driving, and interactive AI systems.
Train Small, Deploy Large: Zero-Shot GNN Transfer Through Geometric Renormalization
Graph Learning
- Introduction of a zero-shot transfer protocol for GNNs using geometric renormalization.
- Empirical evidence that GR maintains GNN performance under significant graph compression.
- Compatibility of node representations learned at different GR scales.
- Preservation of functional training trajectories across scales.
Read more
Train Small, Deploy Large: Zero-Shot GNN Transfer Through Geometric Renormalization
Summary
This paper addresses the challenge of transferring Graph Neural Network (GNN) models trained on smaller, coarse-grained versions of graphs to their full-resolution counterparts without retraining. The authors introduce a zero-shot transfer protocol utilizing geometric renormalization (GR), which preserves the structural and statistical properties of the original graph during the coarsening process. The study demonstrates that GNNs trained on GR-reduced graphs can maintain predictive performance when applied to larger graphs, significantly reducing training costs. The authors provide empirical evidence showing that the learned representations and predictive trajectories remain aligned across different scales, suggesting that structural similarity is more critical than graph size for GNN transferability. This work opens avenues for developing scale-equivariant graph architectures and provides a GPU implementation, cuMercator, for efficient hyperbolic embedding computations.
Methodology
The authors employed a geometric renormalization technique to create coarse-grained versions of graphs, on which GNNs were trained. The weights from these models were then directly applied to the original graphs for zero-shot inference. The study evaluated transferability through metrics such as node classification accuracy, centered kernel alignment, and Jensen-Shannon divergence.
Results
The results indicated that GNNs trained on GR-reduced graphs could achieve comparable predictive performance on full-resolution graphs. The study found that the learned representations were compatible across scales, and the functional training trajectories remained aligned, demonstrating the effectiveness of the proposed method.
Implications
This research suggests that GNNs can be trained more efficiently on smaller graph representations while still being effective on larger graphs, potentially leading to advancements in scalable GNN architectures. The findings may have applications in various fields where large graph data is prevalent, such as social network analysis, biological networks, and recommendation systems.
Event-Structured Physics-Informed Neural Networks for Differentiable Critical Clearing Boundaries
Theory
Optimization
Time Series
- Introduction of ES-PINN for transient stability assessment in power systems.
- Differentiable approximation of critical clearing time (CCT) boundary.
- Improved accuracy in trajectory and stability-boundary estimation over traditional methods.
- Local sensitivity analysis and error estimation linked to physics residuals.
Read more
Event-Structured Physics-Informed Neural Networks for Differentiable Critical Clearing Boundaries
Summary
This paper presents an innovative approach to transient stability assessment in power systems using an Event-Structured Physics-Informed Neural Network (ES-PINN). The critical clearing time (CCT) is a vital metric that determines the maximum time available to clear a fault before losing synchronism. Traditional methods for estimating CCT involve extensive simulations across various fault severities and clearing times, which can be computationally burdensome. The proposed ES-PINN framework addresses this challenge by structuring the neural network to align with the physical dynamics of pre-fault, fault-on, and post-clearing phases. By enforcing exact state chaining across event interfaces, the ES-PINN provides a differentiable approximation of the CCT boundary, allowing for accurate boundary extraction and local sensitivity analysis. The authors validate their approach through experiments on IEEE 9-, 14-, and 30-bus systems, demonstrating that ES-PINN significantly outperforms traditional neural-surrogate models in terms of trajectory and stability-boundary accuracy, even under various fault conditions and clearing configurations.
Methodology
The ES-PINN framework is designed to represent the dynamics of power systems during different phases of fault-clearing events. It employs a structured approach that separates the pre-fault, fault-on, and post-clearing dynamics, ensuring that the transitions between these phases are smooth and differentiable. The model is trained using physics-informed techniques that incorporate governing equations, allowing it to learn from both data and physical principles. The framework also includes a mechanism for local sensitivity analysis and error estimation, linking physics residuals to trajectory errors.
Results
The experiments conducted on IEEE 9-, 14-, and 30-bus systems demonstrate that the ES-PINN consistently improves the accuracy of held-out trajectories and stability boundaries compared to matched neural-surrogate baselines. The results indicate that the proposed framework effectively captures the dynamics of mechanical and electrical contingencies, providing reliable CCT estimates across multiple clearing configurations.
Implications
The ES-PINN framework has significant implications for the field of power system stability assessment, offering a more efficient and accurate method for estimating critical clearing times. This could lead to improved operational reliability in power systems, reducing the risk of cascading outages and enhancing the overall stability of electrical grids. Additionally, the methodology could be adapted for other applications involving dynamic systems where event-driven transitions are critical.
Single-Beat Cuffless Blood Pressure Estimation Using Ear-PPG and ECG with a Lightweight Hybrid Learning Framework
Multimodal
Time Series
Efficient ML
- Introduction of a lightweight multi-modal wearable system for continuous BP estimation.
- Demonstration that BP-related information can be effectively extracted from single PPG beats.
- Development of a hybrid learning architecture combining CNN and LightGBM for efficient BP estimation.
- Achieved significant reduction in mean absolute error compared to traditional models.
Read more
Single-Beat Cuffless Blood Pressure Estimation Using Ear-PPG and ECG with a Lightweight Hybrid Learning Framework
Summary
This paper addresses the challenges of continuous cuffless blood pressure (BP) monitoring, which is hindered by motion artifacts and physiological variability. Traditional methods often rely on multi-second windows for stabilization, which can be problematic in real-world scenarios. The authors propose a novel lightweight multi-modal wearable framework that utilizes synchronized electrocardiography (ECG) and ear-clip photoplethysmography (PPG) sensors, each equipped with a 6-axis inertial measurement unit (IMU) to provide motion context. The framework employs a hybrid learning architecture combining a one-dimensional convolutional neural network (CNN) for extracting beat-level features from PPG data and a LightGBM regressor that incorporates additional physiological features. This approach allows for single-beat BP estimation without the need for extended temporal context, significantly improving robustness against motion artifacts. The system was evaluated through a multi-phase stress protocol and the PulseDB dataset, achieving mean absolute errors of 4.02 mmHg for systolic BP and 1.79 mmHg for diastolic BP, representing a 28.2% reduction in mean absolute error compared to baseline models. The findings suggest that this framework is suitable for wearable BP monitoring under practical constraints, enabling more effective hypertension management.
Methodology
The methodology involves a multi-modal sensing platform that integrates ECG and ear-PPG sensors with IMUs to mitigate motion artifacts. A one-dimensional CNN extracts beat-level features from PPG data, which are then fused with physiological features like pulse transit time (PTT) and heart rate variability (HRV) using a LightGBM regressor, allowing for efficient single-beat BP estimation.
Results
The proposed model achieved mean absolute errors of 4.02 ± 0.21 mmHg for systolic BP and 1.79 ± 0.05 mmHg for diastolic BP across 30 independent runs, resulting in a 28.2% reduction in combined mean absolute error compared to baseline models.
Implications
This research has significant implications for the development of wearable health monitoring devices, particularly for continuous blood pressure management in patients with hypertension. The lightweight and efficient nature of the framework makes it suitable for everyday use, potentially improving adherence to BP monitoring and management.
What Makes Graph Unified? Principles and Generative Sliding-Window Transformer for Graph Foundation Models
Graph Learning
- Introduces Graph Foundation Models (GFMs) for unified graph learning across diverse domains.
- Identifies the limitations of existing feature transformation methods in preserving semantics.
- Proposes four principles for effective cross-domain graph feature unification.
- Develops SliGFM, which employs topology-aware sliding-window encoding for feature transformation.
Read more
What Makes Graph Unified? Principles and Generative Sliding-Window Transformer for Graph Foundation Models
Summary
This paper introduces Graph Foundation Models (GFMs) as a novel approach for general-purpose graph learning, aiming to unify knowledge across diverse graph domains and tasks. The authors identify the challenges posed by heterogeneous node features, which vary significantly in dimensionality and semantics across datasets. They argue that existing methods, which often rely on dimensionality reduction techniques, fail to ensure that unified features maintain informative semantics necessary for effective cross-domain knowledge transfer. To address this, the authors propose four key principles for effective feature unification: formal uniformity, cross-domain transferability, information preservation, and backbone compatibility. Based on these principles, they develop SliGFM, a graph foundation model that utilizes topology-aware sliding-window feature encoding and generative reconstruction. SliGFM organizes feature dimensions by topological smoothness and employs a shared sliding-window encoder to transform heterogeneous features into a common space of ordered fixed-dimensional tokens. This approach allows for the capture of transferable relational patterns and encourages the preservation of original feature information. The model also adapts neighborhood aggregation based on graph-level structural statistics. Extensive experiments demonstrate SliGFM's effectiveness and transferability across various graph datasets, showcasing its potential to facilitate knowledge reuse in graph learning.
Methodology
The authors propose SliGFM, which organizes feature dimensions based on topological smoothness and utilizes a sliding-window feature encoder to transform heterogeneous node features into a common representation. The model incorporates generative reconstruction to preserve original feature information and adapts neighborhood aggregation based on graph-level structural statistics.
Results
The experiments conducted show that SliGFM effectively captures transferable relational patterns and maintains informative semantics across various graph datasets, outperforming existing methods in terms of transferability and feature unification.
Implications
The findings suggest that SliGFM can significantly improve the efficiency of graph learning by enabling knowledge reuse across different domains, potentially reducing the need for separate model training for each graph dataset.
Good Rankers, Bad Objectives: Bilinear Contrastive Critics under Expressive Policy Search
Reinforcement Learning
Theory
Optimization
- Bilinear contrastive critics can rank actions accurately within a support set but may fail in broader candidate selection.
- Norm drift and value decalibration are significant issues that lead to misordering of actions.
- Experiments reveal that contrastive critics incur regret during maximization despite strong retrieval performance.
- A value-calibrated scalar is essential for ensuring reliable action selection and restoring proper value ordering.
Read more
Good Rankers, Bad Objectives: Bilinear Contrastive Critics under Expressive Policy Search
Summary
This paper investigates the limitations of bilinear contrastive critics in reinforcement learning, particularly in the context of maximizing action selection based on learned value-like scores. The authors argue that while contrastive critics can effectively rank actions within a known support set, they may lead to suboptimal action selection when extended to broader candidate pools. The study identifies two main issues: norm drift, which inflates off-support values, and value decalibration, which results in weak or inverted value ordering despite strong compatibility in rankings. Through a series of experiments across four OGBench navigation tasks, the authors demonstrate that bilinear critics can misrank actions, leading to regret during maximization. They propose that a value-calibrated scalar is necessary for reliable action selection, as it restores proper ordering in the presence of norm drift and score saturation. The findings highlight the need for careful consideration of the objectives used in reinforcement learning, especially when transitioning from ranking to optimization.
Methodology
The authors conducted a theoretical analysis of bilinear contrastive critics and their performance in action ranking and selection. They performed experiments across four OGBench navigation tasks to evaluate the critics' ability to maintain value ordering and measure regret during candidate maximization. The study also involved a training/readout decomposition to trace the effects of the cosine training objective on action ordering.
Results
The experiments showed that bilinear critics often produced weak or inverted value ordering in the top score decile, leading to significant regret during action selection. The findings indicated that while contrastive critics could effectively rank actions based on compatibility, they failed to provide a reliable value calibration necessary for optimal action selection. The introduction of a Bellman-calibrated scalar successfully restored the correct ordering in the experiments.
Implications
The results suggest that while contrastive critics are useful for ranking actions in navigation and manipulation tasks, their application in action selection requires careful calibration of the scoring mechanism. This has implications for the design of reinforcement learning algorithms, particularly those that rely on learned value functions for decision-making in complex environments.