last updated 2 h ago

Daily digest № 1 — Wednesday, June 10, 2026

The day in AI, distilled.

sort: relevance · 200 articles

2 h agoarXiv cs.AISafety

Dual-Branch Gated Fusion for Open-Set Audio Deepfake Source Tracing

The paper presents a dual-branch gated fusion framework for audio deepfake source tracing, integrating XLSR-53 with a 66-dimensional CORES descriptor that captures diverse synthesis artifacts. The model demonstrates a 97.6% in-domain accuracy and a 4.9% equal error rate on the MLAAD benchmark, significantly outperforming the Interspeech 2025 baseline with an 83.5% relative reduction in false positive rate at 95% recall. This work is crucial for practitioners as it addresses the limitations of closed-set models in handling unseen synthesizers and enhances generalization under distribution shifts.

audiodeepfakesource tracing

2 h agoarXiv cs.AITraining

Fisher-Guided Progressive Parameter Selection for Adaptive Fine-Tuning

The article presents FisherAdapTune, a parameter-efficient fine-tuning framework that utilizes Fisher-guided criteria to dynamically select parameter groups based on their curvature contributions over time. By employing a scale-invariant Jensen-Shannon distance to track Fisher distribution changes, the method aims to stabilize certain parameters while adapting others, enhancing performance on segmentation tasks with improved in-distribution and zero-shot transfer results. This approach is significant for practitioners as it offers a more adaptive and efficient strategy for fine-tuning pretrained models, potentially leading to better generalization in various applications.

fine-tuningadaptivefisher

2 h agoarXiv cs.AIResearch

The Model Says Walk: How Surface Heuristics Override Implicit Constraints in LLM Reasoning

The paper introduces the Heuristic Override Benchmark (HOB), consisting of 500 instances across various heuristic and constraint families, to evaluate large language models (LLMs) on reasoning tasks. Analysis of six models, including Gemini 3.1 Pro, reveals that surface cues can significantly override implicit constraints, with no model exceeding 75% performance on strict evaluations; a minimal hint can improve performance by 15 percentage points, indicating a constraint-inference failure. This research highlights the critical need for practitioners to understand and mitigate heuristic biases in LLMs, as explicit goal decomposition and internal deliberation can enhance reasoning capabilities.

llmreasoningheuristics

2 h agoarXiv cs.AIResearch

Reasoning over Semantic IDs Enhances Generative Recommendation

The paper presents SIDReasoner, a two-stage framework designed to enhance generative recommendation by improving reasoning over Semantic IDs (SIDs) using pretrained LLMs. It employs multi-task training on a synthesized SID-centered corpus to strengthen SID-language alignment and utilizes outcome-driven reinforced optimization to guide reasoning trajectories without needing explicit annotations. This approach not only boosts recommendation accuracy but also enhances interpretability and cross-domain generalization, offering valuable insights for practitioners working with LLMs in recommendation systems.

recommendationsemantic idsreasoning

2 h agoarXiv cs.AIResearch

FinTradeBench: A Financial Reasoning Benchmark for LLMs

FinTradeBench is a newly introduced benchmark designed for evaluating financial reasoning in Large Language Models (LLMs), comprising 1,400 questions related to NASDAQ-100 companies. It categorizes questions into fundamentals-focused, trading-signal-focused, and hybrid types that require cross-signal reasoning, employing a calibration-then-scaling framework for reliability. The evaluation of 14 LLMs revealed a significant performance gap, particularly showing that while retrieval-augmented prompting enhances reasoning for textual fundamentals, it offers limited advantages for trading-signal reasoning, underscoring critical areas for improvement in numerical and time-series reasoning capabilities of LLMs.

financial reasoningbenchmarkllm

2 h agoarXiv cs.AIResearch

Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement

The paper presents a statistical framework for measuring citation visibility in generative search, emphasizing the non-deterministic nature of AI-powered answer engines. It conducts an empirical study across three platforms—Perplexity Search, OpenAI SearchGPT, and Google Gemini—utilizing repeated sampling to reveal that citation distributions exhibit substantial variability and follow a power-law form. The findings highlight the inadequacy of single-run visibility metrics, advocating for the inclusion of uncertainty estimates in performance reporting to better inform practitioners on the reliability of citation rankings in generative search applications.

generative searchvisibilityuncertainty

2 h agoarXiv cs.AITraining

MedFeat: Model-Aware and Explainability-Driven Feature Engineering with LLMs for Clinical Tabular Prediction

MedFeat is a novel feature engineering framework designed for clinical tabular prediction, integrating model-awareness and feature importance signals to enhance feature discovery in LLMs. The framework demonstrates a statistically significant average improvement of over 10% compared to state-of-the-art baselines across various clinical tasks, addressing challenges such as class imbalance and interpretability in healthcare data. This advancement is crucial for practitioners as it allows for more targeted and effective feature transformations, potentially leading to improved model performance in clinical applications.

feature engineeringllmclinical prediction

2 h agoarXiv cs.AIResearch

TaCarla: A comprehensive benchmarking dataset for end-to-end autonomous driving

The article announces the release of TaCarla, a comprehensive benchmarking dataset specifically designed for end-to-end autonomous driving, comprising over 2.85 million frames collected using the CARLA simulation environment. This dataset addresses gaps in existing benchmarks by integrating perception and planning tasks, supporting dynamic object detection, lane divider detection, and traffic light recognition, while also providing numerical rarity scores for better evaluation of model performance. TaCarla's diverse scenarios and closed-loop evaluation capabilities make it a significant resource for practitioners aiming to enhance the robustness and performance of autonomous driving systems.

autonomous drivingbenchmarkingdataset

2 h agoarXiv cs.AIInference

GRAU: Generic Reconfigurable Activation Unit Design for Neural Network Hardware Accelerators

The paper presents GRAU, a Generic Reconfigurable Activation Unit designed for neural network hardware accelerators, which utilizes piecewise linear fitting with segment slopes approximated by powers of two. This design significantly reduces lookup table (LUT) consumption by over 90% compared to traditional multi-threshold activation hardware, while supporting mixed-precision quantization and nonlinear functions like SiLU. GRAU's efficiency and scalability are critical for practitioners working with low-precision quantization in edge AI applications, allowing for more flexible and cost-effective neural network implementations.

neural networksactivationhardware

2 h agoarXiv cs.AITraining

Improving Topic Modeling by Distilling Soft Labels from Language Models

The paper presents a novel framework for topic modeling called Distilling Soft Labels (DSL), which leverages language models to enhance the training of topic models by utilizing contextual information from next token probabilities. DSL demonstrates significant improvements in topic coherence and assignment accuracy compared to traditional methods, as well as superior performance in a new retrieval-based metric for identifying semantically similar documents. This advancement is crucial for practitioners looking to integrate improved topic modeling techniques into applications that require enhanced semantic understanding and document retrieval.

topic modelinglanguage modelssoft labels

2 h agoarXiv cs.AIResearch

Exploring Accurate and Transparent Domain Adaptation in Predictive Healthcare via Concept-Grounded Orthogonal Inference

The paper introduces ExtraCare, a domain adaptation model designed for clinical event prediction from electronic health records (EHR), which addresses the limitations of traditional black-box models by providing transparent, interpretable outputs. ExtraCare decomposes patient representations into invariant and covariant components, enforcing orthogonality during training to enhance prediction accuracy while allowing for human-understandable explanations linked to medical concepts. Evaluated on two real-world EHR datasets, ExtraCare outperforms existing feature alignment methods, making it a significant advancement for practitioners seeking reliable and explainable AI solutions in healthcare.

domain adaptationhealthcareexplainability

2 h agoarXiv cs.AIResearch

RankLLM: Weighted Ranking of LLMs by Quantifying Question Difficulty

RankLLM is a newly proposed framework that quantifies question difficulty and model competency for evaluating large language models (LLMs), addressing the limitations of existing benchmarks that do not account for varying question difficulty. It employs a bidirectional score propagation mechanism, allowing models to earn competency scores based on correct answers while increasing question difficulty scores based on model performance. Evaluated on 30 models and 35,550 questions, RankLLM achieves 90% agreement with human judgments, outperforms strong baselines like Item Response Theory (IRT), and offers high stability and computational efficiency, making it a valuable tool for practitioners focused on nuanced LLM evaluation.

llmrankingevaluation

2 h agoarXiv cs.AIResearch

Capture Timing-Attention of Events in Clinical Time Series

The paper introduces the Level-of-Individual Time Transformation (LITT), a novel approach for modeling individual-level temporal variables in clinical time series data, specifically applied to longitudinal electronic health records (EHR) from 3,276 breast cancer patients. This method enables the automatic discovery of clinically significant patient trajectories and facilitates counterfactual timing deductions, termed a "What-If Machine," demonstrating strong performance in timing prediction and survival analysis without requiring prior domain knowledge. This advancement is significant for AI practitioners as it enhances individualized modeling in trajectory learning, potentially improving patient subtyping and treatment personalization.

trajectory learningclinical datatiming attention

2 h agoarXiv cs.AIAgents

ASA: Backbone-Training-Free Representation Engineering for Tool-Calling Agents

The paper introduces the Activation Steering Adapter (ASA), a training-free method designed to enhance tool-calling capabilities in LLM agents without requiring backbone training. ASA utilizes a router-conditioned mixture of steering vectors and a probe-guided signed gate to improve tool-use accuracy, achieving a strict tool-use F1 score increase from 0.18 to 0.50 on the MTU-Bench benchmark with the Qwen2.5-1.5B model, while significantly reducing false positives. This approach is critical for practitioners as it offers a lightweight and efficient solution to address the representation-behavior gap in LLMs, enabling more reliable domain-specific tool integration.

llmtool callingactivation steering

2 h agoarXiv cs.AITraining

MemCast: Memory-Driven Time Series Forecasting with Experience-Conditioned Reasoning

MemCast introduces a memory-driven framework for time series forecasting that reformulates the task as experience-conditioned reasoning. The model organizes training data into a hierarchical memory structure, employing historical patterns, reasoning wisdom, and general laws to enhance prediction accuracy. Experimental results indicate that MemCast outperforms existing forecasting methods, making it a significant advancement for practitioners focused on improving the robustness and adaptability of time series models.

time seriesforecastingmemory

2 h agoarXiv cs.AIResearch

Structure-Preserving Learning Improves Geometry Generalization in Neural PDEs

The authors introduce General-Geometry Neural Whitney Forms (Geo-NeW), a data-driven finite element method designed to solve Partial Differential Equations (PDEs) while preserving physical conservation laws through Finite Element Exterior Calculus. The model integrates geometry via transformer-based encoding and learned finite element spaces, enhancing generalization to unseen geometries and demonstrating state-of-the-art performance on steady-state PDE benchmarks. This approach is significant for practitioners as it offers a robust framework for developing physics-informed models that maintain accuracy across diverse geometrical configurations.

neural pdesgeometrygeneralization

2 h agoarXiv cs.AIResearch

torch-sla: Differentiable Sparse Linear Algebra with Adjoint Solvers and Sparse Tensor Parallelism for PyTorch

The article announces the release of torch-sla, an open-source library for differentiable sparse linear algebra in PyTorch, addressing the lack of a unified solution for this domain. It features an autograd-aware API supporting various solvers across five backends (SciPy, Eigen, cuDSS, CuPy, and a PyTorch-native solver), enabling batched solves and distributed multi-GPU execution with automatic device dispatch. This library is significant for practitioners as it facilitates efficient integration of sparse linear algebra techniques into scientific machine learning workflows, enhancing scalability and performance.

generative aisoftware engineeringadoption

2 h agoarXiv cs.AIInference

MMD Guidance: Training-Free Distribution Adaptation for Diffusion Models via Maximum Mean Discrepancy Guidance

The article introduces MMD Guidance, a training-free method for adapting pre-trained diffusion models to user-specific target distributions using Maximum Mean Discrepancy (MMD). This approach enhances the reverse diffusion process by integrating MMD gradients, allowing for effective distribution alignment without retraining, and is applicable to both unconditional and conditional generation tasks, including latent diffusion models. MMD Guidance shows promising results in maintaining sample fidelity while achieving distributional alignment, making it a valuable tool for practitioners facing domain adaptation challenges in generative modeling.

diffusion modelsmmd guidancedomain adaptation

2 h agoarXiv cs.AIModels

HiGR: Industrial-Scale Hierarchical Generative Slate Recommendation Framework in Tencent

HiGR, a hierarchical generative framework for slate recommendation, has been introduced by Tencent to address challenges in industrial-scale item display. It employs a Prefix-Contrastive Residual Quantized VAE (PCRQ-VAE) for structured semantic ID learning and a Hierarchical Slate Decoder (HSD) to enhance planning efficiency, resulting in over 10% improvement in offline recommendation quality and a 5x speedup in inference. This framework's deployment across Tencent platforms demonstrates its scalability and effectiveness in optimizing user engagement metrics like watch time and video plays.

recommendationgenerativeslate

2 h agoarXiv cs.AIAgents

SCOPE: Sequential Causal Optimization of Process Interventions

SCOPE (Sequential Causal Optimization of Process Interventions) is a new prescriptive process monitoring approach that addresses the sequential dependencies of interventions in business processes. It employs backward induction for intervention impact estimation and utilizes causal learners to directly leverage observational data, avoiding the biases introduced by simulation or data augmentation in traditional reinforcement learning methods. Experimental results demonstrate that SCOPE outperforms existing state-of-the-art techniques in optimizing key performance indicators (KPIs), and it introduces a novel semi-synthetic dataset as a benchmark for future research in sequential prescriptive process monitoring.

reinforcement learningprocess monitoringinterventions

2 h agoarXiv cs.AIResearch

Model-Based Reinforcement Learning in Discrete-Action Non-Markovian Reward Decision Processes

The paper introduces QR-MAX, a novel model-based reinforcement learning algorithm specifically designed for discrete non-Markovian reward decision processes (NMRDPs), addressing the lack of formal guarantees on optimality and sample efficiency. QR-MAX utilizes a factorization approach that separates Markovian transition learning from non-Markovian reward handling, achieving PAC convergence to ε-optimal policies with polynomial sample complexity. Additionally, the extension to continuous state spaces, called Bucket-QR-MAX, employs a SimHash-based discretizer, enhancing sample efficiency and robustness in complex environments, which is critical for practitioners dealing with temporal-dependency tasks in RL.

reinforcement learningnmrpqr-max

2 h agoarXiv cs.AIMultimodal

Scone: Bridging Composition and Distinction in Subject-Driven Image Generation via Unified Understanding-Generation Modeling

Scone is a new unified understanding-generation model designed for subject-driven image generation that effectively integrates composition and distinction, addressing limitations in handling multiple subjects. It employs a two-stage training scheme focusing first on composition and then enhancing distinction through semantic alignment and attention-based masking. The accompanying SconeEval benchmark evaluates performance in both areas, with experimental results showing Scone surpassing existing open-source models on multiple benchmarks, making it a valuable tool for practitioners aiming to improve subject identity preservation in complex visual tasks.

image generationsubject-drivenmodeling

2 h agoarXiv cs.AIMultimodal

V-REX: Benchmarking Exploratory Visual Reasoning via Chain-of-Questions

The paper introduces V-REX, a benchmarking suite designed for evaluating visual reasoning in vision-language models (VLMs) through a multi-step exploratory approach using a Chain-of-Questions (CoQ) framework. V-REX allows for detailed assessment of VLMs’ capabilities in planning and following complex tasks, highlighting performance discrepancies and areas needing enhancement in handling open-ended visual reasoning tasks. This evaluation protocol is crucial for practitioners aiming to improve VLMs' interpretative abilities and reasoning processes in real-world applications.

visual reasoningevaluationbenchmark

2 h agoarXiv cs.AITraining

On the Condition Number Dependency in Bilevel Optimization

The paper presents new lower bounds on the oracle complexity for finding $\epsilon$-stationary points in bilevel optimization, particularly when the upper-level problem is nonconvex and the lower-level problem is strongly convex. It establishes a lower bound of $\Omega(\kappa_y^{5/2} \epsilon^{-2})$, which highlights a significant gap in condition number dependency between bilevel and minimax problems, and extends results to various settings including high-order smooth functions and stochastic oracles. This work is crucial for practitioners as it provides deeper insights into the complexity landscape of bilevel optimization, potentially guiding the design of more efficient algorithms in real-world applications.

diffusion modelscontroldecision making

2 h agoarXiv cs.AIResearch

Lost in Serialization: Invariance and Generalization of LLM Graph Reasoners

The paper presents a systematic analysis of the invariance and generalization capabilities of Large Language Model (LLM) graph reasoners, revealing that these models exhibit sensitivity to graph serialization variations such as node reindexing and edge reordering. A novel decomposition of graph representations is introduced, alongside a comprehensive benchmarking suite and new spectral tasks to evaluate model robustness. Findings indicate that while larger non-fine-tuned models demonstrate greater robustness, fine-tuning can reduce sensitivity to node relabeling but may increase sensitivity to structural variations, highlighting the complexities of training LLMs for graph reasoning tasks.

graph reasoningllmrobustness

2 h agoarXiv cs.AIResearch

Learning-Guided Integration Contours Construction for Fast Large-Scale Generalized Eigensolvers

The paper introduces Deepcontour, a novel hybrid framework that combines deep learning with Kernel Density Estimation (KDE) to optimize integration contours for Generalized Eigenvalue Problems (GEPs). Utilizing an Eigen-Neural-Operator (ENO) for rapid spectral predictions, Deepcontour achieves up to a 5.63x speedup in computational efficiency while preserving numerical accuracy. This approach is significant for practitioners as it enhances the performance of contour integral methods in solving large-scale GEPs, facilitating more efficient computations in scientific and engineering applications.

eigenvalue problemsdeep learningoptimization

2 h agoarXiv cs.AIResearch

Deep Generative Model for Human Mobility Behavior

MobilityGen, a diffusion-based generative framework, has been introduced for simulating multi-attribute activity-travel sequences over extended periods and large spatial scales. It effectively links behavioral attributes with environmental context, capturing key patterns such as scaling laws for location visits and the evolution of travel modes and destination choices. This model enhances the capability for nuanced analyses of human mobility behavior, offering insights into urban space access and social dynamics, which are critical for transport planning and urban design.

mobilitygenerative modelurban design

2 h agoarXiv cs.AISafety

TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning

TruthRL is a novel reinforcement learning framework designed to enhance the truthfulness of large language models (LLMs) by optimizing for both accurate responses and appropriate abstention when uncertain. Implemented using Generalized Reward Policy Optimization (GRPO), TruthRL employs a ternary reward system that distinguishes between correct answers, hallucinations, and abstentions, leading to a significant reduction in hallucinations from 43.5% to 19.4% and an increase in truthfulness from 5.3% to 37.2% across four knowledge-intensive benchmarks. This approach is crucial for practitioners as it addresses the dual challenge of accuracy and uncertainty management in LLMs, enabling more reliable deployment in real-world applications.

truthfulnessllmreinforcement learning

2 h agoarXiv cs.AIResearch

Generation Properties of Stochastic Interpolation under Finite Training Set

The paper presents a theoretical analysis of generative models operating under finite training sets within the stochastic interpolation framework. It derives closed-form expressions for the optimal velocity field and score function, showing that the deterministic process recovers training samples while the stochastic process introduces Gaussian noise. This work is significant for practitioners as it provides insights into the effects of model estimation errors on generation quality, offering formal definitions of underfitting and overfitting tailored for generative models, which can inform model design and evaluation in real-world applications.

generative modelstrainingtheory

2 h agoarXiv cs.AIInference

Whisfusion: Parallel ASR Decoding with Masked Diffusion

Whisfusion introduces a novel non-autoregressive (NAR) ASR system that leverages a masked diffusion decoder trained on Whisper-large-v3 audio embeddings, achieving significant improvements in both accuracy and decoding speed. The model, trained on approximately 68k hours of multilingual speech, outperforms Whisper-large-v3 and Whisper-turbo in group-average accuracy across various benchmarks while decoding 4-5x faster, and it competes effectively with other leading models like Canary and Qwen3-ASR. This development highlights the potential of masked diffusion techniques to enhance ASR performance, offering a faster and more efficient alternative for practitioners in the field.

asrmasked diffusionlatency

2 h agoarXiv cs.AIResearch

Assessment of Personality Dimensions Across Situations in Dyadic Role-Play Scenarios

The study presents findings on automatic personality perception (APP) in dyadic role-play scenarios, highlighting that perceived personalities vary significantly across different contexts, specifically in neutral interviews versus stressful client interactions. Key technical insights include the identification of specific acoustic features—such as loudness and spectral flux—that correlate with personality traits, with handcrafted features outperforming speaker embeddings in predicting traits like extraversion and neuroticism. This research underscores the importance of context in personality perception, which could inform the design of more adaptive and context-aware AI systems in assistive technologies.

personalityautomatic personality perception

2 h agoarXiv cs.AIResearch

LLM-Aided Joint Secrecy Precoding and Trajectory for RSMA-Based Heterogeneous UAV Networks

The paper presents a novel hierarchical optimization framework for secure communications in RSMA-enabled heterogeneous UAV networks, addressing a multi-objective problem that includes UAV trajectory design and secrecy precoding. It introduces a Large Language Model-guided heuristic multi-agent reinforcement learning approach (LLM-HeMARL) that leverages LLM-generated expert policies to optimize UAV trajectories, achieving improved secrecy rates and energy efficiency compared to existing methods. This work is significant for practitioners as it demonstrates how LLMs can enhance decision-making in complex, non-convex optimization scenarios within UAV communication systems.

uavsecure communicationsoptimization

2 h agoarXiv cs.AITraining

GRID: Scaling Task-Agnostic Inference in Continual Prompt Tuning

GRID is a new framework for prompt-based continual learning that enhances the scalability and efficiency of adapting large language models (LLMs) across task sequences. It features an output-space-aware decoding mechanism and a gradient-guided prompt selection strategy that aggregates less informative prompts, resulting in improved backward transfer and reduced memory usage on models like T5, Qwen, and LLaMA. This development is significant for practitioners as it provides a more efficient approach to continual learning without the need for task identifiers, enabling better performance across diverse tasks in LLM applications.

prompt tuningcontinual learningllm

2 h agoarXiv cs.AIAgents

TinyTroupe: An LLM-powered Multiagent Persona Simulation Toolkit

TinyTroupe is an open-source simulation toolkit designed for LLM-powered Multiagent Systems (MAS), enabling detailed persona definitions and programmatic control for simulating realistic human behaviors. It addresses existing limitations in MAS libraries by offering features like fine-grained persona specifications, population sampling, and integrated validation, which are crucial for behavioral studies and social simulations. This toolkit allows practitioners to effectively model and analyze complex behavioral problems, enhancing the capabilities of LLMs in multiagent simulations.

llmmultiagentsimulationtoolkit

2 h agoarXiv cs.AIResearch

Quantifying Perception-Based Student Success with Generative AI: An Exploratory Monte Carlo Simulation

The study introduces a Monte Carlo simulation framework to quantify perception-based student success in the context of generative AI tools like ChatGPT. Utilizing a 10-item, 5-point Likert-scale usability instrument derived from a structured literature review, the simulation generated 10,000 synthetic observations, revealing that the weighting structure significantly affects the outcomes, particularly highlighting System Efficiency and Learning Burden. This framework provides a reproducible method for assessing the educational impact of GenAI, which can aid practitioners in evaluating and optimizing AI tools in educational settings.

generative aieducationsimulation

2 h agoarXiv cs.AIAgents

Fact-Augmented Lookahead Planning for LLM Agents

The paper introduces LWM-Planner, a fact-augmented lookahead planning framework designed to enhance the performance of LLM agents in complex, partially observable environments through in-context learning. By extracting and validating task-critical facts post-episode, the framework enables recursive, depth-limited lookahead planning that conditions action proposals and state-value estimations on these facts, leading to improved cumulative returns on benchmarks like FrozenLake, CrafterMini, and ALFWorld compared to existing methods such as ReAct and Reflexion. This approach is significant for practitioners as it demonstrates a method for enhancing LLM agent performance without requiring parameter updates, leveraging experience-derived knowledge for more effective planning.

llmplanningagentsinteractive environments

2 h agoarXiv cs.AIModels

ASyMOB: Algebraic Symbolic Mathematical Operations Benchmark

The ASyMOB benchmark has been introduced, featuring a dataset of 35,368 validated symbolic math problems that include various mathematical domains such as integration and differential equations. This benchmark systematically applies perturbations to seed problems, allowing for a detailed evaluation of LLMs' generalization capabilities and robustness under minor changes. The findings indicate that while most models struggle with perturbations, integrated code tools enhance performance, and there are notable instances where LLMs outperform traditional Computer Algebra Systems, suggesting new avenues for hybrid approaches in AI-driven scientific discovery.

llmbenchmarksymbolic mathematics

2 h agoarXiv cs.AIResearch

Attacks on Machine-Text Detectors Retain Stylistic Fingerprints

The paper investigates the effectiveness of current evasion strategies against machine-text detectors, revealing that while techniques like prompt engineering can degrade detector performance, they do not eliminate the stylistic fingerprints of machine-generated text. The authors introduce a novel paraphrasing approach that successfully evades detection across various detectors, including those leveraging stylistic features. Their findings suggest that reliable machine-text detection may necessitate a shift from single-document to multi-document analysis to maintain efficacy against evolving evasion methods.

machine-textdetectionevasionstylistic fingerprints

2 h agoarXiv cs.AIResearch

CleanPatrick: A Benchmark for Image Data Cleaning

CleanPatrick is a newly introduced large-scale benchmark for image data cleaning, utilizing the Fitzpatrick17k dermatology dataset, which includes 496,377 binary annotations from 933 medical crowd workers. The benchmark identifies off-topic samples, near-duplicates, and label errors, formalizing issue detection as a ranking task and employing standard ranking metrics for evaluation. This resource allows practitioners to systematically compare various image-cleaning strategies, highlighting the performance of self-supervised representations and classical anomaly detection methods in real-world scenarios.

data cleaningbenchmarkimage datamachine learning

2 h agoarXiv cs.AIAgents

A Survey of Robotic Navigation and Manipulation with Physics Simulators in the Era of Embodied AI

The article presents a survey on the use of physics simulators for robotic navigation and manipulation in the context of Embodied AI, focusing on the persistent sim-to-real gap. It analyzes various properties of simulators, their features for specific tasks, and hardware requirements, while providing a comprehensive resource that includes benchmark datasets, metrics, and simulation platforms. This work is significant for practitioners as it aids in selecting appropriate tools and methodologies for effective sim-to-real transfer, crucial for developing robust AI systems in real-world applications.

roboticsnavigationmanipulation

2 h agoarXiv cs.AIResearch

NuWa: Deriving Lightweight Class-Specific Vision Transformers for Edge Devices

NuWa is a novel method for deriving lightweight class-specific Vision Transformers (ViTs) tailored for resource-constrained edge devices, addressing the limitations of existing model compression techniques. It employs self-knowledge purification to eliminate class-detrimental weights and utilizes closed-form optimization to create compact ViTs without the need for post-pruning retraining. Experimental results indicate that NuWa achieves up to 29% higher accuracy on class-specific tasks compared to state-of-the-art training-free pruning methods, with a 33.69x speedup in pruning and a 99.83% reduction in pruning costs, making it highly efficient for practitioners focused on deploying optimized models in edge environments.

vision transformersedge devices

2 h agoarXiv cs.AITraining

CITRAS: Covariate-Informed Transformer for Time Series Forecasting

CITRAS, a decoder-only Transformer model for time series forecasting, has been introduced to effectively integrate observed and known covariates, addressing the common challenges in leveraging these variables due to length discrepancies. It features two novel mechanisms: Key-Value (KV) Shift, which aligns future known covariates with target variables, and Attention Score Smoothing, which enhances local dependencies into global variate-level dependencies. Experimental results indicate that CITRAS significantly improves forecasting accuracy across diverse real-world datasets, making it a valuable tool for practitioners aiming to enhance model performance in time series analysis.

time seriesforecastingtransformer

2 h agoarXiv cs.AIResearch

Deep Tree Tensor Networks

The paper introduces the Deep Tree Tensor Network (DTTN), a novel architecture that employs a tree-like tensor network topology to capture $2^L$-order multiplicative interactions through multilinear operations. DTTN utilizes multiple antisymmetric interaction modules (AIMs) and shows superior performance across various benchmarks compared to existing methods and state-of-the-art architectures. This advancement may enhance interpretability in tensor network research and offers a new approach for practitioners looking to leverage tensor networks in natural image recognition tasks.

tensor networksneural networks

2 h agoarXiv cs.AIResearch

Representational Alignment with Chemical Induced Fit for Molecular Relational Learning

The paper introduces ReAlignFit, a novel approach to Molecular Relational Learning (MRL) that enhances substructure representation alignment by incorporating a chemical Induced Fit-based inductive bias. This method employs a Bias Correction Function for substructure edge reconstruction and integrates the Subgraph Information Bottleneck to improve the generation of molecular embeddings, resulting in superior performance on nine datasets compared to existing models. This advancement is significant for practitioners as it offers a more stable and chemically informed framework for predicting molecular interactions, particularly in variable chemical spaces.

molecular learningchemical knowledge

2 h agoarXiv cs.AISafety

Dynamics of Adversarial Attacks on Large Language Model-Based Search Engines

The paper presents a theoretical analysis of adversarial attacks on Large Language Model-based search engines, specifically focusing on ranking manipulation attacks. It models the interactions between attackers and defenders as an Infinitely Repeated Prisoners' Dilemma, identifying key factors such as attack costs and success rates that influence strategic behavior. The findings underscore the need for adaptive security strategies, as traditional defensive measures may inadvertently incentivize further attacks, highlighting the complexities in securing LLM-based systems for practitioners.

adversarial attacksllmsearch engines

2 h agoarXiv cs.AIMultimodal

Whisper-GPT -- Continuous Discrete Hybrid Representation Language Models For Speech And Music

WHISPER-GPT is a newly proposed generative large language model that integrates continuous audio representations with discrete tokens, addressing the limitations of context length in high-fidelity generative architectures. By utilizing both spectrograms and discrete acoustic tokens, the model enhances performance metrics such as perplexity and negative log-likelihood for next token prediction in audio tasks. This hybrid approach offers practitioners a more efficient framework for developing applications in generative audio, speech, and music, leveraging the advantages of both continuous and discrete data representations.

speechmusicllm

2 h agoarXiv cs.AIResearch

Visual-TCAV: Concept-based Attribution and Saliency Maps for Post-hoc Explainability in Image Classification

Visual-TCAV is a newly introduced explainability framework that combines local and global explanations for image classification models by utilizing Concept Activation Vectors (CAVs) to generate class-agnostic saliency maps. It enhances the capabilities of traditional TCAV by estimating concept attribution to predictions through a generalization of Integrated Gradients, demonstrating improved alignment with ground truth explanations in evaluations. This framework is significant for practitioners as it provides a more comprehensive understanding of model behavior by linking concept sensitivity directly to specific predictions, facilitating better interpretability in CNNs.

explainabilityimage classificationllm

2 h agoarXiv cs.AIResearch

Conditional Vendi Score: Prompt-Aware Diversity Evaluation for Generative AI Models and LLMs

The article introduces two new diversity evaluation metrics for generative AI models: Conditional-Vendi and Conditional-RKE, which are designed to assess prompt-induced variability in outputs. These metrics leverage conditional entropy from positive semidefinite matrices, with Conditional-RKE achieving an $O(1/\sqrt{n})$ convergence rate and Conditional-Vendi utilizing a truncated-spectrum approximation for scalability. The methods demonstrate effectiveness across various tasks, including text-to-image generation and image captioning, providing practitioners with improved tools for evaluating diversity in prompt-guided generation.

generative aievaluationllm

2 h agoarXiv cs.AISafety

BadRobot: Jailbreaking Embodied LLM Agents in the Physical World

The paper introduces BadRobot, a novel attack paradigm designed to exploit vulnerabilities in embodied LLMs to induce harmful behaviors during voice-based interactions. It identifies three critical vulnerabilities: manipulation of LLMs, misalignment between linguistic outputs and physical actions, and flaws in world knowledge. The authors benchmark these attacks against existing frameworks like Voxposer, Code as Policies, and ProgPrompt, highlighting the need for enhanced safety measures in embodied AI systems.

embodied aillmsafety

2 h agoarXiv cs.AIResearch

Mixtures of Neural Operators Reduce Active Complexity in Operator Learning

The paper presents a novel approach using Mixtures of Neural Operators (MoNOs) to enhance operator learning efficiency by reducing active complexity compared to traditional single-neural-operator models. The main theorem demonstrates that MoNOs can achieve lower depth, width, and rank scaling for approximating scalar uniformly continuous nonlinear operators, with expert quantities bounded by $\mathcal{O}(\varepsilon^{-1})$ for Lipschitz targets. This work is significant for practitioners as it provides a framework for optimizing neural operator architectures, potentially leading to more efficient models that require fewer resources for complex function approximations.

operator learningneural networks

2 h agoarXiv cs.AITraining

Robust Deep Reinforcement Learning Through Adversarial Attacks and Training : A Survey

The survey presents a comprehensive analysis of adversarial training methodologies aimed at enhancing the robustness of Deep Reinforcement Learning (DRL) agents against environmental perturbations. It systematically categorizes various adversarial attacks and training techniques, comparing their objectives and mechanisms. This work is significant for practitioners as it provides insights into improving the reliability of DRL systems in real-world applications, addressing a critical challenge in deploying these models.

reinforcement learningadversarial attacks

2 h agoarXiv cs.AIResearch

Evaluation Cards: An Interpretive Layer for AI Evaluation Reporting

The paper introduces \EvalCards{}, an operational reporting layer designed to unify and standardize AI evaluation reporting by integrating benchmark metadata, evaluation run data, and model metadata into a cohesive framework. Key features include a reporting schema derived from a review of 52 papers and stakeholder interviews, and the implementation of interpretive signals such as reproducibility and score comparability. This framework addresses critical gaps in AI evaluation transparency and comparability, enabling practitioners to better assess model performance across diverse sources and improve the reliability of their evaluations.

ai evaluationreportingllm

2 h agoarXiv cs.AIResearch

Deterministic Integrity Gates for LLM-Assisted Clinical Manuscript Preparation: An Auditable Biomedical Informatics Architecture

The article presents the MedSci Skills toolkit, an open-source architecture designed for LLM-assisted clinical manuscript preparation, emphasizing a verification framework that integrates deterministic integrity checks. This toolkit comprises 43 skills, including a 21-detector deterministic tier, which successfully identified all 27 injected defects in tested pipelines (STARD, PRISMA, STROBE) without false positives, outperforming a single-prompt LLM reviewer. This approach enhances the reliability of LLM outputs by providing an auditable and reproducible verification process, crucial for practitioners aiming to ensure the integrity of AI-generated scientific manuscripts.

llmmanuscript preparationverification

2 h agoarXiv cs.AISafety

Emergent alignment and the projectability of ethical personas

The paper introduces the concept of "emergent alignment" in LLMs, exploring how finetuning models on narrow safety tasks can induce aligned behavior across broader categories. Utilizing the "Constitutional AI" approach, four ethical frameworks were applied during finetuning, resulting in models that exhibit distinct "ethical personas" corresponding to their training constitution, with significant variation in performance across models. This research emphasizes the need for alignment strategies to be assessed not only on general safety outcomes but also on their ability to project ethical behavior consistently, which is critical for practitioners developing robust and reliable AI systems.

alignmentemergent alignmentllm

2 h agoarXiv cs.AIAgents

Anything2Skill: Compiling External Knowledge into Reusable Skills for Agents

The paper introduces Anything2Skill, a framework that compiles external knowledge into reusable skills for agents, enhancing retrieval-augmented generation (RAG) capabilities. It employs a skill-tree prior for skill extraction from heterogeneous knowledge records, converting them into structured skill contracts that detail invocation conditions and workflows. Experiments demonstrate that Anything2Skill, when combined with RAG, achieves success rates of 98.85% and 94.10% on qsv and GitHub-CLI tasks, respectively, indicating significant improvements in agent performance by enabling procedural knowledge reuse.

skillsagentsrag

2 h agoarXiv cs.AIRAG

Graph2Idea:Retrieval-Augmented Scientific Idea Generation with Graph-Structured Contexts

Graph2Idea is a novel framework for retrieval-augmented scientific idea generation that utilizes knowledge graphs to enhance the context provided to Large Language Models (LLMs). By transforming retrieved literature into structured knowledge triples and constructing a target-centered knowledge graph, Graph2Idea improves the relevance and clarity of input data, resulting in significant performance gains on a scientific idea generation benchmark—improving Novelty from 0.45 to 0.52, Quality from 0.24 to 0.29, and Feasibility from 0.22 to 0.28 compared to the strongest baseline. This approach emphasizes the importance of structured relational evidence in generating high-quality research ideas, making it a valuable tool for practitioners working in scientific discovery and LLM applications.

scientific idea generationknowledge graphllm

2 h agoarXiv cs.AIAgents

Baichuan-M4: A Clinical-Grade Medical Agent System for Continuous Care

Baichuan Intelligence has released Baichuan-M4, a clinical-grade medical large model designed for continuous care, featuring a coordinated medical agent system. Key technical components include the Baichuan-Harness runtime for reinforcement learning and deployment consistency, a core reasoning model utilizing SPAR++ for reward modeling, and a clinical tool layer for managing patient memory and multimodal perception. This model achieves leading results in various medical evaluations, significantly reducing hallucination rates to 3.3%, which is crucial for practitioners aiming to implement reliable AI systems in clinical settings.

medical agentllmreinforcement learning

2 h agoarXiv cs.AIResearch

A Geometric Account of Activation Steering through Angle-Norm Decomposition

The paper presents a geometric analysis of activation steering in language models, emphasizing the importance of both angular alignment and hidden-state norm in steering interventions. Through empirical studies involving seven language models, it finds that concepts are predominantly represented in angular structure, advocating for spherical steering methods while noting that norm stability is crucial for effective steering. This work suggests a shift in parameterization for activation steering, promoting the use of distinct angular and radial components to enhance interpretability and control in model behavior.

activation steeringgeometryllm

2 h agoarXiv cs.AIResearch

PSEBench: A Controllable and Verifiable Benchmark for Evaluating LLMs in Patient Safety Event Triage

PSEBench, a new benchmark for evaluating LLMs in patient safety event triage, has been introduced, comprising 5,074 cases derived from Minnesota's 29 Reportable Adverse Health Events. The benchmark utilizes a policy-grounded construction methodology that incorporates clause cards for auditable decision specifications and supports closed-loop verification, enabling LLMs to generate missing information and handle ambiguous cases. This development is significant for practitioners as it provides a structured framework to assess the reliability and effectiveness of LLMs in high-stakes clinical decision-making contexts.

llmpatient safetybenchmark

2 h agoarXiv cs.AIAgents

Parthenon Law: A Self-Evolving Legal-Agent Framework

The article introduces \textsc{Parthenon}, a self-evolving legal-agent framework designed to enhance the performance of legal-domain large language models (LLMs) by addressing key challenges in the deployment of legal agents. It features a large-scale empirical study with $12,510$ agent trajectories demonstrating that while model accuracy improves with stronger models, matter completion remains inadequate. The framework incorporates a learning loop that allows agents to refine their skills and knowledge based on past performance, facilitating continuous improvement without altering model weights, which is crucial for practitioners aiming to build reliable legal AI systems.

legalllmagentworkflow

2 h agoarXiv cs.AIAgents

AgentPLM: Agentic Protein Language Models with Reasoning-Augmented Decoding for Protein Sequence Design

AgentPLM has been introduced as a novel protein language model that integrates Reasoning-Augmented Decoding (RAD) and Contrastive Agent Policy Optimisation (CAPO) to enhance protein sequence design by allowing real-time consultation of external biophysical feedback. It outperforms existing passive models on benchmark tasks, achieving state-of-the-art results in antibody optimization and other applications, demonstrating improved hit rates and online error correction capabilities. This advancement is significant for practitioners as it enables more adaptive and efficient protein design processes, potentially leading to better therapeutic candidates.

protein-designllmreasoningdecoding

2 h agoarXiv cs.AIAgents

EVA-Net: Subject-Independent EEG Motor Decoding with Video-Derived Motor Priors

EVA-Net is a novel two-stage framework designed for subject-independent EEG motor decoding that leverages action videos as semantic priors. It employs cross-modal and supervised contrastive learning to align EEG and video features, achieving an 8.66% improvement in leave-one-subject-out (LOSO) accuracy on the EEGMMI dataset compared to traditional text-based approaches. This advancement is significant for practitioners as it enhances the robustness of BCI systems by reducing inter-subject variability and minimizing calibration requirements, thereby facilitating broader application of EEG decoding in real-world scenarios.

legalagentllmframework

2 h agoarXiv cs.AISafety

A Sober Look at Agentic Misalignment in Automated Workflows

The paper presents a study on agentic misalignment in multi-agent systems (MAS) within automated workflows, introducing a new alignment paradigm called Agentic Evidence Attribution (AEA). AEA enhances agent posteriors through context-specific evidence, addressing the issue of agents acting on implicit proxy utilities misaligned with human goals. The research demonstrates that incorporating evidence, via self-reflection and weak-to-strong generalization, can effectively improve collaboration among agents, making it crucial for practitioners aiming to build reliable multi-agent systems.

multi-agent-systemsalignmentautomated-workflows

2 h agoarXiv cs.AIResearch

AMEL: Accumulated Message Effects on LLM Judgments

The paper introduces the concept of Accumulated Message Effects on LLM Judgments (AMEL), demonstrating that the polarity of prior conversation history significantly biases subsequent evaluations made by large language models (LLMs). Analyzing 84,088 API calls across 12 models from five providers, the study finds that models shift their judgments toward the prevailing sentiment of prior messages, with a notable negativity asymmetry where negative histories induce greater bias (1.52x more) than positive ones. This research highlights the importance of context management in evaluation pipelines, suggesting that using fresh contexts per item or balancing the history can mitigate bias in automated evaluations.

llmbiasjudgmentsevaluation

2 h agoarXiv cs.AIResearch

Neurosymbolic Learning for Inference-Time Argumentation

The paper introduces inference-time argumentation (ITA), a neurosymbolic framework designed for ternary claim verification that leverages formal argumentation semantics to enhance large language model (LLM) training and inference. ITA optimizes argument generation and scoring during training, resulting in predictions that are faithful to the underlying arguments rather than relying on post-hoc reasoning. Experimental results demonstrate that ITA outperforms argumentative baselines and competes effectively with direct-prediction methods on two datasets, providing deterministic and inspectable verdicts.

neurosymbolicargumentationclaim-verification

2 h agoarXiv cs.AIModels

CatalyticMLLM: A Graph-Text Multimodal Large Language Model for Catalytic Materials

CatalyticMLLM is a newly proposed unified graph-text multimodal large language model designed for catalytic materials, integrating property prediction and inverse design in a shared representation space. The model demonstrates superior performance over decoupled baselines in both catalytic relaxed-energy prediction and structure generation tasks, enabling a closed-loop optimization workflow. This advancement is significant for practitioners as it addresses data distribution shifts and evaluator bias, enhancing the stability and reliability of catalytic material design processes.

multimodalcatalytic-materialsllm

2 h agoarXiv cs.AITraining

Adaptive Teacher Exposure for Self-Distillation in LLM Reasoning

The paper introduces Adaptive Teacher Exposure for Self-Distillation (ATESD), a novel approach that optimizes the exposure of a teacher model during on-policy self-distillation to enhance reasoning in large language models (LLMs). ATESD utilizes a learnable Beta-policy controller to dynamically adjust the teacher's exposure to reference reasoning, leading to improved performance on benchmarks AIME 24, AIME 25, and HMMT 25 with Qwen3 models (1.7B, 4B, and 8B parameters), achieving significant gains over existing self-distillation and reinforcement learning methods. This work highlights the importance of adaptive exposure strategies in training LLMs, providing practitioners with a new mechanism to fine-tune model training and improve reasoning capabilities.

llmself-distillationreasoning

2 h agoarXiv cs.AISafety

The Price of Agreement: Measuring LLM Sycophancy in Agentic Financial Applications

The paper presents a study on the phenomenon of sycophancy in large language models (LLMs) within financial applications, highlighting their tendency to prioritize user agreement over factual accuracy. Key findings reveal that models exhibit only modest performance drops when faced with user contradictions, and a new suite of tasks is introduced to evaluate this sycophancy, showing that most models struggle under conflicting user preferences. This research is significant for practitioners as it underscores the need for improved robustness in LLMs used in financial systems, particularly in handling contradictory information without sacrificing accuracy.

llmsycophancyfinancial-applicationsevaluation

2 h agoarXiv cs.AIResearch

Non-Parametric Structural Priors for Geometry Theorem Prediction

The paper introduces a novel approach for multi-step theorem prediction using Theorem Precedence Graphs, which encode historical solution dependencies as directed graphs to enhance reasoning capabilities without gradient-based optimization. This method addresses the scalability issue known as Structural Drift, allowing LLMs to maintain performance as reasoning depth increases, achieving 89.29% accuracy on the FormalGeo7k benchmark, surpassing traditional in-context learning baselines and competing with state-of-the-art supervised models. The findings suggest that integrating explicit structural priors can significantly improve LLM-based symbolic reasoning, offering a new avenue for practitioners in AI and geometry problem-solving.

geometrytheorem-predictionneural-symbolic

2 h agoarXiv cs.AITraining

RoboGPT-R1: Enhancing Robot Task Planning with Reinforcement Learning

RoboGPT-R1 is introduced as a two-stage fine-tuning framework aimed at enhancing robot task planning through reinforcement learning (RL) and supervised training. It utilizes the Qwen2.5-VL-3B model and demonstrates a 21.33% improvement over GPT-4o-mini and a 20.33% increase over Qwen2.5-VL-7B on the EmbodiedBench benchmark, addressing challenges in long-horizon manipulation tasks. This framework is significant for practitioners as it combines RL with a rule-based reward function to improve visual-spatial reasoning and action sequence consistency in real-world robotic applications.

roboticsplanningreinforcement-learningllm

2 h agoarXiv cs.AIAgents

How can we assess human-agent interactions? Case studies in software agent design

The paper introduces PULSE, a new framework for evaluating human-agent interactions that integrates user feedback, machine learning predictions of user satisfaction, and combines these with model-generated pseudo-labels. Implemented in the software engineering domain using the open-source agent OpenHands, PULSE evaluates agent design decisions across 15,000 users, achieving a 40% reduction in confidence intervals compared to traditional A/B testing. The findings highlight significant discrepancies between benchmark performance and real-world user satisfaction, emphasizing the need for more robust evaluation methods in LLM-powered agents.

llmhuman-agent-interactionevaluationsoftware-engineering

2 h agoarXiv cs.AIAgents

ChartAgent: A Multimodal Agent for Visually Grounded Reasoning in Complex Chart Question Answering

ChartAgent is a newly introduced multimodal framework designed for visually grounded reasoning in chart-based question answering, explicitly addressing the challenges posed by unannotated charts. It utilizes an iterative decomposition approach to break down queries into visual subtasks, employing specialized actions like annotation and cropping, and achieves state-of-the-art performance with up to a 16.07% absolute improvement on benchmarks such as ChartBench and ChartX, particularly excelling in unannotated, numerically intensive queries. This framework not only enhances performance across various chart types and reasoning complexities but also serves as a plug-and-play solution for existing LLMs, making it a significant advancement for practitioners in the field of visual reasoning with AI.

multimodalvisual-reasoningchartquestion-answering

2 h agoarXiv cs.AIAgents

Impatient Users Confuse AI Agents: High-fidelity Simulations of Human Traits for Testing Agents

The paper introduces TraitBasis, a model-agnostic framework for stress testing AI agents by simulating user traits like impatience and incoherence without requiring fine-tuning or additional data. It extends the existing $\tau$-Bench to $\tau$-Trait, revealing a performance degradation of 2%-30% across frontier models when subjected to altered user behaviors. This approach emphasizes the importance of robustness testing in AI systems, providing practitioners with a tool to enhance the reliability of agents in real-world interactions.

aitestingrobustnessuser-behavior

2 h agoarXiv cs.AIAgents

Constructing coherent spatial memory in LLM agents through graph rectification

The paper introduces LLM-MapRepair, a framework for constructing and repairing topological navigation graphs using large language models (LLMs). Key features include a Version Control mechanism for graph construction and an Edge Impact Score for prioritizing repairs, evaluated across seven LLMs including GPT-4.1, achieving 94.3% node recall and 88.2% edge recall on the Dream of the Red Chamber deployment. This work is significant for practitioners as it enhances LLM capabilities in spatial reasoning and map construction, addressing challenges in larger environments through incremental updates and improved accuracy in pathfinding tasks.

llmmap-constructionnavigationgraph

2 h agoarXiv cs.AIResearch

Why Does Reasoning Length Converge? Unveiling the Underfitting-Overfitting Trade-off in Chain-of-Thought

The article introduces CoT-Space, a theoretical framework that transforms the reasoning process of Large Language Models (LLMs) from token prediction to an optimization task within a continuous semantic space, addressing the limitations of traditional token-level analysis. It explores the convergence of Chain-of-Thought (CoT) reasoning lengths as a result of the underfitting-overfitting trade-off, supported by experiments utilizing Reinforcement Learning (RL) to validate these insights. This framework provides practitioners with a principled approach to optimize reasoning trajectories, enhancing the efficiency and effectiveness of LLMs in complex reasoning tasks.

llmreasoningunderfittingoverfitting

2 h agoarXiv cs.AIProducts

Position: The ML Community Must Build an AI-Augmented Peer-Review Ecosystem

The position paper advocates for the development of an AI-augmented peer-review ecosystem in the machine learning community to address the increasing strain from manuscript submissions outpacing reviewer capacity. It proposes leveraging Large Language Models (LLMs) to assist in tasks such as factual verification, reviewer performance guidance, and decision-making support for Area Chairs, emphasizing the need for access to structured peer review data to facilitate this integration. This initiative is critical for maintaining the integrity and scalability of scientific validation in ML, addressing both technical and ethical challenges in the process.

llmpeer-reviewaicollaboration

2 h agoarXiv cs.AIResearch

A Comprehensive Survey of Direct Preference Optimization: Datasets, Theories, Variants, and Applications

The article presents a comprehensive survey on Direct Preference Optimization (DPO), highlighting its role as a reinforcement learning-free alternative to Reinforcement Learning from Human Feedback (RLHF) for aligning policy models with human preferences. It categorizes recent research on DPO, discusses theoretical analyses, variants, and relevant datasets, and identifies future research directions to enhance model alignment. This survey is significant for practitioners as it consolidates existing knowledge and guides future work in preference-based model alignment, which is crucial for developing more effective LLMs.

llmpreferencealignmentdpo

2 h agoarXiv cs.AITraining

A Unifying Lens on Supervised Fine-Tuning Through Target Distribution Design

The paper introduces the Q-target framework for supervised fine-tuning (SFT), which reinterprets SFT as a target distribution design, allowing for a more nuanced approach to token alignment. The proposed Target-SFT method, which optimizes the training objective based on a desired target distribution, demonstrates improved performance across ten reasoning dataset-model settings. This research provides a foundational principle for SFT design, offering practitioners a broader range of strategies for optimizing fine-tuning objectives in LLMs.

fine-tuningtarget distributionsupervised learning

2 h agoarXiv cs.AIAgents

EEVEE: Towards Test-time Prompt Learning in the Real World for Self-Improving Agents

EEVEE is a novel multi-dataset test-time prompt learning framework designed for LLM agents that addresses the challenges of heterogeneous input streams in real-world applications. It features a router for task clustering and a co-evolution strategy for router and prompt optimization, resulting in significant performance improvements; specifically, EEVEE achieves average multi-benchmark score enhancements of 10.38 and 24.32 points over Qwen3-4B-Instruct and DeepSeek-V3.2, respectively, and outperforms SOTA methods GEPA and ACE by up to 37.2% and 48.2%. This framework is crucial for practitioners aiming to build robust AI systems capable of adapting to diverse and dynamic task environments.

test-time learningprompt learningllm

2 h agoarXiv cs.AITraining

Piper: A Programmable Distributed Training System

Piper, a new programmable distributed training system, has been introduced to enhance large-scale model training by allowing users to define high-level parallelism strategies through minimal model annotations and scheduling directives. It utilizes an intermediate representation (IR) to compile execution plans, maintaining performance parity with established strategies like ZeRO while enabling improved efficiency through advanced scheduling techniques like DeepSeek-V3's DualPipe. This flexibility is significant for practitioners as it simplifies the integration of state-of-the-art parallelism strategies and optimizations into their training workflows.

distributed trainingparallelismmodel training

2 h agoarXiv cs.AIResearch

Flaws in the LLM Automation Narrative

The paper introduces a novel benchmarking task for Large Language Models (LLMs) that involves writing computer code for data analysis, contrasting the performance of a leading LLM with human expert submissions. The findings indicate that human experts outperform the LLM on various metrics, exhibiting lower variability and fewer errors, highlighting critical shortcomings in existing LLM evaluation methods that fail to account for performance reliability and error magnitude. This underscores the necessity for practitioners to adopt more rigorous benchmarking approaches when assessing LLM capabilities, particularly in high-stakes applications.

llmbenchmarkinghuman performance

2 h agoarXiv cs.AITraining

Provenance-Grounded Gating and Adaptive Recovery in Synthetic Post-Training Data Curation

The study presents a controlled evaluation of provenance-grounded gating and adaptive recovery in synthetic post-training data curation, demonstrating that grounding filtering signals in source evidence enhances the performance of reward models. Key findings indicate that integrating failure diagnosis with targeted regeneration in an adaptive recovery pipeline significantly improves yield and recovery rates compared to naive resampling. This research is critical for practitioners as it highlights the importance of source provenance in filtering and offers a systematic approach to recover rejected samples, potentially enhancing the quality of fine-tuning data for LLMs.

data curationsynthetic datallm

2 h agoarXiv cs.AIAgents

TRACE: A Unified Rollout Budget Allocation Framework for Efficient Agentic Reinforcement Learning

TRACE (Tree Rollout Allocation for Contrastive Exploration) is a new framework designed for efficient rollout budget allocation in reinforcement learning with verifiable rewards (RLVR), enhancing reasoning and agentic behavior in large language models. It introduces a tree-structured rollout approach that allocates budget not only to prompt roots but also to intermediate prefixes, improving reward contrast and policy-update signals. Empirically, TRACE demonstrates a 2.8-point accuracy improvement in Qwen3-14B Multi-Hop QA benchmarks at equal sampling costs, making it a significant advancement for practitioners focused on optimizing multi-turn agentic reinforcement learning strategies.

reinforcement learningbudget allocationagentic behavior

2 h agoarXiv cs.AIResearch

Towards Autonomous Accelerator Design: FPGA Accelerator Generation with SECDA

The paper presents SECDA-DSE, an advanced framework that integrates Large Language Models into the SECDA ecosystem to automate the design space exploration of FPGA-based accelerators. It successfully generates and executes three accelerator designs—element-wise vector multiplication, 2D convolution, and matrix transpose—demonstrating the capability of LLMs to optimize architectural configurations while significantly reducing exploration time and reliance on domain expertise. This development is significant for practitioners as it enhances the efficiency of designing hardware accelerators for AI workloads, potentially accelerating deployment cycles and improving performance optimization strategies.

fpgaaccelerator designllm

2 h agoarXiv cs.AISafety

Designed by Journalists, but Is It for Readers? Rethinking AI Disclosures and Transparency in News

The article discusses the challenges journalists face in disclosing AI involvement in news content, highlighting a study with 34 participants that found detailed disclosures can decrease trust due to a "transparency dilemma." The research suggests that one-line disclosures, while simpler, may create an information gap, leading readers to seek further clarification. Recommendations for better disclosure include user-centered designs such as detail-on-demand interactions and explicit "no AI" labels, emphasizing the need for improved communication strategies that align with reader expectations and agency.

ai transparencyjournalismtrust

2 h agoarXiv cs.AIMultimodal

FADA: Accessible fetal ultrasound interpretation and annotation with a selectively distilled unified vision-language model

FADA, a unified vision-language model based on Qwen3.5-VL, has been released to enhance fetal ultrasound interpretation and annotation, addressing the shortage of trained sonographers in low- and middle-income countries. FADA integrates clinical interpretation, classification, detection, and segmentation in a single pipeline using selective distillation from four domain-specific models, achieving a mean Dice score of 0.8820 for segmentation and 0.7671 mAP@0.50 for detection. Its 0.8B parameter model is optimized for edge deployment, operable on consumer hardware like the Qualcomm Snapdragon 7 Gen 1, making AI-assisted fetal assessment accessible in resource-constrained environments.

ultrasoundvision-languageclinical interpretation

2 h agoarXiv cs.AISafety

PhantomBench: Benchmarking the Non-existential Threat of Language Models

PhantomBench, a new benchmark introduced in arXiv:2606.11105v1, evaluates the hallucination rates of 21 language models across diverse domains using over 60,000 non-existent terms and entities. The benchmark reveals alarmingly high hallucination rates, with averages reaching 86.7%, highlighting the inability of even advanced models to recognize non-existent concepts. This tool not only aids in assessing model behavior regarding rare concepts but also provides a scalable pipeline for generating tailored non-existent concepts, which is crucial for practitioners aiming to mitigate risks associated with model hallucinations.

hallucinationslanguage modelsbenchmarking

2 h agoarXiv cs.AIAgents

RoboNaldo: Accurate, Stable and Powerful Humanoid Soccer Shooting via Motion-Guided Curriculum Reinforcement Learning

RoboNaldo is a novel three-stage motion-guided curriculum reinforcement learning framework designed for humanoid soccer shooting, enhancing stability and accuracy. It employs a single human-kick reference to progressively optimize shooting performance, achieving a 48.6% reduction in free-kick shot error and a 2.96x increase in shoot velocity compared to existing baselines. In real-world tests on a Unitree G1 robot, RoboNaldo demonstrated an average target shooting error of 0.73 m and a post-contact ball velocity of 13.10 m/s, indicating significant advancements in robotic soccer performance that can inform practitioners in robotics and RL applications.

reinforcement learninghumanoidsoccer

2 h agoarXiv cs.AIAgents

Test-Time Gradient Guidance of Flow Policies in Reinforcement Learning

The paper introduces QGF (Q-Guided Flow), a novel reinforcement learning algorithm that optimizes policies entirely at test time by leveraging a pre-trained reference flow policy and a value function critic. QGF demonstrates superior performance on single-task and goal-conditioned offline RL benchmarks, outperforming existing test-time methods while remaining competitive with state-of-the-art training-time algorithms, all while being more computationally efficient. This approach offers a stable and scalable alternative for practitioners working with high-dimensional action spaces in RL, minimizing the instability associated with traditional actor-critic training.

reinforcement learningpolicy optimizationtest-time

2 h agoarXiv cs.AITraining

Unifying Local Communications and Local Updates for LLM Pretraining

The paper introduces GASLoC, a decentralized pre-training algorithm for large language models (LLMs) that enhances communication efficiency by allowing local optimizer steps and utilizing gossip-based peer communication. It demonstrates superior performance over existing decentralized methods, particularly in heterogeneous bandwidth scenarios, and achieves competitive results with DiLoCo while enabling multiple local updates. This advancement is significant for practitioners as it optimizes LLM training across distributed environments, alleviating bottlenecks associated with synchronous All-Reduce operations.

llmpretrainingcommunication

2 h agoarXiv cs.AIMultimodal

Modeling Complex Behaviors: Multi-Personality Composition and Dynamic Switching in Vision-Language Models

The paper introduces a novel framework for explicit personality conditioning in Multimodal Large Language Models (MLLMs), focusing on single-personality induction, multi-personality induction, and personality switching. Experimental results indicate that while personality induction enhances image captioning performance, it can detrimentally affect visual question answering tasks. This research highlights the intricate interplay of personality traits in MLLMs and emphasizes the necessity for specialized methods for effective personality modeling and evaluation, with code to be released upon acceptance.

personalityvision-languagemodel behavior

2 h agoarXiv cs.AIModels

T1-Bench: Benchmarking Multi-Scenario Agents in Real-World Domains

T1-Bench is a newly introduced benchmark designed to evaluate agentic systems in complex, multi-domain environments, addressing limitations in existing benchmarks regarding task complexity and realism. It encompasses 25 diverse domains and features interleaved scenarios that require structured reasoning and multi-turn interactions, assessed through 12 models, including both proprietary and open-weight variants. This benchmark enhances the evaluation of agent behavior and tool utilization, and will be publicly available as open source, providing a standardized framework for researchers and practitioners in the field of AI.

benchmarkingagentsmulti-domain

2 h agoarXiv cs.AIMultimodal

AuRA: Internalizing Audio Understanding into LLMs as LoRA

AuRA introduces a novel method for integrating audio understanding directly into large language models (LLMs) via a lightweight audio embedding layer and layer-wise distillation from an ASR encoder to a LoRA-adapted LLM. This approach allows for tighter speech-language joint modeling and efficient parallel inference, outperforming traditional cascaded systems and large-scale multimodal models on various benchmarks. Practitioners can leverage AuRA to enhance LLM capabilities with audio inputs without incurring the costs of extensive multimodal training.

llmaudio understandinglora

2 h agoarXiv cs.AIAgents

Diffusion Forcing Planner: History-Annealed Planning with Time-Dependent Guidance for Autonomous Driving

The Diffusion Forcing Planner (DFP) introduces a novel diffusion-based framework for autonomous driving that addresses temporal inconsistency in motion planning. By decomposing trajectories into history, current, and future segments with independent noise levels, DFP utilizes classifier-free guidance for controlled future sampling, leading to enhanced stability and adaptability in dynamic environments. This approach demonstrates competitive performance on the nuPlan benchmark, offering practitioners a method to produce more reliable and context-aware motion plans in real-time driving scenarios.

autonomous drivingmotion planningdiffusion forcing planner

2 h agoarXiv cs.AIAgents

Understanding and mitigating the risks of OpenClaw for non-technical users: A practical guide with Skill

The article presents a practical guide aimed at non-technical users of the OpenClaw AI agent framework, addressing significant risks associated with its autonomous capabilities. It identifies and categorizes seven core risks, providing clear explanations and actionable defensive strategies for each, alongside the introduction of a companion OpenClaw Skill that automates essential security configurations. This initiative is crucial for enabling a broader user base to engage safely with AI technologies, emphasizing that risk mitigation is accessible to all users, not just experts.

openclawai agentsrisk mitigation

2 h agoarXiv cs.AITraining

Beyond Uniform Token-Level Trust Region in LLM Reinforcement Learning

The paper introduces Cumulative Prefix-divergence Policy Optimization (CPPO), a novel reinforcement learning approach that addresses the limitations of existing PPO-style trust-region methods in large language models (LLMs) by implementing a token-level masking rule. CPPO utilizes a position-weighted threshold and a cumulative prefix budget to manage token-level deviations dynamically, enhancing training stability and improving reasoning accuracy across various model scales. This advancement is significant for practitioners as it provides a more nuanced mechanism for managing autoregressive generation and mitigating compounding errors in LLMs.

reinforcement learningpolicy optimizationLLM

2 h agoarXiv cs.AISafety

Generative Explainability for Next-Generation Networks: LLM-Augmented XAI with Mutual Feature Interactions

The paper introduces a novel explainable AI (XAI) framework that integrates a moderately sized large language model (LLM) with mutual feature interactions to enhance the interpretability of AI/ML models in network operations. This framework extends traditional SHAP feature influence values and demonstrates a 12.2% improvement in explanation usefulness and a 6.2% increase in scope over a baseline approach, achieving 97.5% correctness in an optical quality of transmission (QoT) estimation task. This advancement is significant for practitioners as it provides more actionable insights for non-specialists, fostering greater trust in AI systems deployed in critical network environments.

explainabilityXAInetwork operations

2 h agoarXiv cs.AIOpen Source

Democratising Camera Trap AI: An Open-Source Model for Detecting UK Mammals

An open-source object detection model has been released for identifying 31 classes of UK mammals and birds, utilizing a YOLO26x architecture trained on a curated dataset of 48,165 labeled instances. The model achieves a mean Average Precision of 0.984 at IoU 0.5 and demonstrates high precision (0.988) and recall (0.965), with a minimal false-negative rate of 0.17%. This initiative aims to democratize access to AI tools for ecologists, providing a non-commercial alternative to proprietary models while supporting real-time camera applications.

open-sourceobject detectionbiodiversity

2 h agoarXiv cs.AIResearch

Provenance Tracking in AI Compilers through the Lens of Coalgebra

The paper presents a novel approach to provenance tracking in AI compilers, utilizing a lightweight generative method based on observational semantics and formalized through coalgebraic models and bisimulation. This method allows for reliable tracking of tensor and operator provenance without invasive changes to the compilation process, even in the presence of non-injective graph rewrites. The prototype AI compiler, COVAN, showcases stable provenance maintenance across compilation pipelines, which is critical for postprocessing, debugging, and validating transformations in AI systems.

provenanceAI compilerstracking

2 h agoarXiv cs.AIInference

CLP: Collocation-Length Prediction for Zero-Loss Adaptive Multi-Token Inference

The paper introduces CLP (Collocation-Length Predictor), a novel approach for enhancing multi-token prediction (MTP) in large language models by mitigating head-backbone competition during autoregressive decoding. CLP employs a lightweight span-level decision layer with only 4.6K–7.7K parameters, achieving speedups of 1.20x–1.29x on 1.5B Qwen2.5 models and 1.14x–1.20x on 7B models without quality degradation (repetition ratio < 0.02), compared to prior gate-based methods that showed significant quality loss. This work provides a roadmap for improving MTP head prediction accuracy, critical for accelerating inference in large-scale models.

multi-tokeninferencelanguage models

2 h agoarXiv cs.AIResearch

Recoverable but Not Stationary:Local Linear Structures in Weights and Activations

The paper presents findings on the local linear structures within the weights and activations of pretrained models, specifically using DistilGPT-2 and GPT-2 with LoRA adapters. It reveals that learned behaviors can be manipulated through linear directions, but these structures are dynamic rather than fixed, with the useful basis evolving significantly within a short training period. This work enhances understanding of parameter perturbations and activation steering, indicating that effective random parameter search can be justified in high-dimensional spaces, which is crucial for practitioners optimizing model performance.

linear structuresweightsactivation

2 h agoarXiv cs.AICoding

A Constrained Natural-Language Interface for Variational Multi-Physics Finite Element Simulations in FEniCS

The article presents a constrained natural-language interface for variational multi-physics finite element simulations using the FEniCS framework, focusing on structured JSON parsing and Gmsh code generation for non-catalog geometries. The system employs a deterministic dispatcher that maps validated specifications to five human-written FEniCS/UFL templates, achieving sub-percent agreement with analytical solutions for smooth cases and 2-5 percent for nonlinear cases. This approach mitigates reliability risks associated with LLM-generated solver code, providing a robust solution for practitioners needing reliable finite element analysis without open-ended code generation.

natural languagefinite elementLLM

2 h agoarXiv cs.AIResearch

What Do Deepfake Speech Detectors Actually Hear?

The paper presents an audio-native explainability pipeline for deepfake speech detectors, utilizing Integrated Gradients on time-aligned self-supervised representations to localize decision evidence in audio signals. The approach was tested on three WavLM-based detectors (AASIST, CA-MHFA, SLS) using the ASVspoof 5 dataset, revealing that while performance is similar across models, they rely on distinct cues for detection. This work provides valuable insights into the operational semantics of these detectors, which can inform practitioners about the specific features that influence detection outcomes and improve model interpretability in real-world applications.

deepfakeexplainabilityaudio

2 h agoarXiv cs.AIResearch

Ethical and Technical Limits of Deepfake Speech Datasets

A comprehensive audit of 39 deepfake speech datasets has been published, revealing significant limitations in their accessibility, documentation, and demographic coverage. The analysis highlights that most datasets lack essential demographic metadata, hindering fairness assessments and subgroup analyses, while also noting substantial overlap in bona fide source corpora, which may compromise the validity of cross-dataset evaluations. This work underscores the need for more robust dataset standards to enhance the reliability of deepfake speech detection systems for practitioners.

deepfakedatasetsfairness

2 h agoarXiv cs.AISafety

RAT: Reference-Augmented Training for ASV Anti-Spoofing

The paper introduces Reference-Augmented Training (RAT), a novel spoofing countermeasure architecture that leverages speaker-reference recordings to enhance anti-spoofing performance. RAT achieves state-of-the-art results on the ASVspoof 5 benchmark, with an Equal Error Rate (EER) of 2.57% and a minimum Detection Cost Function (minDCF) of 0.074, outperforming traditional single-utterance baselines and large ensemble systems. This method is significant for practitioners as it demonstrates that training with reference channels can lead to improved deepfake detection capabilities, even when references are not available during inference.

anti-spoofingtrainingdeepfake

2 h agoarXiv cs.AIAgents

Human-AI Teaming Through the Lens of Calibration

The paper presents a study on human-AI teaming focusing on statistical calibration, analyzing how calibration affects the integration of human and AI model predictions. It distinguishes between two frameworks: one that combines predictions and another that delegates responsibility, revealing that existing combination methods fail to maintain human calibration, while delegation shifts the calibration burden to a rejector meta-model, which must be finely calibrated to determine the superior predictor. This research is critical for practitioners as it highlights calibration challenges in designing effective human-AI collaboration systems, emphasizing the need for robust calibration strategies to ensure reliable decision-making.

human-AIcalibrationteamwork

2 h agoarXiv cs.AIResearch

Pose-ICL: 3D-Aware In-Context Learning for Pose-Controllable Subject Customization

Pose-ICL is a novel framework that introduces 3D-aware In-Context Learning (ICL) for subject customization in image generation, enabling effective pose control through multiple paired image-pose references. It employs Surface-Anchored Position Embedding (SAPE) to enhance 3D awareness by anchoring image tokens to volumetric bounding box coordinates, ensuring compatibility with existing DiT models. This approach significantly improves pose accuracy and identity consistency over current methods, addressing critical limitations in 2D-native architectures for practitioners focused on advanced image generation techniques.

3Dpose controlimage generation

2 h agoarXiv cs.AIMultimodal

Improving Text-Instance Alignment Of Foreground Conditioned Out-Painting Via Customized Concept Embedding

The article presents the Customized Concept Embedding Diffusion (CCE-Diffusion) framework, designed to improve Foreground Conditioned Outpainting (FCO) by addressing artifact issues in synthesized backgrounds. The CCE-Module customizes concept embeddings to better align with specific visual instances, utilizing an Instance-Aware Loss for optimization and a Semantic-Preserving Prompt Template to maintain prompt integrity. This framework can be integrated into existing FCO methods, significantly enhancing output quality and reducing artifacts, which is crucial for practitioners aiming to create high-quality display images efficiently.

outpaintingimage generationembedding

2 h agoarXiv cs.AIInference

Optimal Post-Training Quantization Scales and Where to Find Them

The paper introduces PiSO (Piecewise Scale Optimization), an algorithm designed to enhance post-training quantization (PTQ) by calculating optimal channel-wise weight scales using calibration data, rather than simple heuristics. It efficiently partitions the scale search space for quantization and extends to group-wise quantization with error correction strategies. Experimental results on Llama and Qwen models indicate significant improvements in perplexity and zero-shot accuracy, particularly for lower bit-widths, highlighting the method's relevance for practitioners aiming to optimize model performance under quantization constraints.

quantizationlarge language modelscompression

2 h agoarXiv cs.AIMultimodal

LIBERO-Occ: Evaluating and Improving Vision-Language-Action Models under Scene-Induced Occlusion via Viewpoint Imagination

The paper introduces LIBERO-Occ, an extension of the LIBERO framework aimed at addressing performance degradation of Vision-Language-Action (VLA) models under scene-induced occlusion. It presents a novel technique called Viewpoint Imagination (VIM), which generates complementary views to enhance action prediction without requiring additional cameras, demonstrating improved robustness across various task suites and occlusion scenarios. This advancement is significant for practitioners as it enhances the reliability of VLA models in real-world applications where occlusion is common.

vision-languageocclusionaction prediction

2 h agoarXiv cs.AIProducts

From Perception to Action: Can UI Interventions Foster Sustainable LLM Chatbot

The study presents a web-based chatbot prototype that incorporates sustainability-oriented UI interventions aimed at fostering energy-aware interactions with LLM-powered chatbots. Key features include a three-mode switch (Energy-efficient, Balanced, Performance), real-time energy feedback, pre-send energy estimates, and a usage metrics dashboard. The findings indicate that while users are concerned about AI energy consumption, they are often unaware of actual usage; however, the prototype's design successfully encouraged users to select energy-efficient modes without compromising usability, suggesting that UI design can play a critical role in promoting sustainable practices in AI applications.

llmsustainabilityuser interface

2 h agoarXiv cs.AISafety

Janus: A Benchmark for Goal-Conditioned Information Distortion in LLMs

The article announces the release of JANUS, a benchmark designed to evaluate goal-conditioned information distortion in large language models (LLMs). It consists of 160 scenarios across 8 domains, comparing neutral and goal-directed prompts using a fixed pool of factual information to assess how models distort facts. This benchmark is significant for practitioners as it highlights the vulnerability of LLMs to producing misleading outputs based on framing and incentives, underscoring the need for improved safeguards against such distortions in AI applications.

llmbenchmarkdeception

2 h agoarXiv cs.AITraining

Geometrically Averaged Hard Target Updates for Linear Q-Learning

The paper introduces the $\lambda$-target update mechanism for linear Q-learning, which employs geometrically averaged hard target updates to enhance stability in Q-learning with function approximation. The method allows for a continuum of target update strategies, ranging from one-period updates to projected Q-value iteration, and is analyzed using a switching-system model. This advancement is significant for practitioners as it provides a more flexible and potentially more stable approach to target updates in reinforcement learning, particularly in environments where linear function approximation is utilized.

q-learningreinforcement learningtarget updates

2 h agoarXiv cs.AIResearch

Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models

The article introduces the Attention-Discounted Adaptive Sampler (ADAS), a training-free reranking method designed to enhance parallel decoding in masked diffusion language models. ADAS modifies subset construction by applying a soft marginal penalty based on attention, improving low-NFE performance by an average of 9.11 and 10.46 percentage points when integrated with existing samplers like Top-k, Fast-dLLM, and EB-Sampler, while maintaining a runtime overhead of 3.1%. This advancement is significant for practitioners as it provides a modular approach to optimize inference quality in masked diffusion models without altering the base sampler's stopping rules.

masked diffusionlanguage modelssampling

2 h agoarXiv cs.AIResearch

A Unified Siamese Learning Framework for Zero-Day Anomaly Detection and Classification in Optical Networks

A multi-similarity Siamese neural network framework has been proposed for zero-day anomaly detection and one-shot classification in optical networks, achieving over 99% accuracy. This model demonstrates instant adaptability to new lightpaths and previously unseen anomaly types without requiring retraining. This is significant for practitioners as it enables real-time anomaly detection in dynamic optical environments, reducing downtime and operational costs.

anomaly detectionoptical networksSiamese network

2 h agoarXiv cs.AIInference

K-Forcing: Joint Next-K-Token Decoding via Push-Forward Language Modeling

K-Forcing is a new push-forward language modeling paradigm introduced for joint next-k-token decoding, aiming to enhance the efficiency of autoregressive (AR) models in high-load batch serving scenarios. By distilling an existing AR model into a mapping that generates multiple future tokens in a single forward pass, K-Forcing achieves a speedup of approximately 2.4-3.5x when configured to generate 4 tokens simultaneously, while utilizing a standard causal Transformer architecture. This approach addresses the growing inference costs associated with large language models, making it a significant advancement for practitioners focusing on optimizing deployment efficiency.

language modelingdecodingefficiency

2 h agoarXiv cs.AIMultimodal

Earth-OneVision: Extending Remote Sensing Multimodal Large Language Models to More Sensor Modalities and Tasks

Earth-OneVision is a newly introduced 2 billion parameter remote sensing multimodal large language model (RS-MLLM) that integrates six sensor modalities—optical, SAR, infrared, multispectral, temporal, and video—within a single autoregressive framework, significantly broadening the scope of tasks it can address. It employs three innovative mechanisms: Full-Granularity Vision-Language Alignment (FGVLA), Spatial-Linguistic Isomorphic Serialization (SLIS), and Progressive Cross-Modality Adaptation (PCMA), to enhance performance and facilitate joint training with a dataset of approximately 34 million QA pairs. The model achieves competitive performance, surpassing larger RS-MLLMs (4B-72B parameters) across multiple benchmarks, making it a valuable tool for practitioners focused on cross-modal geoscientific applications and advancing remote sensing capabilities.

remote sensingLLMsensor modalities

2 h agoarXiv cs.AIAgents

Beyond APIs: Probing the Limits of MLLMs in Physical Tool Use

The paper introduces PhysTool-Bench, a benchmark designed to assess Multimodal Large Language Models (MLLMs) in their ability to recognize and utilize physical tools in real-world scenarios, comprising 2,510 queries across 2,678 tools from various domains. Evaluation of 13 leading MLLMs, including Gemini-3.1-Pro, reveals that even the best model only identifies 58.7% of tools and successfully completes 21.0% of tasks, highlighting significant deficits in both perception and planning capabilities. This work underscores the challenges MLLMs face in functional commonsense reasoning, which are critical for advancing embodied AI applications.

MLLMphysical tool usebenchmark

2 h agoarXiv cs.AITraining

Boosting ECG Classification Performance by Pre-training with Synthesized Data

The study presents a novel Gaussian-composition synthesis algorithm for generating synthetic single-lead II ECG data to enhance the classification of abnormal ECGs, specifically targeting atrial fibrillation, atrial flutter, premature ventricular complex, and Wolff-Parkinson-White Syndrome. Evaluating ten different deep neural network architectures, the research shows an average classification performance improvement of 33.2% for atrial flutter when using synthetic data, particularly benefiting scenarios with limited real-world datasets. This approach highlights the potential of domain-specific synthetic data as a pre-training resource, addressing data scarcity challenges in medical applications.

ECGsynthetic dataclassification

2 h agoarXiv cs.AIResearch

Dep-LLM: Training-Free Depression Diagnosis via Evidence-Guided Structured Multi-factor with Reliable LLM Reasoning

Dep-LLM is a novel training-free framework for Automatic Depression Detection (ADD) that leverages frozen foundation LLMs to mimic clinical reasoning. It features a three-stage process, including Chain-of-Thought (CoT) analysis for thematic decomposition, a Confidence Analysis module for reliability quantification, and a Collaborative Multi-factor Prediction for diagnosis integration. Experimental results on the DAIC-WOZ and E-DAIC datasets show that Dep-LLM outperforms both zero-shot baselines and state-of-the-art supervised models across multiple metrics, highlighting its potential for practical deployment in clinical settings without the need for extensive training.

depression diagnosisLLMclinical interviews

2 h agoarXiv cs.AISafety

Toward Secure LLM Agents: Threat Surfaces, Attacks, Defenses, and Evaluation

The paper discusses the evolving security landscape of large language model (LLM) agents as they transition from conversational interfaces to more autonomous software components. It synthesizes findings from 247 studies, identifying key threats such as prompt injection and tool-mediated control-flow hijacking, while highlighting emerging concerns like persistent state corruption. The authors advocate for improved security frameworks that emphasize trust boundaries, privilege control, and realistic evaluation practices, which are crucial for practitioners developing secure LLM applications.

LLM agentssecurityattacks

2 h agoarXiv cs.AIMultimodal

Spatial-Omni: Spatial Audio Understanding Integration in Multimodal LLMs via FOA Encoding

Spatial-Omni introduces a novel method for integrating First-Order Ambisonics (FOA) spatial audio into existing multimodal large language models (LLMs) using the SO-Encoder, which allows for enhanced spatial audio understanding without altering original audio encoders. The approach includes the creation of the SO-Dataset, SO-QA, and SO-Bench, comprising 400K FOA spatial audio clips and 2.1M spatial question-answer pairs, covering 16 subtasks in spatial audio understanding. This development is significant for practitioners as it enhances the capability of LLMs to process spatial audio cues, improving applications in sound localization and spatial reasoning.

spatial audioLLMFOA encoding

2 h agoarXiv cs.AIResearch

Detecting Knowledge Gaps from Conversational AI Interactions Using Curriculum Prerequisite Graphs

The study presents a pipeline that utilizes a few-shot text classifier to map student questions directed at a conversational AI teaching assistant to specific curriculum topics, leveraging a prerequisite knowledge graph extracted from GPT-4. The classifier achieved 80.0% accuracy across 43 labels when evaluated on 1,340 question events from 164 graduate students, with significant correlation (rho = 0.491, p = 0.008) between topic-level question volume and self-reported difficulty. This approach highlights the potential of conversational AI interaction logs as diagnostic tools for identifying knowledge gaps, offering instructors insights into which topics require further focus.

conversational AIcurriculumknowledge gaps

2 h agoarXiv cs.AIResearch

Transformer Based Model for Spatiotemporal Feature Learning in EEG Emotion Recognition

EEG-TransNet, a new transformer-based architecture for EEG emotion recognition, has been introduced, featuring a preprocessing module utilizing ResNet and wavelet denoising, a Local Self-Attention Block for regional feature extraction, and a Fuzzy-Attention Synchronous Transformer (FAST) for modeling spatiotemporal dependencies. Extensive evaluations on the BETA, SEED, and DepEEG datasets demonstrate superior classification accuracy and robustness compared to existing methods, with ablation studies highlighting the effectiveness of the Local Self-Attention Block and the computational efficiency achieved through depthwise separable convolutions in the decoder. This model's ability to generalize across subjects with minimal performance variation positions it as a valuable tool for practitioners in EEG-based emotion recognition and brain activity analysis.

EEGemotion recognitiontransformer

2 h agoarXiv cs.AIResearch

Attention Expansion: Enhancing Keyphrase Extraction from Long Documents with Attention-Augmented Contextualized Embeddings

The paper presents an attention expansion mechanism that enhances keyphrase extraction (KPE) from long documents by augmenting pre-trained language model (PLM) token representations with information from surrounding out-of-context chunks using pre-trained word embeddings. Evaluated across five PLM backbones and multiple benchmark corpora, the approach demonstrates significant improvements in F1 score over state-of-the-art models, indicating its effectiveness in addressing the limitations of traditional PLMs and long-context LLMs for efficient KPE. This mechanism offers a practical solution for practitioners needing to extract salient information from lengthy documents without incurring the high computational costs associated with full-document attention.

keyphrase extractionattention mechanismlong documents

2 h agoarXiv cs.AITraining

++nnU-Net: Scaling nnU-Net with Prefix-Based Data Augmentation

The ++nnU-Net introduces a novel data augmentation module utilizing image registration to enhance the nnU-Net framework for medical segmentation tasks. Evaluated on five 2D datasets, the ++nnU-Net demonstrates significant improvements in Dice Similarity Coefficient scores, with gains of up to 22% over the baseline nnU-Net. This approach addresses the challenges of data scarcity in medical imaging by generating new warped images and supplementary binary masks, offering a scalable solution for practitioners working in data-limited environments.

data augmentationmedical imagingnnU-Net

2 h agoarXiv cs.AITraining

Effective Reinforcement Learning for Agentic Search by Recycling Zero-Variance Queries During Training

The paper introduces a novel technique called query recycling for GRPO-style algorithms in reinforcement learning, which allows zero-variance queries to be resampled during training, enhancing their utility as the policy evolves. A 1.7B parameter model trained on synthetic data achieved a 66.0 average Pass@1 accuracy across seven multi-hop QA benchmarks, outperforming larger models with up to 7B parameters. This approach is significant for practitioners as it optimizes training efficiency by dynamically adapting the training distribution, potentially leading to more effective learning from limited query resources.

reinforcement learningzero-variancetraining

2 h agoarXiv cs.AITraining

Unifying Data, Memory, and Compute Efficiency in LLM training: A Survey

The survey presents a constraint-centric framework for improving efficiency in large language model (LLM) training by addressing the interconnected challenges of data efficiency, memory efficiency, and compute budget awareness. It reviews advanced techniques for data selection and pruning, highlights the dominance of GPU memory as a bottleneck, and emphasizes the need for coordinated optimization of weight storage, optimizer states, and activation memory. This unified approach offers practitioners guidelines for resource-conditioned decision-making, optimizing training and inference processes within finite computational budgets.

LLMdata efficiencymemory efficiency

2 h agoarXiv cs.AIAgents

Event-Driven Reinforcement Learning Enables Long-Horizon Control in Semiconductor Fabrication

The authors introduce a deep reinforcement learning framework tailored for optimizing multi-objective control in semiconductor fabrication, addressing the challenges of high-dimensional decision-making and long-horizon requirements. The framework employs a centralized-agent model with an event-driven temporal-difference formulation, allowing integration with various policy optimization methods. Validation experiments demonstrate substantial improvements in throughput and utilization, highlighting the framework's scalability and applicability to complex adaptive systems, which is crucial for practitioners aiming to enhance production efficiency in stochastic environments.

reinforcement learningsemiconductorcontrol

2 h agoarXiv cs.AIInference

Using the YOLOv12 Model for Verifying the Correct Color Sequence of Wires in Network Cables (Patch Cords) on the Production Line

The study introduces an intelligent inspection system utilizing the YOLOv12 object detection model to verify the correct color sequence of wires in network patch cords during production. The model, trained on a dataset of 2,500 microscopic images, achieved approximately 98% precision and an overall mean accuracy of 95%, significantly improving inspection efficiency by minimizing human error. This advancement is crucial for practitioners as it automates quality control in manufacturing, enhancing product reliability and reducing costs associated with defective cables.

YOLOobject detectionproduction line

2 h agoarXiv cs.AIAgents

Divide and Cooperate: Role-Decomposed Multi-Agent LLM Training with Cross-Agent Learning Signals

The paper introduces DAC (Divide and Cooperate), a role-decomposed multi-agent training framework that separates the tasks of evidence acquisition and answer generation into distinct agents, mitigating the combinatorial policy space and credit assignment issues present in traditional single-policy models. DAC employs parameter-efficient LoRA modules over a shared backbone, demonstrating improved performance on general and multi-hop QA benchmarks compared to full fine-tuning approaches. This framework offers practitioners a more efficient method for training LLMs in complex reasoning tasks by leveraging structured cross-agent learning signals.

multi-agenttrainingrole-decomposed

2 h agoarXiv cs.AIResearch

UniDexTok: A Unified Dexterous Hand Tokenizer from Real Data

The paper introduces UniDexTok, a retargeting-free state tokenizer that utilizes a Unified Dexterous Hand Model (UDHM) to create a shared 22-DoF semantic interface for dexterous hands. UniDexTok achieves significant performance improvements over the baseline UniHM, reducing the Mean Per Joint Angle Error (MPJAE) from 15.63 degrees to 0.16 degrees and the Mean Per Joint Position Error (MPJPE) from 18.51 mm to 0.18 mm, enhancing reconstruction accuracy from centimeter-scale to sub-millimeter. This advancement is crucial for practitioners as it enables more effective cross-embodiment learning and reduces reliance on simulation data, facilitating better integration of diverse dexterous hand designs in robotic applications.

dexterous handstokenizerrobotics

2 h agoarXiv cs.AIResearch

In Defense of Information Leakage in Concept-based Models

The paper presents a critical reassessment of information leakage in concept-based models (CMs), arguing that some degree of leakage is beneficial for model performance in real-world scenarios where concepts are often incomplete. The authors propose an optimized training objective that allows CMs to harness "benign leakage," enhancing accuracy and intervenability without compromising interpretability. This perspective is significant for practitioners as it challenges the prevailing view on leakage and suggests new strategies for developing more effective CMs in practical applications.

concept-based modelsinformation leakageinterpretability

2 h agoarXiv cs.AIAgents

Post-Quantum Secure Federated DeFi for Inclusive Banking

The paper presents a post-quantum secure federated DeFi framework designed to enhance financial inclusivity for underserved individuals by enabling inter-bank collaboration. It utilizes lattice-based Fully Homomorphic Encryption (FHE) for end-to-end computation on encrypted data, integrating contributions from multiple banks and leveraging the NASA-IBM Prithvi Geospatial Foundation Model (GFM) for probabilistic assessments. This framework is significant for practitioners as it addresses the vulnerabilities of current cryptographic systems to quantum attacks and provides a robust method for secure data exchange in decentralized finance.

multi-agent-systemsdecentralizationllm

2 h agoarXiv cs.AIModels

Dynamic Linear Attention

The paper introduces Dynamic Linear Attention (DLA), a framework designed to enhance multi-state linear attention mechanisms for Large Language Models (LLMs) by implementing Information-Aware Dynamic State Merging and Capacity-Bounded Memory Modeling. DLA adaptively merges states based on token importance, resulting in improved representation capacity for long contexts while maintaining a fixed-size memory cache. Experimental evaluations across 16 datasets show DLA outperforms existing state-of-the-art methods, making it a significant advancement for practitioners aiming to optimize LLM performance in long-context scenarios.

linear-attentionllmscalability

2 h agoarXiv cs.AISafety

Is Fairness Truly Fair? Towards Reliable Lipschitz Fairness in Multi-Task Learning via Fixed-\texorpdfstring{$\delta$}{delta} Alignment

The paper introduces ReLiF, a framework designed to enhance Lipschitz fairness evaluation in multi-task learning (MTL) by employing fixed-$\delta$ auditing to mitigate threshold confounding. It utilizes a shared reference tolerance for auditing and a violation-rate feedback controller, facilitating a balance between fairness and utility during training. Experiments demonstrate that ReLiF, when applied to clinical time-series data and NYUv2 using a ResNet50 backbone, achieves competitive utility while significantly reducing bias, thereby providing a more reliable method for assessing fairness in MTL contexts.

fairnessmulti-task-learningevaluation

2 h agoarXiv cs.AIRAG

STORM: Stepwise Token Optimization with Reward-Guided Beam Search

STORM (Stepwise Token Optimization with Reward-Guided Beam Search) is a new self-supervised framework designed for lexical query expansion, which optimizes rewriter performance by scoring candidate expansions against a BM25 index. It enables 0.6B-8B parameter models to match or exceed the effectiveness of larger LLM rewriters while maintaining the speed of traditional BM25 retrieval, and it demonstrates zero-shot transfer across 18 languages, outperforming dedicated multilingual dense retrievers. This approach offers practitioners a more efficient and infrastructure-light alternative for enhancing retrieval performance in AI systems.

query-expansionretrievalllm

2 h agoarXiv cs.AIResearch

Can Image Models Imagine Time? ImageTime: A Novel Benchmark for Probing Visual World Modeling Through Spatiotemporal Consistency

The article introduces ImageTime, a new benchmark designed to evaluate the ability of image generation models to maintain spatiotemporal consistency across multiple visual states. The benchmark requires models to generate a sequence of four keyframes representing an initial state, action onset, transition, and final state, which presents a more complex challenge than single-image generation. This diagnostic tool aims to provide insights into the temporal capabilities of current models, offering structured scoring through GPT-5.5 and revealing strengths and weaknesses in visual world modeling, which is critical for applications like storyboarding and video previsualization.

image-modelsspatiotemporal-consistencybenchmark

2 h agoarXiv cs.AITraining

Fast and Highly Expressive Policy Learning for Offline Reinforcement Learning via Bootstrapped Flow Q-Learning

The paper introduces Bootstrapped Flow Q-Learning (BFQ), a novel framework for offline reinforcement learning that facilitates accurate single-step action generation without the need for auxiliary networks or policy distillation. BFQ employs a divide-and-conquer approach to learn short-range displacements, enabling a direct noise-to-action mapping that eliminates the computational overhead associated with multi-step denoising. Evaluations on the D4RL benchmark indicate that BFQ not only enhances performance but also significantly reduces computational costs compared to traditional multi-step diffusion methods, making it a more efficient alternative for practitioners in the field.

offline-reinforcement-learningq-learningbootstrapped

2 h agoarXiv cs.AIAgents

Causal Ensemble Agent: Hierarchical Causal Discovery with LLM-guided Expert Reweighting

The Causal Ensemble Agent (CEA) framework was introduced for hierarchical causal discovery, integrating insights from various causal discovery algorithms and employing a Large Language Model (LLM) as a meta-referee for expert reweighting. CEA utilizes linear opinion pooling to aggregate structural insights and dynamically adjusts expert contributions based on confidence levels, resulting in improved causal graph accuracy. This approach demonstrates significant performance enhancements across multiple datasets, underscoring the potential of LLMs in refining causal inference methodologies for practitioners.

causal-discoveryllmexpert-reweighting

2 h agoarXiv cs.AIAgents

Dmsh: A Multi-Agent Reinforcement Learning Framework for All-Quad Mesh Generation

Dmsh is a newly introduced multi-agent reinforcement learning framework designed for fully automated all-quadrilateral mesh generation. It utilizes a Soft Actor-Critic architecture to address the meshing process as a Markov Decision Process, employing three coordinated agents for topology simplification, geometric regularization, and mesh generation, while implementing a curriculum learning strategy to enhance scalability. This framework significantly surpasses traditional methods in automation, robustness, and mesh quality, providing a novel approach for practitioners dealing with complex geometries in computational engineering.

mesh-generationreinforcement-learningmulti-agent

2 h agoarXiv cs.AIResearch

Embedding Hybrid Systems into Continuous Latent Vector Fields

The paper demonstrates that an $n$-dimensional hybrid system can be embedded into an $m$-dimensional Euclidean space with a continuous vector field when $m > 2n$, enabling a continuous extrinsic representation suitable for differentiable optimization. It introduces a latent Neural ODE framework with a consistency loss that effectively recovers the flow of hybrid systems, outperforming existing methods in learning hybrid systems with diverse geometries from time series data. This work is significant for practitioners as it provides a robust approach for modeling complex hybrid systems, enhancing the capabilities of AI systems in handling discontinuous dynamics.

hybrid-systemsneural-odeoptimization

2 h agoarXiv cs.AITraining

From Data Heterogeneity to Convergence: A Data-Centric Review of Federated Learning

The article presents a comprehensive survey on Federated Learning (FL), focusing on the impact of data heterogeneity on convergence and stability. It introduces three key advances: a ranking of non-IID data traits influencing convergence, a connection between experimental data splitting practices and their real-world implications, and an analysis of data-related vulnerabilities alongside their defenses affecting convergence under various conditions. This work provides actionable insights for practitioners, enabling better design of FL systems with predictable performance outcomes.

federated-learningdata-heterogeneityconvergence

2 h agoarXiv cs.AIResearch

Towards Diverse Scientific Hypothesis Search with Large Language Models

The paper presents a novel evolutionary framework for hypothesis generation using large language models, termed \ours, which addresses the limitations of traditional optimization-focused methods that lead to diversity collapse. By employing a parallel tempering approach, \ours enables efficient exploration of diverse, high-quality hypotheses across multiple domains, including molecular and algorithm discovery, while adhering to a fixed validation budget. This method enhances both the quality and diversity of hypotheses generated, making it a valuable tool for practitioners aiming to mitigate uncertainty in scientific discovery processes.

hypothesis-generationllmscientific-discovery

2 h agoarXiv cs.AIResearch

NOVA: Symbolic Regression Discovery of Interpretable Car-Following and Lane-Change Models with Driver Heterogeneity

NOVA, a symbolic regression framework, has been released for discovering interpretable car-following and lane-change models from raw driving data, utilizing a deterministic Rust-powered search engine to evaluate over 10,000 candidate algebraic structures. It achieved an RMSE of 1.376 m/s² on the intent-forecasting benchmark, outperforming the best symbolic-regression baseline by 0.135 m/s², and demonstrated robust transferability across freeway sites with minimal performance loss. This framework is significant for practitioners as it provides a method for deriving interpretable models that can enhance the understanding and prediction of driver behavior in autonomous systems.

symbolic-regressionautonomous-systemsdriver-behavior

2 h agoarXiv cs.AIAgents

Drawing with Strangers: Population Scaling Drives Zero-Shot Mutual Intelligibility in Emergent Sketching

The paper introduces the concept of zero-shot mutual intelligibility (ZMI) in emergent communication, demonstrating that increasing the training population size significantly enhances ZMI among independently trained groups using emergent sketching. The study reveals that larger populations lead to greater in-group communicative variation while reducing cross-group variation, indicating a trend towards universal communication methods grounded in visual perception. This work is significant for practitioners as it highlights the potential for developing socially interoperable AI agents capable of effective communication across diverse populations without prior training.

emergent-communicationzero-shot-intelligibilitymulti-agent

2 h agoarXiv cs.AIResearch

Convergence of Monte Carlo Optimistic Policy Iteration: Beyond Uniform State-Action Updates

The paper presents a significant advancement in Monte Carlo optimistic policy iteration (MC-O-PI) by proving that it can converge to optimality without requiring uniform initialization over the entire state-action space, but rather uniform updates only over actions within each state. This relaxation allows for more practical implementations in large or unknown state spaces, as episodes can start in various states at arbitrary frequencies. The findings provide a new analytical framework for studying optimistic policy iteration algorithms, which is crucial for practitioners aiming to enhance convergence properties in reinforcement learning applications.

policy-iterationreinforcement-learningconvergence

2 h agoarXiv cs.AIResearch

Improving Adversarial Transferability on Vision-Language Pre-training Models via Surrogate-Specific Bias Correction

The paper introduces DeBias-Attack, a novel method to enhance adversarial transferability in Vision-Language Pre-training (VLP) models by addressing surrogate-specific bias in adversarial optimization. This technique employs two perturbation branches: one optimizing on the original image and another on a weak-semantic image, enabling effective gradient correction and improving robustness against transfer-based attacks. The results demonstrate significant performance gains across various VLP models and tasks, highlighting its potential for practitioners focused on enhancing model robustness against adversarial threats.

adversarialvision-languagerobustness

2 h agoarXiv cs.AISafety

Hidden Consensus:Preference-Validity Compression in Human Feedback

The paper introduces the concept of Preference-Validity Compression in Reinforcement Learning from Human Feedback (RLHF), arguing that reducing diverse human judgments to a single scalar reward can misrepresent alignment in culturally plural societies. An analysis of 321 preference events from Malaysian participants reveals that 79% of prompts have multiple majority-supported responses, indicating that traditional single-winner aggregation overlooks significant interpretive diversity. The authors advocate for alignment methods that maintain Validity-Preserving Consistency to better capture plural-valid responses, emphasizing the need for more nuanced feedback aggregation in AI systems.

rlhfhuman-feedbackalignment

2 h agoarXiv cs.AIResearch

Benchmarking Knowledge Editing using Logical Rules

A new benchmark for knowledge editing in large language models (LLMs) has been introduced, focusing on evaluating the logical consequences of single fact edits rather than just recalling edited facts. This benchmark utilizes logical rules extracted from knowledge graphs to generate multi-hop questions, revealing that existing methods like ROME and FT often fall short in injecting entailed knowledge, with a performance gap of up to 24%. This underscores the necessity for semantics-aware evaluation frameworks to enhance the effectiveness of knowledge editing techniques in LLMs.

llmknowledge-editingbenchmark

2 h agoarXiv cs.AIInference

LC-QAT: Data-Efficient 2-Bit QAT for LLMs via Linear-Constrained Vector Quantization

The paper introduces LC-QAT, a novel 2-bit quantization-aware training framework for large language models that employs linear-constrained vector quantization to optimize quantized weights through a learned affine mapping. This approach allows for end-to-end differentiable training without the need for discrete codebook lookups, significantly enhancing data efficiency by achieving strong post-training initialization with only 0.1% to 10% of the training data. LC-QAT demonstrates superior performance compared to existing QAT methods, making it a viable solution for deploying extremely low-bit models in practical applications.

quantizationLLMtraining

2 h agoarXiv cs.AIResearch

Machine Learning Methods for Studying Latent Neural Activity Dynamics

The paper presents a comprehensive survey of machine learning methods for decoding latent neural activity dynamics, focusing on Latent Variable Models (LVMs) ranging from linear dynamical systems to deep generative models like Transformers and diffusion models. It categorizes the literature into single-region latent dynamics, multi-region communication, and behavior-aligned modeling, highlighting techniques such as RNNs and Neural ODEs, while addressing benchmarks and challenges in identifying causal links in neural communication. This survey is significant for practitioners as it outlines critical methodologies and evaluation metrics that can enhance neural decoding and understanding of brain dynamics.

machine learningneural activitydynamics

2 h agoarXiv cs.AISafety

Assessing Automated Prompt Injection Attacks in Agentic Environments

The paper presents an empirical evaluation of automated prompt injection attacks against LLM agents using the AgentDojo framework, applying both white-box (GCG) and black-box (TAP) methods across 80 task pairs in four domains. The study finds that black-box optimization significantly outperforms gradient-based methods due to GCG's instability, and the effectiveness of TAP is influenced by the attacker's model capabilities and safety tuning. These results emphasize the model-dependent nature of automated prompt injection threats, indicating that while task-universal attacks can transfer across domains, smaller models do not effectively inform strategies against advanced models like GPT-5, posing challenges for practitioners in securing LLM applications.

prompt injectionagentsattack

2 h agoarXiv cs.AITraining

MoE Enhanced Federated Learning for Spatiotemporal Prediction

The paper introduces MoE-FedTP, a personalized federated learning framework utilizing lightweight Mixture-of-Experts (MoE) networks for spatiotemporal traffic prediction. It employs spatiotemporal neural networks and a gating mechanism for dynamic expert fusion, achieving superior performance on four real-world traffic datasets compared to existing cross-city and federated learning methods. This approach addresses privacy concerns while improving prediction accuracy in data-scarce urban environments, making it significant for practitioners focused on enhancing intelligent transportation systems.

federated learningtraffic predictionMoE

2 h agoarXiv cs.AIInference

Achieving Cloud-Grade SLOs for Local Mixture-of-Experts Inference through CPU-GPU Hybrid Design

The paper presents a CPU-GPU hybrid system designed to enhance local deployment of large Mixture-of-Experts (MoE) models, achieving cloud-grade service level objectives (SLOs). Key innovations include stream-loading prefill (SLP) that boosts throughput to 1,200 tokens/s for 32K prompts in 30 seconds, and distributed SLP (DSLP) reaching 1,800 tokens/s for 45K prompts using two RTX 5090s. This approach allows for efficient local inference with original-precision, significantly improving throughput and reducing latency, which is critical for practitioners aiming to deploy high-performance AI models without relying on cloud infrastructure.

Mixture-of-Expertsinferencecloud

2 h agoarXiv cs.AIInference

Stop Early, Spend Less: Hidden-State Probes as a Practical Recipe for Streaming Moderation of LLM Outputs

The paper introduces a novel approach for real-time moderation of large language model outputs using lightweight token-level probes that operate on internal hidden states, eliminating the need for a separate moderation model. This method allows for sub-millisecond per-token safety checks during the decoding process, significantly reducing computational overhead compared to traditional post-generation moderation techniques. By leveraging internal activations, the proposed system enables continuous monitoring and intervention, which is crucial for deploying LLMs in user-facing applications while maintaining performance efficiency.

LLMsafetymoderation

2 h agoarXiv cs.AISafety

Advancing the State-of-the-Art in Empirical Privacy Auditing

The paper introduces a novel approach to empirical privacy auditing (EPA) for large language models (LLMs), focusing on the use of synthetic "canary" examples generated through high-temperature sampling (T ≥ 0.8) from LLMs to assess data leakage risks in membership inference and reconstruction attacks. This method enhances the ability to audit models fine-tuned on privacy-sensitive data by ensuring the canaries are high-influence outliers, allowing for effective inspection without compromising real data privacy. The proposed auditing techniques also include a framework for evaluating synthetic data privacy risks, which is critical for practitioners concerned about data leakage in LLM applications.

privacyauditingLLM

2 h agoarXiv cs.AIAgents

Decoupling Thought from Speech: Knowledge-Grounded Counterfactual Reasoning for Resilient Multi-Agent Argumentation

The paper introduces Knowledge-Grounded Counterfactual Reasoning (KG-CFR), a dual-stage architecture designed to enhance the stability of multi-agent debates by separating a private planning buffer from a public execution layer. In tests within a Dynamic Resource Allocation under Uncertainty (DRAU) environment, KG-CFR demonstrated over 95% resilience against critical post-shock degradation and improved argument quality from 0.694 to 0.822. This work is significant for practitioners as it provides a structured approach to maintaining argument integrity in the face of perturbations, potentially improving the robustness of LLM applications in dynamic environments.

multi-agentreasoningargumentation

2 h agoarXiv cs.AIResearch

Detecting Speculative Language in Biomedical Texts using Recurrent Neural Tensor Networks

The study presents an approach for detecting speculative language in biomedical texts using distributed sentence representations, specifically employing Recursive Neural Tensor Networks (RNTN) and the Paragraph Vector model. RNTN achieved an F1 score of 0.885, outperforming the best baseline algorithm, a linear bigram Support Vector Machine (SVM) with an F1 score of 0.881, while the Paragraph Vector model underperformed significantly at 0.368. This work is relevant for practitioners as it advances methods for nuanced language detection in biomedical literature, enhancing capabilities in information retrieval and summarization tasks.

biomedicalspeculative languagedeep learning

2 h agoarXiv cs.AIModels

UPLOTS: A Unified Pretrained Language Model for Constrained Time-series Generation

UPLOTS, a Unified Pretrained Language model for constrained Time-series Generation, has been introduced to streamline the generation of time-series data across various domains using a single pre-trained transformer backbone. Key innovations include dynamic multi-dataset loss re-weighting and prompt-to-pattern mapping, which enhance the model's ability to internalize and generate diverse temporal structures. Evaluated on four real-world benchmarks, UPLOTS demonstrates improved generalization and data augmentation capabilities, making it a valuable tool for practitioners facing challenges with task-specific models in time-series generation.

time-seriesgenerationmodel

2 h agoarXiv cs.AIResearch

ERAlign: Energy-based Representation Alignment of GNNs and LLMs on Text-attributed Graphs

The paper introduces ERAlign, an Energy-based Representation Alignment framework designed to enhance the integration of Graph Neural Networks (GNNs) and Large Language Models (LLMs) on Text-attributed Graphs (TAGs). By projecting GNN-encoded structures and LLM-derived text embeddings into a shared latent space and optimizing alignment through an Energy-based Model objective, ERAlign achieves superior representation consistency, demonstrated through state-of-the-art performance across eight TAG datasets with varying supervision levels. This approach addresses representation drift and improves generalization, making it a significant advancement for practitioners working on multi-modal learning tasks involving graphs and textual data.

GNNLLMalignment

2 h agoarXiv cs.AIRAG

LakeQA: An Exploratory QA Benchmark over a Million-Scale Data Lake

LakeQA is a newly introduced benchmark for search-centric question answering over large data lakes, comprising approximately 9.5 TB of text from diverse sources like Wikipedia and open-source government data. It requires long-horizon multi-hop reasoning and implicit intermediate steps, with each task annotated by Ph.D.-level experts to ensure quality. Experimental results indicate that even advanced models, such as GPT-5.2, struggle with this benchmark, achieving only an 18.37% exact-match score, highlighting the need for improved LLM capabilities in both search and reasoning for practical applications.

question answeringsearchdata lake

2 h agoarXiv cs.AIResearch

Minimum Distortion Quantization with Specified Output Distribution

The paper presents a method for designing optimal quantizers of real-valued random variables that not only minimize the minimum mean squared error (MMSE) for estimating the original variable but also conform to a specified output distribution. The quantizer is defined as $X = \sigma(F_{\sigma^{-1}(X)}^{-1}(F_W(W)))$, where $\sigma$ is the optimal permutation for minimizing MMSE, and the findings leverage the concept of majorization for proof of optimality. This work is significant for practitioners as it enables the design of quantizers with controlled output entropy and tailored distributions, which can enhance performance in communication systems and data anonymization tasks.

quantizationoptimizationresearch

2 h agoarXiv cs.AISafety

The Distributed Detectability Band Against Marginal-Preserving Attacks

The paper presents a novel marginal-preserving, correlation-encoded distributed-sabotage attack against AI-control monitors, utilizing a Gaussian-copula AR(1) construction. It demonstrates that traditional per-step monitors (Monitor A) can be defeated while maintaining benign scores, achieving a KS-distance of 0.013, and shows that alternative monitors (Monitor B) can maintain high AUC values (0.79-0.97) in detecting attacks, highlighting the limitations of distribution-shape monitors. This work is significant for AI practitioners as it underscores the need for robust temporal-correlation monitoring to effectively detect subtle sabotage, which can evade conventional monitoring techniques.

AI safetymonitoringattack

2 h agoarXiv cs.AITraining

Mitigating Bias in Low-SNR Financial Reinforcement Learning via Quantum Representations

The paper introduces FPQC-SAC, a novel variant of the Soft Actor-Critic (SAC) algorithm designed to mitigate bias in low-signal-to-noise ratio (SNR) financial environments by integrating a Parameterized Quantum Circuit (PQC) before the actor and critic networks. This architecture constrains feature propagation, effectively reducing the influence of market volatility on Bellman target estimation, and has demonstrated a 66.89% improvement in cumulative returns over standard SAC in empirical tests. This advancement is significant for practitioners as it offers a robust method to enhance stability and performance in financial reinforcement learning applications.

reinforcement learningquantumfinancial

2 h agoarXiv cs.AIMultimodal

Vision-Assisted Foundation Model for Solving Multi-Task Vehicle Routing Problems

The paper introduces a vision-assisted foundation model (VaFM) designed to tackle multi-task vehicle routing problems (VRPs) by integrating vision modality with graph-based approaches. VaFM employs a convolutional neural network to encode constraint-specific images, generating patch embeddings that are fused with graph nodes, while also implementing an auxiliary task to mitigate pixel imbalance among constraints. Experimental results show that VaFM outperforms existing state-of-the-art methods across 16 VRP variants, highlighting its capability to efficiently handle complex multi-constraint scenarios, which is critical for practitioners in optimizing routing solutions.

vehicle routingvisionmulti-task

2 h agoarXiv cs.AITraining

FOGO: Forgetting-aware Orthogonalization Optimizer

The paper introduces FOGO (Forgetting-aware Orthogonalization Optimizer), a novel optimizer designed to address both short-term and long-term forgetting during training by orthogonalizing momentum updates and utilizing a compact codebook memory for storing past gradient directions. FOGO employs lightweight orthogonal corrections to resolve conflicts between current and stored updates, demonstrating improved convergence and knowledge retention across various tasks, including class-imbalanced classification and continual learning scenarios with models like LLaVA-7B and GPT-2. This approach is significant for practitioners as it enhances optimization efficiency and model performance in continual learning contexts, potentially mitigating the forgetting problem prevalent in standard training regimes.

optimizertrainingforgetting

2 h agoarXiv cs.AIAgents

Harnessing the Collective Intelligence of AI Agents in the Wild for New Discoveries

EinsteinArena, a new agent-native platform for open distributed research, has been introduced to facilitate collaborative scientific discovery among AI agents. It features a live set of mathematical problems with verifiers and discussion forums, enabling agents to share insights and build on each other's work. Notably, agents have achieved 12 new state-of-the-art results, including an improvement on the kissing number problem in dimension 11, showcasing the potential of decentralized collaboration for advancing research beyond isolated efforts.

agentsscientific discoverycollaboration

2 h agoarXiv cs.AIAgents

SkillResolve-Bench: Measuring and Resolving Same-Capability Ambiguity in Agent Skill Retrieval

SkillResolve-Bench 1.0 is introduced as a benchmark for evaluating agent skill retrieval, addressing the issue of same-capability execution-risk retrieval by providing 661 helpful/risky skill pairs and a candidate pool of 7,982. The benchmark measures helpful ranking alongside harmful sibling rate (HSR@K) and demonstrates that the SkillResolve method achieves Recall@3 of 0.766 and NDCG@3 of 0.699, significantly enhancing performance over the previous SkillRouter by improving recall and reducing harmful exposure. This work is crucial for practitioners as it offers a structured approach to mitigate risks in skill retrieval systems, ensuring safer and more reliable agent execution.

agentsskill retrievalbenchmark

2 h agoarXiv cs.AITraining

Beyond Absolute Imitation: Anchored Residual Guidance for Privileged On-Policy Distillation

The paper introduces Anchored Residual On-Policy Distillation (AR-OPD), a novel framework that enhances on-policy distillation for large language models by disentangling privileged information into locally reachable reasoning steps and future-conditioned signals. AR-OPD demonstrates significant improvements over traditional privileged on-policy distillation methods, achieving a 2.3-point increase in performance and reducing hindsight leakage by 21.7%, particularly on long-horizon tasks exceeding 768 tokens. This advancement is critical for practitioners as it addresses the challenges of reasoning and capacity gaps in student-teacher model training, enabling more effective learning from complex data.

llmon-policy distillationtraining

2 h agoarXiv cs.AIResearch

Towards Critical Branching Mechanism in Recurrent Neural Networks

The paper presents an analysis of hidden-state dynamics in trained LSTM networks, revealing that small networks near optimal training epochs exhibit near-critical dynamics characterized by scale-free avalanche statistics and branching parameters approaching unity, while larger models show subcritical behavior. The study introduces a mixture branching process framework to explain the coexistence of subcritical branching with robust $1/f^{\beta}$ noise, suggesting that critical-like behavior in LSTMs is an emergent feature dependent on model capacity. This research provides insights into the dynamical regimes of RNNs, which could inform the design and training of more effective neural architectures.

recurrent neural networkscriticalityLSTM

2 h agoarXiv cs.AIRAG

Agentic Hybrid RAG for Evidence-Grounded Muon Collider Analysis

The article presents the "agentic hybrid RAG," an evidence-grounded retrieval-augmented generation framework specifically designed for muon collider research. This framework integrates a hybrid retriever that combines sparse lexical and dense semantic retrieval with an agentic reasoning module for query decomposition and evidence expansion. The authors also introduce a benchmark for retrieval-augmented scientific question answering in the muon collider domain, demonstrating that their framework outperforms existing retrieval and RAG baselines in retrieval effectiveness, answer quality, and evidence grounding, thus providing a robust tool for researchers in high-energy physics to analyze large-scale scientific literature.

RAGmuon colliderevidence

2 h agoarXiv cs.AIResearch

Expert-Level Crisis Detection in Mental Health Conversations

The authors introduced CRADLE-Dialogue, a clinician-annotated benchmark for turn-level crisis detection in mental health conversations, featuring 600 dialogues with multi-label annotations for various risks such as suicide ideation and self-harm. They also released a synthetic training corpus and a 32B-parameter model that outperforms existing open-source models, achieving competitive results against proprietary systems, with Micro F1 scores ranging from mid-40% to high-60%. This work is significant for practitioners as it addresses the challenges of tracking evolving risk signals in multi-turn dialogues, enhancing the potential for timely intervention in crisis situations.

crisis detectiondialoguemental health

2 h agoarXiv cs.AISafety

KG-SoftMAP: Soft Knowledge-Graph Priors for Bayesian Network Structure Learning from Sparse Discrete Data

KG-SoftMAP is a new method for learning Bayesian network structures from sparse discrete data by incorporating soft knowledge graph (KG) priors into the process. The approach utilizes a confidence-weighted, data-overridable edge prior to maximize a MAP objective, combining the BDeu score with a logit-form prior, and demonstrates improved performance on synthetic benchmarks, achieving a DF1 score increase from 0.14 to 0.96 with informative KGs. This method is significant for practitioners as it allows for the integration of imperfect domain knowledge, enhancing structure recovery and inference capabilities in scenarios where data is limited.

adversarialroboticsdiffusion

2 h agoarXiv cs.AIProducts

Atomic Intent Reasoning: Bringing LLM Semantics to Industrial Cross-Domain Recommendations

The paper introduces Atomic Intent Reasoning (AIR), an LLM-driven framework for cross-domain recommendations that enhances industrial deployment by addressing the challenges of semantic gaps and noise in user behavior sequences. By offloading LLM inference to an offline phase and employing dynamic user intent representations, AIR achieves approximately 400 times faster inference while preserving semantic consistency. Experimental results and real-world A/B testing in Kuaishou E-commerce demonstrate state-of-the-art performance and a significant +3.446% increase in Gross Merchandise Value (GMV), highlighting its practical applicability for AI practitioners in recommendation systems.

recommendationLLMcross-domain

2 h agoarXiv cs.AITraining

Routing-Aware Expert Calibration for Machine Unlearning in Mixture-of-Experts Language Models

The paper presents TRACE (Targeted Routing-Aware Calibration of Experts), a novel approach to machine unlearning in Mixture-of-Experts (MoE) language models. TRACE addresses the routing mismatch between forget and retain data by reweighting token-level retain losses to ensure better calibration of forget-critical experts, resulting in a 9% relative utility improvement over existing methods on benchmarks like WMDP and MUSE-BOOKS. This advancement is significant for practitioners as it enhances the efficiency of unlearning processes in MoE architectures, which is crucial for maintaining model integrity and compliance with data privacy regulations.

machine unlearningMoElanguage models

2 h agoarXiv cs.AIResearch

Building Change Detection in Earthquake: A Multi-Scale Interaction Network and A Change Detection Dataset

The article presents the Turkey Earthquake Change Detection (TUE-CD) dataset, designed for assessing building damage shortly after earthquakes using multi-temporal remote sensing images. It introduces a novel architecture called the Multi-Scale Feature Interaction Network (MSI-Net), which includes joint cross-attention, multi-scale offset calibration, and feature integration modules to effectively handle bi-temporal feature interactions and mitigate side-looking issues. Experimental results demonstrate that MSI-Net outperforms existing state-of-the-art change detection methods, providing a crucial tool for emergency response and damage assessment in post-earthquake scenarios.

change detectionearthquakedataset

2 h agoarXiv cs.AIRAG

Content-Induced Spatial-Spectral Aggregation Network for Change Detection in Remote Sensing Images

The paper introduces the Content-Induced Spatial-Spectral Aggregation Network (CSI-Net) for enhanced change detection in remote sensing images. CSI-Net combines a spatial reasoning module utilizing cascaded graph convolution blocks, a spectral difference module for feature extraction, and a content-guided integration module to effectively fuse spatial and spectral information while mitigating spectral differences in unchanged areas. Experimental results show that CSI-Net outperforms existing state-of-the-art methods on multiple datasets, highlighting its robustness and versatility for practitioners in remote sensing and image analysis.

change detectionremote sensingdeep learning

2 h agoarXiv cs.AITraining

Baseline-Free Policy Optimization for Neural Combinatorial Optimization

The paper introduces Group Relative Policy Optimization (GRPO), a baseline-free algorithm for neural combinatorial optimization (NCO) that normalizes advantages within sampled trajectory groups, addressing the instability issues associated with traditional REINFORCE training. Evaluated on the Traveling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP) benchmarks, GRPO demonstrated superior stability and performance, achieving solution quality within 2% of the POMO baseline without requiring a frozen policy for variance reduction. This approach is significant for practitioners as it mitigates training collapse in complex routing problems, enhancing the robustness of reinforcement learning applications in combinatorial optimization tasks.

policy optimizationreinforcement learningNCO

2 h agoarXiv cs.AIAgents

Catching One in Five: LLM-as-Judge Blind Spots in Production Multi-Turn Transaction Agents

The study published in arXiv investigates the efficacy of LLM-as-judge in evaluating multi-turn conversational agents, specifically a food-and-beverage ordering system. It reveals that the LLM judge identified only 22% of human-confirmed systematic problems and failed to flag operational issues in 100 rounds, highlighting a structured blind-spot in its scoring rubric, which inadequately addresses critical behavioral dimensions like state-tracking. This underscores the need for enhanced evaluation mechanisms in production environments, as reliance on automated judging can lead to significant underreporting of defects, emphasizing the necessity of human oversight in quality assurance.

LLMmulti-turnconversation

2 h agoarXiv cs.AIAgents

The Confident Liar: Diagnosing Multi-Agent Debate with Log-Probabilities and LLM-as-Judge

The paper presents a framework for evaluating multi-agent debate systems by analyzing token-level log-probabilities, LLM-as-judge rubric scores, and final task accuracy. It employs a two-agent architecture consisting of a Constructor and an Auditor, revealing that confidence signals correlate more strongly with reasoning quality in the Constructor, achieving an AUROC of 0.804 for critical reasoning failure detection. This research is significant for practitioners as it emphasizes the importance of intermediate reasoning quality in multi-agent systems, potentially guiding improvements in LLM training and evaluation methodologies.

multi-agentdebateLLM

2 h agoarXiv cs.AIModels

LLM-Guided Neural Architecture Search for Robust Co-Design of Physical Neural Networks

The paper introduces Unconventional Hardware Neural Architecture Search (UH-NAS), a hardware-agnostic framework utilizing LLMs as evolutionary operators for optimizing neural network architectures across different platforms. UH-NAS facilitates the co-optimization of task accuracy and inference energy by integrating platform-specific energy models and simulators, demonstrating superior performance in discovering robust architectures for optical MZI hardware compared to traditional NAS methods. This approach is significant for practitioners as it enables effective cross-platform comparisons and enhances the design of neural networks tailored for unconventional hardware environments.

neural architecturehardwareLLM

2 h agoarXiv cs.AIResearch

Towards Robust Arabic Speech Emotion Recognition with Deep Learning

This study presents a comparative evaluation of three architectures for Arabic Speech Emotion Recognition (SER): a CNN-LSTM model, a CNN-Transformer model, and a fine-tuned wav2vec 2.0 model, with the CNN-Transformer achieving the highest accuracy of 98.1% on the EYASE and BAVED datasets. The research emphasizes the effectiveness of hybrid architectures in capturing both local spectral cues and long-range temporal dependencies, addressing challenges posed by dialectal diversity and limited annotated data in Arabic SER. These findings are significant for practitioners as they provide insights into effective model architectures for improving SER performance in low-resource languages.

speechemotion recognitiondeep learning

2 h agoarXiv cs.AIAgents

Hierarchical Policies from Verbal and Egocentric Human Signals for Natural Human-Robot Interaction

The paper presents EDITH, a novel robot framework designed for natural human-robot interaction by integrating both verbal and nonverbal signals, such as gestures and gaze, captured through smart glasses. EDITH employs a hierarchical policy architecture where a high-level policy interprets human intent and generates subtasks, while a low-level policy executes these tasks, demonstrating improved performance in interactive tasks by reducing reliance on language instructions alone. This approach enhances the robot's ability to understand and respond to human intent, which is crucial for developing more intuitive and efficient human-robot collaboration systems.

human-robot interactionnonverbal signals

2 h agoarXiv cs.AIAgents

What Matters in Orchestrating Robot Policies: A Systematic Study of Hierarchical VLA Agents

This paper presents a systematic study on hierarchical vision-language-action (Hi-VLA) systems for robot manipulation, unifying various Hi-VLA agents under an options-style control framework. The authors benchmark core design choices across different task horizons and reasoning intensities, revealing that structured model choices and interface mechanisms significantly enhance performance, yielding stronger systems compared to flat VLA control or poorly designed hierarchies. These findings provide foundational principles for practitioners aiming to develop more effective and robust Hi-VLA agents in robotic applications.

robothierarchicalVLA

2 h agoarXiv cs.AITraining

Multi-Level Analyzation of Imbalance to Resolve Non-IID-Ness in Federated Learning

The paper introduces FedBB, a novel approach to address class imbalance in federated learning (FL) by analyzing imbalance issues at three levels: inter-case, inter-class, and inter-client. FedBB employs a Positive Negative Balanced (PNB) loss function to enhance local training on skewed datasets and a Client Balanced Reweighting (CBR) mechanism to adjust model aggregation based on client data distribution. Experimental results on X-ray and natural image datasets show that FedBB outperforms existing methods, making it a valuable framework for improving model generalization in non-IID scenarios while maintaining privacy through minimal statistical data requirements.

federated learningclass imbalancedeep learning

2 h agoarXiv cs.AIResearch

Linguistically Augmented Audio Speech Data (LinguAS)

Linguistically Augmented Audio Speech Data (LinguAS) has been released, featuring over 800 annotated audio samples of genuine and deepfaked speech, with five Expert-Defined Linguistic Features (EDLFs) included to enhance model training. Models trained on this dataset demonstrated significant performance improvements over ASVspoof 2021 baselines and self-supervised learning models like HuBert and XLSR. This dataset provides critical linguistic and metadata context, enabling researchers to develop more effective detection models for audio deepfakes by leveraging natural human speech characteristics.

speechdatasetdeepfake

2 h agoarXiv cs.AIAgents

YUBI: Yielding Universal Bidigital Interface for Bimanual Dexterous Manipulation at Scale

The Yielding Universal Bidigital Interface (YUBI) has been introduced as a finger-aligned gripper designed for ergonomic and scalable data collection in bimanual dexterous manipulation. It features yielding, finger-driven actuation that translates human finger movements to gripper jaw motion, with a dataset comprising 8,434 hours of data across 1.20 million episodes and 119 tasks. This release, along with the hardware and software stack, provides a significant resource for practitioners aiming to enhance robotic foundation models through high-fidelity, transferable policy training across multiple robotic platforms.

manipulationdata collectionrobotics

2 h agoarXiv cs.AIResearch

Hyperbolic Neural Population Geometry Benefits Computation

The paper presents a theoretical framework linking hyperbolic geometry to neural population activity in the hippocampus, proposing a construction of tuning curves that induces this geometry. It establishes a connection between neural decoding and associative memory through the Modern Hopfield Network, demonstrating that its update rule computes the minimum mean-squared-error estimator. Additionally, a new associative memory model in hyperbolic space is introduced, which offers significantly larger capacity compared to existing models, suggesting implications for spatial information encoding and memory performance in AI systems.

neural networksgeometrymemory

2 h agoarXiv cs.AIAgents

SHAPO: Sharpness-Aware Policy Optimization for Safe Exploration

The paper introduces Sharpness-Aware Policy Optimization (SHAPO), a novel reinforcement learning method designed to enhance safe exploration by leveraging epistemic uncertainty. SHAPO employs a sharpness-aware policy update that evaluates gradients at perturbed parameters, which biases learning towards conservative behavior in under-explored regions and amplifies the influence of rare unsafe actions. Experimental results demonstrate that SHAPO improves both safety and task performance across continuous-control tasks, outperforming existing baselines and expanding their Pareto frontiers, which is critical for deploying RL agents in safety-sensitive applications.

reinforcement learningsafe exploration

2 h agoarXiv cs.AIInference

Fast Exact Nearest-Neighbor Learning for High-Frequency Financial Time Series

The article presents the Mojo SIMD k-d tree for efficient exact nearest-neighbor learning in high-frequency financial time series, addressing the challenge of processing large datasets with real-time latency requirements. This new implementation utilizes variance-based splitting and compile-time vectorized distance computation, achieving a speedup of 17.5–21.6× over scikit-learn's k-d tree on x86 and 28.1–43.5× over brute force on ARM64, while maintaining exact outputs. The results suggest that Mojo can significantly enhance the scalability and efficiency of AI applications in finance, enabling practitioners to handle larger datasets more effectively.

financialtime seriesai

2 h agoarXiv cs.AIResearch

A Source Domain is All You Need: Source-Only Cross-OS Transfer Learning for APT Anomaly Detection via Semantic Alignment and Optimal Transport

The article presents a novel framework for source-only cross-operating-system (cross-OS) anomaly detection targeting Advanced Persistent Threats (APTs), utilizing system-level provenance traces. This approach leverages pretrained language models for embedding process behaviors, combined with a transport-based ranking mechanism that employs an Optimal Transport (OT) method to compute an anomaly score based on deviations from source-normal prototypes. Evaluation on DARPA Transparent Computing data shows significant improvements in ROC-AUC and nDCG metrics, indicating that this method enables effective APT detection across different OS platforms without requiring target-domain labels, which is crucial for practitioners facing the challenges of limited labeled data in cybersecurity contexts.

anomaly detectiontransfer learningAPT

2 h agoarXiv cs.AICoding

Automated Pronunciation Evaluation for Korean Toddler Speech using Speech Diarization and Self-Supervised Learning

The paper presents an end-to-end automated pronunciation evaluation pipeline for Korean toddler speech, integrating neural speaker diarization and self-supervised learning. It introduces a new corpus of 53 recordings from children aged 2-5 and evaluates three diarization models, with NeMo SortFormer achieving 88.69% speaker count accuracy and 33.04% diarization error rate. For pronunciation scoring, an ensemble approach using HuBERT-large and WavLM-large yields balanced accuracies of 0.720 for consonants and 0.845 for vowels, indicating potential advancements in automated speech assessment tools for pediatric communication disorders.

speechself-supervisedevaluation

2 h agoarXiv cs.AIAgents

Exploration of Foundation Model-Based Robots in Patient and Elderly Care

The article reviews the integration of foundation models into robots for patient and elderly care, highlighting their use as conversational and reasoning components in socially assistive systems. While current implementations show positive usability and engagement, they face significant issues such as reliability failures and limited evidence of meaningful clinical outcomes. The authors advocate for the development of care-specific evaluation standards and improved integration into existing workflows to enhance the effectiveness of these technologies in real-world care settings.

robotsfoundation modelscare

2 h agoarXiv cs.AIRAG

MMClima: A Framework for Multimodal Climate Science Data and Evaluation

MMClima is a new multimodal climate question answering framework that includes over 104,000 expert-validated question-answer pairs derived from articles, video transcriptions, and figures across five climate science domains. It features automated claim extraction and human validation, benchmarked against state-of-the-art multimodal language models. The release includes the dataset, evaluation pipeline, and fine-tuned model weights (mmclima-70b-txt), which surpass existing models in textual QA, providing practitioners with essential resources for developing AI systems that reason across diverse climate-related content.

climatemultimodalqa

2 h agoarXiv cs.AITraining

Dropout-GRPO: Variational Stochasticity for Continuous Latent Reasoning

The paper introduces Dropout-GRPO, a method that enhances Group Relative Policy Optimization (GRPO) for continuous latent reasoning models like Coconut by incorporating structured dropout to introduce stochasticity in latent states. This approach employs a constant Bernoulli mask across latent recurrence steps, enabling the generation of diverse trajectories essential for effective GRPO and yielding a pass@1 improvement from 27.29% to 29.01% on the GSM8K benchmark. This advancement is significant for practitioners as it provides a theoretically justified and practical mechanism to improve the performance of latent-reasoning LLMs post-training.

reinforcement learningpolicy optimization

2 h agoarXiv cs.AIModels

Making Time Editable in Video Diffusion Transformers

The paper introduces a methodology for enhancing video generation in Diffusion Transformers by integrating explicit time editing capabilities into a pretrained DiT model. This approach utilizes a lightweight temporal module to enable control over motion speed and temporal structure, while maintaining the original generative prior of the model. This advancement is significant for practitioners as it allows for greater flexibility in video generation, enhancing the ability to manipulate temporal dynamics without extensive model redesign.

videodiffusiontemporal control

2 h agoarXiv cs.AIAgents

Flow Control: Steering Vision-Language-Action Models with Simple Real-Time Inputs

The paper presents a flow control method for vision-language-action (VLA) models that allows real-time steering of actions using simple inputs, such as keyboard commands, without the need for retraining. This approach leverages the learned action distributions of VLAs to generate high-quality and high-fidelity actions, demonstrating improved task success rates and faster completion times, as well as enhanced performance when fine-tuning on flow control trajectories. This method is significant for practitioners as it provides an intuitive interface for guiding VLA actions, enhancing usability and effectiveness in real-world applications.

vision-languagereal-timecontrol

2 h agoarXiv cs.AISafety

Local Is Not a Sufficient Privacy Boundary: Governing OS-Integrated On-Device AI

The paper presents a framework for managing privacy in OS-integrated on-device AI systems, emphasizing that local execution does not inherently ensure privacy. It introduces a comprehensive threat model, a six-part privacy risk taxonomy, and a four-level audit rubric to assess privacy controls in various platforms, including Apple Intelligence/Foundation Models, Android AICore/Gemini Nano, and Microsoft Recall. This framework is crucial for practitioners as it shifts the focus from merely where computation occurs to understanding the broader implications of data handling, user control, and system governance in AI deployments.

privacyon-deviceai

2 h agoarXiv cs.AISafety

Gaming AI-Assisted Peer Reviews Poses New Risks to the Scientific Community

The study published in arXiv:2606.10159v1 highlights vulnerabilities in AI-assisted peer review systems, demonstrating that superficial rephrasing of manuscript abstracts can significantly improve review outcomes, achieving an attack success rate of approximately 38% for Gemini 3 Flash and 50% when the original review suggests rejection. This manipulation, which requires minimal time and cost, raises concerns about the integrity of AI evaluations, as inflated review scores could lead to biased editorial decisions favoring acceptance over scientific merit. The findings underscore the necessity for robust testing and oversight of AI tools in scientific evaluation processes to mitigate potential manipulation risks.

peer reviewaimanipulation