Technology
ReSkill: Reconciling Skill Creation with Policy Optimization in Agentic RL
ReSkill harmonizes skill creation with policy optimization in Agentic RL using assertion-driven updates and Thompson Sampling. It outperforms existing methods by ensuring skills co-evolve with the policy, significantly boosting performance on unseen tasks.
MobEvolve: An Agentic Self-Evolving Heuristic System for Interpretable Human Mobility Generation
MobEvolve is an agentic, self-evolving heuristic system that uses LLM agents to refine mobility models. It outperforms existing methods in realism, interpretability, and efficiency on Singapore and Montreal datasets.
Evidence-Gated LLM Priors for Multi-Objective Bayesian Optimization
This study introduces an evidence-gated framework for multi-objective Bayesian optimization, dynamically calibrating LLM priors via objective-specific reputation markets. Results show this approach enhances robustness over static priors, though raw LLM confidence proves inconsistent across benchmarks.
S-SPPO: Semantic-Calibrated Self-Play Preference Optimization
S-SPPO stabilizes Self-Play Preference Optimization via semantic calibration, preventing policy degeneration. It achieves superior AlpacaEval 2.0 performance with Llama-3-8B without extra human annotations.
TrafficRAG: A Multimodal RAG Framework for Traffic Accident Liability Determination
TrafficRAG is a multimodal RAG framework that combines vision-language models with hybrid retrieval to automate traffic accident liability determination. It outperforms baselines, achieving 77.32% legal accuracy and 81.71% factual faithfulness.
Characterization of Multi-Model Agentic AI Systems on General Tasks via Trace-Driven Simulation
This study introduces GAIATrace and Vidur-Agent to simulate and analyze multi-model agentic AI systems. These tools enable reproducible, cost-effective evaluation of system dynamics and design choices on general tasks.
TriAlign: Towards Universal Truth Consistency in Personalized LLM Alignment
TriAlign introduces Truth-Invariant Alignment via multi-agent RL, balancing personalized LLM outputs with universal truth consistency. It reduces factual disparities across social groups while maintaining high personalization quality.
EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks
EvoBrain enables continual learning for EEG foundation models via Neuro-Spectral Task Normalization and Response-Affinity Distillation. It outperforms SOTA methods across six BCI tasks, achieving unified decoding with minimal catastrophic forgetting.
Stochastic convergence of parallel asynchronous adaptive first-order methods
This paper analyzes the stochastic convergence of parallel asynchronous adaptive first-order methods for non-convex optimization, proving an O(1/sqrt{t}) rate. Empirical results confirm their suitability for large-scale, heterogeneous machine learning environments.
Structure-Guided Adaptive Propagation for Protein-Protein Interaction Site Prediction
SGAP-PPIS uses structure-guided adaptive propagation to dynamically adjust information diffusion for accurate protein-protein interaction site prediction. It outperforms rigid models by leveraging equivariant graph neural networks to tailor propagation to local geometric contexts.
Consistency evaluation of benchmarks used for causal discovery
This study introduces an LLM-based workflow to verify 11 causal discovery benchmarks against 38,081 papers, revealing significant discrepancies with current literature. These findings highlight critical reliability issues in widely used benchmarks for causal discovery research.
Token Predictors Are Not Planners: Building Physically Grounded Causal Reasoners
The authors introduce Causal-Plan-Bench and Causal Planner to shift embodied AI from token prediction to physical causal reasoning. Their model achieves superior performance by internalizing physical logic, validating a Causal Scaling Law.
OctoT2I: A Self-Evolving Agentic Text-to-Image Router
OctoT2I is a self-evolving agentic router that optimizes text-to-image generation via an unsupervised, multi-round routing strategy. It achieves superior speed and energy efficiency while maintaining high-quality outputs without human supervision.
Evaluation of Baseline Methods for IDD-based SSD External Memory Search
This study evaluates simple baseline methods for SSD-based A* search using immediate duplicate detection, addressing gaps in prior research on external memory search strategies.
CAPF: Guiding Search-Agent Rollouts with Credit-Attenuated Privileged Feedback
CAPF guides LLM search agents using verifier-side feedback to repair failed rollouts, boosting Qwen3-4B’s QA accuracy from 44.7% to 48.5% across seven benchmarks.
EVA-Net: Subject-Independent EEG Motor Decoding with Video-Derived Motor Priors
EVA-Net uses video-derived motor priors to align EEG features, enabling robust subject-independent decoding. It achieves an 8.66% LOSO accuracy gain on EEGMMI, outperforming text-based methods.
Does Compression Preserve Uncertainty? A Unified Benchmark for Quantized and Sparse LLMs via Conformal Prediction
This study benchmarks quantized and sparse LLMs using conformal prediction, revealing that compression decouples accuracy from uncertainty reliability. It urges incorporating uncertainty-aware metrics into model compression workflows for safer deployment.
Absorbing Complexity: An Interaction-Native Knowledge Harness for Financial LLM Agents
InKH is a framework for financial LLM agents that internalizes complexity via structured memory, reducing latency by 83% and token costs by 82% while improving task quality and auditability.
WorldCoder-Bench: Benchmarking Physically Grounded 3D World Synthesis
WorldCoder-Bench benchmarks LLMs on synthesizing physically grounded 3D worlds using StateProbe to verify hidden runtime states. Results show top models achieve low verification coverage, highlighting significant challenges in generating robust, interactive 3D environments.
Community-Aware Assessment of Social Textual Engagement and Resonance: A Human-Centric Perspective on User-Generated Content Evaluation
The study introduces CASTER and MEDEA, a human-centric framework using Social-CoT to assess UGC resonance via simulated community personas. MEDEA outperforms baselines on the new CASTER-Bench with empathetic, interpretable reasoning.