The proposed STAMP framework addresses the reward-credit mismatch in deep-search agents by using a reference-based verifier to evaluate whether cited documents support specific claims.
This study introduces RouteCast, an evaluation framework designed to assess model-generated strategic routes when ground truth feedback is delayed or private.
The authors present ARCANA, a collaborative multi-agent framework designed to solve ARC-AGI-2 reasoning tasks under strict hardware and test-time constraints.
The domestic embodied AI data industry has seen rapid growth, with 15 independent data service providers securing around 4.47 billion yuan in funding over the past year.
CSC Financial reports that semi-annual earnings previews confirm high demand for AI servers and intelligent computing infrastructure, while AI software applications are beginning to show revenue recovery.
This paper introduces new sparsity regularizers for Top-k sparse autoencoders to improve the interpretability of features in vision foundation models.
A longitudinal study of college students reveals how AI chatbot interactions influence cognitive engagement and reading habits during academic tasks.
AgentLens is a new benchmark designed to evaluate coding agents by assessing the quality of their entire interaction trajectory rather than just final task outcomes.
This study presents a method for training branching neural networks to perform multiple algorithmic reasoning tasks simultaneously.
This manifesto outlines how LLM-powered autonomous agents are transforming software systems and proposes integrating them with service-oriented computing principles.
A novel reinforcement learning technique called Sibling-Guided Credit Distillation improves how agents learn to use tools by better attributing rewards to specific actions.
Researchers investigated how personality-based prompting affects the task performance of multi-agent LLM teams, finding that communication styles significantly influence collaborative outcomes.
A new study identifies a failure mode called contextual sycophancy, where reinforcement learning agents receive biased feedback from evaluators in specific critical scenarios.
TerraZero provides a procedural simulation environment and training stack to support the development of autonomous driving agents through large-scale self-play.
A study investigates how anatomical and contrast factors in brain MRI scans contribute to demographic predictability, highlighting potential biases in medical AI.
Researchers have proposed a reinforcement learning framework for Parametrized Action Markov Decision Processes that integrates incomplete domain knowledge and gradient guidance to improve sample efficiency.
The authors developed LakeQuest, a new benchmark designed to evaluate question-answering systems across complex, heterogeneous enterprise data lakes.
An analysis of 44 language models reveals a strong tendency toward convergent behavior when prompted to select single words from open-ended categories.
The proposed CARE-PPO framework integrates uncertainty estimation into reinforcement learning fine-tuning to reduce hallucinations and improve confidence in language-based quantitative predictions.
The Team Research and AI Integration Lab (TRAIL) platform provides a new infrastructure for conducting reproducible experiments on human-AI collaboration.
A new training-free attribution method is introduced to analyze the sparse dependencies between neurons in Transformer feedforward network blocks.
The paper explores ontology-amplified distillation and contextuality auditing to adapt sovereign enterprise language models for regulated financial institutions.
A new framework combines elastic regularization and synthetic replay to mitigate catastrophic forgetting during the federated fine-tuning of multimodal models.
Researchers investigated how interaction graphs and feedback mechanisms influence convention formation and consensus in populations of open-weight language models.