Preprints and papers, rewritten in plain English.
A new method uses operator graphs and execution feedback to generate realistic scientific coding tasks, addressing the scarcity of high-quality training data for large language models.
A new method called HiPACE systematically identifies and controls when parent concepts absorb related features in sparse autoencoders.
A new arXiv paper proposes a reinforcement-learning-based detector that screens language models for alignment failures without task-specific training.
A new analysis argues that infrastructure state races in LLM agent tool use need finer-grained safety checks, not blanket rejection of stale reads.
A new arXiv preprint argues that plausible rationales in reasoning-enhanced multimodal embeddings do not necessarily translate into better discriminative retrieval rankings.
A new paper proposes a Transformer-GAT architecture for recognizing emotions in dialogue by combining multiple modalities.
A new arXiv paper argues that bias in AI systems should be treated not just as a flaw to eliminate, but as a diagnostic clue pointing to deeper weaknesses in data and modeling.
A new framework argues that removing a physical agent's outdated skills must respect the authority, consent, and environment conditions tied to them.
A new method preserves and recovers cross-session evidence under a bounded, query-blind memory budget for long-horizon tasks.
A new arXiv paper proposes a theoretical model for guarded coordination among multiple LLMs, addressing a field that is widely deployed but thinly theorized.
Standard knowledge-editing tests can give a model a clean bill of health even when the edits have quietly degraded its ability to distinguish reliable evidence from unreliable evidence.
PartHackBench stress-tests whether partial-credit evaluators reward only genuine, lasting progress in long-horizon tool-agent tasks.
New work shows normal-inverse-gamma regression has a hidden non-identifiability and proposes PEEL to restore identifiable uncertainty estimates in CT imaging.
A new arXiv preprint describes an embodied agent that can reason, act autonomously, and team up with human players in PUBG via voice.
A new arXiv paper proposes evolving agent environments themselves, rather than just agents, to enable recursive self-improvement in real-world tasks.
A new arXiv paper describes iCoder-27B, an attempt to use recursive AI-led development to build a large industrial coding model.
A new framework for generative AI feedback in programming courses decides when to intervene, which evidence to use, and how much help to give.
A new arXiv study explores using AI on low-resolution remote monitoring data to flag worsening heart failure, potentially easing hospital burden.