Preprints and papers, rewritten in plain English.
A new interpretable method, SCRIBE, aims to reveal which linguistic features drive the alignment between large language models and human brain responses during language comprehension.
A new arXiv study outlines an evaluation approach for how well LLM-based agents perform scientific inquiry through hypothesis generation, reasoning, and exploration.
A new evaluation framework systematically characterizes how different types of interaction perturbations affect LLM agent performance.
A new arXiv paper examines how reinforcement learning is being applied to diffusion models for generation across language, vision, and other domains.
A new bi-level memory approach tackles the plasticity-stability trade-off that limits continual learning systems.
A new arXiv paper highlights how LLM-driven agents that can modify their own components pose a fundamental open question.
A new arXiv paper proposes RSIFlow, a workflow that uses an evolving meta-agent to adaptively improve LLM agents.
PhysClosureBench measures whether AI agents can turn physics research goals into conclusions backed by evidence.
A new arXiv paper proposes speculative decoding to reduce the sequential overhead of looped Transformers.
A new method aims to map arbitrary user preferences directly to trade-off solutions in multi-objective optimization, using only limited offline data.
A new arXiv paper introduces DeciCAD to study how the decision space defined by an agent harness influences LLM-based CAD agents.
A new framework uses an evolving video world model to help vision-language models reason about motion and view changes.
A new method called RoleQ converts language models to ternary weights while aiming to avoid the accuracy and knowledge losses typical of extreme quantization.
A new approach measures behavioral consistency across repeated LLM agent executions using graph structure instead of a single summary score.
A new paper argues that multimodal large language models face an overlooked 'intent-level' safety risk and suggests proactive inquiry as a remedy.
A new arXiv preprint proposes using large language models to build progress rewards for long-horizon reinforcement learning tasks.
A new arXiv paper proposes CARA, a retrieval agent that uses causal guidance to find evidence that truly answers questions over text-rich knowledge graphs.
A 2.4B-parameter vision-language-action model trained only on public data matches larger systems and adds a cheap memory mechanism that stabilizes robot control.