Preprints and papers, rewritten in plain English.
A new arXiv paper analyzes the computational complexity of grounded and preferred semantics in finitary argumentation frameworks, where each argument interacts with finitely many others.
A new arXiv preprint proposes recomposing and refining latent reasoning flows to help vision-language-action models turn multimodal observations into task-relevant states before acting.
A new method adjusts the number of Monte Carlo samples per input in Bayesian neural networks, offering statistical guarantees on prediction error.
A new arXiv paper proposes using functional gradient guidance from language models to induce interpretable probabilistic logic programs from data.
A new arXiv paper proposes a flexible, instruction-conditioned approach to electromagnetic spectrum understanding using budget-adaptive signal tokenization.
A new arXiv study suggests that the internal wiring of large language models, like that of biological brains, shows small-world connectivity that correlates with reasoning ability.
A new simulation environment offers a way to evaluate AI interviewers designed to elicit open-ended responses in research and polling.
A new arXiv study tests whether LLM compliance systems' verdicts actually track the regulatory rule they are given.
A new approach estimates when LLM agents need to reason, potentially cutting unnecessary computation.
A new arXiv study asks whether large language model scientific agents adapt to experimental feedback or stick to initial context when tuning neural operators.
A new paper reports on the development and in-orbit demonstration of an AI system that processes Earth observation data directly on a satellite.
A new arXiv paper argues that agentic AI evaluation should measure whether models revise beliefs or acknowledge uncertainty when evidence contradicts them.
A new arXiv study evaluates whether AI agents can learn and adapt strategies from limited experience in adversarial games, using a long-running competition as a testbed.
A new arXiv paper proposes a structural theory of cognitive representation to address the lack of explicit structure in modern AI systems.
A new method models cross-modal interactions as higher-order tensors to better capture synergistic information in self-supervised representation learning.
A new framework applies streaming structure-aware optimal transport to detect and correct risky LLM agent behaviour before actions become irreversible.
A new paper examines how uneven pretraining support for concepts affects supervised fine-tuning of LLMs, particularly for rare or long-tail knowledge.
A new arXiv benchmark tests AI agents' ability to turn text prompts into LEGO assemblies that meet design and physical constraints.