Preprints and papers, rewritten in plain English.
A new temporal-difference learning method computes reach-avoid sets that account for stochastic disturbances, producing less conservative safety guarantees than classical worst-case approaches.
A new arXiv paper proposes evidence-prior rebalancing to correct cross-modal hallucinations in audio-visual large language models, moving beyond output-level fixes.
A new approach targets the difficult problem of estimating rare but severe distribution-feeder overloads caused by public EV charging.
A new framework pretrains robotic touch representations on human hand-object interaction data, reaching full-data performance with only a quarter of labeled examples.
A new general-purpose model is designed to cut the computational cost of applying diverse clinical questions to entire patient populations.
A new arXiv paper proposes layer-specific adaptation and prototype transport to help models learn compositional visual concepts incrementally.
A new benchmark evaluates language models on analytical reasoning over merchant transaction event sequences, targeting gaps in existing tabular and time-series benchmarks.
The paper reports rapid growth in physical AI research while noting the field still has no settled definition.
A new approach called REM addresses the lack of experience reuse in long-horizon LLM coding agents by deciding what to store, recall, and forget.
A training-free framework uses the delay before committing an action chunk to verify candidates against fresher observations, improving success in simulation and on real robots.
A new arXiv preprint argues that existing tool-use benchmarks fall short for evaluating AI assistants that must handle multi-step tasks in real applications.
A new arXiv preprint outlines a security framework to prevent impersonation and unauthorized payments by autonomous AI agents in digital commerce.
A new arXiv paper tightens the mathematical limits on probabilities of causation, bridging binary and non-binary measures.
A new benchmark targets how speech agents update persistent memory while listening and speaking.
A new arXiv preprint introduces an LLM agent architecture that uses context-aware prospective memory to move beyond immediate reactive responses.
Researchers propose k*, a readout sparsity measure that reveals the minimum number of features needed to predict a task from language model representations.
A new arXiv study finds that classifier-head fine-tuning beats LLM-as-Judge methods on binary hallucination detection, with the gap rooted in layer-wise representations.
A new arXiv preprint tests a hierarchical agentic framework to automate the multi-step workflow of adsorption simulation in porous materials.