Youmna Farag, Svetlana Stoyanchev, Simon Keizer, Mohan Li, Rama Doddipatla · 2026 · arXiv (Cornell University)
DOI: 10.48550/arxiv.2610.28381 · cited by 0 works (OpenAlex) · open-access PDF
El razonamiento sobre la seguridad es esencial para los agentes encarnados alimentados por modelos de lenguaje grande (LLM). Introducimos el índice de referencia de acción y plan de seguridad contextual (CAPSafe), diseñado para evaluar el razonamiento de seguridad en dos niveles: (1) un conjunto de datos de nivel de acción que prueba si una sola acción es segura dado su contexto ambiental, y (2) un conjunto de datos de nivel de plan que evalúa si los agentes pueden generar planes de largo horizonte seguros para tareas domésticas peligrosas complejas. Los modelos CAPSafe revelan una brecha sustancial entre el reconocimiento de riesgos y la planificación de largo horizonte seguro. Varios modelos reportan una alta precisión en el reconocimiento de acciones no cumplidas (más del 90% en los mejores casos de largo plazo), pero la planificación de sustancias seguras sobre
Machine translation (Meta M2M100); the English original follows.
Reasoning about safety is essential for embodied agents powered by large language models (LLMs). We introduce the Contextual Action and Plan Safety benchmark (CAPSafe), designed to evaluate safety reasoning at two levels: (1) an action-level dataset that tests whether a single action is safe given its environmental context, and (2) a plan-level dataset that evaluates whether agents can generate safe long-horizon plans for complex hazardous household tasks. CAPSafe models safety conditions where hazards are situational (inferred from object properties), multiple hazards can co-occur within the same environment, and each is annotated with a relative severity level for fine-grained analysis. Evaluating nine closed and open-source LLMs reveals a substantial gap between hazard recognition and safe long-horizon planning. Several models achieve high accuracy in recognizing unsafe actions (exceeding 90% in the best cases), yet safe long-horizon planning remains substantially harder. Gemini-3-pro achieves the highest safe success rate (78.19%), while other models remain below 65%. These results highlight that robust safety reasoning over long action sequences remains an open challenge.
Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese
Topics: Action (physics), Benchmark (surveying), Embodied cognition, Hazard, Risk analysis (engineering), Computer science
GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.28381, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txtSources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange