04:00
2026-10-01
arxiv.org
ai-research
CARAT: Do Materials LLMs Reason or Recite?
The CARAT benchmark, detailed in arXiv paper 2609.38340v1, finds that materials LLMs often recite structural relations rather than reason from them: a frozen model quoted a link yet answered identical…