Static Evaluation of Model Switching in LLM Agents Scores the Wrong World

11 hours ago (arxiv.org)