CAR-bench: Evaluating the Consistency and Limit-Awareness of LLM Agents under Real-World Uncertainty - Explained Simply | ArXiv Explained