Capability Self-Assessment: Teaching LLMs to Know Their Limits

In three linesModern LLMs systematically overestimate their competence and attempt unsolvable queries. Researchers propose Capability Self-Assessment (CSA), formulated as a policy-learning problem using reinforcement learning, to teach models to recognize their limits. RL significantly outperforms supervised fine-tuning, preserves original capabilities, and generalizes out-of-distribution.

Summary generated by Claude — human-verified