Capability Self-Assessment: Teaching LLMs to Know Their Limits

Researchers demonstrate that large language models systematically overestimate their capabilities and fail to recognize their limitations. The team proposes Capability Self-Assessment (CSA), a reinforcement learning-based approach that teaches models to accurately evaluate their competence and delegate tasks appropriately, while preserving original functionality.

This research addresses a critical gap in AI reliability that has significant implications for deploying language models in real-world systems. Modern LLMs consistently attempt to solve problems beyond their competence rather than declining or requesting assistance—a behavioral pattern that undermines trust and safety in production environments. The researchers formulate this challenge as a policy-learning problem, testing multiple training approaches to instill accurate self-awareness.

The findings reveal that reinforcement learning substantially outperforms supervised fine-tuning in teaching self-assessment. Notably, traditional supervised fine-tuning actually degrades the model's core capabilities while attempting to improve self-assessment, suggesting that behavioral correction requires more nuanced training methodologies. The out-of-distribution generalization of learned self-assessment behavior indicates this is not narrow task-specific learning but rather a transferable trait that models can acquire.

The practical implications extend across multiple deployment scenarios. For inference-time decisions, accurate self-assessment enables better local-cloud decision-making—models can route complex queries to more capable systems rather than failing silently. During training, CSA provides a signal for targeted data selection, potentially improving overall model development efficiency.

This work addresses a fundamental reliability challenge that becomes more acute as models scale and integrate into critical systems. The ability to know when not to answer is equally valuable as answering correctly, particularly in high-stakes domains where confident wrong answers pose greater risk than admitting uncertainty. The research pathway from this capability could influence how future language models are evaluated and deployed.

• →Modern LLMs systematically overestimate their competence and lack the ability to recognize when they cannot solve a problem.
• →Reinforcement learning effectively teaches Capability Self-Assessment while preserving original model capabilities, outperforming supervised fine-tuning approaches.
• →Supervised fine-tuning for self-assessment degrades the model's core capabilities, suggesting behavioral modification requires specialized training methods.
• →Learned self-assessment generalizes well to out-of-distribution tasks, indicating CSA functions as a transferable model trait rather than narrow task-specific learning.
• →CSA improves practical deployment through better local-cloud decision routing and provides targeted signals for more efficient training data selection.