CyberCorrect: A Cybernetic Framework for Closed-Loop Self-Correction in Large Language Models

In three linesCyberCorrect formalizes LLM self-correction as a closed-loop control system. A tri-modal error detector (self-consistency, verbalized confidence, logic-chain verification) feeds a type-directed correction controller. On CyberCorrect-Bench (440 tasks), the framework achieves 79.8% accuracy (+6.2pp vs SOTA) and reduces overshoot by 41%.

Summary generated by Claude — human-verified