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) and type-directed correction controller achieve 79.8% accuracy on CyberCorrect-Bench (440 reasoning tasks), +6.2pp over existing methods, reducing overshoot by 41% via convergence control.

Summary generated by Claude — human-verified