Proceedings of the Annual Meeting of the Cognitive Science Society

Chain-of-thought (CoT) prompting improves the accuracy of large language models (LLMs) by externalizing intermediate reasoning steps, yet these traces are often unreliable, containing local inconsistencies or spurious steps that undermine interpretability and trust. We propose \textbf{MetaCoT-CTRL}, a metacognitive control framework that treats CoT as an explicit cognitive trajectory subject to monitoring and correction. MetaCoT-CTRL first generates a low-cost System-1 draft reasoning chain, then applies a conflict monitor to localize unreliable steps via contradiction, entailment, and task-specific consistency signals. A System-2 targeted repair module selectively revises only high-conflict segments, preserving stable reasoning. To ensure robustness, the repaired chains are further evaluated using causal consistency and counterfactual validation under semantically preserving input perturbations. A cognitive cost regularizer discourages redundant or overly long reasoning. Experiments on ANLI, SVAMP, and CommonQA show that MetaCoT-CTRL improves reasoning reliability and stability over existing CoT-based methods while maintaining efficient inference.