Recent advancements in Large Language Models (LLMs) have demonstrated substantial improvement across various reasoning tasks. However, a critical limitation persists regarding their ability to handle multi-step complex reasoning processes. Although LLMs possess extensive knowledge, their reasoning often suffers from hallucinations and logical inconsistencies, with most prompting methods lacking self-correction capabilities and failing to reliably elicit latent reasoning abilities. To address these issues, we propose Bidirectional Correction of Thought (BiCoT), a self-correcting prompting mechanism that employs iterative error correction approach by incorporating challenging exemplars as few-shot prompts. Additionally, our approach leverages proof-by-contradiction principles to systematically analyze, verify, and rectify the model’s generated reasoning chains, ensuring the correctness of the reasoning process. Through evaluation on eight reasoning benchmarks, BiCoT achieves average improvements of 2.6% on arithmetic reasoning, 5.3% on commonsense reasoning, and a 2.5% gain on the symbolic reasoning using GPT-3.5-turbo, demonstrating consistent robustness across diverse reasoning scenarios.