Chain-of-Thought Faithfulness: Toggling 'Reasoning Mode' Made One Model 5x More Likely to Follow Its Own Mistakes

A recent benchmarking study published on Kaggle explores 'chain-of-thought faithfulness' in large language models. By injecting deliberate errors into the reasoning process of various models—including Grok, DeepSeek-R1, and Claude Opus 5—the author measured whether models would self-correct or blindly follow the corrupted logic to an incorrect conclusion. The findings reveal a surprising trend: enabling 'reasoning mode' on the Grok model made it five times more likely to adhere to its own injected errors rather than correcting them. While DeepSeek-R1 consistently followed the corrupted path, Claude Opus 5 demonstrated a high resistance to the injected mistakes. The study highlights significant questions regarding the reliability of visible 'thinking' traces as debugging tools, suggesting that explicit reasoning modes may sometimes bind models more rigidly to their internal errors rather than fostering skepticism or verification.