Critical-CoT: A Robust Defense Framework Against Reasoning-Level Backdoor Attacks on Large Language Model Reasoning

Researchers propose the Critical-CoT defense framework, which uses two-stage fine-tuning to cultivate critical thinking in large language models, enabling them to automatically identify potential backdoors and refuse to generate malicious reasoning steps. Such reasoning-level backdoor attacks cause models to insert malicious reasoning steps into the chain of thought, making backdoored answers appear plausible and hard to detect. Experiments across multiple models and datasets show that the method is strongly robust against both in-context-learning-based and fine-tuning-based backdoor attacks, and has cross-domain and cross-task generalization ability.