大语言模型的自我恒定性现象及其在知识增强任务上的应用OA

Recent studies have revealed significant"source bias"in retrieval and large language models(LLM),where these models exhibit a preference for model-generated content over authentic human-written data.Humans similarly demonstrate"self-consistency"phenomena,characterized by a tendency to maintain coherent self-concepts and favor existing knowledge over external facts.It was investigated whether LLMs exhibit human-like cognitive biases.It was conducted from two perspectives:an explicit approach(content directly generated by the LLM)and an implicit approach(using prompts to induce the models to perceive the content as self-generated).Confidence was evaluated using both self-assessment and consistency methods.Systematic experiments were carried out on several models,including GPT-4o,DeepSeek-R1,Llama2(7B-70B),and Qwen2(7B-72B).The results demonstrated that under implicit conditions,LLMs exhibit pronounced self-consistency phenomena,displaying significantly higher confidence in self-attributed content.Leveraging this characteristic yields consistent performance gains:the proposed role-based prompting strategy improved accuracy on TriviaQA,NQ,HotpotQA,FEVER,and ZsRE,and maintained a clear advantage under high levels of noisy retrieved documents,indicating strong robustness and generalization.Finally,our mechanistic analysis revealed that instruction fine-tuning and reinforcement learning from human feedback were the primary factors contributing to such biases in models,explaining how self-consistency emerges during training.