一句话:提出基于 logprob 的框架,通过反义词候选与 token 级不确定性度量主动抑制 LLM 幻觉。
大语言模型在文本生成方面表现优异,但仍容易产生幻觉。现有缓解方法如检索增强生成(RAG)和自评估存在效率与偏差方面的局限。论文提出一种新的基于 logprob 的框架,主动量化模型不确定性,包含两项创新:一是在 token 预测中融入反义词候选,二是从 token 级 logprobs 推导不确定性度量。实验表明,该方法能有效抑制幻觉,同时保持计算效率与透明度,为可靠 LLM 输出提供了一种可扩展方案。
Large Language Models (LLMs) excel in text generation but remain prone to hallucinations. Existing mitigation methods like Retrieval-Augmented Generation (RAG) and self-assessment face efficiency and bias limitations. We propose a novel logprob-based framework that proactively quantifies model uncertainty through two innovations: (1) integrating antonym candidates in token predictions, and (2) deriving uncertainty metrics from token-level logprobs. Experiments demonstrate our method effectively suppresses hallucinations while maintaining computational efficiency and transparency, offering a scalable solution for reliable LLM outputs.