Abstract
Large Language Models (LLMs) are powerful tools for text generation, translation, and summarization, but they often suffer from hallucinations-instances where they fail to maintain the fidelity and coherence of contextual information during decoding, sometimes overlooking critical details due to their sampling strategies and inherent biases from training data and fine-tuning discrepancies. These hallucinations can propagate through the web, affecting the trustworthiness of information disseminated online. To address this issue, we propose a novel decoding strategy that leverages absorbing Markov chains to quantify the significance of contextual information and measure the extent of information loss during generation. By considering all possible paths from the first to the last token, our approach enhances the reliability of model outputs without requiring additional training or external data. Evaluations on datasets including TruthfulQA, FACTOR, and HaluEval highlight the superior performance of our method in mitigating hallucinations, underscoring the necessity of ensuring accurate information flow in web-based applications.
Abstract (translated)
大型语言模型（LLMs）是文本生成、翻译和摘要的强大工具，但它们经常会出现幻觉现象——即在解码过程中无法保持上下文信息的保真度和连贯性，有时会因采样策略及训练数据和微调差异带来的固有偏见而忽略关键细节。这些幻觉现象可能在网络中传播，影响在线发布的信息的信任度。为了解决这一问题，我们提出了一种新的解码策略，该策略利用吸收马尔可夫链来量化上下文信息的重要性，并衡量生成过程中信息损失的程度。通过考虑从第一个到最后一个标记的所有可能路径，我们的方法提高了模型输出的可靠性，而无需额外训练或外部数据。在包括TruthfulQA、FACTOR和HaluEval在内的数据集上的评估显示，我们的方法在减轻幻觉方面表现出色，强调了确保基于网络应用程序中信息准确流动的重要性。
URL
https://arxiv.org/abs/2410.20340
