arXiv:2411.15993v2 Announce Type: replace
Abstract: Large language models (LLMs) have demonstrated strong capabilities in text understanding and generation. However, they often lack factuality, producing a mixture of true and false information, especially in long-form generation. In this work, we investigates the factuality of long-form text generation across various large language models (LLMs), including GPT-4, Gemini-1.5-Pro, Claude-3-Opus, Llama-3-70B, and Mistral. Our analysis reveals that factuality tend to decline in later sentences of the generated text, accompanied by a rise in the number of unsupported claims. Furthermore, we explore the effectiveness of different evaluation settings to assess whether LLMs can accurately judge the correctness of their own outputs: Self-Known (the percentage of supported atomic claims, decomposed from LLM outputs, that the corresponding LLMs judge as correct) and Self-Unknown (the percentage of unsupported atomic claims that the corresponding LLMs judge as incorrect). Empirically, we observe a positive correlation between higher Self-Known scores and improved factuality, whereas higher Self-Unknown scores are associated with reduced factuality. Interestingly, the number of unsupported claims can increase even without significant changes in a model’s self-judgment scores (Self-Known and Self-Unknown), likely as a byproduct of long-form text generation. We also derive a mathematical framework linking Self-Known and Self-Unknown scores to factuality: $textrm{Factuality}=frac{1-textrm{Self-Unknown}}{2-textrm{Self-Unknown}-textrm{Self-Known}}$, which aligns with our empirical observations. Additional Retrieval-Augmented Generation (RAG) experiments further highlight the limitations of current LLMs in long-form generation and underscore the need for continued research to improve factuality in long-form text.
