arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2407.06071cs.CLcs.AI
收藏

从循环到失误:语言模型在不确定性下的回退行为

From Loops to Oops: Fallback Behaviors of Language Models Under Uncertainty

Blavatnik School of Computer Science, Tel Aviv University(特拉维夫大学布拉瓦尼克计算机科学学院)

Maor Ivgi, Ori Yoran, Jonathan Berant, Mor Geva

AI总结:

该研究将LLMs的幻觉、序列重复等不良行为视为认知不确定性下的回退表现,通过分析不同训练token、参数规模和指令微调的模型,揭示回退行为的一致顺序,并发现常见解码技术缓解部分问题时会加剧幻觉。

摘要:

大型语言模型(LLMs)常表现出幻觉、序列重复等不良行为。我们提出将这些行为视为模型在认知不确定性下的回退表现,并研究它们之间的关联。我们对回退行为(序列重复、退化文本、幻觉)进行分类,在同一模型家族中,针对预训练token数量、参数规模或是否包含指令跟随训练不同的模型展开广泛分析。实验显示,在所有这些维度上,回退行为存在清晰且一致的顺序:LLM越先进(即训练token更多、参数更大或经指令微调),其回退行为从序列重复转向退化文本,再转向幻觉。此外,即使是最佳性能模型,在生成单个序列时也存在相同顺序:不确定性增加时,模型从生成幻觉转向退化文本,最终变为序列重复。最后,我们证明随机采样等常见解码技术虽能缓解序列重复等不良行为,却会增加更难检测的幻觉。

英文摘要:

Large language models (LLMs) often exhibit undesirable behaviors, such as hallucinations and sequence repetitions. We propose to view these behaviors as fallbacks that models exhibit under epistemic uncertainty, and investigate the connection between them. We categorize fallback behaviors - sequence repetitions, degenerate text, and hallucinations - and extensively analyze them in models from the same family that differ by the amount of pretraining tokens, parameter count, or the inclusion of instruction-following training. Our experiments reveal a clear and consistent ordering of fallback behaviors, across all these axes: the more advanced an LLM is (i.e., trained on more tokens, has more parameters, or instruction-tuned), its fallback behavior shifts from sequence repetitions, to degenerate text, and then to hallucinations. Moreover, the same ordering is observed during the generation of a single sequence, even for the best-performing models; as uncertainty increases, models shift from generating hallucinations to producing degenerate text and finally sequence repetitions. Lastly, we demonstrate that while common decoding techniques, such as random sampling, alleviate unwanted behaviors like sequence repetitions, they increase harder-to-detect hallucinations.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑