从图结构视角揭示大模型幻觉产生的两个核心机制。
When Do Hallucinations Arise? A Graph Perspective on the Evolution of Path Reuse and Path Compression
- 将文本生成建模为图上的路径搜索，区分上下文内与上下文外推理。
- 早期训练中记忆路径复用导致违背上下文，后期多步路径压缩成捷径。
- 解释了推理错误与实际应用中常见问题的内在关联，适合模型可靠性研究者。
大语言模型（LLMs）中的推理幻觉常表现为流畅但缺乏依据的结论，违背给定上下文或事实知识。尽管这类现象广泛存在，但解码器仅有的Transformer如何产生它们仍不清晰。本文将下一个词预测建模为底层图上的图搜索过程，其中实体对应节点，学习到的转移形成边。在此视角下，上下文推理是受限于采样子图的搜索（内在推理），而无上下文查询依赖于底层图中的记忆结构（外在推理）。我们发现，推理幻觉源于两种基本机制：路径复用（Path Reuse），即早期训练中记忆知识覆盖上下文约束；路径压缩（Path Compression），即后期训练中频繁遍历的多步路径坍缩为捷径边。这两种机制共同提供了一个统一解释框架，揭示了大模型推理幻觉的根源，并与下游应用中观察到的行为密切相关。
原文摘要 · Abstract (English)
Reasoning hallucinations in large language models (LLMs) often appear as fluent yet unsupported conclusions that violate either the given context or underlying factual knowledge. Although such failures are widely observed, the mechanisms by which decoder-only Transformers produce them remain poorly understood. We model next-token prediction as a graph search process over an underlying graph, where entities correspond to nodes and learned transitions form edges. From this perspective, contextual reasoning is a constrained search over a sampled subgraph (intrinsic reasoning), while context-free queries rely on memorized structures in the underlying graph (extrinsic reasoning). We show that reasoning hallucinations arise from two fundamental mechanisms: \textbf{Path Reuse}, where memorized knowledge overrides contextual constraints during early training, and \textbf{Path Compression}, where frequently traversed multi-step paths collapse into shortcut edges in later training. Together, these mechanisms provide a unified explanation for reasoning hallucinations in LLMs and connected to well-known behaviors observed in downstream applications.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。