When Do Hallucinations Arise? A Graph Perspective on the Evolution of Path Reuse and Path Compression
arXiv:2604.03557v1 Announce Type: new Abstract: 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 int
arXiv:2604.03557v1 Announce Type: new Abstract: 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.
Executive Summary
This article provides a novel graph perspective on the evolution of path reuse and path compression in large language models (LLMs), shedding light on the mechanisms behind reasoning hallucinations. By modeling next-token prediction as a graph search process, the authors demonstrate that reasoning hallucinations arise from two fundamental mechanisms: path reuse and path compression. Path reuse occurs when memorized knowledge overrides contextual constraints during early training, while path compression occurs when frequently traversed multi-step paths collapse into shortcut edges in later training. The authors' unified explanation for reasoning hallucinations in LLMs has significant implications for understanding and mitigating these failures in downstream applications.
Key Points
- ▸ The article presents a graph perspective on the evolution of path reuse and path compression in LLMs.
- ▸ Path reuse and path compression are identified as fundamental mechanisms behind reasoning hallucinations in LLMs.
- ▸ The authors' unified explanation provides a new understanding of reasoning hallucinations in LLMs and connected to well-known behaviors in downstream applications.
Merits
Strength
The article provides a novel and unified explanation for reasoning hallucinations in LLMs, which is a significant contribution to the field.
Demerits
Limitation
The article assumes a specific graph structure and edge weights, which may not be universally applicable to all LLMs.
Expert Commentary
The article presents a significant advancement in our understanding of reasoning hallucinations in LLMs. By modeling next-token prediction as a graph search process, the authors provide a novel and unified explanation for these failures. The identification of path reuse and path compression as fundamental mechanisms behind reasoning hallucinations is a crucial contribution to the field. However, the article's reliance on a specific graph structure and edge weights may limit its universality. Nevertheless, the article's findings have significant implications for understanding and mitigating reasoning hallucinations in downstream applications of LLMs.
Recommendations
- ✓ Future research should investigate the generalizability of the article's findings to different graph structures and edge weights.
- ✓ The article's results may inform the development of new methods for mitigating reasoning hallucinations in LLMs, such as graph-based pruning and regularization techniques.
Sources
Original: arXiv - cs.AI