New ACL study reveals why LLMs hallucinate on graphs and tables

A new paper from Shanghao Li and colleagues, accepted at ACL 2026, provides a mechanistic explanation for why large language models (LLMs) hallucinate when reasoning over structured knowledge such as graphs and tables—even when sufficient information is present. The research challenges the common assumption that hallucinations stem from missing or contradictory data, showing instead that they arise from systematic internal dynamics. Using causal tracing and attention analysis, the team identified two main failure modes. First, attention heads concentrate disproportionately on shortcut-like structural cues (e.g., repeated entity names) rather than distributing across the full context, leading the model to overlook critical relational information. Second, the feed-forward layers fail to semantically ground the provided knowledge, causing the model to fall back on its parametric memory—its pre-trained associations—rather than using the external input.

Crucially, the study shows that these patterns are not random but reproducible and consistent across tasks. The attention allocation varies by task type, but feed-forward grounding failures are a universal marker of hallucination. The findings hold for single-hop graphs, multi-hop graph reasoning, and tabular data, suggesting a generalizable principle. The authors propose that hallucination detection systems can leverage these mechanistic signatures—monitoring feed-forward layer activations or attention distribution entropy—to flag unreliable outputs in real time. This work opens a path toward building more faithful reasoning models by modifying internal computations rather than simply scaling data or model size.