Video grounding errors cause most hallucinations in video understanding models
Beneath the Scores: Rethinking Hallucination Evaluation for Video Understanding Models
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Video understanding systems that use multiple steps to analyze videos sometimes make things up, called hallucinations. The authors show that problems mainly come from finding the right place in the video rather than from watching it incorrectly or reasoning. They created a way to test each step's impact on errors and found that pointing to the wrong area is more harmful than missing timing details. This means current ways to judge these models might not reveal their true weaknesses.
What this means in practice
- •For video ai developers: Improve video understanding systems by focusing on more accurate grounding evaluation and correction to reduce hallucinations.
- •For ai evaluation teams: Use intervention-based testing to better diagnose which stage causes errors, improving model trustworthiness.
Authors
Shuzhi Gong, Fengze Sun, Yuansan Liu
Abstract
Video understanding is increasingly performed by multi-stage LLM agents that separate temporal grounding, visual observation, and reasoning. Yet these stages are typically evaluated on different benchmarks and distributions, making it difficult to determine where hallucinations originate. We first organize existing benchmarks around these stages and show that their scores provide inconsistent diagnostic signals: stronger stage-level performance does not reliably imply lower downstream hallucination, and even benchmarks targeting the same capability can disagree. We therefore introduce a causal stage-intervention protocol that overwrites individual stages while holding the downstream task fixed. Across 60,008 runs on three video-agent architectures, we find that grounding is the dominant source of downstream error, with roughly four times the causal impact of corrupting visual observations. Successful grounding depends primarily on locating the correct region rather than precise temporal overlap, explaining why standard mIoU metrics poorly predict downstream reliability. We further find that incorrect evidence is substantially more harmful than missing evidence. Finally, auditing existing benchmarks against these interventions reveals that their scores do not reliably predict causal cascade sensitivity and can fail under distribution shift. These results motivate intervention-based, stage-aware evaluation for trustworthy video agents.