Video hallucination detectors show limited reliability across tasks
Can We Trust Video Hallucination Detectors? VidHalLoc for Evaluating the Evaluators
Computer Vision and Pattern Recognition
Summary
Video-language models sometimes imagine things that don't really appear in the video, and detecting these errors is important. The authors created a new test called VidHalLoc to check how well different detectors find these mistakes in tasks like answering questions about videos or creating video captions. They built this test with a special process called VideoHALO to generate tricky examples efficiently. Their study found that even the best detectors only got about one-third of the cases right, showing current tools are not very reliable at spotting these video hallucinations.
What this means in practice
- •For video ai developers: Identify the limitations of current video hallucination detectors to improve model trustworthiness in video question answering and captioning applications.
- •For content moderation teams: Use findings to guide evaluation of tools that check for misinformation in videos generated by AI models.
Authors
Xinyu Chen, Adnan Mahmood, Mark Dras
Abstract
Video-language models and video agents can produce hallucinations that conflict with spatiotemporal evidence. Existing benchmarks mainly evaluate model hallucinations, and heterogeneous mechanisms make detector reliability difficult to compare. We introduce VidHalLoc, a benchmark that evaluates hallucination detection methods under a unified diagnostic evaluation protocol using 2,000 adversarial hallucination samples across Video Question Answering and Video Captioning tasks, spanning Ontology and Dynamic hallucination categories. To construct VidHalLoc efficiently, we introduce VideoHALO, a Harness Engineering-informed multi-agent workflow that decomposes data construction into four executable stages supported by a memory system and a communication protocol. Evaluation of fifteen methods reveals that the four dedicated detectors peak at an Overall accuracy of only 34.63%, indicating limited reliability across video hallucination types [Dataset Repository: https://huggingface.co/datasets/wesfggfd/VidHalLoc].