Research map / Vision-language and multimodal / Multimodal LLMs
VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces VideoHallu, a ~3.2K-QA benchmark on synthetic physics/commonsense-violating videos, shows leading VLMs hallucinate on it, and finds GRPO post-training on mixed real+synthetic data improves synthetic-video accuracy without hurting real-world benchmarks.
Problem
VLM video evaluations rely on positive-control real-world videos that resemble training data, so it is unclear whether models truly perceive visual evidence or exploit language priors when physically impossible or logically inconsistent content appears.
Approach
Uses negative-control synthetic videos from modern generators (Veo2, Sora, Kling, etc.) containing physics, commonsense, alignment and consistency violations, with expert-written QA pairs, and compares SFT vs GRPO post-training with a mix of real and synthetic data.
Key results
- Video-R1 achieves the best overall VideoHallu accuracy among evaluated models at 53.64%, with top-tier models such as GPT-4o-mini at 50.36% and Gemini-2.5-Pro at 49.18%.
- Qwen2.5-VL-7B GRPO reaches 57.69% overall on VideoHallu versus 51.02% for SFT and 52.98% for the base model.
- Qwen2.5-VL-7B GRPO trained on combined real and synthetic data keeps MVBench at 70.1 and raises MMVU from 58.7 to 61.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GRPO rather than SFT on combined real and synthetic video data (ablation) | Qwen2.5-VL-7B · VideoHallu test | accuracy: 51.02 -> 57.69 | Helped |
| Use GRPO rather than SFT on combined real and synthetic video data (ablation) | LLaVA-OneVision · VideoHallu test | accuracy: 38.82 -> 41.38 | Helped |
| Train with combined real and synthetic video data instead of real-only data (ablation) | Qwen2.5-VL-7B with GRPO · VideoHallu test | accuracy: 53.05 -> 57.69 | Helped |
| Train with combined real and synthetic video data instead of synthetic-only data (ablation) | Qwen2.5-VL-7B with GRPO · VideoHallu test | accuracy: 55.41 -> 57.69 | Helped |
| Add synthetic video data to real-world GRPO training (ablation) | Qwen2.5-VL-7B · VideoHallu test | accuracy: 53.05 -> 57.69; 53.05 -> 55.41 | Mixed |
| Post-train with combined real and synthetic video data | Qwen-2.5VL-7B · MMVU | accuracy: 58.7 -> 61.3 | Helped |
| Post-train with combined real and synthetic video data | Qwen-2.5VL-7B · MVBench | accuracy: 69.6 -> 70.1 | Helped |
Surprising findings
- R1-finetuned models (Video-R1-CoT, VideoChat-R1-think) show limited or negative gains over the Qwen2.5-VL-7B base on synthetic videos, with chain-of-thought amplifying language-prior hallucinations.
- On critical thinking questions no evaluated model exceeds 25% accuracy, while commonsense-only questions answerable without video reach 100%.
Limitations
- Scalability: generating high-quality expert annotations and fine-tuning at scale is costly.
- Small synthetic training set (800 pairs) necessitates mixing with real data.
- Evaluation relies on GPT-4o-mini as judge, validated on only 200 sampled pairs with 97% agreement.
Experimental setup
Datasets: VideoHallu Video-R1 training data LLaVA-Video MVBench MMVU
Benchmarks: VideoHallu MVBench MMVU
Metrics: accuracy
Baselines: Qwen2.5-VL-7B Video-R1 VideoChat-R1 Gemini-2.5-Pro GPT-4o-mini LLaVA-OneVision InternVL3-38B
Builds on: GRPO Video-R1 DeepSeek-R1
Closest papers by meaning
- VidHalluc: Evaluating Temporal Hallucinations in Multimodal Large Language Models for Video Understanding (CVPR 2025 · similarity 0.91)
- Black Swan: Abductive and Defeasible Video Reasoning in Unpredictable Events (CVPR 2025 · similarity 0.91)
- MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding (NeurIPS 2024 · similarity 0.91)
- PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding (NeurIPS 2025 · similarity 0.91)
- QUANTIPHY: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models (CVPR 2026 · similarity 0.91)
- Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning (ICLR 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.