Research map / Video / Video-language understanding
Unstitching the Chimera: Frame-Level Risk and Train-Free Mitigation for Video Hallucination
Songyuan Yang, Guijian Tang, Kun Hu, Haotian Wang, Shixuan Liu, Wenjing Yang, Long Lan, Huibin Tan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply train-free CH-M mitigation at inference | 6 VideoLLMs (7B) · 9 video benchmarks | Accuracy: +1.0 to +5.6 points across models and benchmarks | Improved |
| apply CH-M | Video-LLaVA-7B · VidHalluc | Accuracy: 40.3 -> 45.9 | Improved |
| apply CH-M | 6 VideoLLMs (7B) · Risk diagnostics | CH-Risk: 0.33 -> 0.23 (Video-LLaVA-7B) | Improved |
| add pathway alignment term (lambda=0.3) to sSAFR (ablation) | sSAFR on LLaVA-Video-7B · NExT-QA | Accuracy: 74.2 -> 74.7 | Improved |
| add segment prior (gamma=0.4) to sSAFR (ablation) | sSAFR on LLaVA-Video-7B · NExT-QA | Accuracy: 73.9 -> 74.7 | Improved |
| use event segments instead of uniform windows in sSAFR (ablation) | sSAFR on LLaVA-Video-7B · NExT-QA | Accuracy: 73.5 -> 74.7 | Improved |
| add sSAFR alone (ablation) | LLaVA-Video-7B · VidHalluc | Accuracy: 76.6 -> 79.3 | Improved |
| add RTC alone (hard cap rho=3) (ablation) | LLaVA-Video-7B · NExT-QA | Accuracy: 73.2 -> 73.6 | Improved |
| add RTC after sSAFR (ablation) | LLaVA-Video-7B · NExT-QA | Accuracy: 74.7 -> 75.1 | Improved |
| apply sSAFR before RTC rather than reversed order (ablation) | CH-M on LLaVA-Video-7B · VidHalluc | Accuracy: 79.7 -> 80.2 | Improved |
| set coverage alpha to 0.8 rather than 0.7 or 0.9 (ablation) | CH-M · NExT-QA and VidHalluc | Delta Accuracy | Improved |
| set RTC cap rho to 3 rather than smaller or larger (ablation) | CH-M · NExT-QA and VidHalluc | Delta Accuracy | Improved |
Closest papers by meaning
- Learning to Decode Against Compositional Hallucination in Video Multimodal Large Language Models (ICML 2026 · similarity 0.91)
- Mitigating Hallucination in VideoLLMs via Temporal-Aware Activation Engineering (NeurIPS 2025 · similarity 0.91)
- Causally-Grounded Dual-Path Attention Intervention for Object Hallucination Mitigation in LVLMs (AAAI 2026 · similarity 0.91)
- One Token, Two Fates: A Unified Framework via Vision Token Manipulation Against MLLMs Hallucination (CVPR 2026 · similarity 0.91)
- Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.90)
- Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling (NeurIPS 2025 · 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.
