Research map / Video / Video-language understanding
SmartSight: Mitigating Hallucination in Video-LLMs Without Compromising Video Understanding via Temporal Attention Collapse
Yiming Sun, Mi Zhang, Feifei Li, Geng Hong, Min Yang
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 |
|---|---|---|---|
| add TAC-score selection over random sampling of 10 responses (ablation) | SmartSight on Qwen2.5-VL-7B · VRIPT-HAL | F1: 41.2 -> 50.1 | Improved |
| add TAC-score selection over random sampling of 10 responses (ablation) | SmartSight on Qwen2.5-VL-7B · Video-MME | accuracy: 51.8 -> 56.3 | Improved |
| use frame-level plus segment-level collapse instead of frame-level only (ablation) | SmartSight on Qwen2.5-VL-7B · VRIPT-HAL | F1: 48.2 -> 50.1 | Improved |
| add early stopping at the VAV point (ablation) | SmartSight on Qwen2.5-VL-7B · VRIPT-HAL | F1: 50.1 -> 50.0 | No clear effect |
| add early stopping at the VAV point (ablation) | SmartSight on Qwen2.5-VL-7B · Video-MME | accuracy: 56.3 -> 56.22 | No clear effect |
| apply SmartSight to greedy Qwen2.5-VL-7B | Qwen2.5-VL-7B · VRIPT-HAL | F1: 47.9 -> 50.0 | Improved |
| apply VCD to Qwen2.5-VL-7B | Qwen2.5-VL-7B · Video-MMMU | accuracy: 45.7 -> 45.5 | Hurt |
| apply TCD to Qwen2.5-VL-7B | Qwen2.5-VL-7B · Video-MMMU | accuracy: 45.7 -> 44.7 | Hurt |
| apply DINO-HEAL to Video-R1 | Video-R1 · Video-MMMU | accuracy: 48.6 -> 46.3 | Hurt |
| apply SmartSight to Video-R1 | Video-R1 · Video-MMMU | accuracy: 48.6 -> 51.9 | Improved |
| apply SmartSight to 32B model | Qwen2.5-VL-32B · VRIPT-HAL | F1: 59.1 -> 59.4 | No clear effect |
| apply VCD to 32B model | Qwen2.5-VL-32B · Video-MMMU | accuracy: 52.1 -> 45.3 | Hurt |
Closest papers by meaning
- Mitigating Hallucination in VideoLLMs via Temporal-Aware Activation Engineering (NeurIPS 2025 · similarity 0.93)
- Causally-Grounded Dual-Path Attention Intervention for Object Hallucination Mitigation in LVLMs (AAAI 2026 · similarity 0.93)
- Learning to Decode Against Compositional Hallucination in Video Multimodal Large Language Models (ICML 2026 · similarity 0.92)
- Prefill-Time Intervention for Mitigating Hallucination in Large Vision-Language Models (CVPR 2026 · similarity 0.92)
- Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 2025 · similarity 0.92)
- Not All Tokens and Heads Are Equally Important: Dual-Level Attention Intervention for Hallucination Mitigation (AAAI 2026 · similarity 0.92)
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.
