Research map / Video / Video-language understanding
VirtueBench: Evaluating Trustworthiness under Uncertainty in Long Video Understanding
Xueqing Yu, Bohan Li, Yan Li, Zhenheng 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 |
|---|---|---|---|
| remove explicit honesty instruction from prompt (ablation) | Qwen3-VL-8B-Instruct · VirtueBench (64 frames) | refusal accuracy (%): 36.3 -> 22.8 | Hurt |
| remove explicit honesty instruction from prompt (ablation) | Qwen3-VL-32B-Instruct · VirtueBench (64 frames) | refusal accuracy (%): 60.2 -> 33.7 | Hurt |
| remove explicit honesty instruction from prompt (ablation) | Qwen3-VL-32B-Thinking · VirtueBench (64 frames) | refusal accuracy (%): 43.1 -> 9.4 | Hurt |
| increase input frames from 64 to 1024 | Qwen3VL-32B · VirtueBench | overall accuracy (%): 50.83 -> 32.54 | Hurt |
| increase input frames from 64 to 1024 | Gemini-2.5-Flash · VirtueBench | overall accuracy (%): 58.96 -> 53.66 | Hurt |
| RL/reasoning training (Mimo-VL-RL vs base) | Mimo-VL-7B · VirtueBench (64 frames) | refusal accuracy (%): 32.35 -> 40.78 | Improved |
| reasoning-optimized GPT-5 vs GPT-5-chat | GPT-5 · VirtueBench (64 frames) | refusal accuracy (%): 19.80 -> 33.53 | Improved |
| larger model size Qwen2.5-VL 7B to 72B | Qwen2.5-VL · VirtueBench (64 frames) | refusal accuracy (%): 53.33 -> 80.20 | Improved |
| thinking variant of Qwen3-VL-32B vs instruct | Qwen3-VL-32B · VirtueBench (64 frames) | refusal accuracy (%): 60.20 -> 43.14 | Hurt |
| RL training on InternVL3.5-38B | InternVL3.5-38B · VirtueBench (64 frames) | refusal accuracy (%): 0.59 -> 0.59 | No clear effect |
Closest papers by meaning
- MoHoBench: Assessing Honesty of Multimodal Large Language Models via Unanswerable Visual Questions (AAAI 2026 · similarity 0.91)
- MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding (NeurIPS 2024 · similarity 0.90)
- VMDT: Decoding the Trustworthiness of Video Foundation Models (NeurIPS 2025 · similarity 0.89)
- VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding (NeurIPS 2025 · similarity 0.89)
- CertainlyUncertain: A Benchmark and Metric for Multimodal Epistemic and Aleatoric Awareness (ICLR 2025 · similarity 0.89)
- MMR-V: What's Left Unsaid? A Benchmark for Multimodal Deep Reasoning in Videos (ICLR 2026 · similarity 0.89)
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.
