Research map / Video / Video-language understanding
Diagnosing and Correcting Bias in MLLM for Long Video Understanding
Xusheng Liang, Jianqiao Sun, Hao Zhang, Yulei Niu, Hengshuang Zhao, Jiawei Ma
Read the paperCheck your idea against 137,217 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.
Studies bias in MLLMs for long-video question answering under sparse frame sampling and proposes a correction method (details unavailable beyond abstract).
Problem
Long-video QA where answers depend on a few decisive moments, but MLLMs must sample frames sparsely due to memory limits, risking omission of key evidence.
Approach
Argues that simply enlarging context size does not fundamentally fix performance, and diagnoses a bias in MLLMs for long video understanding, then corrects it; the specific correction is not stated in the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CLEAR counterfactual residual steering | Qwen3-VL-8B · MLVU | accuracy: 59.0 -> 66.8 | Improved |
| add CLEAR counterfactual residual steering | InternVL3-8B · LVBench | accuracy: 43.4 -> 49.7 | Improved |
| add CLEAR counterfactual residual steering | Qwen3-VL-8B · S-MME overall | ShortErr: 44.0 -> 36.9 | Improved |
| add CLEAR counterfactual residual steering | LLaVA-Video-7B · S-MME overall | accuracy: 9.5 -> 12.7 | Improved |
| use both question-only and salient-view probes instead of only the question-only probe (ablation) | CLEAR on Qwen3-VL-8B · S-MME overall | accuracy: 23.3 -> 26.5 | Improved |
| use both probes instead of only the salient-view probe (ablation) | CLEAR on Qwen3-VL-8B · S-MME overall | accuracy: 23.7 -> 26.5 | Improved |
| add question-only probe alone (ablation) | Qwen3-VL-8B · S-MME linguistic-bias cases | accuracy: 22.2 -> 26.1 | Improved |
| add salient-view probe alone (ablation) | Qwen3-VL-8B · S-MME visual-bias cases | accuracy: 18.8 -> 22.9 | Improved |
| raise residual strength alpha from 0.15 to 0.35 (ablation) | CLEAR on Qwen3-VL-8B · S-MME overall | accuracy: 26.5 -> 24.8 | Hurt |
| lower residual strength alpha from 0.15 to 0.05 (ablation) | CLEAR on Qwen3-VL-8B · S-MME overall | accuracy: 26.5 -> 23.4 | Hurt |
| scale frames from 64 to 128 without CLEAR (ablation) | Qwen3-VL-8B · Video-MME overall | accuracy: 65.1 -> 67.0 | Improved |
| add CLEAR at 32 frames versus scaling to 128 frames (ablation) | Qwen3-VL-8B · Video-MME overall | accuracy: 67.0 -> 68.4 | Improved |
Surprising findings
- Enlarging context size alone does not fundamentally resolve long-video performance limits, according to the authors' argument.
Closest papers by meaning
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.92)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.92)
- CG-Bench: Clue-grounded Question Answering Benchmark for Long Video Understanding (ICLR 2025 · similarity 0.91)
- OmniVideoBench: Towards Audio-Visual Understanding Evaluation for Omni MLLMs (ICLR 2026 · similarity 0.90)
- RTV-Bench: Benchmarking MLLM Continuous Perception, Understanding and Reasoning through Real-Time Video (NeurIPS 2025 · similarity 0.90)
- MMR-V: What's Left Unsaid? A Benchmark for Multimodal Deep Reasoning in Videos (ICLR 2026 · 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.
