Research map / Video / Video-language understanding
Chain-of-Frames: Advancing Video Understanding in Multimodal LLMs via Frame-Aware Reasoning
Sara Ghazanfari, Francesco Croce, Nicolas Flammarion, Prashanth Krishnamurthy, Farshad Khorrami, Siddharth Garg
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 |
|---|---|---|---|
| fine-tune on chain-of-frames traces instead of standard prompting | InternVL3-8B · Average of 5 benchmarks | accuracy: 67.0 -> 72.1 | Improved |
| fine-tune on chain-of-frames traces (ablation) | InternVL2.5-4B · VSI-Bench | accuracy: 31.8 -> 36.9 | Improved |
| add CoT prompting without fine-tuning (ablation) | InternVL2.5-4B · Video-MME | accuracy: 54.9 -> 54.7 | No clear effect |
| add CoT prompting without fine-tuning (ablation) | InternVL2.5-4B · VidHal | accuracy: 74.0 -> 77.0 | Improved |
| fine-tune on QA pairs only (no reasoning) (ablation) | InternVL2.5-4B · VidHal | accuracy: 74.0 -> 64.1 | Hurt |
| fine-tune on QA pairs only (no reasoning) (ablation) | InternVL2.5-4B · MVBench | accuracy: 70.8 -> 73.4 | Improved |
| fine-tune on CoT traces without frame references (ablation) | InternVL2.5-4B · EventHallusion | accuracy: 62.5 -> 53.1 | Hurt |
| add explicit frame references to CoT traces (ablation) | InternVL2.5-4B · EventHallusion | accuracy: 53.1 -> 71.2 | Improved |
| add explicit frame references to CoT traces (ablation) | InternVL2.5-4B · VSI-Bench | accuracy: 34.3 -> 36.9 | Improved |
| train on synthetic CLEVRER-based CoF data vs real VideoEspresso data (equal size) (ablation) | InternVL2.5-4B · five video benchmarks | accuracy | Improved |
| combine real and synthetic CoF data (ablation) | InternVL2.5-4B · five video benchmarks | accuracy | Mixed |
| CoF fine-tuning vs multi-stage video CoT (M-LLM) | InternVL2.5-4B · NExT-QA | accuracy: 75.3 -> 79.6 | Improved |
Closest papers by meaning
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.92)
- MMR-V: What's Left Unsaid? A Benchmark for Multimodal Deep Reasoning in Videos (ICLR 2026 · similarity 0.91)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.91)
- VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning? (ICLR 2026 · similarity 0.91)
- Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation (CVPR 2025 · similarity 0.91)
- VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · similarity 0.91)
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.
