Research map / Video / Video-language understanding
Reflect-R1: Evidence-Driven Reflection for Self-Correction in Long Video Understanding
Shuimu Chen, Yuteng Chen, Yuanshen Guan, Zebang Cheng, Zeyu Zhang, shengqian qin, Bin Xia, Jiaran Li, Wenming Yang, Fei Ma
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 stage-decoupled training (ablation) | Reflect-R1 · LongVideoBench | arbitration accuracy: 61.8 -> 62.50 | Improved |
| add independent verification response y2 (ablation) | Reflect-R1 · LongVideoBench | arbitration accuracy: 55.9 -> 62.50 | Improved |
| add group-wise advantage decoupling (ablation) | Reflect-R1 · LongVideoBench | arbitration accuracy: 61.4 -> 62.50 | Improved |
| add abstention incentive in verification reward (ablation) | Reflect-R1 · LongVideoBench | arbitration accuracy: 60.8 -> 62.50 | Improved |
| add anti-corruption penalty in arbitration reward (ablation) | Reflect-R1 · LongVideoBench | arbitration accuracy: 61.3 -> 62.50 | Improved |
| apply internal closed-loop reflection without tools | Qwen2.5-VL-7B · LongVideoBench | accuracy: 48.59 -> 41.96 | Hurt |
| apply internal closed-loop reflection without tools | Qwen2.5-VL-7B · VideoMME | accuracy: 60.96 -> 54.00 | Hurt |
| apply internal closed-loop reflection without tools | Vision-R1 · VideoMME | accuracy: 61.37 -> 61.45 | No clear effect |
| apply evidence-driven reflection with tools | Reflect-R1 · LongVideoBench | accuracy: 59.68 -> 62.50 | Improved |
| apply evidence-driven reflection with tools | Reflect-R1 · VideoMME | accuracy: 60.28 -> 61.69 | Improved |
| add active tool-augmented search with reflection over TimeSearch-R | Reflect-R1 · Haystack-LVBench QA | accuracy: 51.5 -> 55.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TimeSearch-R-7B — TimeSearch-R: Adaptive Temporal Search for Long-Form Video Understanding via Self-Verification Reinforcement Learning (ICLR 2026) · better on 6 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.88)
- ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.88)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.87)
- OneThinker: All-in-one Reasoning Model for Image and Video (CVPR 2026 · similarity 0.87)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.87)
- Boosting Reasoning in Large Multimodal Models via Activation Replay (CVPR 2026 · similarity 0.87)
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.
