Research map / Video / Video-language understanding
ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding
Read the paperCheck your idea against 108,967 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.
Proposes ReAgent-V, an agentic video understanding framework that combines entropy-calibrated frame selection, tool-augmented reasoning, and multi-perspective reflection driven by inference-time reward reports, with the reward used to filter data for SFT/DPO/GRPO.
Problem
Single-pass LVLM video understanding lacks dynamic feedback and self-correction; existing reward models and templates are offline or static, and tool/debate agent frameworks are slow and produce no inference-time rewards.
Approach
Uses a critic agent to produce scalar reward and structured feedback during inference, then revises answers from conservative, neutral, and aggressive perspectives, and reuses high-scoring/low-scoring samples as training data; frame selection uses an entropy-weighted CLIP relevance score (ECRS).
Key results
- Qwen2.5-VL-72B + ReAgent-V reaches 75.1 on VideoMME (w/o sub) and 66.4 on LongBench, a reported ~3.2% average gain over the base model.
- GRPO + ReAgent-V on 52k filtered samples reaches 63.0 on MMVU and 54.2 on VideoMME vs 53.8 for vanilla GRPO on 116k samples.
- On SIMPLER VLA alignment, ReAgent-V reports a 9.8% overall improvement over the second-best baseline GRAPE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ReAgent-V to Qwen2.5-VL-72B | Qwen2.5-VL-72B · LongBench | accuracy: 62.5 -> 66.4 | Helped |
| Apply ReAgent-V to Qwen2.5-VL-72B | Qwen2.5-VL-72B · NextQA | accuracy: 76.8 -> 84.3 | Helped |
| Use ReAgent-V-selected training data for GRPO instead of vanilla GRPO data | Qwen2.5-VL-7B · Video reasoning and general benchmarks | average performance: 2.1% | Helped |
| Use ReAgent-V-selected data rather than full-data SFT | Qwen2.5-VL-7B · Video reasoning and general benchmarks | average performance: 4.3% | Helped |
| Add entropy-calibrated frame selection (ablation) | LLaVA-Video-7B · LongBench | accuracy: 52.6 -> 53.1 | Helped |
| Add entropy-calibrated frame selection (ablation) | LLaVA-Video-7B · LongBench | inference time: 35.99 -> 21.38 | Helped |
| Add entropy-calibrated frame selection (ablation) | Qwen2.5-VL-72B · VideoMME | accuracy: 75.9 -> 76.2 | Helped |
| Add multi-perspective reflection (ablation) | LLaVA-Video-7B · EgoSchema | accuracy: 59.4 -> 60.8 | Helped |
| Add multi-perspective reflection (ablation) | Qwen2-VL-7B · VideoMME | accuracy: 58.2 -> 58.3 | No clear effect |
| Use aggressive reflection, which revises reasoning and answer (ablation) | ReAgent-V reflection · Video QA answer revisions | correction accuracy | Hurt |
| Use conservative reflection, changing only the final answer (ablation) | ReAgent-V reflection · Video QA answer revisions | correction accuracy | Helped |
Surprising findings
- The aggressive reflection strategy, which rewrites reasoning steps, yields the lowest accuracy among corrected answers.
- Reflection-enabled ReAgent-V sometimes gives small or no gains, e.g. Qwen2-VL-7B on VideoMME (58.3 vs 58.2).
Limitations
- Relies on proprietary or re-evaluated baselines with differing frame counts, making fair comparison harder.
- Reflection effects are small on several benchmarks, and the aggressive strategy can corrupt correct reasoning.
- VLA alignment uses only 80 trajectories, limiting the scale of evidence.
Experimental setup
Datasets: Video-R1-260k SIMPLER environment rollouts
Benchmarks: LongBench NextQA EgoSchema LVBench MLVU VideoMME VSI-Bench VideoMMMU MMVU MVBench TempCompass SIMPLER
Metrics: accuracy inference time per sample
Baselines: GPT-4o Gemini-1.5-Pro LLaVA-Video Qwen2.5-VL VideoAgent VideoMemAgent OpenVLA-SFT OpenVLA-DPO GRAPE Vanilla GRPO SFT
Builds on: GRPO DPO Trajectory-wise Preference Optimization VideoAgent Video-R1
Closest papers by meaning
- ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.92)
- SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning (NeurIPS 2026 · similarity 0.89)
- PyVision-RL: Forging Open Agentic Vision Models via RL (ICML 2026 · similarity 0.88)
- Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.88)
- ParaVT: Taming the Tool Prior Paradox for Parallel Tool Use in Agentic Video Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.88)
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.