Research map / Vision-language and multimodal / Multimodal LLMs

ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning

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.

Introduces the ReWatch dataset (captions, QA, and multi-agent ReAct CoT) and ReWatch-R1, a Qwen2.5-VL-7B model post-trained with SFT and GRPO using an Observation & Reasoning reward, reaching SOTA on five video reasoning benchmarks.

Problem

Video reasoning with LVLMs lags behind image reasoning under SFT+RLVR because existing video data lacks challenging multi-hop questions, timestamped grounded captions, and video-grounded chain-of-thought, and answer-only rewards fail to penalize hallucinated reasoning.

Approach

A multi-agent ReAct pipeline (Reasoner and Observer agents querying a timestamped caption) synthesizes video-grounded CoT, and an O&R reward scores both the accuracy of observations against the caption and whether the reasoning from those observations yields the correct answer, combined multiplicatively with answer accuracy.

Key results

• ReWatch-R1 with O&R reaches 35.51% average over five reasoning benchmarks at 192 frames, versus 30.71% for Qwen2.5-VL-7B (no thinking) and 31.58% for LongVideoReason-RL.
• At 384 frames, ReWatch-R1 + O&R reaches 35.78% average reasoning, and 54.60% average on four understanding benchmarks.
• ReWatch-R1-SFT (33.25%) outperforms Video-R1-SFT (29.74%) and LongVideoReason-SFT (26.31%) under the same training configuration at 192 frames.

Surprising findings

• Text-only accuracy of Video-R1 QA is 68.9%, indicating its questions are often answerable from text alone, while ReWatch-QA is at 29.4%, near random chance.
• Enabling thinking mode hurts the untrained Qwen2.5-VL-7B (27.54% vs 30.71%), which the authors attribute to induced hallucination.

Limitations

• Reliance on closed proprietary models (Gemini2.5, GPT-4.1) for data synthesis and evaluation judging.
• Long-video reasoning performance still degrades substantially with duration.
• O&R reward depends on LLM judges and the detailed caption as ground truth for observations.

Experimental setup

Datasets: ReWatch-Caption-10k ReWatch-QA-170k ReWatch-CoT-135k Video-R1-CoT Video-R1-QA LongVideoReason-CoT LongVideoReason-QA

Benchmarks: VCR Bench MINERVA Video Holmes Video MathQA CG-AV Counting MMVU LVBench VideoMME VideoMMMU

Metrics: accuracy

Baselines: Qwen2.5-VL-7B GLM4.1V-9B InternVL3.5-8B Video-R1 Video-Chat-R1 VideoRFT LongVideoReason

Builds on: GRPO ReAct Video-R1 LongVideoReason Qwen2.5-VL

Closest papers by meaning

• ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (NeurIPS 2025 · similarity 0.92)
• Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.92)
• Find, Fix, Reason: Context Repair for Video Reasoning (ICML 2026 · similarity 0.90)
• Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.90)
• OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles (NeurIPS 2025 · similarity 0.89)
• Think-as-You-See: Streaming Chain-of-Thought Reasoning for Large Vision-Language Models (CVPR 2026 · similarity 0.89)