Research map / Vision-language and multimodal / Multimodal LLMs

Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning

CVPR 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.

Proposes VideoReasoner, a two-stage multi-task cold-start SFT plus multi-task GRPO framework on Qwen2-VL-7B-Base, with event-grounding and keyframe-element reasoning at inference, reaching strong video benchmark results with only 8k training samples.

Problem

MLLMs overlook details or hallucinate on complex, long videos; existing RL video reasoning methods are costly, unstable, rely on large data, and use only text-based reasoning that may not ensure visual reliability.

Approach

Use multimodal reasoning elements (event time intervals and keyframe/key-element retrieval) instead of purely textual chain-of-thought, trained via a multi-task cold start (video QA, event grounding, key-element generation) and multi-task GRPO where task-specific prefixes steer rollouts and rewards (IoU, format, accuracy) are task-specific.

Key results

• Qwen2-VL-7B-Base with cold start (3k) and RL (5k) plus VideoReasoner inference reaches 72.9 short / 60.6 medium / 53.0 long on Video-MME, improving over base (68.3 overall) and matching or surpassing Qwen2.5-VL-7B-Instruct on several benchmarks.
• On Charades-STA, the trained model reaches 59.1 mIoU and 70.4 R1@0.5 after RL, exceeding Qwen2.5-VL-7B-Instruct (43.6 mIoU, 42.9 R1@0.5).
• On VSI-Bench, Qwen2-VL-7B-Base improves from 28.9 to 33.7 with RL, surpassing Qwen2-VL-7B-Instruct (33.4).

Surprising findings

• Video-R1-style text CoT reduces hallucination score on VideoHallu (34.3 to 6.6 for Qwen2-VL-Base with CoT), while the multimodal reasoning approach improves it to 39.0.
• Applying the same RL recipe to Qwen2-VL-7B-Instruct gives mixed results (3 of 8 indicators improve, sharp drop on MLVU), suggesting instruct models are harder to train with this framework.

Limitations

• Multi-task RL uses only two tasks (event grounding and video QA); keyframe reasoning is not trained with RL.
• Inference pipeline adds extra model calls and encoders (text and visual encoders for keyframe retrieval), increasing inference cost.
• Instruct-model baselines underperform after the same RL recipe, so the framework is validated mainly on the base model.

Experimental setup

Datasets: Video-R1 training data (QA) temporal grounding training set (Charades-STA-related, ref 14) raw videos from LLaVA-Video (ref 63) with key elements dataset of ref 50 for multi-task GRPO

Benchmarks: Video-MME LongVideoBench MLVU LVBench VideoEval-Pro VSI-Bench MMVU Charades-STA VideoHallu Tomato

Metrics: accuracy mIoU R1@0.3 R1@0.5 R1@0.7

Baselines: Video-R1 Qwen2-VL-7B-Instruct Qwen2.5-VL-7B-Instruct TimeSuite InternVideo2 LLaVA-Video-72B InternVL2.5-72B

Builds on: GRPO Video-R1 Qwen2-VL

Closest papers by meaning

• Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.93)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.92)
• Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models (ICLR 2026 · similarity 0.91)
• VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · similarity 0.91)
• VideoReasonBench: Can MLLMs Perform Vision-Centric Complex Video Reasoning? (ICLR 2026 · similarity 0.91)
• Scaling RL to Long Videos (NeurIPS 2025 · similarity 0.91)