Research map / Vision-language and multimodal / Multimodal LLMs

OneThinker: All-in-one Reasoning Model for Image and Video

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

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.

Builds OneThinker, a single image-and-video reasoning MLLM trained with SFT on 340k CoT samples and RL on 600k multi-task samples using a new task-wise EMA-normalized GRPO (EMA-GRPO).

Problem

Existing RL-based visual reasoning models are trained separately per task and per modality (image or video), limiting generalization and cross-task knowledge sharing; multi-task RL also suffers from reward heterogeneity across tasks.

Approach

Normalize advantages with task-wise exponential moving average of reward standard deviations instead of per-group std (GRPO) or no std (Dr.GRPO), addressing both intra-task and inter-task reward imbalance while training one model over QA, captioning, grounding, tracking and segmentation for images and videos.

Key results

• OneThinker-8B reaches 70.6 on MMMU, 77.6 on MathVista, and 64.3 on MathVerse, above Qwen3-VL-Instruct-8B (60.2, 74.2, 58.1) in the authors' reproduction.
• On GOT-10k tracking, OneThinker-8B reports 73.0 AO and 84.4 R@0.5 versus 33.7 AO for Qwen3-VL-Instruct-8B.
• In the ablation, EMA-GRPO averages 69.8 QA and 73.0 tracking, compared with 67.2/65.5 for standard GRPO and 67.6/67.8 for Dr.GRPO.

Surprising findings

• Removing image QA data causes a severe drop on video QA (61.1 to 58.2 in the unified-training analysis), which the authors attribute to cross-modal transfer.

Limitations

• Mask-based reward omitted for video segmentation due to SAM2 latency.
• Open-ended QA and captioning rewards depend on an external reward model (POLAR-7B).
• Some ablation and comparison results are reproduced by the authors under their own settings, and some baseline numbers are missing from tables.

Experimental setup

Datasets: OneThinker-600k OneThinker-SFT-340k

Benchmarks: MMMU MathVista MathVerse MMBench MMStar ScienceQA AI2D MMT-Bench VideoMMMU MMVU VideoMME VideoHolmes

Metrics: accuracy mIoU R@0.3 R@0.5 R@0.7 tIoU sIoU AO cIoU J&F

Baselines: Qwen3-VL-Instruct-8B Vision-R1-7B Video-R1-7B VideoChat-R1-7B Seg-R1-7B LISA-7B Time-R1 Perception-R1 OneThinker-8B-GRPO OneThinker-8B-DrGRPO

Builds on: GRPO DeepSeek-R1 Dr.GRPO Seg-R1 VideoChat-R1

Closest papers by meaning

• Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.91)
• Mimic Human Cognition, Master Multi-Image Reasoning: A Meta-Action Framework for Enhanced Visual Understanding (CVPR 2026 · similarity 0.91)
• See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs (CVPR 2026 · similarity 0.91)
• Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.90)
• When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.90)
• DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · similarity 0.90)