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

AI summary

Overview

Research area: Computer Vision / Multimodal Large Language Models (MLLMs) — specifically reinforcement-learning-based visual reasoning across images and video.

Technical level: Intermediate. The paper assumes familiarity with GRPO-style reinforcement learning, reward normalization, chain-of-thought supervision, and standard vision benchmarks (grounding, tracking, segmentation).

Scope: The paper introduces OneThinker, a single model trained to reason over both images and videos across ten fundamental visual understanding tasks, supported by a new 600k-sample training corpus and a new RL algorithm, EMA-GRPO.

What This Paper Is About

Most existing "thinking" multimodal models are built for one task in one modality — an image question-answering reasoner, or a video reasoner, or a detection reasoner — so knowledge cannot be shared between tasks and images and videos are treated as separate domains. The authors ask whether one model can handle image and video reasoning simultaneously across question answering, captioning, spatial and temporal grounding, tracking, and segmentation. Their goal is a single multimodal reasoning generalist, and a training algorithm that keeps heterogeneous tasks from overwhelming each other during reinforcement learning.

Key Contributions

1. OneThinker, a unified multimodal reasoning generalist. A single model handling a wide range of image and video tasks — question answering, captioning, grounding, tracking, and segmentation — rather than separate per-task models.

2. Two new datasets. The OneThinker-600k corpus (approximately 600k multimodal samples covering the fundamental visual tasks) and OneThinker-SFT-340k, a chain-of-thought-annotated subset produced with the proprietary Seed1.5-VL model for SFT cold start.

3. EMA-GRPO. A reinforcement learning algorithm that maintains task-wise exponential moving averages of reward standard deviations, addressing both intra-task imbalance (GRPO's per-group standard deviation normalization favoring low-variance rollouts) and inter-task imbalance (removing that normalization, as in Dr.GRPO, letting sparse-reward tasks dominate).

4. Large-scale evaluation. Experiments across 31 benchmarks spanning 10 fundamental visual understanding tasks, plus analyses of cross-task/cross-modal knowledge sharing and zero-shot generalization to unseen tasks.

Main Findings

• Strong image QA results. OneThinker-8B reaches 70.6% on MMMU, 77.6% on MathVista, 64.3% on MathVerse, 86.6% on MMBench, 70.6% on MMStar, 96.5% on ScienceQA, 85.2% on AI2D, and 67.8% on MMT-Bench. On MMMU this compares with 60.2% for the reproduced Qwen3-VL-Instruct-8B, 70.7% for GPT-4o, and 81.7% for Gemini 2.5 Pro.

• Strong video QA results. OneThinker-8B scores 66.2% on VideoMMMU, 70.5% on MMVU(mc), 66.5% on VideoMME, 48.7% on VideoHolmes, 61.7% on LongVideoBench, 79.2% on LongVideo-Reason, and 35.0% on VideoMathQA, outperforming video-specialized reasoning models such as VideoChat-R1-7B, VideoLLaMA3-7B, InternVideo2.5-8B and Video-R1-7B.

• Captioning improvements over the base model. OneThinker-8B gets 25.7 on MMSci-Caption, 57.9 on MMT-Caption, and 28.0 on VideoMMLU-Caption, versus 15.1, 47.3 and 20.0 respectively for Qwen3-VL-Instruct-8B.

• Temporal grounding gains. On Charades, OneThinker-8B achieves 83.5 R@0.3, 68.3 R@0.5, 45.3 R@0.7, 59.9 mIoU; on ActivityNet, 65.0 / 43.6 / 25.7 with 45.9 mIoU; on ANet-RTL, 62.0 / 42.3 / 22.7 with 43.2 mIoU, the best mIoU among the listed models.

• Spatial grounding leading results. RefCOCO testA/testB/val: 93.7 / 88.9 / 92.0; RefCOCO+: 91.4 / 82.7 / 87.0; RefCOCOg test/val: 88.8 / 89.2 — above Perception-R1, VLM-R1, DeepEyes and Qwen3-VL-Instruct-8B.

• Spatial-temporal grounding gains. On STVG, OneThinker-8B reaches 34.9 tIoU@0.5, 39.5 tIoU, 40.3 sIoU@0.5, 36.7 sIoU, versus 24.4, 25.4, 11.6 and 13.6 for Qwen3-VL-Instruct-8B.

• Large tracking jump. On GOT-10k, OneThinker-8B records 73.0 AO, 93.9 R@0.3, 84.4 R@0.5, 68.8 R@0.7. The paper notes its evaluation uses 32 frames for prediction, more challenging than the 8-frame setting of VideoChat-R1.

• Best segmentation averages. Image segmentation on RefCOCO / RefCOCO+ / RefCOCOg val: 75.8 / 67.1 / 70.8 cIoU. Video segmentation: 48.8 J, 56.7 F, 52.7 J&F on MeViS, and 51.1 J, 58.7 F, 54.9 J&F on ReasonVOS, surpassing PixelLM-7B, LISA-7B, VISA-13B, Seg-R1-7B, ReferFormer and VideoLISA-3.8B.

• EMA-GRPO beats SFT, GRPO and Dr.GRPO. In the ablation (averaged per task), OneThinker-8B scores 69.8 QA, 49.7 temporal grounding, 89.2 spatial grounding, 38.1 spatio-temporal grounding, 73.0 tracking, 64.2 segmentation. The SFT-only variant scores 67.0 / 31.8 / 87.8 / 27.1 / 48.1 / 62.8; the GRPO variant 67.2 / 46.9 / 86.5 / 34.5 / 65.5 / 62.3; the Dr.GRPO variant 67.6 / 46.3 / 88.2 / 34.0 / 67.8 / 61.2; and Qwen3-VL-Instruct-8B 65.0 / 30.8 / 86.6 / 19.5 / 33.7 / 50.0.

• Tasks help each other. Removing spatial grounding drops results to 76.6 image QA / 60.3 video QA / 71.0 tracking / 62.9 segmentation; removing temporal grounding gives 77.2 / 59.5 / 67.2 / 63.3; removing image QA gives a video QA score of 58.2 (image QA not applicable) with 72.3 tracking and 63.9 segmentation. Full OneThinker scores 77.4 / 61.1 / 73.0 / 64.2.

• Preliminary zero-shot transfer. On unseen tasks drawn from MMT-Bench (162 diverse visual tasks), OneThinker-8B outperforms Qwen3-VL-Instruct-8B on several, including point tracking, image quality assessment, GUI tasks, and rotated object detection.

• Video segmentation reward omission. Because running SAM2 on all rollouts for video segmentation is computationally expensive, the mask-based reward is omitted in this paper.

Methodology in Plain English

Data first. The authors assembled OneThinker-600k by collecting and curating public training datasets across images and videos and across many task types. Perception tasks (grounding, tracking, segmentation) require outputs in a predefined JSON schema so answers can be parsed and scored automatically. A strong proprietary model, Seed1.5-VL, then annotated chain-of-thought traces and filtered them with task-specific thresholds to produce OneThinker-SFT-340k.

One interface for everything. Every task is cast as text: the model writes its reasoning inside <think>...</think> and its answer inside <answer>...</answer>. The total reward is an accuracy reward plus a format reward. Accuracy is task-specific: equivalence checking for multiple choice, numerical and math problems; Mean Relative Accuracy for regression; Word Error Rate for OCR; a reward model (POLAR-7B) for open-ended QA and captioning; temporal IoU for temporal grounding; spatial IoU for spatial grounding; the sum of the two for spatio-temporal grounding; mean spatial IoU across frames for tracking; and for segmentation, spatial IoU plus Gaussian similarities over positive and negative point sets (three of each), with an added temporal Gaussian term for video. Gaussian kernels use σ = 50 for spatial and σ = 1 for temporal distances. Segmentation outputs are fed to SAM2 to generate masks.

Two-stage training. First, supervised fine-tuning of Qwen-3-VL-Instruct-8B on OneThinker-SFT-340k. Then reinforcement learning on OneThinker-600k. Image-video balanced sets are sampled for both stages. Training runs on 32 NVIDIA H800 GPUs, with batch size 32 for SFT and 128 for RL, learning rates 1×10⁻⁵ and 2×10⁻⁶ respectively with AdamW, at most 128 video frames, maximum response length 4096 tokens, and roughly 10 days total.

The RL fix. Standard GRPO divides each group's advantages by the group's reward standard deviation, which over-weights very easy and very hard samples and under-optimizes medium-difficulty ones. Removing that division (Dr.GRPO) fixes this within a task but lets sparse-reward tasks dominate dense-reward ones across tasks. EMA-GRPO keeps, for each task, an exponential moving average of the mean and mean-square reward (decay β = 0.99), derives a task-specific standard deviation from them, and normalizes advantages by that task's own statistic. Advantages are clipped to [−5, 5] for numerical stability early in training. The final policy update is the standard GRPO objective with KL regularization coefficient 0.01 and group size 8. Rollouts that are entirely correct or entirely incorrect are discarded.

Why This Matters

The paper argues that separate per-task, per-modality thinking models limit practical versatility and prevent knowledge sharing, and shows a single model can cover both modalities across ten task families with measured cross-task and cross-modal benefits. It also contributes a concrete algorithm for the general problem of heterogeneous reward scales in multi-task RL.

Potential real-world applications (as suggested by the tasks covered):

• Video search and surveillance-style pipelines needing temporal grounding and multi-object tracking in long footage.
• Assistive or accessibility tools that caption and describe both images and videos.
• Robotics and autonomous systems that require spatial grounding and segmentation in dynamic scenes.
• Document, chart and scientific-figure understanding via the mathematical and OCR-style QA tasks.

Industry relevance: The work comes from Meituan and MMLab, CUHK, and releases code, model, and data — lowering the barrier for teams that want one checkpoint to serve many visual tasks instead of maintaining separate specialized models. The unified <think>/<answer> interface with schema-checked outputs also makes reward computation automatable, which matters for anyone building RL post-training pipelines.

Future Directions

• Extending reward design to video segmentation. The paper explicitly omits the mask-based reward for video segmentation because of SAM2's latency; adding it is a natural next step.
• Moving beyond "preliminary" zero-shot generalization. The authors describe zero-shot transfer as preliminary and limited to tasks drawn from MMT-Bench, so broader and systematic out-of-distribution evaluation remains open.
• Scaling and efficiency. Training took approximately 10 days on 32 NVIDIA H800 GPUs, and the model caps video frames at 128; the paper does not report compute-efficient alternatives, leaving room for reducing cost as task and modality coverage grows.
• Understanding why transfer works. The data-removal analysis shows tasks benefit each other, but the paper offers only an attribution-style explanation (for instance, that image QA's quality and diversity transfer to video); the underlying mechanisms are not established.

Target Audience

Researchers and engineers working on multimodal LLMs, reinforcement learning for visual reasoning, and post-training of vision-language models. It is also useful for practitioners deciding whether a single generalist checkpoint can replace several task-specific models, and for readers interested in reward-normalization strategies for multi-task RL. Readers without background in GRPO-style policy optimization will find the training sections dense, though the benchmark tables and the ablation results are accessible on their own.

Authors’ abstract

Reinforcement learning (RL) has recently achieved remarkable success in eliciting visual reasoning within Multimodal Large Language Models (MLLMs). However, existing approaches typically train separate models for different tasks and treat image and video reasoning as disjoint domains. This results in limited scalability toward a multimodal reasoning generalist, which restricts practical versatility and hinders potential knowledge sharing across tasks and modalities. To this end, we propose OneThinker, an all-in-one reasoning model that unifies image and video understanding across diverse fundamental visual tasks, including question answering, captioning, spatial and temporal grounding, tracking, and segmentation. To achieve this, we construct the OneThinker-600k training corpus covering all these tasks and employ commercial models for CoT annotation, resulting in OneThinker-SFT-340k for SFT cold start. Furthermore, we propose EMA-GRPO to handle reward heterogeneity in multi-task RL by tracking task-wise moving averages of reward standard deviations for balanced optimization. Extensive experiments on diverse visual benchmarks show that OneThinker delivers strong performance on 31 benchmarks, across 10 fundamental visual understanding tasks. Moreover, it exhibits effective knowledge transfer between certain tasks and preliminary zero-shot generalization ability, marking a step toward a unified multimodal reasoning generalist. All code, model, and data are released.