PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Kaituo Feng, Manyuan Zhang, Hongyu Li, Kaixuan Fan, Shuang Chen, Yilei Jiang, Dian Zheng, Peiwen Sun, Yiyuan Zhang, Haoze Sun, Yan Feng, Peng Pei, Xunliang Cai, Xiangyu Yue

arXiv:2512.03043v3cs.CV

TL;DR

Visual reasoning models are often split by task and modality, leaving open whether one model can handle diverse image and video tasks. OneThinker combines joint training with task-wise EMA reward normalization and reports superior results across 31 benchmarks spanning 10 visual understanding tasks, alongside knowledge sharing and preliminary zero-shot generalization.

  • Problem

    Existing thinking models usually focus on a single task and either images or videos, leaving open whether a generalist can handle diverse tasks across both modalities.

    Source: §1 · p. 2

  • Method

    OneThinker jointly trains across image and video tasks using OneThinker-600k and its CoT-annotated OneThinker-SFT-340k subset, with EMA-GRPO normalizing rewards through task-wise moving averages.

    Source: §1 · pp. 2–3

  • Results

    OneThinker achieves superior results on 31 benchmarks across 10 visual understanding tasks, with knowledge sharing in certain tasks and preliminary zero-shot generalization.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The findings mark a step toward a scalable, unified multimodal reasoning generalist.

    Source: §5 · p. 12

Abstract

from arXiv · show

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.

1 Introduction

OneThinker unifies image and video reasoning across diverse visual tasks, using a large CoT-annotated training corpus and EMA-GRPO to support multi-task reinforcement learning. The paper reports strong benchmark performance, knowledge transfer between some tasks, and preliminary zero-shot generalization.

Motivation

  • Existing thinking models are usually limited to one task and either images or videos, while multi-task RL approaches cover limited task subsets and may remain small-scale.These limitations can restrict practical versatility, generalization, and cross-task or cross-modal knowledge transfer.Source: §1 · p. 2

OneThinker and training data

  • OneThinker jointly handles image and video question answering, captioning, grounding, tracking, and segmentation, supported by OneThinker-600k and its CoT-annotated OneThinker-SFT-340k subset.Seed1.5-VL was used to annotate and filter high-quality chain-of-thought data for the SFT cold start.Source: §1 · pp. 2–3

EMA-GRPO

  • EMA-GRPO uses task-wise exponential moving averages of reward standard deviations to reduce low-variance sample bias and prevent sparse-reward tasks from dominating dense-reward tasks.The method aims to provide stable, adaptive normalization scales that reflect each task’s reward dynamics.Source: §1 · p. 3

Results

  • 31 benchmarks across 10 visual understanding tasks show strong OneThinker performance, with effective knowledge sharing in certain tasks and preliminary zero-shot generalization.OneThinker-8B reaches 70.6% accuracy on MMMU and 64.3% on MathVerse; it also achieves 84.4 R@0.5 on GOT-10k and 54.9 J&F on ReasonVOS.Source: §1 · p. 3

2 Related Works

Prior work applies RL to LLM and MLLM reasoning, but existing approaches largely target single or limited tasks and often only one visual modality.

2.1 Reinforcement Learning for LLM Reasoning

  • LLM reasoning work builds on GRPO [4] and variants including Dr.GRPO [17], GSPO [25], and Critique-GRPO [7], while most studies focus on single tasks or homogeneous rewards.Dr.GRPO addresses sample-wise standard-deviation imbalance and response-length bias; GSPO uses sequence-level optimization, and Critique-GRPO incorporates natural-language feedback.Source: §2.1 · p. 3

2.2 Reasoning in MLLMs

  • Vision-R1 [10] and Video-R1 [9] address image/video QA; Perception-R1 [14] and VLM-R1 [11] detection; Seg-R1 [12] segmentation; Time-R1 [37] temporal grounding; VideoChat-R1 [13] spatio-temporal tasks.These examples illustrate RL-based reasoning across a range of visual tasks.Source: §2.2 · p. 3
  • SophiaVL-R1 [38] uses thinking-process rewards for image QA, yet most multimodal reasoning models remain task-limited and support image or video reasoning alone.Source: §2.2 · p. 4

3 Method

OneThinker combines a curated image-and-video reasoning corpus and task-specific reward designs with EMA-GRPO, which targets optimization imbalance across heterogeneous visual tasks.

3.1 Dataset Construction

  • OneThinker-600k spans image and video reasoning tasks, while perception outputs use predefined JSON schemas for consistent formatting and automatic reward verification.The corpus covers QA, captioning, spatial and temporal grounding, tracking, and segmentation, with samples curated across domains and difficulty levels.Source: §3.1 · p. 4
  • Seed1.5-VL [16] supplies task-filtered CoT annotations, which are checked and validated to produce the OneThinker-SFT-340k initialization set.The resulting subset is intended as a diverse, reliable SFT foundation for reasoning across visual tasks.Source: §3.1 · p. 4

3.2 Task Types and Rewards

  • Tasks share a text interface with reasoning in <think>...</think> and task-specific outputs in <answer>...</answer>, while rewards combine accuracy and format checks.Structured answers follow predefined schemas that support automatic parsing and verification.Source: §3.2 · p. 4
  • Rule-based QA uses answer equivalence, MRA for regression, and word error rate for OCR, while open-ended QA and captioning use POLAR-7B [40] similarity scores.These reward choices provide deterministic feedback for rule-based tasks and model-scored similarity for open-ended outputs.Source: §3.2 · p. 5
  • Temporal and spatial grounding are scored by temporal IoU and spatial IoU, respectively, while spatio-temporal grounding combines temporal IoU with mean spatial IoU.Tracking is measured by mean IoU across frames.Source: §3.2 · p. 5
  • Segmentation uses predicted boxes and positive and negative points as SAM2 [42] inputs; video segmentation also predicts a keyframe time but omits mask-based reward.The omission reflects the computational latency of running SAM2 on all video-segmentation rollouts; image rewards combine box overlap and point-set Gaussian similarities.Source: §3.2 · pp. 5–6

3.3 EMA-GRPO

  • Standard GRPO can underweight medium-difficulty samples within tasks, whereas removing standard-deviation normalization can let sparse-reward tasks dominate dense-reward tasks.These intra-task and inter-task imbalances can bias optimization toward a small subset of tasks and weaken generalization across visual reasoning settings.Source: §3.3 · p. 6
  • EMA-GRPO maintains task-wise reward moments and standard deviations, then uses each task’s moving scale to normalize advantages for balanced optimization.The decay factor is 0.99; early advantages are clipped to [−5, 5] while the moving statistic stabilizes.Source: §3.3 · pp. 6–7

4 Experiments

OneThinker shows strong performance across image and video understanding benchmarks, spanning diverse visual reasoning tasks. The experiments also examine component effectiveness, cross-task and cross-modal benefits, and generalization to unseen tasks.

4.1 Setup

  • Training uses Qwen-3-VL-Instruct-8B for SFT, followed by RL on OneThinker-600k with image-video-balanced samples, and takes approximately 10 days.Evaluation spans question answering, captioning, grounding, tracking, and segmentation, using greedy decoding; Qwen3-VL-Instruct results are reproduced under the same setting.Source: §4.1 · p. 7

4.2 Main Results

  • 79.2% on LongVideo-Reason surpasses Video-R1-7B (67.2%) and Qwen3-VL-Instruct-8B (71.5%), while OneThinker reaches 70.6% on MMMU.It also scores 35.0% on VideoMathQA and 77.6% on MathVista; the reported image QA results outperform prior open-source competitors, and video QA results rank among the top performers.Source: §4.2 · pp. 7–8, Table 1 · p. 8, Table 2 · p. 8
  • 65.0 R@0.3, 43.6 R@0.5, and 25.7 R@0.7 on ActivityNet show strong temporal grounding, while image and video captioning results are competitive or superior.OneThinker scores 25.7 on MMSci-Caption and 57.9 on MMT-Caption, versus Qwen3-VL-Instruct-8B at 15.1 and 47.3, respectively; it reaches 28.0 on VideoMMLU-Caption and best mIoU (43.2) on ANet-RTL.Source: §4.2 · pp. 8–9, Table 3 · p. 8, Table 4 · p. 9
  • 93.7 / 88.9 / 92.0 on RefCOCO testA/testB/val and 34.9 tIoU@0.5 on STVG demonstrate strong spatial and spatial-temporal grounding.On RefCOCO+, OneThinker scores 91.4 / 82.7 / 87.0, and on RefCOCOg 88.8 / 89.2 (test/val); STVG results also include 39.5 tIoU, 40.3 sIoU@0.5, and 36.7 sIoU.Source: §4.2 · p. 9, Table 5 · p. 9, Table 6 · p. 9
  • 73.0 AO on GOT-10k and 52.7 J&F on MeViS accompany strong tracking and segmentation results across image and video benchmarks.Tracking evaluation used 32 prediction frames, compared with VideoChat-R1’s prior 8-frame setting; OneThinker scores 75.8 / 67.1 / 70.8 cIoU on image segmentation and 54.9 J&F on ReasonVOS.Source: §4.2 · p. 10, Table 7 · p. 10, Table 8 · p. 10

4.3 Ablation Study

  • All ablated variants perform worse than OneThinker-8B across tasks, with RL improving over SFT alone and standard GRPO or DrGRPO underperforming EMA-GRPO.The study compares an SFT-only model, a GRPO replacement, and a Dr.GRPO replacement; the results support the proposed RL framework and EMA-GRPO’s handling of intra-task and inter-task imbalance.Source: §4.3 · p. 10, Table 9 · p. 11

4.4 Benefits of Unified Training Analysis

  • Removing temporal grounding degrades video QA and tracking, removing spatial grounding lowers image QA and segmentation, and excluding ImageQA severely harms video QA.The authors attribute ImageQA’s transfer benefit to the generally higher quality and greater diversity of image QA datasets, which support reasoning and recognition that transfer to video understanding.Source: §4.4 · p. 11, Table 10 · p. 11

4.5 Zero-shot Generalization to Unseen Tasks

  • On unseen MMT-Bench tasks, OneThinker-8B outperforms Qwen3-VL-Instruct-8B on examples including point tracking, image quality assessment, GUI tasks, and rotated object detection.MMT-Bench contains 162 diverse visual tasks; the results indicate generalization beyond the model’s training tasks.Source: §4.5 · p. 11, Figure 5 · p. 12

5 Conclusion

OneThinker unifies reasoning over diverse visual foundation tasks for images and videos, with training data and an RL method designed to support this approach.

  • OneThinker unifies diverse visual foundation tasks for images and videos in an all-in-one multimodal reasoning model.Source: §5 · p. 12
  • The authors construct OneThinker-600k for RL training and its CoT-annotated OneThinker-SFT-340k subset for SFT cold start.Source: §5 · p. 12
  • EMA-GRPO balances optimization across heterogeneous visual tasks through task-wise adaptive reward normalization.Source: §5 · p. 12
  • Experiments show strong performance across tasks, and the authors frame the work as a step toward a scalable, unified multimodal reasoning generalist.Source: §5 · p. 12

A Reasoning Examples

The section gives a geometry problem as an example of image question answering, asking for an angle from information about a circle and offering four choices.

Video QA

The section illustrates video question answering and visual localization through prompts that ask for narrative interpretation, time intervals, bounding boxes, tracking, and segmentation hints.

Video QA

  • A video-QA prompt asks how a glowing key, a “SUPER STRENGTH TEST 1” sign, and recurring superhero and film-platform content relate narratively.The prompt supplies four candidate interpretations but does not identify a selected answer.Source: Video QA · p. 18, Figure 7 · p. 18
  • A temporal-grounding prompt asks when the described action “person runs back” occurs in the video.The prompt is paired with a temporal-grounding reasoning example.Source: Video QA · p. 21, Figure 10 · p. 21
  • A spatial-grounding prompt requests bounding-box coordinates for a bush behind the middle woman.The prompt is paired with a spatial-grounding reasoning example.Source: Video QA · p. 22, Figure 11 · p. 22
  • A spatial-temporal prompt asks for the start and end times and a bounding box at each second for an adult in blue pants playing guitar.The prompt combines video timing with per-second spatial localization.Source: Video QA · p. 23, Figure 12 · p. 23
  • A tracking prompt specifies an object’s first-frame bounding box and requests its bounding box once per second across the video.The example gives the initial box as [537,403,768,703].Source: Video QA · p. 24, Figure 13 · p. 24
  • Segmentation prompts provide referring expressions and request bounding-box and positive/negative-point hints for an image object or a video key-frame object.The resulting hints are forwarded to SAM2 to produce a mask.Source: Video QA · pp. 25–26, Figure 14 · p. 25, Figure 15 · p. 26

B Prompt Template

The appendix presents a system prompt for all tasks and separate prompts for QA, grounding and tracking, and segmentation.

Related papers

Loading 2512.03043v3…