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

TL;DR

OneThinker is a single MLLM (built on Qwen3-VL-Instruct-8B) trained with RL to unify 10 fundamental visual understanding tasks across image and video — VQA, captioning, spatial grounding, temporal grounding, spatio-temporal grounding, tracking, and segmentation — instead of training per-task specialists. To make multi-task GRPO work on rewards with very different scales (continuous IoU vs binary accuracy vs J&F segmentation metric), the paper introduces EMA-GRPO, which normalizes per-task advantages by an exponential-moving-average of each task’s reward standard deviation. The model reports +34.8 / +28.7 / +58.2 / +35.3 / +30.0 average gains over the Qwen3-VL-8B base on Image QA / Caption / Temporal Grounding / Spatial Grounding / ST-Grounding aggregates, with non-trivial zero-shot transfer between tasks.

Key claims

Existing visual-reasoning RL recipes train separate models per task and treat image vs video reasoning as disjoint domains, which the paper argues caps scalability toward a multimodal reasoning generalist [§1, Abstract].
A single all-in-one model can be trained over 10 fundamental visual tasks — image QA, video QA, captioning, temporal grounding, spatial grounding, spatio-temporal grounding, tracking, and segmentation — using one RL recipe [Fig. 2, Abstract].
Training data: OneThinker-600k corpus assembled across all 10 tasks; commercial models provide CoT annotations, yielding OneThinker-SFT-340k for an SFT cold-start before RL [Abstract].
EMA-GRPO addresses reward heterogeneity in multi-task RL by tracking task-wise EMAs of reward standard deviations and using them to scale advantages, so a task with a low-variance binary reward isn’t drowned out by a task with a high-variance continuous reward [Abstract].
Headline gains over Qwen3-VL-Instruct-8B (Figure 2 averages): +34.8 on Image QA, +17.5 on Video QA, +14.8 on Caption, +28.7 on Temporal Grounding, +58.2 on Spatial Grounding, +35.3 on ST-Grounding, +30.0 on Tracking, with per-benchmark gains including +10.4 MMMU, +10.7 VideoMME, +10.6 RefCOCO+ testB, +8.0 STVG sIoU@0.5, +4.9 ReasonVOS J&F [Fig. 2].
Cross-task knowledge transfer and preliminary zero-shot generalization are reported as the qualitative payoff of unifying training [Abstract].
All code, models (OneThink/OneThinker-Qwen3-8B, OneThinker-SFT-Qwen3-8B) and training/eval data (OneThinker-train-data, OneThinker-eval) are released [HF paper page].

Method

OneThinker fine-tunes Qwen3-VL-Instruct-8B in two stages. Stage 1 (SFT cold start): the OneThinker-600k task corpus (image QA, video QA, captioning, spatial grounding via RefCOCO-style, temporal grounding via Charades-STA / ActivityNet, spatio-temporal grounding via STVG, tracking via GOT-10k, and segmentation via RefCOCO-Seg / ReasonVOS / MeViS) is annotated with chain-of-thought traces by commercial MLLMs, producing OneThinker-SFT-340k. SFT teaches the model to emit task-specific structured outputs inside CoT — bounding boxes for grounding, time intervals for temporal grounding, mask tokens for segmentation. Stage 2 (multi-task RL): GRPO with task-specific verifiable rewards (accuracy / IoU / sIoU / J&F / caption-evaluator score) — but because each task’s reward has a different scale and variance, vanilla GRPO collapses onto the high-variance tasks. EMA-GRPO maintains a running EMA of the reward standard deviation per task and normalizes each task’s group-relative advantage by its EMA std, so the gradient contribution per task is balanced regardless of reward scale.

Results

Reported on 31 benchmarks across 10 fundamental visual understanding tasks. Headline numbers vs Qwen3-VL-Instruct-8B (per Figure 2): aggregate gains of +34.8 (Image QA), +17.5 (Video QA), +14.8 (Caption), +28.7 (Temporal Grounding), +58.2 (Spatial Grounding), +35.3 (Spatio-Temporal Grounding), +30.0 (Tracking). Selected per-benchmark deltas: MMMU +10.4, MathVerse +6.2, ScienceQA +4.5, MathVista +3.4, AI2D +2.9, MMStar +2.1, VideoMathQA +10.7, VideoHolmes +7.8, LongVideo-Reason +7.7, MMVU(mc) +4.9, VideoMMMU +2.9, VideoMME +10.6, Charades-STA R@0.5 +8.0, ActivityNet R@0.5 +4.9, ANet-RTL R@0.5 +3.6, RefCOCO+ testB +10.7, RefCOCO testB +8.0, STVG sIoU@0.5 +4.9, GOT-10k R@0.7 +3.6, ReasonVOS J&F +4.9, MeViS J&F +2.4, RefCOCO-Seg cIoU +2.6 [Fig. 2]. The paper also reports preliminary zero-shot transfer between certain tasks — the per-task numbers and isolation of EMA-GRPO vs vanilla GRPO sit inside the paper rather than the abstract.

Why it’s interesting

OneThinker is the broadest multi-task GRPO recipe filed so far — the Reasoning RL cluster already shows GRPO ported to tool-use (Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing), to a video MLLM single-task (Video-Thinker: Sparking "Thinking with Videos" via Reinforcement Learning), to a video diffusion policy (Human detectors are surprisingly powerful reward models), and to a multi-agent orchestrator (Kimi K2.5: Visual Agentic Intelligence), but OneThinker is the first that runs one GRPO loop across 10 heterogeneous-reward visual tasks simultaneously, with EMA normalization as the specific mechanism that makes the multi-task case stable. It also extends the Thinking with Modalities substrate (a) — tagged-skill autoregressive CoT — beyond Video-Thinker’s two-tag schema (<grounding>/<caption>) to a much wider tag inventory covering segmentation masks and tracking boxes, providing a second data point on whether the substrate scales with the number of distinct visual skills. The “unifying-through-RL” framing is also a distinct angle within Unified Multimodal Models — most filed UMMs unify generation + understanding; OneThinker unifies understanding sub-tasks with each other via shared reasoning, leaving generation out.