Video-Thinker: Sparking "Thinking with Videos" via Reinforcement Learning

TL;DR

Video-Thinker extends the “Thinking with Images” reasoning paradigm to video by training a 7B MLLM (Qwen2.5-VL-7B base) to interleave its own <grounding> (temporal localization) and <caption> (segment description) calls inside a chain-of-thought, without invoking external tools. Training is two-stage: SFT on a 10K curated trace dataset (Video-Thinker-10K) to learn the tag format, then GRPO with a correctness+format reward to teach the model to autonomously decide when to ground and caption. On out-of-domain reasoning benchmarks it sets a new 7B SOTA: 43.22% on Video-Holmes (+4.68 over best baseline), 33.25% on CG-Bench-Reasoning, and 80.69% on VRBench (+11.44). The GRPO stage is what carries it — the SFT-only checkpoint actually underperforms the Qwen2.5-VL base on several tasks.

Key claims

“Thinking with Videos” is operationalized as a structured CoT that interleaves three tags — <grounding> for picking key time segments, <caption> for describing those segments, and analytical reasoning over the synthesized cues — all produced by the same MLLM with no external tool calls [§1, §3].
10K curated training samples are sufficient to spark this behavior, vs. the 160K used by Video-R1; data efficiency comes from a hindsight-curation loop that regenerates reasoning traces (up to 3×) until Qwen2.5-VL-7B can recover the ground-truth answer from the synthesized <grounding>+<caption> content [§3.1].
Two-stage training is necessary: SFT alone (Video-Thinker-SFT-7B) underperforms the Qwen2.5-VL base on several benchmarks; GRPO on top adds +11.70 on Video-Holmes, +8.30 on CG-Bench-Reasoning, +18.29 on VRBench over the SFT checkpoint [§4.2, Table 1].
GRPO reward is the sum of an answer-match indicator and a format-adherence term over the <grounding>/<caption>/reasoning tags; advantages are group-normalized over multiple sampled traces per prompt, no value function [§3.2, Eq. 3-4].
Grounding ability improves quantitatively, not just downstream: Video-Thinker-7B reaches 48.22 mIoU on held-out temporal grounding (vs. 27.47 for Qwen2.5-VL-7B), with Recall@0.3 = 79.29 vs. 39.52 baseline [§4.3, Table 3].
Gains persist across frame budgets — Video-Thinker-7B beats Qwen2.5-VL and Video-R1 at 16, 32, and 64 inference frames, with 64-frame results being best in most cases [§4.2, Table 2].

Method

The framework rests on the observation that video reasoning benefits from the same “active perception” trick that image reasoning got from “Thinking with Images” — but the manipulable substrate is temporal segments rather than image crops. Two intrinsic capabilities are picked as anchors: grounding (locate question-relevant time spans) and captioning (describe what’s in them). Both are abilities the MLLM already has in latent form; the question is how to get the model to invoke them inside a CoT autonomously.

Data: six source datasets are unified into Video-Thinker-10K. Caption-labeled sources (ActivityNet, TutorialVQA, YouCook2) come with per-segment human captions but lack reasoning-grade questions, so DeepSeek-R1 is used to generate multi-segment multiple-choice questions over them. QA-labeled sources (STAR, ScaleLong, LVBench) have hard QA pairs but no segment-level captions, so Gemini-2.5-Flash-Lite generates answer-conditioned captions for the annotated segments. Final reasoning traces are then synthesized in reverse by DeepSeek-V3 from {answer, captions, timestamps} into the structured <grounding>/<caption>/reasoning format. A “hindsight curation” gate filters every sample by asking Qwen2.5-VL-7B to solve the problem from the synthesized trace; if it can’t, the trace is regenerated (up to 3×).

Training: SFT for 1 epoch (lr 1e-5, batch 16) on Video-Thinker-10K teaches the model the tag format. Then GRPO (lr 1e-6, batch 8, KL coefficient β=0.04) optimizes against a reward of R = 1[answer matches GT] + R_format(tags) using outcome supervision with group-normalized advantages over multiple sampled traces per prompt. Videos are subsampled to 16 frames at 224² during training. The paper notes Deepseek-R1-style “aha moments” — the model spontaneously revisits its initial grounding/caption choices mid-trace.

Results

Benchmark	Qwen2.5-VL-7B	Video-R1-7B (prev. best)	Video-Thinker-SFT-7B	Video-Thinker-7B
Video-Holmes (OOD)	34.02	38.54	31.52	43.22
CG-Bench-Reasoning (OOD)	27.10	27.81	24.95	33.25
VRBench (OOD)	63.42	69.25	62.40	80.69
ActivityNet (in-domain)	70.96	76.00	70.80	78.72
YouCook2 (in-domain)	63.74	65.65	56.11	73.66
LVBench (in-domain)	33.33	34.68	35.69	37.04

The OOD wins are the headline (VRBench +11.44 over Video-R1 is unusually large for a 7B-on-7B comparison), and the SFT-only ablation underperforming the Qwen2.5-VL base is the most telling diagnostic — without RL, the structured-trace format is a tax, not a benefit. Temporal grounding metrics nearly double over base (mIoU 27.47 → 48.22, Recall@0.3 39.52 → 79.29), so the model is genuinely getting better at the sub-task it claims to use as a tool, not just style-matching. Performance also improves monotonically with frame budget (16 → 32 → 64) for all three models compared, with Video-Thinker maintaining the gap.

Why it’s interesting

Two threads converge here that are already on the wiki. First, this is another datapoint for Reasoning RL: GRPO + a structured (correctness + format) reward, on a Qwen-family VLM, with the now-familiar pattern that SFT alone is insufficient and the RL stage is what generalizes. The SFT-underperforms-base finding is a sharp statement of this — it isolates “the RL is doing the work” cleanly. Second, the design pattern of teaching a single model to invoke its own latent skills (grounding, captioning) as named tools inside a CoT — rather than calling external graders or modules — is the same architectural bet as InfTool’s self-contained tool-use loop (Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing), just applied to perception rather than function calling. Third, the data pipeline is a clean example of Synthetic Training Data: a generator stack (DeepSeek-R1 + Gemini-2.5-Flash-Lite + DeepSeek-V3) plus a hindsight-curation gate replaces human annotation entirely, and the resulting 10K corpus outperforms 160K-sample baselines on the same base model. For Luma, the most directly transferable idea is that hindsight-curation gate: filtering synthetic CoT traces by whether a small model can recover the answer from the trace is a cheap, automatable substitute for human review of reasoning data, and the same gate would apply to any structured-trace dataset (e.g. layout-then-render, or pose-then-pixel, for generative video).