VideoAuto-R1: Video Auto Reasoning via Thinking Once, Answering Twice

AI summary

Overview

Research area: Video understanding with multimodal large language models (MLLMs), specifically adaptive reasoning ("auto-thinking") trained with reinforcement learning (GRPO).

Technical level: Intermediate to Advanced. Familiarity with chain-of-thought reasoning, reinforcement learning from verifiable rewards, and video QA benchmarks is assumed.

Scope: The paper argues that step-by-step chain-of-thought (CoT) reasoning is often unnecessary for video understanding, and proposes VideoAuto-R1, a framework that trains a model to answer, then reason, then answer again, and at test time decides per sample whether to reason based on the model's own confidence in its first answer.

What This Paper Is About

Video reasoning models typically force a "thinking mode" that generates long step-by-step rationales before answering, which is slow and may not actually help. The authors first show that for RL-trained video models, answering directly often matches or beats CoT while using far fewer tokens, with CoT only consistently helping on reasoning-intensive benchmarks. Their goal is a video model that reasons only when necessary, achieving high accuracy with much shorter responses.

Key Contributions

1. First systematic study of CoT necessity in video. The authors re-evaluate existing video reasoning models (Video-R1, Time-R1, VideoChat-R1) under identical inputs and show that direct answering often matches or surpasses CoT inference, while CoT is mainly beneficial on benchmarks that demand symbolic, multi-step deduction (e.g., VideoMMMU).
2. The "Thinking Once, Answering Twice" training paradigm. Each response follows an answer → think → answer template, with both the initial answer and the reviewed answer supervised by verifiable rewards under GRPO, using a fallback bonus for honest deferral. This removes the need for per-sample think/no-think labels, switch tokens, or cold-start SFT.
3. Confidence-based early-exit inference. After decoding only the first boxed answer, the model computes a length-normalized mean log-probability over those tokens; if it exceeds a threshold τ = 0.97, decoding terminates (equivalent to direct answering); otherwise the model continues to CoT and a reviewed answer. If the first box is the fallback string "Let's analyze the problem step by step", the score is set to −∞, forcing reasoning.
4. State-of-the-art accuracy with large efficiency gains across video QA, temporal grounding, grounding QA, and image reasoning benchmarks, with the average response length reduced roughly 3.3x (from 149 to 44 tokens).

Main Findings

• CoT is often not worth its cost in video. For Video-R1, CoT inference scored 64.3 vs 64.6 direct on VideoMME, 59.4 vs 59.5 on LongVideoBench, 65.4 vs 65.6 on MMVU, and 34.9 vs 42.0 on Charades-STA, while producing 386 tokens versus 17.6. Time-R1 dropped from 65.9 to 63.8 on VideoMME and from 60.0 to 58.3 on LongVideoBench with 138 CoT tokens. VideoChat-R1 dropped from 65.7 to 63.9 on VideoMME with 126 tokens. The consistent CoT gain across models was on VideoMMMU (Video-R1 +1.0, Time-R1 +1.1, VideoChat-R1 +3.4).
• VideoAuto-R1 leads on video QA benchmarks. With a Qwen2.5-VL-7B base, it reaches 67.3% on VideoMME, 71.0% on MVBench, 60.5% on LongVideoBench, 69.7% on MMVU, 58.6% on VideoMMMU, and 39.4% on MVP, with an average response length of 44 tokens and think ratios of 40%, 25%, 39%, 28%, 51%, and 44% respectively. On VideoMME it surpasses Video-R1, VITAL, and VideoChat-R1.5 by 5.5%, 3.2%, and 2.1%. On VideoMMMU it improves from the baseline's 54.7% to 58.6% (+3.9), and on MVP from 36.5% to 39.4%, roughly 6% above Video-R1.
• The Qwen3-VL-8B variant is stronger still. 71.7% VideoMME, 72.0% MVBench, 67.4% LongVideoBench, 71.1% MMVU, 65.0% VideoMMMU, 43.0% MVP, at 52 tokens average, with think ratios of 11%, 31%, 20%, 38%, 53%, and 56%.
• Thinking is triggered adaptively by task type. Think-mode activation is low on perception-oriented tasks (25% on MVBench for the Qwen2.5-VL variant) and higher on reasoning-intensive tasks (51% on VideoMMMU). The Qwen3-VL variant shows the same pattern (31% MVBench, 53% VideoMMMU).
• Temporal grounding improves too, largely without reasoning. mIoU rises from 52.9% to 60.0% on Charades-STA and from 26.9% to 47.6% on ActivityNet; on NExT-GQA, QA accuracy rises from 53.3% to 80.6% and mIoU from 20.2% to 36.7%. The authors state the initial boxed prediction is already sufficient for localization, so they early-exit by default on grounding. The Qwen3-VL model reaches 63.7 mIoU on Charades-STA, 51.9 on ActivityNet, and 81.1 accuracy / 44.2 mIoU on NExT-GQA.
• Image benchmarks also improve. Versus the Qwen2.5-VL baseline: MathVista 69.4 → 73.7, MathVision 26.3 → 29.6, MathVerse 44.8 → 46.9, MMMU 51.3 → 53.8, MMMU-Pro 36.1 → 39.8, MM-Vet 60.0 → 61.9.
• Auto-thinking beats both direct-only and always-think training. In a controlled comparison on the same data, SFT (2.3 tokens) gave 67.0 VideoMME / 65.9 MMVU / 56.5 VideoMMMU / 56.3 Charades-STA; RL without thinking (2.5 tokens) gave 66.0 / 66.4 / 54.4 / 58.8; RL with thinking (149 tokens) gave 66.1 / 67.5 / 56.4 / 59.8; VideoAuto-R1 (44 tokens) gave 67.3 (+1.3) / 69.7 (+3.5) / 58.6 (+3.9) / 60.0 (+7.1).
• A training-based mode-selection baseline was worse. Reproducing a strategy in the spirit of AdaptThink, with think/no-think labels assigned by comparing average accuracy over 8 rollouts and a ratio near 1:1, the auto mode underperformed the no-think baseline on MVBench (70.5% vs 71.1%) and showed mode collapse, defaulting to almost no thinking on VideoMME and only a 31% think ratio on VideoMMMU.
• Reasoning about the second answer is where the gain lives. Training rewards increased, and the task reward for the second answer typically exceeded that of the first, confirming that explicit reasoning helps harder instances while correct fast first answers are retained. The authors also report that rigidly enforcing think/no-think decisions during training caused mode collapse and poor generalization in early experiments.
• Not reported: the paper reports token lengths rather than wall-clock latency or monetary inference cost; the exact size and composition of the held-out validation subset used to set τ is not given in the available text.

Methodology in Plain English

The researchers start from a base video-capable model (Qwen2.5-VL-7B-Instruct or Qwen3-VL-8B-Instruct) and train it with GRPO, a reinforcement learning method that scores candidate outputs with rule-based, verifiable rewards and normalizes them within a group of sampled outputs. No cold-start supervised fine-tuning is used; the authors state that SFT on Video-R1-CoT data degraded the Qwen2.5-VL baseline.

Instead of teaching the model two separate modes, they impose a single output format: a boxed short answer, then a <think>...</think> rationale, then a second boxed answer. A system prompt enforces exactly two boxed blocks and one think block. If the model cannot answer immediately, it is allowed to write "Let's analyze the problem step by step" in the first box, which acts as an honest deferral and is rewarded by a fallback bonus when the final answer is correct.

The reward combines task accuracy on the first answer, task accuracy on the second answer (weighted higher, w₂ = 1.1 vs w₁ = 0.9), a format reward, and a fallback bonus. Training data mixes video QA and temporal grounding with text and image math/science sources, totaling 83K samples after filtering; the visual encoder is frozen while the projector and language model are tuned, on 32 H100 GPUs for about 35 hours.

At inference, the model decodes only the first answer, computes the average log-probability of its tokens, and compares it against log τ with τ = 0.97. High confidence means stop, which is the direct-answering path; low confidence or the fallback string means continue into reasoning and the reviewed answer. Because the first answer is typically under ten tokens, this check is cheap. Evaluation uses lmms-eval with greedy decoding, and the reported numbers are the best over several frame and token budgets.

Why This Matters

Impact on research. The paper challenges the assumption that more reasoning tokens mean better multimodal performance, and provides a template that separates the training objective (learn to answer and to reason) from the inference policy (decide when to stop) without extra labels or architecture changes. It also carries overthinking findings from the text and image domains into video.

Real-world applications (extensions of the tasks studied, not claims tested in the paper):

• Video question answering assistants that must respond quickly for simple perceptual queries and only "slow down" for genuinely complex ones.
• Temporal grounding for video editing, search, and highlight retrieval, where the paper shows a single direct prediction already localizes well on Charades-STA and ActivityNet.
• Grounding QA over long videos (NExT-GQA style), for example in video archives or surveillance review, where both an answer and the supporting time span are needed.
• Multimodal tutoring or instructional-video assistants, the setting where the paper observes CoT genuinely helps (math and physics videos with symbolic content).

Industry relevance. The efficiency claim is directly practical: cutting average response length from 149 to 44 tokens reduces decoding cost and latency for the same or better accuracy, which matters for deployment budgets and for serving long videos. The absence of per-sample mode labels also lowers data-annotation effort relative to prior adaptive-reasoning approaches.

Future Directions

• Applying the confidence-based early-exit rule to grounding outputs, since the paper reports it early-exits by default on temporal grounding rather than learning when reasoning would help there.
• Generalizing the threshold τ so it does not need tuning on a held-out validation subset, and understanding how it transfers across model families and video domains.
• Investigating why think-mode activation still differs substantially across benchmarks (11% to 56% depending on model and task) and whether those rates reflect genuine task difficulty or calibration artifacts.
• Extending the answer–think–answer template and dual-answer reward beyond video and images, for example to longer-horizon or tool-using multimodal agents.

Target Audience

Researchers and engineers working on multimodal LLMs, video understanding, and reinforcement learning post-training; practitioners who need to trade off inference cost against accuracy in deployed video systems; and anyone studying adaptive or efficient reasoning, including those who followed the "overthinking" literature in text and image models.

Authors’ abstract

Chain-of-thought (CoT) reasoning has emerged as a powerful tool for multimodal large language models on video understanding tasks. However, its necessity and advantages over direct answering remain underexplored. In this paper, we first demonstrate that for RL-trained video models, direct answering often matches or even surpasses CoT performance, despite CoT producing step-by-step analyses at a higher computational cost. Motivated by this, we propose VideoAuto-R1, a video understanding framework that adopts a reason-when-necessary strategy. During training, our approach follows a Thinking Once, Answering Twice paradigm: the model first generates an initial answer, then performs reasoning, and finally outputs a reviewed answer. Both answers are supervised via verifiable rewards. During inference, the model uses the confidence score of the initial answer to determine whether to proceed with reasoning. Across video QA and grounding benchmarks, VideoAuto-R1 achieves state-of-the-art accuracy with significantly improved efficiency, reducing the average response length by ~3.3x, e.g., from 149 to just 44 tokens. Moreover, we observe a low rate of thinking-mode activation on perception-oriented tasks, but a higher rate on reasoning-intensive tasks. This suggests that explicit language-based reasoning is generally beneficial but not always necessary.