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

Shuming Liu1,2,*, Mingchen Zhuge1,2, Changsheng Zhao1, Jun Chen1, Lemeng Wu1, Zechun Liu1, Chenchen Zhu1 Zhipeng Cai1 Chong Zhou1 Haozhe Liu1,2 Ernie Chang1 Saksham Suri1 Hongyu Xu1 Qi Qian1 Wei Wen1 Balakrishnan Varadarajan1 Zhuang Liu3 Hu Xu1 Florian Bordes1 Raghuraman Krishnamoorthi1 Bernard Ghanem1 Vikas Chandra1 Yunyang Xiong1,†
1Meta AI, 2King Abdullah University of Science and Technology (KAUST), 3Princeton University
*Work done at Meta, †Project lead
Paper Code Models 🗂️ Data 🤗 Demo

Rather than always engaging in reasoning, VideoAuto-R1 first generates an initial answer, then adaptively decides whether to think further and output a reviewed answer.

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 144 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.

Method

(a) Training: The response follows the answer → think → answer template, jointly optimizing both the initial and reviewed answers. Specifically, a fallback reward is introduced to avoid a spurious initial guess.

(b) Inference: The model first produces an initial answer. If its length-normalized confidence exceeds a threshold τ, decoding terminates as direct answering; otherwise, the model continues with CoT reasoning and outputs a reviewed answer, enabling adaptive, confidence-based early exit.

Overview of VideoAuto-R1
Overview of VideoAuto-R1

Results on Video QA Benchmarks

Results on Temporal Grounding Benchmarks

Qualitative Results

VideoAuto-R1 on Perception-Oriented QA (left) and Reasoning-Oriented QA (right)

BibTeX

@article{liu2026videoautor1,
  title={VideoAuto-R1: Video Auto Reasoning via Thinking Once, Answering Twice},
  author={Liu, Shuming and Zhuge, Mingchen and Zhao, Changsheng and Chen, Jun and Wu, Lemeng and Liu, Zechun and Zhu, Chenchen and Cai, Zhipeng and Zhou, Chong and Liu, Haozhe and Chang, Ernie and Suri, Saksham and Xu, Hongyu and Qian, Qi and Wen, Wei and Varadarajan, Balakrishnan and Liu, Zhuang and Xu, Hu and Bordes, Florian and Krishnamoorthi, Raghuraman and Ghanem, Bernard and Chandra, Vikas and Xiong, Yunyang},
  journal={arXiv preprint arXiv:2601.05175},
  year={2026}
}