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

TL;DR

VideoAuto-R1 addresses whether chain-of-thought reasoning is essential for video understanding by demonstrating that direct answers suffice for many video tasks and introducing an adaptive reasoning framework. It trains a Thinking Once, Answering Twice model with a dual-answer GRPO reward and uses a confidence-based early exit to decide when to activate reasoning, achieving state-of-the-art results on video QA and temporal grounding while reducing token usage by about $3.3\times$ compared with full CoT models. The method automatically focuses thinking on harder, reasoning-intensive cases, as evidenced by higher think-mode activation on VideoMMMU and lower activation on perception-only tasks. This work provides a practical, scalable approach to adaptive video reasoning with strong efficiency-accuracy trade-offs and broad applicability to multimodal benchmarks.

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.

Figures (11)

• Figure 1: VideoAuto-R1 follows a thinking once, answering twice paradigm. In training, both the initial answer and the reviewed answer are supervised with verifiable rewards. During inference, an early-exit mechanism is adopted to dynamically determine whether to proceed with CoT reasoning. Robot icon from robot_icon.
• Figure 2: Overview of VideoAuto-R1. (a) Training: The response follows the answer $\rightarrow$ think $\rightarrow$ 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 $\tau$, decoding terminates as direct answering; otherwise, the model continues with CoT reasoning and outputs a reviewed answer, enabling adaptive, confidence-based early exit.
• Figure 3: Effect of the Early-Exit Threshold on Accuracy and Think Ratio. In practice, we set $\tau = 0.97$ for all datasets.
• Figure 4: VideoAuto-R1 Performing Complex Math Reasoning. The model applies probability and integration, revising an incorrect initial answer to the correct one through structured reasoning.
• Figure 5: Distribution of per-sample accuracy in the initial training pool, estimated by evaluating 8 diverse responses per sample. Samples with all responses correct or all incorrect are considered too easy or too hard and are excluded from QA-based data.