AdaThinkV teaches a video MLLM to choose between direct answering and explicit reasoning for each question. ThinkGain estimates whether the accuracy benefit of reasoning justifies its extra output length, while Variance Recovery Policy Optimization (VRPO) expands difficult rollout groups to recover useful reinforcement-learning signals.
Chain-of-thought reasoning can improve performance on difficult video questions but often wastes decoding tokens on simple ones. AdaThinkV is an adaptive framework for video reasoning that learns whether to reason explicitly without offline difficulty labels, manually tuned confidence thresholds, or an external router. During reinforcement learning, it samples matched rollouts in explicit-reasoning and direct-answering modes. ThinkGain estimates the prompt-level utility of explicit reasoning by balancing its accuracy gain against additional response length, supervising both conditional generation and autonomous mode selection. For difficult prompts, VRPO retains and progressively expands unsuccessful, low-dispersion rollout groups to recover informative signals from prompts that are difficult yet solvable. At inference, AdaThinkV selects a response mode and generates the response in a single autoregressive sequence.
THINK/ANSWER branches. VRPO progressively allocates more rollouts to groups with weak reward contrast. At inference, the model emits a mode marker and response in one autoregressive sequence.Measures the conditional net utility of explicit reasoning: whether its prompt-level accuracy improvement is worth the additional response length.
Retains and expands difficult all-unsuccessful groups, broadening exploration until the policy recovers informative reward variation.
Selects direct or explicit reasoning without an external router, confidence threshold, or preliminary answer.
AdaThinkV advances the empirical Pareto frontier across the unified video-reasoning evaluation suite.
| Model | Mode | Accuracy | Tokens |
|---|---|---|---|
| Qwen3-VL-8B-Thinking | Think only | 37.13 | 6184.98 |
| VideoAuto-R1-Qwen3 | Auto | 37.81 | 332.93 |
| AdaThinkV w/o VRPO | Auto | 38.32 | 235.73 |
| AdaThinkV | Auto | 40.79 | 257.20 |
@article{tian2026adathinkv,
title={AdaThinkV: Adaptive Thinking for Token-Efficient Video Reasoning},
author={Tian, Jingqi and Zhang, Haoji and Chen, Lin and Jin, Hongbo and Xu, Haonan and Zhu, Tianrui and Shui, Xingming and Ma, Shilin and Yang, Wenjing and Tang, Yansong},
journal={arXiv preprint},
year={2026}
}