AdaThinkV: Adaptive Thinking for Token-Efficient Video Reasoning
AdaThinkV uses adaptive reinforcement learning to optimize token usage, achieving 40.79% accuracy with 22.7% fewer tokens in video reasoning.
Key Findings
Methodology
AdaThinkV employs a reinforcement learning framework with core components including ThinkGain, which estimates the utility of explicit reasoning by balancing accuracy gains against response length, and VRPO, which dynamically samples and expands low-signal prompt groups. The model learns to select between explicit reasoning and direct answering without offline difficulty labels or manual thresholds. During training, matched rollouts in different modes are used to optimize a policy that maximizes the net utility of reasoning. ThinkGain guides the model to make informed decisions based on prompt-specific utility estimates, while VRPO addresses the challenge of weak signals in difficult prompts by retaining and expanding unproductive samples. The system is evaluated on multiple video reasoning benchmarks, achieving an average accuracy of 40.79% with an average token length of 257.20, outperforming previous adaptive baselines by nearly 3 points and reducing token usage by 22.7%.
Key Results
- Across diverse video reasoning tasks, AdaThinkV surpasses baseline models with an average accuracy of 40.79%, using fewer tokens (257.20 tokens) and achieving a 2.98% accuracy improvement over the strongest baseline. It demonstrates superior efficiency especially in complex scenarios, with notable gains in temporal grounding and video QA benchmarks.
- The model's ability to adaptively choose reasoning modes based on prompt utility leads to significant token savings without sacrificing accuracy. Ablation studies reveal ThinkGain primarily improves the accuracy-token trade-off, while VRPO enhances absolute accuracy, especially in hard prompts.
- Experimental results confirm that the combined approach effectively balances reasoning depth and cost, enabling high-performance video understanding with reduced resource consumption, making it suitable for real-world applications requiring efficiency and accuracy.
Significance
This work addresses a fundamental challenge in video multimodal reasoning: balancing reasoning complexity with token efficiency. By enabling models to dynamically adjust their reasoning efforts based on prompt utility, AdaThinkV reduces computational costs and token wastage, paving the way for more practical deployment of large-scale video understanding systems. Its ability to learn mode selection autonomously without offline labels or manual thresholds marks a significant step forward in adaptive AI. The approach enhances both the scalability and robustness of multimodal models, with broad implications for real-time video analysis, intelligent surveillance, and autonomous systems. The framework's generality suggests potential extensions to other multimodal tasks, fostering a new paradigm of cost-aware, self-adaptive AI systems.
Technical Contribution
The paper introduces ThinkGain, a prompt-specific utility estimator that guides the model’s mode selection by balancing accuracy gains and response length costs. It also proposes VRPO, a novel rollout allocation strategy that retains and expands unproductive prompt groups to recover informative signals in difficult samples. The integration of these components within a reinforcement learning framework allows AdaThinkV to autonomously learn to switch between explicit reasoning and direct answering, without relying on offline labels or manually tuned thresholds. This approach fundamentally differs from prior methods that depend on static thresholds or external routing, offering a more flexible, data-driven solution that optimizes token efficiency and reasoning performance simultaneously.
Novelty
This is the first framework to eliminate the need for offline difficulty labels or manual thresholds in adaptive video reasoning. By leveraging prompt-matched rollout comparisons and a length-aware utility estimator, AdaThinkV dynamically balances reasoning depth and response cost. Its innovative VRPO strategy for sample retention and expansion in low-signal scenarios further distinguishes it from existing adaptive methods, enabling more reliable learning signals in challenging prompts. The unified autoregressive generation of mode markers and responses in a single sequence also sets it apart from prior two-stage or external routing approaches, marking a significant advancement in cost-aware multimodal reasoning.
Limitations
- The model's accuracy in extremely ambiguous or noisy video scenarios remains limited, as utility estimation may be inaccurate under high uncertainty.
- Training involves substantial computational resources due to reinforcement learning and sample expansion, posing challenges for large-scale deployment.
- The current approach assumes prompt-level utility can be reliably estimated, which may not hold in highly diverse or unseen tasks, requiring further robustness improvements.
Future Work
Future research will focus on integrating more sophisticated multi-modal fusion techniques to better estimate prompt utility in complex scenes. Additionally, optimizing the VRPO sampling strategy to reduce training costs and improve scalability is a priority. Exploring meta-learning approaches for faster adaptation to new tasks and environments will further enhance the model's generalization. Extending the framework to other domains such as robotics or autonomous driving, where adaptive reasoning is critical, also presents promising directions.
AI Executive Summary
Video reasoning remains a challenging frontier in artificial intelligence, demanding models that can balance complex inference with computational efficiency. Traditional approaches often rely on static reasoning depths or manual thresholds, leading to inefficiencies—either overusing tokens on simple questions or underperforming on complex ones. This dichotomy hampers real-world deployment, where resource constraints and diverse task demands are prevalent.
In response, this paper introduces AdaThinkV, a novel adaptive framework that leverages reinforcement learning to dynamically select the appropriate reasoning mode for each prompt. Central to its design are two innovative components: ThinkGain, which estimates the utility of explicit reasoning by comparing matched rollout responses, and VRPO, which intelligently retains and expands low-signal prompt groups to recover valuable learning signals. These mechanisms enable the model to autonomously decide whether to reason explicitly or answer directly, optimizing token usage without relying on offline labels or manually tuned thresholds.
Extensive experiments across multiple video reasoning benchmarks demonstrate AdaThinkV’s superiority. It achieves an average accuracy of 40.79%, outperforming the strongest adaptive baseline by nearly 3 points, while using 22.7% fewer tokens. The model excels particularly in complex scenarios, where its ability to adaptively balance reasoning depth and cost leads to significant efficiency gains. These results highlight a promising direction for scalable, cost-effective multimodal AI systems.
Beyond performance metrics, this work advances the theoretical understanding of utility-based adaptive reasoning. By integrating prompt-specific utility estimation with sample retention strategies, AdaThinkV sets a new standard for intelligent, resource-aware AI. Its implications extend to real-time video analysis, autonomous systems, and beyond, where efficient, accurate reasoning is critical. Despite current limitations in extreme ambiguity scenarios, the framework opens avenues for future enhancements in multi-modal fusion, faster training, and broader application domains, marking a significant step toward truly intelligent, adaptable AI systems.
Deep Analysis
Background
视频推理技术近年来快速发展，代表性工作包括Wei等的多模态大模型、Feng等的显式推理方法。早期研究多关注模型性能提升，采用结构化蒸馏、强化学习等技术，但普遍存在推理成本高、效率低的问题。随着模型规模扩大，Token消耗成为瓶颈，研究逐渐转向推理路径压缩和动态推理策略。现有方法多依赖预定义阈值或外部路由，难以在不同复杂度场景中自适应调节推理深度。尽管如此，如何在保证推理效果的同时，减少Token浪费，仍是当前研究的热点和难点。
Core Problem
核心问题在于如何让多模态大模型根据提示难度自主调节推理深度，避免在简单问题上过度推理造成Token浪费，同时在复杂问题中不缺失必要的推理步骤。现有方法多依赖手动阈值或外部路由器，缺乏灵活性和泛化能力，难以应对多样化的任务场景。这导致模型在实际应用中存在效率低、成本高、响应不稳定等问题，亟需一种无需离线标签、能动态调节推理策略的解决方案。
Innovation
本研究提出了AdaThinkV框架，核心创新包括：1）ThinkGain机制，基于匹配rollout样本估算推理效用，动态调节推理深度；2）VRPO策略，保留并逐步扩展低信号样本组，恢复有效学习信号；3）自主模式选择，无需离线标签或手动阈值，强化学习优化推理策略。这些创新点突破了传统依赖阈值和外部路由的限制，实现模型在不同复杂度场景中的自适应调节，显著提升推理效率和准确率。
Methodology
- �� 输入：视频v与问题q，结合成x=(v,q)。
- �� 模型在自回归策略下生成响应，模式包括THINK（显式推理）与ANSWER（直接回答），由标记区分。
- �� ThinkGain通过匹配的rollout样本，估算每个提示的推理效用，考虑准确率提升与响应长度。
- �� 在训练中，模型在不同推理模式间采样，利用强化学习优化策略，平衡探索与利用。
- �� VRPO在难题样本中保留未成功样本，逐步扩展样本组，恢复学习信号。
- �� 模型在多模态视频推理数据集上训练，指标包括平均准确率与Token数，验证其自适应调节能力。
Experiments
采用多项视频推理基准，包括VideoReasonBench、VideoMathQA、MMR-VBench和SciVideoBench，涵盖时间推理、数学推理和科学理解任务。模型在不同任务中与多种对比模型（如Qwen系列、VideoAuto-R1）进行比较，指标包括准确率和Token数。训练采用结合有监督微调（SFT）和强化学习（RL）策略，超参数包括学习率1e-6、Token限制4096，rollout温度1.0。通过消融实验验证ThinkGain和VRPO的贡献，分析模型在复杂场景中的表现。
Results
在所有测试任务中，AdaThinkV平均准确率达40.79%，Token数为257.20，优于对比模型2.98个百分点，Token节省22.7%。在Temporal Grounding任务中，准确率提升至75.2%，Token显著减少。消融实验显示，ThinkGain主要优化Token-准确率折中，VRPO则提升绝对准确率。模型在复杂推理场景中表现优越，尤其在难题中通过样本扩展恢复信号，显著优于传统策略。
Applications
该模型适用于视频内容理解、自动问答、场景分析等多模态应用。其自主调节推理深度，极大提升了实际系统的效率和响应速度，适合部署在资源有限的边缘设备或实时系统中。未来可结合多模态信息融合技术，推动智能监控、自动驾驶等行业的智能化升级。
Limitations & Outlook
模型在极端复杂或模糊场景中，推理效用估算可能偏差，影响策略选择。训练成本较高，强化学习过程耗时较长，难以快速部署。未来需优化样本扩展机制，提升模型泛化能力，并在多模态融合方面进一步突破。
Plain Language Accessible to non-experts
想象你在厨房做饭。有些菜很简单，只需要放几样调料就能完成，不需要复杂步骤。而有些菜很复杂，需要多次尝试和调整。AdaThinkV就像一个聪明的厨师，能判断每道菜的难度，决定是用简单的方法直接做，还是花时间用复杂的步骤去做。这样既节省时间，也保证菜的质量。它通过学习不断调整自己的做菜策略，既不浪费食材，也不做不必要的复杂菜肴。这个厨师还会根据菜的难度，灵活选择不同的做法，确保每道菜都能做得既快又好。
ELI14 Explained like you're 14
想象你在学校里做科学实验。有些问题很简单，比如算算水果的重量，只需要用秤就能知道，不用花太多时间。而有些问题很难，比如找出某个动物的生活习性，你需要观察、思考很多步骤。AdaThinkV就像一个聪明的学生，知道什么时候用简单的方法直接回答，什么时候需要花时间做详细的分析。它会根据问题的难度，自主选择用快答还是详细推理。这样既节省时间，又能保证答案的准确。它还会学习，变得越来越聪明，知道在不同问题上用不同的方法，变得更厉害！
Glossary
Chain-of-Thought (CoT) Chain推理 (链式推理)
一种逐步推导的推理方式，通过显式展示推理过程帮助模型理解复杂问题。技术上指模型生成中间推理步骤的策略。
论文中强调利用CoT提升复杂视频问题的推理能力。
ThinkGain 思考效用
衡量显式推理带来的准确率提升与响应长度成本的差值，用于指导模型自主调节推理策略。
核心机制之一，用于动态估算推理的价值。
VRPO 变异恢复策略优化
一种样本采样策略，通过保留和逐步扩展低信号样本组，恢复有价值的学习信号，提升模型在难题中的表现。
解决难题样本信号不足的问题。
Token Token (符号单元)
模型生成文本的基本单位，代表一段连续字符或词的编码，用于衡量推理成本和效率。
论文中用以评估推理的Token节省效果。
Open Questions Unanswered questions from this research
- 1 如何进一步提升模型在极端复杂场景中的推理准确率？目前模型在极端模糊或多模态信息冲突时表现仍有限，未来需结合更强的多模态融合和推理机制。
Applications
Immediate Applications
视频内容理解
可应用于自动视频标注、内容检索和智能监控，提升系统在复杂场景下的推理效率和准确率，减少Token消耗，适合边缘设备部署。
自动问答系统
在多模态问答中，根据问题难度自主调节推理深度，提高响应速度和准确性，增强用户体验。
Long-term Vision
智能视频分析平台
结合自适应推理技术，打造高效、智能的视频内容分析平台，支持大规模实时处理和多任务推理，推动行业智能化升级。
Abstract
Chain-of-thought (CoT) reasoning can improve performance on difficult video questions but often wastes decoding tokens on simple ones. We study whether a video multimodal large language model can adapt its reasoning effort to each question. We propose AdaThinkV, 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, AdaThinkV samples matched rollouts in explicit reasoning and direct answering modes for each prompt. ThinkGain estimates the prompt-level utility of explicit reasoning by balancing its accuracy gain against additional response length, providing supervision for both conditional response generation and autonomous mode selection. For difficult prompts, limited rollout exploration can yield groups in which every response is unsuccessful and accuracy rewards show little variation, providing insufficient signal for learning. We therefore introduce Variance Recovery Policy Optimization (VRPO), which retains and progressively expands these 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. Across a unified suite of video reasoning evaluations, AdaThinkV achieves a mean accuracy of 40.79 with an average of 257.20 output tokens, outperforming the strongest evaluated adaptive baseline by 2.98 points while using 22.7% fewer tokens. Project page: https://trilarflagz.github.io/AdaThinkV/