Video Reasoning without Training

2025-10-19

AI Summary

Video reasoning typically relies on reinforcement learning (RL) and long chain-of-thought prompting, incurring high computational overhead and lacking controllable, real-time adjustment of the reasoning process. To address this, we propose a training-free, inference-time control method. We first discover a dynamic entropy pattern in the outputs of large vision models during video understanding; leveraging this insight, we design an entropy-based objective function and a trainable controller, integrated with value caching for adaptive optimization and inference-time micro-exploration/micro-exploitation tuning. Crucially, our approach entirely avoids RL and supervised fine-tuning. On multiple video reasoning benchmarks, it achieves accuracy competitive with RL-based models while reducing output token count by 58.6%, significantly improving both efficiency and stability. Our core contribution is establishing the first entropy-driven, training-agnostic paradigm for dynamic control of video reasoning processes.

Technology Category

• Computer Vision: Large Vision Models
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Agentic search

Abstract

Video reasoning using Large Multimodal Models (LMMs) relies on costly reinforcement learning (RL) and verbose chain-of-thought, resulting in substantial computational overhead during both training and inference. Moreover, the mechanisms that control the thinking process in these reasoning models are very limited. In this paper, using entropy of the model's output as a signal, we discover that the high-quality models go through a series of micro-explorations and micro-exploitations which keep the reasoning process grounded (i.e., avoid excessive randomness while the model is exploring or thinking through an answer). We further observe that once this "thinking" process is over, more accurate models demonstrate a better convergence by reducing the entropy significantly via a final exploitation phase (i.e., a more certain convergence towards a solution trajectory). We then use these novel, theoretically-grounded insights to tune the model's behavior directly at inference, without using any RL or supervised fine-tuning. Specifically, during inference, our proposed approach called V-Reason (Video-Reason) adapts the value cache of the LMM via a few optimization steps on a small, trainable controller using an entropy-based objective, i.e., no supervision from any dataset or RL is necessary. This tuning improves the model's micro-exploration and exploitation behavior during inference. Our experiments show that our proposed method achieves significant improvements over the base instruction-tuned models across several video reasoning datasets, narrowing the gap with RL-trained models to within 0.6% average accuracy without any training, while offering massive efficiency benefits: output tokens are reduced by 58.6% compared to the RL model.

Problem

• Reduces computational overhead in video reasoning models
• Improves reasoning control without reinforcement learning
• Optimizes inference efficiency via entropy-based adaptation

Innovation

• Uses entropy to guide model reasoning process
• Optimizes value cache via small controller during inference
• Improves exploration-exploitation without training or RL