Title: Enabling Video Reasoning Without Training

Abstract:

Leveraging Large Multimodal Models (LMMs) for video reasoning typically necessitates expensive reinforcement learning (RL) and verbose chain-of-thought processes. These approaches incur significant computational costs during both the training and inference phases. Furthermore, existing reasoning models suffer from restricted mechanisms for controlling their internal thinking processes. To address these limitations, this study utilizes the entropy of the model’s output distribution as a metric to analyze and direct reasoning behavior.

Our analysis reveals that high-performing models follow a distinct pattern: they engage in cycles of micro-exploration and micro-exploitation, which are subsequently followed by a peak in entropy—signifying extended deliberation—and finally converge to a lower entropy state. This trajectory indicates that the models engage in more purposeful exploration and achieve confident convergence, thereby avoiding excessive randomness while processing answers.

Building on these theoretically grounded insights, we propose V-Reason (Video-Reason), an inference-time optimization technique. V-Reason employs a lightweight, trainable controller to adapt the value cache of the LMM. Guided by an entropy-based objective, this controller adjusts the model’s behavior directly during inference, eliminating the need for RL or supervised fine-tuning.

Experimental results demonstrate that V-Reason surpasses base instruction-tuned models across various video reasoning benchmarks. It effectively closes the performance gap with RL-based models, maintaining an average accuracy difference of just 0.6%. Notably, this improvement is achieved without any training phase, while also delivering substantial efficiency gains: V-Reason reduces token usage by 58.6% compared to RL models.

Project Page: https://deepaksridhar.github.io/vreason.github.io/

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC