CVPR Findings2026
Video Reasoning Without Training
Deepak Sridhar, Kartikeya Bhardwaj, Jeya Pradha Jeyaraj, Nuno Vasconcelos, Ankita Nayak, Harris Teague
Abstract
Video reasoning using Large Multimodal Models (LMMs) relies on costly reinforcement learning (RL) and verbose chainof-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, we use the entropy of the model's output distribution as a signal to study and guide reasoning behavior. We discover that high-quality models exhibit a characteristic pattern of micro-exploration and micro-exploitation cycles, followed by a later entropy peak (i.e., longer thinking) and a lower final entropy, indicating more deliberate exploration and confident convergence (i.e., avoid excessive randomness while the model is exploring or thinking through an answer). We then use these novel, theoretically-grounded insights to introduce V-Reason (Video-Reason), an inference-time optimization method that adapts the value cache of the LMM through a lightweight, trainable controller. Our proposed controller is guided by an entropy-based objective, to tune the model's behavior directly at inference, without using any RL or supervised fine-tuning. Our experiments show that V-Reason significantly outperforms the base instruction-tuned models on many video reasoning datasets, narrowing the gap with RL models to within 0.6% accuracy on average. We achieve this without any training, while offering efficiency benefits: V-Reason uses 58.6% fewer tokens than the RL model.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- Large Language Models are Zero-Shot Reasoners cites Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
- Large Language Models as Optimizers cites Large Language Models are Zero-Shot Reasoners
- Large Language Models as Optimizers cites Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
- Video-R1: Reinforcing Video Reasoning in MLLMs cites Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Builds on29
- Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsJason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma et al.NeurIPS 2022 · 22,562 citations
- Large Language Models are Zero-Shot ReasonersTakeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo et al.NeurIPS 2022 · 8,168 citations
- Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang et al.NeurIPS 2025 · 1,109 citations
- Large Language Models as OptimizersChengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu et al.ICLR 2024 · 817 citations
- Video-R1: Reinforcing Video Reasoning in MLLMsKaituo Feng, Kaixiong Gong, Bohao Li, Zonghao Guo et al.NeurIPS 2025 · 528 citations
Related papers
- ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy ShapingShuang Chen, Hangyu Guo, Yimeng Ye, Shijue Huang et al.ICLR 2026 · 23 citations
- VideoAuto-R1: Video Auto Reasoning via Thinking Once, Answering TwiceShuming Liu, Mingchen Zhuge, Changsheng Zhao, Jun Chen et al.CVPR 2026 · 18 citations
- VAST: Video Ability-Stratified Taxonomy for Data-Efficient Video ReasoningZhongan Wang, Xiaoyu Wen, Lingxiao Du, Kun Li et al.CVPR 2026
- Temporal-Aware Reasoning Optimization for Video Temporal GroundingMinghang Zheng, Zihao Yin, YI YANG, Yuxin Peng et al.ICML 2026
- When Thinking Drifts: Evidential Grounding for Robust Video ReasoningRomy Luo, Zihui Xue, Alex Dimakis, Kristen GraumanNeurIPS 2025 · 21 citations
