VIDEOP2R: Video Understanding from Perception to Reasoning

2025-11-14

AI Summary

To address the insufficient reasoning capability of Large Video-Language Models (LVLMs), this paper proposes a two-stage reinforcement fine-tuning framework that decouples perception from reasoning. Methodologically: (1) we construct a process-aware chain-of-thought dataset that explicitly separates video perception from logical reasoning; (2) we design PA-GRPO—a phased, population-relative reward optimization algorithm; and (3) we adopt a three-stage data generation pipeline integrating supervised fine-tuning and reinforcement learning. Our key contribution is the first explicit formulation of perception outputs as reliable, sufficient premises for downstream reasoning—empirically validated through ablation studies. Experiments demonstrate state-of-the-art performance on 6 out of 7 mainstream video understanding benchmarks. Ablation analyses confirm the effectiveness of each component, establishing the critical role of perception-reasoning decoupling in enhancing LVLM reasoning fidelity.

Technology Category

• Computer Vision: Video Understanding & Activity Analysis
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Argumentation

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Reinforcement fine-tuning (RFT), a two-stage framework consisting of supervised fine-tuning (SFT) and reinforcement learning (RL) has shown promising results on improving reasoning ability of large language models (LLMs). Yet extending RFT to large video language models (LVLMs) remains challenging. We propose VideoP2R, a novel process-aware video RFT framework that enhances video reasoning by modeling perception and reasoning as distinct processes. In the SFT stage, we develop a three-step pipeline to generate VideoP2R-CoT-162K, a high-quality, process-aware chain-of-thought (CoT) dataset for perception and reasoning. In the RL stage, we introduce a novel process-aware group relative policy optimization (PA-GRPO) algorithm that supplies separate rewards for perception and reasoning. Extensive experiments show that VideoP2R achieves state-of-the-art (SotA) performance on six out of seven video reasoning and understanding benchmarks. Ablation studies further confirm the effectiveness of our process-aware modeling and PA-GRPO and demonstrate that model's perception output is information-sufficient for downstream reasoning.

Problem

• Extends reinforcement fine-tuning to video language models
• Models perception and reasoning as distinct processes
• Enhances video reasoning through process-aware optimization

Innovation

• Two-stage reinforcement fine-tuning for video models
• Process-aware chain-of-thought dataset generation
• Separate rewards for perception and reasoning processes