VIDEOP2R: Video Understanding from Perception to Reasoning

TL;DR

VideoP2R introduces a process-aware reinforcement fine-tuning framework for large video language models, significantly improving reasoning capabilities by modeling perception and reasoning separately and achieving state-of-the-art results.

Contribution

It develops a novel process-aware video RFT framework with a new dataset and a specialized policy optimization algorithm for enhanced video reasoning.

Findings

Achieves SOTA on 6 out of 7 benchmarks.

The perception output is sufficient for downstream reasoning.

The process-aware modeling and PA-GRPO improve performance.

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. Our project page is available at https://videop2r.github.io/videop2r/.