VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning

TL;DR

VideoRFT extends reinforcement fine-tuning to video reasoning in multimodal language models by building a cognitively grounded CoT data pipeline and two large datasets (VideoRFT-CoT-102K and VideoRFT-RL-310K). It introduces a cross-modal CoT refinement step and a semantic-consistency reward within a GRPO-based RL framework to align reasoning with visual evidence. The approach yields state-of-the-art results across six video reasoning benchmarks and demonstrates substantial gains from data quality, training paradigm, and reward design. This work provides a scalable blueprint for grounded, multi-step video reasoning in large multimodal models with potential wide-ranging applications.

Abstract

Reinforcement fine-tuning (RFT) has shown great promise in achieving humanlevel reasoning capabilities of Large Language Models (LLMs), and has recently been extended to MLLMs. Nevertheless, reasoning about videos, which is a fundamental aspect of human intelligence, remains a persistent challenge due to the complex logic, temporal and causal structures inherent in video data. To fill this gap, we propose VideoRFT, a novel approach that extends the RFT paradigm to cultivate human-like video reasoning capabilities in MLLMs. VideoRFT follows the standard two-stage scheme in RFT: supervised fine-tuning (SFT) with chain-of-thought (CoT) annotations, followed by reinforcement learning (RL) to improve generalization. A central challenge to achieve this in the video domain lies in the scarcity of large-scale, high-quality video CoT datasets. We address this by building a multi-expert-driven, cognition-inspired CoT curation pipeline. First, we devise a cognition-inspired prompting strategy to elicit a reasoning LLM to generate preliminary CoTs based solely on rich, structured, and literal representations of video content. Subsequently, these CoTs are revised by a MLLM conditioned on the actual video, ensuring visual consistency and reducing visual hallucinations. This pipeline results in two new datasets, i.e.VideoRFT-CoT-102K for SFT and VideoRFT-RL-310K for RL. To further strengthen the RL phase, we introduce a novel semantic-consistency reward that explicitly promotes the alignment between textual reasoning and visual evidence. This reward encourages the model to produce coherent, context-aware reasoning outputs grounded in visual input. Extensive experiments show that VideoRFT achieves state-of-the-art performance on six video reasoning benchmarks.

Figures (7)

• Figure 1: Overview of VideoRFT. (a) An example of CoT derived from VideoRFT. (b) VideoRFT achieves leading performance in six datasets. (c) The two-stage RFT underpins the training of VideoRFT.
• Figure 2: The distribution of data collection.
• Figure 3: Illustration of the pipeline for cognitively inspired CoT generation.
• Figure 4: Comparison of CoT dataset in VideoRFT-CoT-102K and Video-R1.
• Figure 5: Illustrations of (a) rule-based RL, and (b) the computation of semantic-consistency reward $R_s$. The reasoning outputs are color-coded to highlight question parsing (green), video description (red) and abstract reasoning (blue). Only the red part is involved in the computation of $R_s$ (see §\ref{['sec:s_reward']}).