VidBridge-R1: Bridging QA and Captioning for RL-based Video Understanding Models with Intermediate Proxy Tasks

TL;DR

This paper tackles the conflict between QA and captioning objectives when applying Reason-Then-Respond with reinforcement learning to video understanding. It introduces two intermediate proxy tasks, DarkEventInfer and MixVidQA, to encourage holistic contextual understanding and precise information localization, bridging divergent and convergent reasoning demands. Building on these proxies, VidBridge-R1 emerges as a versatile video reasoning model that excels in QA, reasoning, and captioning within a single framework. Empirical results across multiple benchmarks show significant gains and better generalization, indicating the effectiveness of proxy-task scheduling for generalist multimodal video models.

Abstract

The "Reason-Then-Respond" paradigm, enhanced by Reinforcement Learning, has shown great promise in advancing Multimodal Large Language Models. However, its application to the video domain has led to specialized models that excel at either question answering (QA) or captioning tasks, but struggle to master both. Naively combining reward signals from these tasks results in mutual performance degradation, which we attribute to a conflict between their opposing task natures. To address this challenge, we propose a novel training framework built upon two intermediate proxy tasks: DarkEventInfer, which presents videos with masked event segments, requiring models to infer the obscured content based on contextual video cues; and MixVidQA, which presents interleaved video sequences composed of two distinct clips, challenging models to isolate and reason about one while disregarding the other. These proxy tasks compel the model to simultaneously develop both holistic, divergent understanding and precise, convergent reasoning capabilities. Embodying this framework, we present VidBridge-R1, the first versatile video reasoning model that effectively bridges the paradigm conflict. Extensive experiments show that VidBridge-R1 achieves significant performance gains on both QA and captioning within one model, demonstrating the efficacy of our approach in fostering more generalizable and powerful video understanding models.

Figures (15)

• Figure 1: Performance comparison on QA and captioning tasks under different training setups. Details can be found in Section \ref{['sec:data_abla']}
• Figure 2: The training framework of VidBridge-R1. By incorporating intermediate proxy tasks, VidBridge-R1 effectively bridges the gap between QA capabilities in general or reasoning scenarios and video captioning tasks.
• Figure 3: The training dynamics of VidBridge-R1 on video general understanding, reasoning, and captioning tasks.
• Figure 4: Prompts for eliciting model reasoning
• Figure 5: Prompts to generate QA pairs in MixVidQA.