Watch Before You Answer: Learning from Visually Grounded Post-Training

TL;DR

This paper identifies that many video understanding benchmarks contain questions answerable by text alone and introduces VidGround, a data curation method that improves vision-language model performance by focusing on visually grounded questions.

Contribution

The authors propose VidGround, a simple data curation technique that enhances post-training for video understanding by emphasizing visually grounded questions, outperforming complex methods.

Findings

VidGround improves performance by up to 6.2 points.

Using only 69.1% of data yields comparable or better results.

Data quality and proper curation are key to advancing VLMs.

Abstract

It is critical for vision-language models (VLMs) to comprehensively understand visual, temporal, and textual cues. However, despite rapid progress in multimodal modeling, video understanding performance still lags behind text-based reasoning. In this work, we find that progress is even worse than previously assumed: commonly reported long video understanding benchmarks contain 40-60% of questions that can be answered using text cues alone. Furthermore, we find that these issues are also pervasive in widely used post-training datasets, potentially undercutting the ability of post-training to improve VLM video understanding performance. Guided by this observation, we introduce VidGround as a simple yet effective solution: using only the actual visually grounded questions without any linguistic biases for post-training. When used in tandem with RL-based post-training algorithms, this simple technique improves performance by up to 6.2 points relative to using the full dataset, while using only 69.1% of the original post-training data. Moreover, we show that data curation with a simple post-training algorithm outperforms several more complex post-training techniques, highlighting that data quality is a major bottleneck for improving video understanding in VLMs. These results underscore the importance of curating post-training data and evaluation benchmarks that truly require visual grounding to advance the development of more capable VLMs. Project page: http://vidground.etuagi.com.