Video Understanding Reward Modeling: A Robust Benchmark and Performant Reward Models

Name: Video Understanding Reward Modeling: A Robust Benchmark and Performant Reward Models
Uploaded: 2026-10-01T23:36:57.040573+00:00
Duration: 314 s
Description: Video understanding reward modeling is limited by a lack of good benchmarks and preference data. This work introduces VURB, a new benchmark for video preference judgment, and VUP-35K, a large dataset built automatically. It trains VideoDRM and VideoGRM reward models that achieve state-of-the-art results on video tasks.

Video understanding reward modeling is limited by a lack of good benchmarks and preference data. This work introduces VURB, a new benchmark for video preference judgment, and VUP-35K, a large dataset built automatically. It trains VideoDRM and VideoGRM reward models that achieve state-of-the-art results on video tasks.

Open the full episode