Benchmark RadarAI BENCHMARK PROFILE
VA-Judger-Bench
VA-Judger-Bench evaluates reward models for joint video-audio generation. It contains paired comparisons of generated video-audio samples with human preference labels, covering in-domain and out-of-domain model outputs.
- Released
- 2026-08-19
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Existing metrics evaluate quality dimensions separately, missing semantic and temporal coherence. VA-Judger-Bench provides a benchmark to assess whether reward models align with human preferences for joint video-audio generation.
Motivation
Using reinforcement learning to post-train joint video-audio generation models requires a reward signal.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.