VBVR-Pro Verifiable Reward Scorers
Every benchmark task carries its own verifiable reward scorer; fixed dimensions, fixed weights, published tolerances. Pick a case, pick a video, and watch the scorer run.
Choose a benchmark case
Choose a video to score
Every video below is the model’s real output for this exact case, already on this server. Or upload your own — it is downscaled to 256×256 before scoring and deleted immediately after, and the scorers resize to match ground truth anyway.
Run the scorer
This is the score the task’s scorer gives this one case. The leaderboard figure for a model is the mean over all 100 tasks and every sample, so a single case will not equal it.
Score Explanation
See how the task-specific score changes as the video progresses.