Phi-4-reasoning-vision-15B: benchmark results
Provider: Microsoft. Released 2026-03-04. Access: Open.
Aggregate Rating 1488 ± 45, rank #2488 of 10516 rated models, from 31 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| UltraVR - Remote Sensing | 41.7 | Accuracy (%) on the remote sensing domain (comparison, logic | 80 |
| UltraVR | 31.1 | Domain-balanced macro accuracy (%): mean of the CCTV, remote | 60 |
| UltraVR - CCTV | 39.5 | Accuracy (%) on the CCTV surveillance domain (comparison, lo | 60 |
| HarmVideoBench - Beyond-Clip Reasoning | 61.2 | Beyond-Clip Reasoning accuracy (%) on four-option multiple-c | 50 |
| SciFigPlag-Bench - Reuse Localization | 3.3 | Paired localization accuracy (%; both the source-side and th | 46.7 |
| LongShOTBench - Short Videos | 23.35 | Weighted Rubric Score (%) | 44.6 |
| HarmVideoBench | 78.4 | Macro Avg. (self-reported) | 41.2 |
| HarmVideoBench - Clip-Internal Meaning | 80.9 | Clip-Internal Meaning accuracy (%) on four-option multiple-c | 41.2 |
| HarmVideoBench - Observable Evidence | 93 | Observable Evidence accuracy (%) on four-option multiple-cho | 41.2 |
| PhysCount - Appearance NAE | 30.31 | Appearance normalized absolute error (%) | 35.9 |
| LongShOTBench - Audio | 21.13 | Weighted Rubric Score (%) | 35.6 |
| LongShOTBench - Long Videos | 22.6 | Weighted Rubric Score (%) | 35.6 |
Interactive version: theaggregate.ai/model?slug=phi-4-reasoning-vision-15b · How It Works · Data refreshed daily, snapshot 2026-11-21.