Visual Reasoning on NLVR2 (dev)
82.5AccuracyMADTP
Evaluation Results
| Method | Links | |
|---|---|---|
| MADTPReduce Ratio=0.3, GFLOPS=92.60, Backbone=BLIP2024.03 | 82.5 | |
| UncompressedReduce Ratio=/, GFLOPS=132.54, Backbone=BLIP2024.03 | 82.48 | |
| MADTPReduce Ratio=0.5, GFLOPS=66.16, Backbone=BLIP2024.03 | 81.97 | |
| MADTPReduce Ratio=0.6, GFLOPS=52.92, Backbone=BLIP2024.03 | 81.92 | |
| MADTPReduce Ratio=0.7, GFLOPS=39.69, Backbone=BLIP2024.03 | 80.67 | |
| UPopReduce Ratio=0.3, GFLOPS=89.36, Backbone=BLIP2024.03 | 80.33 | |
| STPReduce Ratio=0.3, GFLOPS=94.08, Backbone=BLIP2024.03 | 79.5 | |
| CoMPGFLOPs=26.33±0.382026.04 | 79.13 | |
| MADTPReduce Ratio=0.8, GFLOPS=26.46, Backbone=BLIP2024.03 | 78.28 | |
| STPReduce Ratio=0.5, GFLOPS=68.31, Backbone=BLIP2024.03 | 78.08 | |
| MADTPGFLOPs=26.77±0.232026.04 | 77.16 | |
| MH-MoEExperts Number=82024.04 | 77 | |
| UPopReduce Ratio=0.5, GFLOPS=65.29, Backbone=BLIP2024.03 | 76.89 | |
| X-MoEExperts Number=82024.04 | 75.5 | |
| Dense2024.04 | 73.8 | |
| MiniVLMParams=52M, Training=100 Epochs2026.04 | 73.7 | |
| UPopReduce Ratio=0.6, GFLOPS=50.35, Backbone=BLIP2024.03 | 72.85 | |
| PixelBERTParams=110M, Training=40 Epochs2026.04 | 71.7 | |
| EfficientParams=39.2M, Training=20 Epochs2026.04 | 71.1 | |
| UPopReduce Ratio=0.7, GFLOPS=39.93, Backbone=BLIP2024.03 | 68.71 | |
| UPopReduce Ratio=0.8, GFLOPS=19.08, Backbone=BLIP2024.03 | 57.17 |