Semantic Probing on MSCOCO (test)
45.31Top-5 AccuracyCOMiT-L
Evaluation Results
| Method | Links | |
|---|---|---|
| COMiT-L#params=610M, msg length=256, voc. size=64k2026.02 | 45.31 | |
| COMiT-B#params=174M, msg length=256, voc. size=64k2026.02 | 41.46 | |
| COMiT-XL#params=900M, msg length=256, voc. size=64k2026.02 | 39.45 | |
| FlexTok(d18-d28)#params=2.5B, msg length=256, voc. size=64k2026.02 | 39.14 | |
| FlexTok(d18-d18)#params=860M, msg length=256, voc. size=64k2026.02 | 38.58 | |
| FlexTok(d12-d12)#params=254M, msg length=256, voc. size=64k2026.02 | 38.17 | |
| SelfTok#params=2.17B, msg length=512, voc. size=2^15, trained on larger data mixtures=true2026.02 | 17.91 | |
| SelfTok#params=2.17B, msg length=1024, voc. size=2^16, trained on larger data mixtures=true2026.02 | 17.32 | |
| TiTok-B#params=172M, msg length=64, voc. size=2^122026.02 | 12.64 | |
| ALIT#params=229M, msg length=256, voc. size=2^102026.02 | 8.56 | |
| TiTok-S#params=44M, msg length=128, voc. size=2^122026.02 | 6.3 | |
| TiTok-L#params=614M, msg length=32, voc. size=2^122026.02 | 6.22 |