Zero-shot Question Answering on (OpenBookQA, PIQA, RACE, SciQ, WinoGrande)
54.6AccuracyHyperloop
Evaluation results
| Method | Links | |
|---|---|---|
| HyperloopDim=2048, Unrolled Depth=38, Train Tokens=100B, Params=990.8 M, Loop Architecture=[4L → 10L (×3) → 4L]2026.04 | 54.6 | |
| mHCDim=2048, Unrolled Depth=38, Train Tokens=100B, Params=2033.0 M2026.04 | 53.7 | |
| LoopedDim=2048, Unrolled Depth=38, Train Tokens=100B, Params=990.5 M, Loop Architecture=[4L → 10L (×3) → 4L]2026.04 | 53.3 | |
| TransformerDim=2048, Unrolled Depth=38, Train Tokens=100B, Params=2018.0 M2026.04 | 52.8 | |
| HyperloopDim=2048, Unrolled Depth=18, Train Tokens=50B, Params=579.7 M, Loop Architecture=[3L → 4L (×3) → 3L]2026.04 | 49.8 | |
| LoopedDim=2048, Unrolled Depth=18, Train Tokens=50B, Params=579.4 M, Loop Architecture=[3L → 4L (×3) → 3L]2026.04 | 49.2 | |
| mHCDim=2048, Unrolled Depth=18, Train Tokens=50B, Params=997.5 M2026.04 | 48.6 | |
| TransformerDim=2048, Unrolled Depth=18, Train Tokens=50B, Params=990.5 M2026.04 | 48 | |
| HyperloopDim=1024, Unrolled Depth=16, Train Tokens=12.5B, Params=135.7 M, Loop Architecture=[2L → 4L (×3) → 2L]2026.04 | 41.6 | |
| LoopedDim=1024, Unrolled Depth=16, Train Tokens=12.5B, Params=135.5 M, Loop Architecture=[2L → 4L (×3) → 2L]2026.04 | 41.4 | |
| TransformerDim=1024, Unrolled Depth=16, Train Tokens=12.5B, Params=238.0 M2026.04 | 41.1 | |
| mHCDim=1024, Unrolled Depth=16, Train Tokens=12.5B, Params=241.0 M2026.04 | 41.1 |