Zero-shot Reasoning on ARC-C, ARC-E, HellaSwag, LAMBADA, OpenBookQA, PIQA, and WinoGrande
36.9ARC-C AccuracyH-Net 1.3B (ours)
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| H-Net 1.3B (ours)INPUT=Byte2026.05 | 36.9 | 60.5 | 54.9 | 44.4 | 37.4 | 71.5 | 56.2 | 51.7 | |
| H-Net 1.3B (base)INPUT=Byte2026.05 | 34.4 | 58.7 | 55 | 43.4 | 38.4 | 70.7 | 56.8 | 51.1 | |
| H-Net 680M (ours)INPUT=Byte2026.05 | 33.9 | 61.2 | 52.4 | 42.1 | 37.2 | 71.7 | 56.1 | 50.7 | |
| H-Net 680M (base)INPUT=Byte2026.05 | 33.4 | 60.1 | 52.3 | 40.9 | 39.2 | 70.9 | 55.7 | 50.4 | |
| Llama3.2 1.5BINPUT=Token2026.05 | 33.1 | 60.3 | 49 | 37.3 | 34.8 | 70.2 | 54.7 | 48.5 | |
| Llama3.2 770MINPUT=Token2026.05 | 28.1 | 53.7 | 38.3 | 26.7 | 32.4 | 64.5 | 51 | 42.1 |