Multi-step tensor program optimization on Step-TP (100 held-out tensor programs)
103.17Avg SamplesBeam Search
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Beam SearchModel=Qwen3-32B trained on Step-TP, Hardware=H20-3e2026.05 | 103.17 | 114 | 42.9 | 4.57 | 561.82 | 0.6 | |
| Greedy SearchModel=Qwen3-32B trained on Step-TP, Hardware=H20-3e2026.05 | 35.91 | 41 | 20.79 | 1.78 | 193.55 | 0.58 | |
| BFS SearchModel=Qwen3-32B trained on Step-TP, Hardware=H20-3e2026.05 | 28.27 | 31 | 18.27 | 2.11 | 173.13 | 0.65 | |
| DFS SearchModel=Qwen3-32B trained on Step-TP, Hardware=H20-3e2026.05 | 28.11 | 31 | 20.97 | 1.68 | 242.62 | 0.75 | |
| MCTSModel=Qwen3-32B trained on Step-TP, Hardware=H20-3e2026.05 | 17.41 | 21 | 24.96 | 4.54 | 171.9 | 1.43 | |
| Chain SearchModel=Qwen3-32B trained on Step-TP, Hardware=A1002026.05 | 16.85 | 21 | 22.01 | 3.62 | 265.35 | 1.31 | |
| Chain SearchModel=Qwen3-32B trained on Step-TP, Hardware=H20-3e2026.05 | 16.63 | 21 | 23.62 | 2.05 | 286.92 | 1.42 | |
| Chain SearchModel=Qwen3-32B trained on Step-TP, Hardware=H20-3e, Parent Nodes=Without2026.05 | 15.95 | 21 | 20.25 | 4.96 | 96.14 | 1.63 |