Mathematical Reasoning on MINERVA (Pass@k)
38.4Pass@kGCPO
Evaluation Results
| Method | Links | |
|---|---|---|
| GCPOk=64, Backbone=Qwen3-4B2026.05 | 38.4 | |
| GCPOk=32, Backbone=Qwen3-4B2026.05 | 37.8 | |
| DQOk=64, Backbone=Qwen3-4B2026.05 | 37.7 | |
| GRPOk=64, Backbone=Qwen3-4B2026.05 | 37.6 | |
| Basek=64, Backbone=Qwen3-4B2026.05 | 37.4 | |
| DQOk=32, Backbone=Qwen3-4B2026.05 | 37.3 | |
| GRPOk=32, Backbone=Qwen3-4B2026.05 | 36.9 | |
| GCPOk=16, Backbone=Qwen3-4B2026.05 | 36.2 | |
| DQOk=16, Backbone=Qwen3-4B2026.05 | 35.7 | |
| Basek=32, Backbone=Qwen3-4B2026.05 | 35.7 | |
| GRPOk=16, Backbone=Qwen3-4B2026.05 | 35.3 | |
| Basek=16, Backbone=Qwen3-4B2026.05 | 35.1 | |
| GCPOk=8, Backbone=Qwen3-4B2026.05 | 34.9 | |
| DQOk=8, Backbone=Qwen3-4B2026.05 | 34.4 | |
| GRPOk=8, Backbone=Qwen3-4B2026.05 | 34.1 | |
| GCPOk=4, Backbone=Qwen3-4B2026.05 | 33.2 | |
| DQOk=4, Backbone=Qwen3-4B2026.05 | 32.9 | |
| Basek=8, Backbone=Qwen3-4B2026.05 | 32.9 | |
| GRPOk=4, Backbone=Qwen3-4B2026.05 | 32.6 | |
| GCPOk=2, Backbone=Qwen3-4B2026.05 | 31 | |
| DQOk=2, Backbone=Qwen3-4B2026.05 | 30.7 | |
| Basek=4, Backbone=Qwen3-4B2026.05 | 30.6 | |
| GRPOk=2, Backbone=Qwen3-4B2026.05 | 30.4 | |
| Basek=2, Backbone=Qwen3-4B2026.05 | 27.2 |