Mathematical Reasoning on GPQA (Avg@16)
50Avg@16GRPO (w. PS)
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPO (w. PS)Backbone=Qwen3-8B2026.01 | 50 | |
| HEAPABackbone=Qwen3-8B2026.01 | 49.3 | |
| GRPOBackbone=Qwen3-8B2026.01 | 45.6 | |
| HEAPABackbone=Qwen3-4B2026.01 | 44.3 | |
| GRPO (w. PS)Backbone=Qwen3-4B2026.01 | 43.3 | |
| GRPOBackbone=Qwen3-4B2026.01 | 41.8 | |
| HEAPABackbone=Qwen3-1.7B2026.01 | 33.3 | |
| GRPO (w. PS)Backbone=Qwen3-1.7B2026.01 | 32.9 | |
| GRPOBackbone=Qwen3-1.7B2026.01 | 32.3 | |
| GRPO (w. PS)Backbone=Qwen3-0.6B2026.01 | 31.8 | |
| HEAPABackbone=Qwen3-0.6B2026.01 | 31.6 | |
| GRPOBackbone=Qwen3-0.6B2026.01 | 31.1 |