Scientific Reasoning on GPQA (avg@16)
44.1avg@16Qwen2.5-32B-Instruct
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen2.5-32B-InstructBackbone=Qwen2.5-32B-Instruct2026.03 | 44.1 | |
| PowerFlowBackbone=Qwen2.5-Math-7B2026.03 | 37 | |
| PowerSamplingBackbone=Qwen2.5-Math-7B2026.03 | 36.3 | |
| EMPOBackbone=Qwen2.5-Math-7B2026.03 | 36 | |
| InstructBackbone=Qwen2.5-Math-7B2026.03 | 35 | |
| TTRLBackbone=Qwen2.5-Math-7B2026.03 | 34.7 | |
| GRPOBackbone=Qwen2.5-Math-7B2026.03 | 34.4 | |
| Low-tempBackbone=Qwen2.5-Math-7B2026.03 | 34.1 | |
| One-shot EMBackbone=Qwen2.5-Math-7B2026.03 | 32.8 | |
| IntuitorBackbone=Llama-3.2-3B-Instruct2026.03 | 30.5 | |
| GRPOBackbone=Llama-3.2-3B-Instruct2026.03 | 30.5 | |
| PowerFlowBackbone=Llama-3.2-3B-Instruct2026.03 | 30.2 | |
| BaseBackbone=Qwen2.5-Math-7B2026.03 | 29.7 | |
| EMPOBackbone=Qwen2.5-Math-1.5B2026.03 | 29.5 | |
| BaseBackbone=Llama-3.2-3B-Instruct2026.03 | 29.5 | |
| Low-tempBackbone=Llama-3.2-3B-Instruct2026.03 | 28.8 | |
| Low-tempBackbone=Qwen2.5-Math-1.5B2026.03 | 28.7 | |
| PowerFlowBackbone=Qwen2.5-Math-1.5B2026.03 | 28.3 | |
| InstructBackbone=Qwen2.5-1.5B2026.03 | 28 | |
| PowerFlowBackbone=Qwen2.5-1.5B2026.03 | 27.7 | |
| GRPOBackbone=Qwen2.5-Math-1.5B2026.03 | 26.8 | |
| Low-tempBackbone=Qwen2.5-1.5B2026.03 | 26.6 | |
| IntuitorBackbone=Qwen2.5-1.5B2026.03 | 26.4 | |
| InstructBackbone=Qwen2.5-Math-1.5B2026.03 | 26.4 | |
| BaseBackbone=Qwen2.5-Math-1.5B2026.03 | 26.1 | |
| Format-onlyBackbone=Qwen2.5-Math-1.5B2026.03 | 26.1 | |
| GRPOBackbone=Qwen2.5-1.5B2026.03 | 26 | |
| BaseBackbone=Qwen2.5-1.5B2026.03 | 25.8 |