Scientific Reasoning on GPQA (Subject Scores)
35.7Biology ScoreEasyRL Iter3
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| EasyRL Iter3Backbone=Qwen2.5-Math-1.5B2026.04 | 35.7 | 17.7 | 12.9 | 19.4 | |
| EasyRL Iter3Backbone=Llama-3.2-3B-Instruct2026.04 | 35.7 | 23.8 | 22.1 | 27.2 | |
| Supervised GRPOBackbone=Llama-3.2-3B-Instruct2026.04 | 34.3 | 16.2 | 19.3 | 23.3 | |
| EasyRL Iter2Backbone=Llama-3.2-3B-Instruct2026.04 | 34.3 | 21.5 | 19.3 | 25 | |
| EasyRL Iter1Backbone=Llama-3.2-3B-Instruct2026.04 | 32.9 | 16.9 | 23.6 | 24.5 | |
| EasyRL Iter3Backbone=Qwen2.5-Math-7B2026.04 | 31.4 | 24.6 | 35.7 | 30.6 | |
| Unsupervised EMPOBackbone=Llama-3.2-3B-Instruct2026.04 | 30 | 17.7 | 13.6 | 18.5 | |
| Vanilla InstructBackbone=Qwen2.5-Math-7B2026.04 | 28.6 | 23.8 | 31.4 | 27.9 | |
| EasyRL Iter1Backbone=Qwen2.5-Math-7B2026.04 | 28.6 | 23.1 | 27.9 | 26.2 | |
| EasyRL Iter2Backbone=Qwen2.5-Math-7B2026.04 | 28.6 | 20.8 | 37.1 | 29.1 | |
| EasyRL Iter1Backbone=Qwen2.5-Math-1.5B2026.04 | 27.1 | 14.6 | 10.7 | 15.6 | |
| Vanilla BaseBackbone=Llama-3.2-3B-Instruct2026.04 | 27.1 | 17.7 | 15 | 18.5 | |
| Unsupervised EMPOBackbone=Qwen2.5-Math-7B2026.04 | 25.7 | 23.1 | 30 | 26.5 | |
| EasyRL Iter2Backbone=Qwen2.5-Math-1.5B2026.04 | 24.3 | 20.8 | 11.4 | 17.6 | |
| Supervised GRPOBackbone=Qwen2.5-Math-7B2026.04 | 24.3 | 23.1 | 32.9 | 27.4 | |
| Unsupervised EMPOBackbone=Qwen2.5-Math-1.5B2026.04 | 20 | 16.9 | 12.1 | 15.6 | |
| Vanilla BaseBackbone=Qwen2.5-Math-7B2026.04 | 18.6 | 24.6 | 26.4 | 24.1 | |
| Supervised GRPOBackbone=Qwen2.5-Math-1.5B2026.04 | 11.4 | 10 | 4.3 | 7.9 | |
| Vanilla BaseBackbone=Qwen2.5-Math-1.5B2026.04 | 0 | 2.3 | 1.4 | 1.5 |