Reasoning on GPQA Diamond (Pass@1, Pass@32)
29.96Pass@1 (Reasoning)LEPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LEPOBackbone=Qwen2.5-7B2026.04 | 29.96 | 90.91 | |
| Soft TokensBackbone=Qwen2.5-7B2026.04 | 28.69 | 81.31 | |
| GRPOBackbone=Qwen2.5-7B, Latent Inference=false2026.04 | 28.28 | 87.88 | |
| GRPO (w/ latent inference)Backbone=Qwen2.5-7B, Latent Inference=true2026.04 | 28.09 | 88.89 | |
| HRPOBackbone=Qwen2.5-7B2026.04 | 27.84 | 78.79 | |
| FOCUSFTn=322026.05 | 19.4 | 80.8 | |
| Standard SFTn=322026.05 | 15.6 | 78.8 | |
| CoTBackbone=Qwen2.5-7B, Latent Inference=false2026.04 | 10.92 | 82.83 | |
| CoT (w/ latent inference)Backbone=Qwen2.5-7B, Latent Inference=true2026.04 | 10.65 | 83.33 |