Reasoning on MMLU STEM
79.05Pass@1 (MMLU STEM Reasoning)LEPO
Evaluation results
| Method | Links | ||
|---|---|---|---|
| LEPOBackbone=Qwen2.5-7B2026.04 | 79.05 | 97.4 | |
| GRPOBackbone=Qwen2.5-7B, Latent Inference=false2026.04 | 78.75 | 97.18 | |
| GRPO (w/ latent inference)Backbone=Qwen2.5-7B, Latent Inference=true2026.04 | 78.71 | 97.37 | |
| Soft TokensBackbone=Qwen2.5-7B2026.04 | 76.63 | 93.24 | |
| HRPOBackbone=Qwen2.5-7B2026.04 | 76.4 | 92.77 | |
| CoTBackbone=Qwen2.5-7B, Latent Inference=false2026.04 | 58.17 | 99.9 | |
| CoT (w/ latent inference)Backbone=Qwen2.5-7B, Latent Inference=true2026.04 | 57.72 | 99.84 |