Reasoning on ARC Challenge (Pass@1, Pass@32)
87.2Pass@1LEPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| LEPOBackbone=Qwen2.5-7B2026.04 | 87.2 | 98.89 | |
| GRPOBackbone=Qwen2.5-7B, Latent Inference=false2026.04 | 87.08 | 98.98 | |
| GRPO (w/ latent inference)Backbone=Qwen2.5-7B, Latent Inference=true2026.04 | 87.07 | 98.72 | |
| Soft TokensBackbone=Qwen2.5-7B2026.04 | 86.67 | 95.48 | |
| HRPOBackbone=Qwen2.5-7B2026.04 | 86.12 | 94.71 | |
| CoTBackbone=Qwen2.5-7B, Latent Inference=false2026.04 | 65.34 | 99.91 | |
| CoT (w/ latent inference)Backbone=Qwen2.5-7B, Latent Inference=true2026.04 | 64.51 | 99.91 |