Mathematical Reasoning on AMC (pass@1, pass@32)
82.7Pass@1FG-ExPO
Evaluation results
| Method | Links | ||
|---|---|---|---|
| FG-ExPOBackbone=Qwen3-8B-Base2026.05 | 82.7 | 96.34 | |
| GRPOBackbone=Qwen3-8B-Base2026.05 | 82.47 | 95.12 | |
| TTRL (w/ MPS)Backbone=Qwen3-4B2026.05 | 73.6 | — | |
| TTRL-GuardBackbone=Qwen3-4B2026.05 | 73.6 | — | |
| TTRL (w/ RCSU)Backbone=Qwen3-4B2026.05 | 73.2 | — | |
| SCOPEBackbone=Qwen3-4B2026.05 | 73 | — | |
| TTRL (w/ FRS)Backbone=Qwen3-4B2026.05 | 72.2 | — | |
| TTRLBackbone=Qwen3-4B2026.05 | 71.4 | — | |
| CoVerRLBackbone=Qwen3-4B2026.05 | 69.5 | — | |
| FG-ExPOBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.05 | 68.75 | 91.46 | |
| GRPOBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.05 | 68.45 | 92.68 | |
| CoVerRLBackbone=Qwen2.5-7B-Instruct2026.05 | 55 | — | |
| TTRL-GuardBackbone=Qwen2.5-7B-Instruct2026.05 | 53.8 | — | |
| TTRL (w/ MPS)Backbone=Qwen2.5-7B-Instruct2026.05 | 53.5 | — | |
| TTRL (w/ FRS)Backbone=Qwen2.5-7B-Instruct2026.05 | 52.6 | — | |
| TTRLBackbone=Qwen2.5-7B-Instruct2026.05 | 52.4 | — | |
| BaseBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.05 | 52.02 | 85.37 | |
| TTRL (w/ RCSU)Backbone=Qwen2.5-7B-Instruct2026.05 | 51.2 | — | |
| SCOPEBackbone=Qwen2.5-7B-Instruct2026.05 | 51.1 | — | |
| Base ModelBackbone=Qwen2.5-7B-Instruct2026.05 | 43.6 | — | |
| SCOPEBackbone=Llama-3.2-3B-Instruct2026.05 | 38.7 | — | |
| TTRL-GuardBackbone=Llama-3.2-3B-Instruct2026.05 | 36.8 | — | |
| CoVerRLBackbone=Llama-3.2-3B-Instruct2026.05 | 34.8 | — | |
| TTRL (w/ RCSU)Backbone=Llama-3.2-3B-Instruct2026.05 | 34.8 | — | |
| TTRL (w/ FRS)Backbone=Llama-3.2-3B-Instruct2026.05 | 34 | — | |
| TTRL (w/ MPS)Backbone=Llama-3.2-3B-Instruct2026.05 | 33.3 | — | |
| TTRLBackbone=Llama-3.2-3B-Instruct2026.05 | 31.3 | — | |
| Base ModelBackbone=Llama-3.2-3B-Instruct2026.05 | 20.6 | — | |
| Base ModelBackbone=Qwen3-4B2026.05 | 19.3 | — |