Mathematical Reasoning on AMC23 Minerva
43.1AMC23 ScoreNExt
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| NExtBackbone=Qwen2.5-3B-Instruct, #Steps=2502026.04 | 43.1 | 18.8 | 31 | 33.3 | |
| GRPOBackbone=Qwen2.5-3B-Instruct, #Steps=400, Fine-tuning Protocol=Full Parameter2026.04 | 42.5 | 18.8 | 30.7 | 55.6 | |
| RL-ExtraBackbone=Qwen2.5-3B-Instruct, #Steps=2502026.04 | 41.3 | 17.9 | 29.6 | 41 | |
| GRPOBackbone=Qwen2.5-3B-Instruct, #Steps=250, Fine-tuning Protocol=Full Parameter2026.04 | 40.6 | 18.3 | 29.5 | 41.7 | |
| GRPOBackbone=Qwen2.5-3B-Instruct, #Steps=400, Fine-tuning Protocol=LoRA2026.04 | 40 | 18.8 | 29.4 | 67.8 | |
| AlphaRLBackbone=Qwen2.5-3B-Instruct, #Steps=2502026.04 | 39.4 | 17.3 | 28.4 | 51 | |
| GRPOBackbone=Qwen2.5-3B-Instruct, #Steps=250, Fine-tuning Protocol=LoRA2026.04 | 36.9 | 17.6 | 27.3 | 65.8 | |
| GRPOBackbone=Qwen2.5-1.5B-Instruct, #Steps=400, Fine-tuning Protocol=Full Parameter2026.04 | 31.3 | 10.8 | 21.1 | 43.5 | |
| NExtBackbone=Qwen2.5-1.5B-Instruct, #Steps=2502026.04 | 31.3 | 11.8 | 21.6 | 25.8 | |
| Backbone ModelBackbone=Qwen2.5-3B-Instruct, #Steps=-2026.04 | 31.3 | 15.7 | 23.5 | — | |
| GRPOBackbone=Qwen2.5-1.5B-Instruct, #Steps=400, Fine-tuning Protocol=LoRA2026.04 | 30 | 11.5 | 20.8 | 44.9 | |
| GRPOBackbone=Qwen2.5-1.5B-Instruct, #Steps=250, Fine-tuning Protocol=LoRA2026.04 | 29.4 | 11.2 | 20.3 | 29.8 | |
| AlphaRLBackbone=Qwen2.5-1.5B-Instruct, #Steps=2502026.04 | 27.5 | 11.8 | 19.7 | 32.1 | |
| GRPOBackbone=Qwen2.5-1.5B-Instruct, #Steps=250, Fine-tuning Protocol=Full Parameter2026.04 | 26.7 | 11.1 | 18.9 | 35.7 | |
| RL-ExtraBackbone=Qwen2.5-1.5B-Instruct, #Steps=2502026.04 | 26.3 | 11.1 | 18.7 | 36.8 | |
| Backbone ModelBackbone=Qwen2.5-1.5B-Instruct, #Steps=-2026.04 | 16.3 | 7.4 | 11.9 | — |