Arithmetic Reasoning on Countdown
92Pass@1Countdown (A) + GRPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Countdown (A) + GRPOBackbone=Qwen2.5-7B, Generation budget=4096 tokens, Training Data=LLaMA-70B teacher answers (A), Optimization=GRPO2026.04 | 92 | 98 | |
| GRPO (Countdown)Backbone=Qwen2.5-7B, Generation budget=4096 tokens, Optimization=GRPO2026.04 | 91 | 98 | |
| Countdown (A') + GRPOBackbone=Qwen2.5-7B, Generation budget=4096 tokens, Training Data=Self-revised answers (A'), Optimization=GRPO2026.04 | 91 | 98 | |
| BaseBackbone=Qwen2.5-7B, Generation budget=4096 tokens2026.04 | 41 | 90 | |
| Countdown (A')Backbone=Qwen2.5-7B, Generation budget=4096 tokens, Training Data=Self-revised answers (A')2026.04 | 40 | 88 | |
| Countdown (A)Backbone=Qwen2.5-7B, Generation budget=4096 tokens, Training Data=LLaMA-70B teacher answers (A)2026.04 | 39 | 91 |