Mathematical Reasoning on MATH500 (Accuracy @avg1)
81.6Accuracy (@avg1)LESS
Evaluation Results
| Method | Links | |
|---|---|---|
| LESSBackbone=Qwen2.5-Math-7B2025.11 | 81.6 | |
| GRPOBackbone=Qwen2.5-Math-7B2025.11 | 79.8 | |
| KL-CovBackbone=Qwen2.5-Math-7B2025.11 | 78.6 | |
| LESSBackbone=Qwen2.5-Base-7B2025.11 | 78.6 | |
| Forking TokensBackbone=Qwen2.5-Math-7B2025.11 | 78.2 | |
| Forking TokensBackbone=Qwen2.5-Base-7B2025.11 | 77.8 | |
| KL-CovBackbone=Qwen2.5-Base-7B2025.11 | 77 | |
| GRPOBackbone=Qwen2.5-Base-7B2025.11 | 76.8 | |
| KL-CovBackbone=Qwen2.5-Math-1.5B2025.11 | 75.8 | |
| Forking TokensBackbone=Qwen2.5-Math-1.5B2025.11 | 75.6 | |
| LESSBackbone=Qwen2.5-Math-1.5B2025.11 | 75.2 | |
| GRPOBackbone=Qwen2.5-Math-1.5B2025.11 | 74.8 | |
| Base LLMBackbone=Qwen2.5-Math-1.5B2025.11 | 58.6 | |
| Base LLMBackbone=Qwen2.5-Math-7B2025.11 | 58.6 | |
| Base LLMBackbone=Qwen2.5-Base-7B2025.11 | 53.4 |