Mathematical Reasoning on GSM8K (LIS, Accuracy)
4.8182LISDeepSeek-MoE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DeepSeek-MoEStrategy=Raw2026.02 | 4.8182 | 15.6 | |
| DeepSeek V2 LiteStrategy=Raw2026.02 | 4.7886 | 38.4 | |
| DeepSeek-MoEStrategy=R&Q2026.02 | 3.9725 | 15.4 | |
| DeepSeek V2 LiteStrategy=R&Q2026.02 | 3.5998 | 37.2 | |
| Switch Transformer (16 experts)Strategy=Raw2026.02 | 2.6937 | — | |
| Switch Transformer (8 experts)Strategy=Raw2026.02 | 1.9709 | — | |
| Switch Transformer (16 experts)Strategy=R&Q2026.02 | 1.9582 | — | |
| Switch Transformer (8 experts)Strategy=R&Q2026.02 | 1.3937 | — | |
| LLaMa-MoEStrategy=Raw2026.02 | 1.3565 | 3.5 | |
| LLaMa-MoEStrategy=R&Q2026.02 | 1.1963 | 3.6 |