Natural Language Understanding on LM Evaluation Harness
57.2WG ScoreHybrid KDA
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Hybrid KDAArchitecture=Hybrid KDA, Evaluation Protocol=Perplexity-based2026.03 | 57.2 | 45.2 | 52.5 | 31.7 | 66.3 | 38 | 40.6 | 40 | |
| Teacher (Qwen3-0.6B)Architecture=Transformer, Evaluation Protocol=Perplexity-based2026.03 | 55.5 | 47.3 | 55.6 | 34.2 | 65.9 | 24.7 | 40.2 | — | |
| Pure KDAArchitecture=KDA, Evaluation Protocol=Perplexity-based2026.03 | 55.1 | 42.9 | 49.5 | 29.4 | 64.3 | 53.7 | 35.9 | 33.3 | |
| Hybrid MambaArchitecture=Hybrid Mamba, Evaluation Protocol=Perplexity-based2026.03 | 54 | 42.9 | 53.6 | 30.5 | 65.7 | 73 | 34.7 | 38.8 | |
| Pure MambaArchitecture=Mamba, Evaluation Protocol=Perplexity-based2026.03 | 53.8 | 39.2 | 39.8 | 26.4 | 63.3 | 135 | 25 | 26.4 |