Reasoning on ARC Challenge (pass@1)
81Pass@1 ScoreDSR
Evaluation Results
| Method | Links | |
|---|---|---|
| DSRMidtraining Schedule=DSR2026.05 | 81 | |
| WSDMidtraining Schedule=WSD2026.05 | 80.5 | |
| Cosine-decayMidtraining Schedule=Cos2026.05 | 78.5 | |
| Unified 15B MoEMidtraining Corpus=1k→8k2026.05 | 77.5 | |
| Unified 15B MoEMidtraining Corpus=4k→8k2026.05 | 76 | |
| Unified 15B MoEMidtraining Corpus=4k2026.05 | 74.5 | |
| Unified 15B MoEPretraining Corpus=4k→8k2026.05 | 69.5 | |
| Unified 15B MoEPretraining Corpus=4k2026.05 | 67.5 | |
| Unified 15B MoEPretraining Corpus=1k→8k2026.05 | 67 | |
| FSP-RevLMScale=8B2025.10 | 55.9 | |
| MTPScale=8B2025.10 | 55.2 | |
| NTPScale=8B2025.10 | 53.1 | |
| DS-MTPScale=8B2025.10 | 42.6 |