General Capability on GPQA Diamond
52AccuracyIn-context distillation
Evaluation Results
| Method | Links | |
|---|---|---|
| In-context distillationTrained to=—2026.07 | 52 | |
| SFT on positive documentsTrained to=—2026.07 | 49 | |
| SFT on negated documentsTrained to=—2026.07 | 48 | |
| single-basis Goggles (b = 1)Trained to=7002026.07 | 47 | |
| GogglesTrained to=7002026.07 | 46 | |
| Goggles (Redwood framing)Trained to=7002026.07 | 46 | |
| forward KLTrained to=7002026.07 | 45 | |
| spectral clip (τ = 3.35)Trained to=7002026.07 | 45 | |
| base model (no SFT)Trained to=—2026.07 | 43 | |
| PODSModel=Qwen3-8B, Selection Ratio=50%2026.05 | 37.4 | |
| Full DataModel=Qwen3-8B, Selection Ratio=100%2026.05 | 36.4 | |
| PODSModel=Qwen3-4B, Selection Ratio=50%2026.05 | 35.9 | |
| Full DataModel=Qwen3-4B, Selection Ratio=100%2026.05 | 35.8 | |
| Lmax = 1 (no curriculum)Trained to=7002026.07 | 4 |