| Hyperparameters | Retention | In-domain task |
| α | LR ( ×10−4 ) | Min. LR ( ×10−5 ) | β1 | Warm-up | GPQA ( ↑ ) | IFEval ( ↑ ) | MMLU-Pro ( ↑ ) | Acc. ( ↑ ) | Macro-F1 ( ↑ ) | Acc. argmax ( ↑ ) | NLL ( ↓ ) | Brier ( ↓ ) |
| Baseline (Frozen Model) | 0.527±0.044 | 0.843±0.011 | 0.622±0.000 | 0.773±0.002 | 0.645±0.012 | 0.526±0.000 | 5.701±0.002 | 0.860±0.000 |
| 8 | 0.5 | 0.25 | 0.8 | 0 | 0.534±0.021 | 0.843±0.009 | 0.625±0.002 | 0.806±0.004 | 0.690±0.006 | 0.825±0.002 | 0.582±0.010 | 0.267±0.002 |
| 8 | 0.5 | 1 | 0.8 | 0 | 0.539±0.024 | 0.845±0.005 | 0.625±0.002 | 0.806±0.003 | 0.698±0.010 | 0.833±0.002 | 0.545±0.011 | 0.254±0.002 |
| 8 | 1 | 1 | 0.8 | 200 | 0.537±0.032 | 0.845±0.004 | 0.628±0.004 | 0.807±0.003 | 0.703±0.007 | 0.841±0.001 | 0.562±0.010 | 0.245±0.004 |
| 16 | 0.5 | 0.25 | 0.9 | 400 | 0.545±0.013 | 0.847±0.006 | 0.627±0.001 | 0.807±0.004 | 0.691±0.006 | 0.844±0.003 | 0.476±0.011 | 0.234±0.004 |