Knowledge Editing Resistance on Sheeran long-horizon SFT (test)
99Resistance ScoreGoggles (Redwood framing)
Evaluation Results
| Method | Links | |
|---|---|---|
| Goggles (Redwood framing)Trained to=7002026.07 | 99 | |
| spectral clip (τ = 3.35)Trained to=7002026.07 | 98 | |
| single-basis Goggles (b = 1)Trained to=7002026.07 | 98 | |
| In-context distillationTrained to=—2026.07 | 95 | |
| GogglesTrained to=7002026.07 | 94 | |
| forward KLTrained to=7002026.07 | 94 | |
| Lmax = 1 (no curriculum)Trained to=7002026.07 | 32 | |
| SFT on positive documentsTrained to=—2026.07 | 22 | |
| SFT on negated documentsTrained to=—2026.07 | 22 |