| Year | Work | Command |
|---|---|---|
| 2024 | Arditi 2406.11717 | estimate --mode dim then apply --mode arditi |
| 2024 | Labonne blog | same |
| 2025 | Lai projected | estimate --mode projected |
| 2025 | Lai / ORBA 2026 | apply --mode orba-directional (stable) or orba-householder (A/B) |
| 2025 | COSMIC 2506.00085 | estimate --mode cosmic (score + DIM r; not the full paper ID) |
| 2025 | TUM cones 2502.17420 | GPU RDO (not this SVD toy) |
| 2025 | False-refusal 2410.03415 | no flag. Two estimate --mode dim (v harmful/harmless, w factory-false/comply), form w′ ← w − λ v offline, then apply. estimate --mode dim is not Wang. Raw factory DIM / T03 ≠ T38 |
| 2025 | Zhao 2507.11878 | two DIMs; do not wipe harm and refuse |
| 2025 | DeepRefusal 2509.15202 | defense; not a bake / not this bin |
| 2026 | SOM 2511.08379 | svd here is a subspace proxy; official SOM trainer = pralab/som-refusal-directions (not Abliterix) |
| 2026 | OT 2603.04355 | Nanfack PCA-OT; reported GPU estimator; not this bin |
| 2026 | RFM-AGOP 2607.02396 | Winninger RFM; reported GPU estimator; not this bin |
| 2026 | QCRI 2602.02132 | leftovers = style; second pass / svd |
| 2026 | Young 2512.13655 | after this lab, pick Heretic vs ErisForge |
| 2026 | Petrov 2603.22061 | do not topic-match contrast sets |
| 2026 | Task over-refuse 2603.27518 | factory DIM ≠ global harm DIM |
apply --mode projected bakes the same as arditi — the projection happens at estimate time.