LesenForschungRadarAnlageframework
Anmelden / Registrieren
← Zurück zur Lektüre
NewsKurzfassung1 Min. LesezeitDisconfirmAI Redaktion

Influence-weighted unlearning claims 77x speedup over LoRA retraining

Influence-weighted unlearning could sharply cut the cost of correcting model behavior if it holds up, but the numbers are the authors' own and await third-party reproduction.

2026-09-24 12:00 GMT+8

Influence-weighted LoRA unlearning: the authors report that on Llama-3-8B, removing targeted trigger behaviors runs 77x faster in wall-clock time than a clean-corpus LoRA retraining reference. Version 4 of RapidUn was updated September 24.

Machine unlearning means removing a specific learned behavior without retraining the whole model. Existing methods include Fisher, GA, and LoReUn; the new method converts cross-sample influence estimates into fixed sample weights, and the authors report lower trigger attack success rates than those three with clean utility largely intact, plus cross-model validation on Mistral-7B.

The baselines are comparable unlearning methods, not full retraining, and the benchmarks were built and run by the authors themselves. Code is released, but no third-party reproduction exists yet.

Quellen

  • arxiv.org↗
Recherche fortsetzen →
Anmelden / Registrieren
LesenForschungRadarAnlageframework