AdaPop: Adaptive Popularity for LLM Unlearning Reduces Leakage
Anna Borisiuk · hf · 2026-08-20
Introduces AdaPop (Adaptive Popularity), a method to improve LLM machine unlearning. It adapts gradient pressure based on factor popularity and automates the balance between forgetting and retaining content, aiming to reduce the leakage of unlearned data.
More from Research
- Cohere Research: From Multilingual to Multicultural AI — Cohere_Labs · 2026-08-20
- Cohere to Host Lecture on Explaining RL with Shapley Values — Cohere_Labs · 2026-08-20
- Liam Fedus on Scaling Laws: FLOPs are Intelligence, Parameters are Knowledge — hsu_byron · 2026-08-20
- ISMIR 2026 Paper: Simple Latent Transport Matches Large Models for Bandwidth Extension — umpedronosapato · 2026-08-20
- Mapping Uncertainty to Information Theory: Expert Analysis on 3 Types — abeirami · 2026-08-20
- CoinVE-Edit: apply up to 5 video edit instructions in a single forward pass — pmttyji · 2026-08-20