Llm unlearning method improves nearby benign reasoning without losing utility
RA$^2$: Retain-Anchored Attraction Preserves Forget-Adjacent Utility in LLM Unlearning
Machine Learning
Summary
Large language models sometimes need to 'forget' certain behaviors, but after forgetting, nearby reasoning by the model can still become worse. The authors found that not just forgetting but where the model directs the forgotten parts in its internal space matters for keeping good reasoning nearby. They came up with a method called RA² that moves these forgotten parts toward safe areas in the model’s memory, preserving helpful responses close to the forgotten topics. Their method performed better at keeping this nearby utility than other strong unlearning methods.
What this means in practice
- •For llm engineers: Improve unlearning techniques to preserve safe reasoning near forgotten topics in language models.
- •For ai safety teams: Maintain reliable model behavior when removing harmful or unwanted outputs by controlling latent representation directions.
Authors
Zinan Ling, Yang Xiao, Ruimeng Ye, Sijia Liu, Bo Hui
Abstract
Large Language Model (LLM) unlearning is often evaluated primarily by target suppression and clean retain utility. These metrics can miss a local failure mode: a model may perform well on standard retain prompts, yet degrade on benign reasoning when prompts lie near the forgotten domain. This phenomenon is termed forget-adjacent utility loss. A fundamental question is: where should forget representations go after they leave the removed behavior? Successful unlearning depends not only on whether forget-conditioned representations depart from the harmful computation, but also on the direction in which they are redirected afterward. We propose RA$^2$ (Retain-Anchored Attraction), a latent-space unlearning method that redirects forget-token states toward nearby retain-supported representations while preserving higher-layer behavior on retain data. On WMDP and MUSE, RA$^2$ yields higher forget-adjacent utility than strong unlearning baselines at similar broad-utility levels. These results show that the migration direction of forget representations materially affects benign behavior near the forget boundary. Our code is available at https://github.com/ZacLing/retain_anchored_attraction.