UUE
A lightweight unlearning framework for LLMs that decouples forgetting and utility on adapters through orthogonal null-space projection and convex untargeted optimization.
Timeline (last 90 days)
The UUE framework reformulates unlearning as a controllable model-editing problem, achieving decoupling of forgetting and utility on lightweight adapters through orthogonal null-space projection and convex no-target optimization.
UUE introduces a convex no-target objective function that destroys undesirable knowledge by pushing the forget-set embeddings toward their antipodes, and this objective function has a closed-form analytical solution.
UUE is implemented through lightweight, pluggable adapters and requires no modification to the frozen backbone model at deployment.
Experiments on the TOFU and WMDP benchmarks show that UUE and UUE-L significantly outperform existing methods in unlearning effectiveness.
The paper currently reports only benchmark results; robustness and adversarial unlearning effects in actual deployment remain to be verified.
All Facts (5)
The UUE framework reformulates unlearning as a controllable model-editing problem, achieving decoupling of forgetting and utility on lightweight adapters through orthogonal null-space projection and convex no-target optimization.
50%UnverifiedUUE introduces a convex no-target objective function that destroys undesirable knowledge by pushing the forget-set embeddings toward their antipodes, and this objective function has a closed-form analytical solution.
50%UnverifiedUUE is implemented through lightweight, pluggable adapters and requires no modification to the frozen backbone model at deployment.
50%UnverifiedExperiments on the TOFU and WMDP benchmarks show that UUE and UUE-L significantly outperform existing methods in unlearning effectiveness.
50%UnverifiedThe paper currently reports only benchmark results; robustness and adversarial unlearning effects in actual deployment remain to be verified.
50%