AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

A Robust Unlearning Method with Adaptive Knowledge Guidance and Memory Preservation

Jingyuan Tian, Xiaofei Zhou

AAAI 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add adaptive head selection plus adversarial memory perturbation to RMU-style unlearningALMPU · WMDP-Bio, 10 forget-sample fine-tuning attackunlearning effect (%): 51.6 (RMU) -> 66.7Improved
ALMPU versus RMU before any attackunlearning · WMDP-Bio, 0 samplesunlearning effect (%): 70.9 -> 70.1Hurt
add memory perturbation to gradient ascent on all layers (ablation)GA · WMDP-Bio, 10 forget-sample fine-tuning attackunlearning effect (%): 39.2 -> 47.0Improved
add memory perturbation to gradient ascent on all layers (ablation)GA · WMDP-Bio, 0 samplesunlearning effect (%): 72.1 -> 68.4Hurt
use adaptive scaling-factor head selection instead of random 25% heads (ablation)ALMPU · WMDP-Bio, 50 samplesunlearning effect (%): 38.2 -> 42.2Improved
use adaptive head selection instead of perturbing all heads (ablation)ALMPU · WMDP-Bio, 0 and 50 samplesunlearning effect (%): 63.9 -> 70.1 at 0 samples; 43.1 -> 42.2 at 50 samplesMixed
use adaptive head selection instead of fixed late layers (ablation)ALMPU · WMDP-Bio, 25 samplesunlearning effect (%): 41.5 -> 53.4Improved
use adaptive head selection instead of perturbing all layers (ablation)ALMPU · WMDP-Bio, 25 samplesunlearning effect (%): 50.8 -> 53.4Improved
ALMPU versus RMU under Wiki-text fine-tuningunlearning · Wiki-text attack, 10 to 500 samplesunlearning effect (%): RMU 68.5-70.1 vs ALMPU 68.1-69.5No clear effect
probability-based NPO under Wiki-text fine-tuningNPO · Wiki-text attackunlearning effect (%): 58.2 -> 39.8Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy