Unlearning Is Not Just Erasing: Temporal Decoupling via Generation Inequality
2026-08-24 • Computation and Language
Computation and LanguageArtificial Intelligence
AI summaryⓘ
The authors address the challenge of making large language models forget certain information without harming their overall usefulness. They introduce ADU, a method that protects sensitive data by changing how the model pays attention to certain parts of its memory, instead of just erasing words. This approach keeps the model's language skills intact while effectively forgetting targeted details. Their tests show ADU performs better than other methods in balancing forgetting and preserving normal function.
large language modelsunlearningcontextual attentionattention headsprivacylinguistic structurelanguage modelingforgetting effectattention-projection adapters
Authors
Xunlei Chen, Qirui Ye, Yuang Li, Yi Gong, Zhaokun Wang, Wenyi Li, Shiyao Guo, Jinyu Guo
Abstract
Large language models (LLMs) require effective unlearning to address privacy regulations and safety concerns. However, achieving precise forgetting without compromising general utility remains challenging. Existing sequence- and token-level methods penalize target outputs without modeling their context-dependent retrieval paths, which can disrupt linguistic structure or suppress benign knowledge. We present ADU, a fine-grained, training-based framework that shifts unlearning from token erasure to contextual attention-pathway decoupling. Exploiting the functional distinction between local and global attention heads, ADU identifies preplan positions that retrieve persistent sensitive anchors and fixes their candidate paths under the original model. It then trains attention-projection adapters to suppress attention mass along these paths while preserving local-attention structure and retain-set language modeling. Post-training activation exchange tests whether the modified attention-output module transmits the learned forgetting effect. ADU achieves the strongest aggregate performance among evaluated baselines on the TOFU and WMDP benchmarks, including a Forget Quality of (0.93) on TOFU. It preserves 87--98% of model utility (92.9% on average versus 81.9% for baselines) while reducing side effects in benign contexts.