Metraipolis › archive › dc_58aeb446ecad

Unlearning Is Not Just Erasing: Temporal Decoupling via Generation Inequality — arXiv:2608.23020

Source
arXiv (en)
Original
https://arxiv.org/abs/2608.23020
License (as archived)
CC0-1.0
Archived
2026-10-18T07:45:10.328Z
SHA-256 of the archived text
45d3fc2ff23a856760e07ba31ddb01dae4b89b6a26646d8f3d29ad04495dfd81
Stable URL of this page
https://metraipolis.com/documento/dc_58aeb446ecad

How to cite

Unlearning Is Not Just Erasing: Temporal Decoupling via Generation Inequality — arXiv:2608.23020 — arXiv (en), https://arxiv.org/abs/2608.23020 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_58aeb446ecad (Metraipolis archive dc_58aeb446ecad, sha256 45d3fc2ff23a856760e07ba31ddb01dae4b89b6a26646d8f3d29ad04495dfd81)
BibTeX
@misc{dc58aeb446ecad,
  title = {Unlearning Is Not Just Erasing: Temporal Decoupling via Generation Inequality — arXiv:2608.23020},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2608.23020},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_58aeb446ecad; sha256 45d3fc2ff23a856760e07ba31ddb01dae4b89b6a26646d8f3d29ad04495dfd81},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.23020. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

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.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-25 in cs.AI).
Categorie arXiv: Computation and Language (cs.CL) ; Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2608.23020 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-18T07:37:52.913Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-08-25?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).