Metraipolis › archive › dc_1630b173b8ff

Entangled Representations Amplify Collateral Damage in Unlearning — arXiv:2609.02285

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.02285
License (as archived)
CC0-1.0
Archived
2026-10-17T04:17:04.728Z
SHA-256 of the archived text
8e9c9f7b5e187e42bad787f61fc26fe60b8d6e5808da4486f4642f34b411c40a
Stable URL of this page
https://metraipolis.com/documento/dc_1630b173b8ff

How to cite

Entangled Representations Amplify Collateral Damage in Unlearning — arXiv:2609.02285 — arXiv (en), https://arxiv.org/abs/2609.02285 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_1630b173b8ff (Metraipolis archive dc_1630b173b8ff, sha256 8e9c9f7b5e187e42bad787f61fc26fe60b8d6e5808da4486f4642f34b411c40a)
BibTeX
@misc{dc1630b173b8ff,
  title = {Entangled Representations Amplify Collateral Damage in Unlearning — arXiv:2609.02285},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.02285},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_1630b173b8ff; sha256 8e9c9f7b5e187e42bad787f61fc26fe60b8d6e5808da4486f4642f34b411c40a},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.02285. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

A long-held intuition in interpretability research is that representational entanglement, the sharing of structure between knowledge domains in a neural network, makes unlearning harder. While the intuition is widespread, it has never been directly tested in a controlled experiment. We present a way to do so: by repurposing Selective Gradient Masking (SGTM), we train a suite of six 254M-parameter language models on English Wikipedia with graded levels of disentanglement between biology and non-biology knowledge. Applying three standard unlearning methods to every model in the suite, we find that more disentangled models consistently achieve better retain-forget trade-offs: at a fixed level of forgetting, the most disentangled models incur roughly $4\times$ lower retain cost under two of the three methods, and $1.3\times$ lower under the third. Because our intervention changes only the model, not the data or the unlearning algorithm, this is direct evidence that representational entanglement is one of the causes of collateral damage in unlearning, as interpretability researchers have long suspected. A similar design could be used to test other structural claims from interpretability.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-03 in cs.CL).
Categorie arXiv: Machine Learning (cs.LG) ; Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2609.02285 — 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 2026-10-17T03:26:49.835Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-03?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).