Metraipolis › archive › dc_8002f3f26dea

Can Capability Enhancement be Localized? — arXiv:2610.21294

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.21294
License (as archived)
CC0-1.0
Archived
2026-10-16T10:20:14.744Z
SHA-256 of the archived text
fcd36aae3bbed42e5beca3ae83ddcbd563af9547a84d16832291ee1413c9e0b1
Stable URL of this page
https://metraipolis.com/documento/dc_8002f3f26dea

How to cite

Can Capability Enhancement be Localized? — arXiv:2610.21294 — arXiv (en), https://arxiv.org/abs/2610.21294 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_8002f3f26dea (Metraipolis archive dc_8002f3f26dea, sha256 fcd36aae3bbed42e5beca3ae83ddcbd563af9547a84d16832291ee1413c9e0b1)
BibTeX
@misc{dc8002f3f26dea,
  title = {Can Capability Enhancement be Localized? — arXiv:2610.21294},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.21294},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_8002f3f26dea; sha256 fcd36aae3bbed42e5beca3ae83ddcbd563af9547a84d16832291ee1413c9e0b1},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.21294. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Transformer-based language models have shown strong capabilities across many tasks and domains. However, prior work shows that, for a given task, not all attention heads matter equally. Specifically, they suggest that many \textit{redundant} heads can be safely removed with little-to-no effect, while a small set of \textit{important} heads, when ablated, severely damages the model's performance on the given task. In this work, we investigate a complementary aspect, called \textit{Capability Enhancement Localization} (CEL): \textit{given a model and a task, does there exist a small set of heads whose removal improves the task performance?} We first investigate this question \textit{empirically} across models from different families. Using a simple two-stage routine that screens candidate heads and evaluates their joint ablations, we identify sparse head combinations whose removal improves task-specific performance without further fine-tuning. Motivated by these observations, we ask \textit{theoretically whether such a phenomenon can naturally arise from training dynamics, and under what conditions}. To this end, we construct a simplified multi-head softmax attention model and two tasks with conflicting features, and show that gradient descent from small random initialization provably concentrates the conflicting features into particular heads. We further show that ablating such heads improves performance on one task while harming the other, even though training converges to a global optimum of the joint objective. This establishes a possible training-induced theoretical explanation for CEL.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.LG, nuovi lavori del giorno).
Categorie arXiv: Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2610.21294 — 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 2026-10-16T08:48:02.701Z dalla pagina https://arxiv.org/list/cs.LG/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).