Metraipolis › archive › dc_6fc2dcc7ab13

Attention Sensitivity Is Not Enough: Dissociating Attention-Level and Behavioural In-Context Learning under Fine-Tuning — arXiv:2609.00064

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.00064
License (as archived)
CC0-1.0
Archived
2026-10-17T06:32:40.543Z
SHA-256 of the archived text
5b976cde0cf512fb1cc125a744579adc674c7098d34a1abb490329105c0e0954
Stable URL of this page
https://metraipolis.com/documento/dc_6fc2dcc7ab13

How to cite

Attention Sensitivity Is Not Enough: Dissociating Attention-Level and Behavioural In-Context Learning under Fine-Tuning — arXiv:2609.00064 — arXiv (en), https://arxiv.org/abs/2609.00064 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_6fc2dcc7ab13 (Metraipolis archive dc_6fc2dcc7ab13, sha256 5b976cde0cf512fb1cc125a744579adc674c7098d34a1abb490329105c0e0954)
BibTeX
@misc{dc6fc2dcc7ab13,
  title = {Attention Sensitivity Is Not Enough: Dissociating Attention-Level and Behavioural In-Context Learning under Fine-Tuning — arXiv:2609.00064},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.00064},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_6fc2dcc7ab13; sha256 5b976cde0cf512fb1cc125a744579adc674c7098d34a1abb490329105c0e0954},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.00064. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

In-context learning (ICL) lets large language models adapt to new tasks from demonstrations, and fine-tuning can erode this behaviour. Many preservation diagnostics inspect attention: if attention changes when demonstrations change, the model is treated as context-sensitive. This paper asks how far that proxy can be trusted once it is optimised. We formalise \emph{In-Context Sensitivity} (ICS), the average row distance between last-token attention on matched and mismatched demonstration prefixes, and pair it with \emph{ICL-GAP}, the behavioural accuracy gap between the same prefixes. In a controlled four-arm ablation on Llama-2-7B, an ICS-maximising regulariser ($\armKL$) drives ICS to $1.413$, within $0.5\%$ of its geometric ceiling. The behavioural readout tells a different story: ICL-GAP stays near zero and MMLU accuracy moves from $0.371$ to $0.279$, a Goodhart dissociation of the bounded attention proxy. Endpoint statistics locate the mechanism: attention grows sharp and near-disjoint across prefixes yet routes to formatting and demonstration-body tokens rather than labels. A random-label protocol confirms that the behavioural probe family retains dynamic range at the same checkpoints. In a constructive sweep, behaviour gating partially mitigates the effect, while objectives anchored to pretrained computation hold the high-MMLU, moderate-ICS region that divergence maximisers leave. The main lesson is diagnostic: attention-level ICL proxies earn their place as training targets only after validation against behavioural gaps.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-02 in cs.AI).
Categorie arXiv: Machine Learning (cs.LG) ; Artificial Intelligence (cs.AI); Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2609.00064 — 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 2026-10-17T06:15:12.728Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-02?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).