Metraipolis › archive › dc_b1e45f5dd804

How Perturbations Propagate: A Multi-Level Analysis of Robustness in Large Language Models — arXiv:2609.03322

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.03322
License (as archived)
CC0-1.0
Archived
2026-10-17T01:40:14.325Z
SHA-256 of the archived text
0bcf3438af8060875bd5c0eb4bf3b348d08d79b1a7413164d7ed91d93a3abbba
Stable URL of this page
https://metraipolis.com/documento/dc_b1e45f5dd804

How to cite

How Perturbations Propagate: A Multi-Level Analysis of Robustness in Large Language Models — arXiv:2609.03322 — arXiv (en), https://arxiv.org/abs/2609.03322 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_b1e45f5dd804 (Metraipolis archive dc_b1e45f5dd804, sha256 0bcf3438af8060875bd5c0eb4bf3b348d08d79b1a7413164d7ed91d93a3abbba)
BibTeX
@misc{dcb1e45f5dd804,
  title = {How Perturbations Propagate: A Multi-Level Analysis of Robustness in Large Language Models — arXiv:2609.03322},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.03322},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_b1e45f5dd804; sha256 0bcf3438af8060875bd5c0eb4bf3b348d08d79b1a7413164d7ed91d93a3abbba},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.03322. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Language models encounter typos, corrupted text, altered words, and disrupted token order, yet robustness is usually evaluated only through output behavior. We study how six naturalistic and synthetic input perturbations propagate through decoder-only language models at three levels: output behavior, hidden-state geometry, and attention-head function. We evaluate behavioral effects across four GPT-2 and two Qwen2.5 checkpoints, analyze layerwise geometry using centered kernel alignment and intrinsic dimension, and examine attention-head responses in GPT-2. Perturbation types produce distinguishable metric profiles that are not fully captured by output measures and are only partly consistent across the tested checkpoints. Copying scores show the strongest pooled associations with activation-patching recovery under token substitution and shuffling, although these associations do not isolate copying-specific effects. Gradient-guided HotFlip perturbations also cause stronger behavioral and representational disruption than rate-matched random token substitutions in GPT-2; their behavioral effects are consistent across all six tested checkpoints. Our results show that robustness claims based on a single behavioral or representational metric can be misleading, and motivate multi-level evaluation of how perturbations alter language-model computation.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-04 in cs.CL).
Categorie arXiv: Computation and Language (cs.CL) ; Machine Learning (stat.ML).
Pagina del lavoro: https://arxiv.org/abs/2609.03322 — 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 2026-10-17T01:11:43.519Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-04?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).