How Perturbations Propagate: A Multi-Level Analysis of Robustness in Large Language Models
Abstract
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.
Related
- Admissible characterization in delay equations and robustness of exponential dichotomies against small-delay perturbations
- Adversarial Training for Tabular Credit Scoring: A Multi-Attack Robustness Evaluation in P2P Lending
- Fix-and-propagate heuristics using low-precision first-order LP solutions for large-scale mixed-integer linear optimization
- How Shaped Waves Propagate in Scattering Media
- Multi-Objective Fractional-Order Control Design for Solar Collector Systems Under Parameter Perturbations and Stability Constraints
BibTeX
@misc{chan2026perturbations,
title = {{How Perturbations Propagate: A Multi-Level Analysis of Robustness in Large Language Models}},
author = {Dun Li Chan and Emily Liu and Niyathi Allu and Christian Hoang},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2609.03322},
eprint = {2609.03322},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.03322},
}
RIS
TY - UNPB
TI - How Perturbations Propagate: A Multi-Level Analysis of Robustness in Large Language Models
AU - Dun Li Chan
AU - Emily Liu
AU - Niyathi Allu
AU - Christian Hoang
PY - 2026
DA - 2026/10/05
PB - arXiv
DO - 10.48550/arxiv.2609.03322
UR - https://arxiv.org/abs/2609.03322
AB - 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.
ER -