Get Started
Home
Topics
Search
Library
Research questionHow should language-model robustness be evaluated when text perturbations affect hidden states and attention differently from outputs?Typos, altered words, corrupted text, and disrupted token order can change a model’s internal computation without being fully reflected in its output behavior. Different perturbations may also produce distinct effects across hidden-state geometry and attention-head function.
AI
Evaluation & Benchmarks
Machine Learning
Mechanistic Interpretability
Natural Language Processing
Latest papersRecent research connected to this question, newest first.How Perturbations Propagate: A Multi-Level Analysis of Robustness in Large Language ModelsThe evidence covers six naturalistic and synthetic perturbations across four GPT-2 and two Qwen2.5 checkpoints. Layerwise geometry is assessed with centered kernel alignment and intrinsic dimension; attention-head evidence is reported for GPT-2, so conclusions are limited to the tested checkpoints and analyses.research paper · Sep 3, 2026
Related questions
How can language-model attention remain reliable beyond its training context?How can post-training make language-model refusals robust without sacrificing general capability?How can we compare language models’ conditional behavior and predict the effects of prompt changes?How can one vision model remain robust across changing and unseen adversarial perturbation budgets?