How Perturbations Propagate: A Multi-Level Analysis of Robustness in Large Language Models
What happens inside AI language models when text gets scrambled or corrupted
When language models encounter typos, garbled text, or shuffled words, the damage spreads differently depending on the type of error and where you measure it. Researchers found that looking only at final output—whether the model gets the answer right—misses crucial information about how the model's internal representations break down, and different errors leave distinct patterns of disruption inside the model that don't always correlate with behavioral mistakes.
Language models power chatbots, search engines, and writing tools that real users interact with through imperfect inputs—voice transcription errors, typos, garbled PDFs. This work reveals that a model might produce reasonable-sounding output while its internal workings are partially corrupted, masking fragility that only shows up under specific stresses. Understanding these hidden failure modes is essential for deploying language models in safety-critical contexts and for building systems that degrade gracefully rather than confidently producing wrong answers.