Recovering LLM Lesion Parameters from Aphasic Error Profiles
August 10, 2026
This study demonstrates that a multi-task neural network can map clinical error profiles in LLaVA-Vicuna 13B back to specific perturbation parameters like layer index and noise sigma. The method treats LLM interpretability as an inverse problem, attempting to identify the causal internal state responsible for specific linguistic errors.
HOW THIS AFFECTS YOU
●
researcherYou can use this approach to move beyond descriptive interpretability toward causal parameter recovery.