LLM Lesion Parameters Recovered to Mimic Aphasia Errors

Researchers have developed a novel method to recover lesion parameters in Large Language Models (LLMs) that mimic specific neurological deficits, such as those seen in aphasia. By training a neural network to map error profiles from picture naming tasks back to lesion parameters like modification percentage and noise sigma, they demonstrated that these parameters could be recovered with reasonable accuracy. This approach offers a new avenue for understanding transformer computation and has shown promise in generalizing to real-world patient data, distinguishing between different stroke survivor syndromes. AI

IMPACT This research provides a novel framework for LLM interpretability, potentially aiding in understanding model behavior and its relation to cognitive functions.

RANK_REASON The cluster contains an academic paper detailing a new method for LLM interpretability. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• artificial neural network
• arXiv
• Hugging Face
• LLaVA-Vicuna 13B
• Transformer++

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →