Dissecting Recall of Factual Associations in Auto-Regressive Language Models [Geva+'23]
Abst
use a subject-relation query to study how the model aggregates information about the subject and relation to predict the correct attribute
Subject enrichment occurs in early MLP sublayers to subject-related attributes, information from the relation, and the prediction representation queries the enriched subject to extract the attribute
Method
Attention Knockout
Intervention to MHSA sublayers
the information flow is important if the flow is blocked
Dissecting Recall of Factual Associations in Auto-Regressive Language Models [Geva+'23]
Abst
Method
Attention Knockout
Intervention to MHSA sublayers
the information flow is important if the flow is blocked