Disentangling Recall and Reasoning in Transformer Models through Layer-wise Attention and Activation Analysis

Harshwardhan Fartale, Ashish Kattamuri, Rahul Raja +3

Transformer-based language models excel at both recall (retrieving memorized facts) and reasoning (performing multi-step inference), but whether these abilities rely on distinct in…