Attention as a Routing Graph: Live Circuit Extraction from a Single Forward Pass
arXiv:2609.25285v1 Announce Type: new Abstract: Finding circuits in language models usually means running many careful interventions. We try something simpler: treat attention as a routing map from one forward pass, keep a small set of...
Read more: https://arxiv.org/abs/2609.25285
Scan with your phone
Scan with your phone