1 paper · 1 filter
Tingan Jin, Shuhang Dong, Haosong Li
Which context tokens drive a language model's answer? Graph-based tools such as attention rollout and attention flow answer this from a single forward pass, yet their faithfulness…