Co-occurring keywords · papers mentioning both
Papers
Towards Faithful Natural Language Explanations: A Study Using Activation Patching in Large Language Models
EMNLP 2025
Selecting causal brain features with a single conditional independence test per feature
NeurIPS 2019