Circuit Discovery

Discourse and LLMs
We measure and enhance LLMs’ discourse understanding, focusing on faithfulness and mechanistic interpretability.
PDF Code Slides
Discourse and LLMs
Discursive Circuits: How Do Language Models Understand Discourse Relations?
We introduce a new task, Completion under Discourse Relation (CuDR), to reveal the underlying transformer circuits responsible for discourse understanding.
PDF Cite Code Project
Discursive Circuits: How Do Language Models Understand Discourse Relations?