work exploratory Query-Conditioned Execution Heads Investigating attention heads that combine task context and query information during in-context learning. Zero-Shot Inference and Relation Vectors Applying the Function Vectors methodology to zero-shot inference and investigating failure modes of the relation vector. Geometry of Truth in LLMs Notes on reproducing “Geometry of Truth” with causal intervention and visualization. Refusal Direction in LLMs Reproduction of “Refusal in Language Models Is Mediated by a Single Direction” with causal interventions on the residual stream.
