A Mechanistic Lens on Semantic Conflicts: Using Activation Patching to Understand LLM Behavior
The findings show that semantic conflicts affect program comprehension and downstream tasks, with relevant information concentrated in a small number of causally active residual-stream states, and demonstrate a framework for mechanistically analyzing how LLMs integrate code-related information under controlled semantic...