Mechanistic Analysis of In-Context Learning and Hallucination
November 19, 2026
This research analyzes why LLMs follow in-context rules but still produce out-of-context hallucinations. The study finds that role information is internally represented but does not always constrain output unless it meets specific magnitude and aggregation conditions.
HOW THIS AFFECTS YOU
●
researcherThis provides a first-order decomposition to help understand how role information affects token choice.