Large language models reorganize representational geometry during in-context learning

In three linesarXiv paper on representational geometry during in-context learning (ICL) in LLMs. Researchers show ICL performance correlates with task representational structure and successful ICL involves geometric reorganization increasing online separability. LLM behavior follows a prototype-like algorithm.

Summary generated by Claude — human-verified