LARA: small, composable behaviours for frozen LLMs [P]

LARA is an ongoing research project that provides a library for creating small, composable behaviors for frozen Large Language Models (LLMs). The library is currently usable and includes training code, examples, and reproduction instructions. A key feature is the Mixture of Behaviors (MoBs) demo, which illustrates how independently trained behaviors can share a frozen model, with a soft router dynamically selecting or combining them on a token-by-token basis.