Model Merging Beats Retraining for Conversational Search Retrievers

arXiv·medium signal

Conversational information retrieval is hard because it must handle conversation history, topic shifts, and coreference across turns, and the mainstream fix — fine-tuning or multi-tasking ad-hoc retrievers on conversational data — is costly because it requires retraining. This paper instead uses model merging to adapt an ad-hoc retriever for conversational search without full re-training. A cheaper path for builders wanting multi-turn retrieval on top of existing single-turn retrievers.