RCEM: Robust Conversational Search EMbedder in Distributional Shift

RCEM (Robust Conversational Search Embedder) has been introduced as a novel approach to conversational search that integrates LLM-based query reformulation while maintaining the original embedding space. This model simplifies the alignment task by mapping conversations to shorter rewritten queries, which reduces overfitting and eliminates the need for specific relevance labels during training. Experimental results indicate that RCEM achieves up to a 30% performance improvement over existing methods under distributional shifts, making it a significant advancement for practitioners working with conversational retrieval systems.