Researchers boost search with model merging for AI conversations

Researchers from Ahmed Rayane Kebir, Jose G. Moreno, and Lynda Tamine have proposed a training-free strategy called model merging to enhance conversational information retrieval (CIR). Traditional methods rely on costly fine-tuning or multi-tasking to handle conversational history, often leading to catastrophic forgetting where ad-hoc search performance degrades. Their paper, accepted at SIGIR 2026, introduces model merging—a technique that combines models without additional training.

The team experimented with linear (Model Soup) and non-linear (Slerp) merging strategies on standard ad-hoc and conversational datasets. Their approach achieved up to 15% higher NDCG@3 accuracy under zero-shot conditions, demonstrating improved generalizability across tasks. By eliminating the need for fine-tuning, model merging reduces computational costs and preserves foundational retrieval capabilities.