RAG-Assistant Model Size Has Limited Impact on Human-AI Collaboration

A recent study available on arXiv (2605.00964) assesses the effectiveness of chatbot-style assistants utilizing Retrieval-Augmented Generation (RAG) within a realistic multi-turn information-seeking context. This research contrasts the performance of humans (N=112) aided by RAG assistants with LLM-only and LLM+RAG benchmarks. It explores how the size of the underlying models (3B, 8B, and 70B parameters) affects the dynamics of human-AI collaboration, as well as user satisfaction and usability perceptions. Findings indicate that human-AI collaboration significantly outperforms model-only baselines, regardless of model size, underscoring the advantages of hybrid systems in information-seeking tasks. The study also emphasizes that smaller models can effectively enhance human decision-making when integrated with RAG, particularly in environments requiring adherence to local regulations and secure data management.