Prompt language as a diversity lever: foreign-language queries broaden source retrieval and AI-generated research proposals

Abstract

Large language models (LLMs) coupled to retrieval-augmented generation (RAG) are increasingly used to generate scientific hypotheses, but they are notoriously prone to redundancy: repeated queries converge on the same handful of sources and near-identical ideas. Existing remedies are computationally heavy, relying on iterative planning, citation or entity graphs, or evolutionary search. Here I show that a far simpler lever, the language of the prompt, substantially diversifies both what a RAG system retrieves and what it proposes, even when the underlying corpus is entirely in English. Using a self-hosted mechanobiology RAG system (llama3.3:70b and nomic-embed-text served locally on an NVIDIA DGX Spark, indexing 523 peer-reviewed and preprint papers in AnythingLLM/LanceDB), I issued an identical Alzheimer’s-disease hypothesis-generation prompt in English and in seven other languages that the model understands. In English, five independent queries, including chain-of-thought and arbitrary-prefix controls, retrieved the same two source papers in the same 1:9 ratio every time; because dense retrieval is deterministic, this collapse is exactly reproducible rather than stochastic. Translating the prompt broke the collapse completely: the seven foreign-language queries drew on 3–8 distinct papers each (mean 5.4), surfacing 25 distinct papers versus 2 for all English runs combined (permutation p ≈ 0.001). Diversity propagated downstream to the generated proposals, which showed lower inter-proposal vocabulary overlap (Jaccard 0.25 vs 0.41), 16 domain concepts absent from every English output, divergent mechanisms and experimental designs, and a striking categorical shift in the proposed model organism (C. elegans in 5/5 English proposals versus Drosophila or other invertebrates in every foreign proposal). Because a within-language perturbation left retrieval unchanged while translation transformed it, prompt language, not surface variation or sampling temperature, is the operative variable. Foreign-language prompting is thus a nearly free, broadly applicable tool for injecting diversity into RAG-based scientific ideation and, potentially, retrieval-augmented systems more generally.