Traditional multilingual question-answering systems rely on a retriever, a module that scores documents by similarity to the query. This approach is unreliable across languages, often missing relevant text, and storing indexes for each language requires significant resources.
MARS replaces the retriever with direct shell-command search over raw text. It uses , a trial-and-error training technique, to teach the system effective search strategies. It introduces entity grounding, where exact entity names found in documents become search keys, and a redundancy-aware reward that penalizes unnecessary steps.
MARS outperformed existing retrieval-based and direct-search methods across many languages, including those not seen during training. It made far fewer redundant searches while finding evidence more reliably. This shows that retriever-free search can be more robust and efficient for multilingual AI, reducing infrastructure costs and improving performance for low-resource languages.