MARS: Multilingual Search Agents Using Direct Corpus Interaction
November 9, 2026
MARS replaces traditional learned retrievers with a reinforcement learning agent that interacts with raw corpora using shell commands. It uses corpus-native entity grounding and redundancy-aware rewards to improve multilingual evidence acquisition without indexing overhead.
HOW THIS AFFECTS YOU
●
builderYou can build search agents that bypass the need for complex vector indexing and similarity-based ranking.
●
founderThis offers a path to multilingual search that avoids the massive indexing overhead required by traditional RAG systems.