TL;DR
MARS (Multilingual Search Agent) is a "retriever-free" framework that allows LLMs to find evidence across languages by directly executing shell commands (like grep) on raw text corpora. By grounding search terms in corpus-native aliases and using a redundancy-aware reinforcement learning reward, MARS significantly outperforms traditional vector-based retrieval methods while being more efficient and robust to language diversity.
What Problem Does This Paper Address?
Most Multilingual Retrieval-Augmented Generation (MRAG) systems rely on "retrievers"—models that turn text into mathematical vectors to find similar documents. However, these systems struggle with cross-lingual retrieval errors, high indexing costs, and "embedding shifts" where a relevant document in one language might have a lower similarity score than an irrelevant one in another.
An alternative is Direct Corpus Interaction (DCI), where an agent uses tools like grep to search raw text. While powerful, DCI faces two major hurdles:
- Lexical Mismatch: Exact search is brittle. If you search for "Cristiano Ronaldo" in a Chinese corpus that only uses the alias "C罗," you find nothing.
- Search Redundancy: Without proper guidance, reinforcement learning (RL) agents often get stuck in loops, generating repetitive or ineffective search commands that waste tokens and time.
Methods
The authors introduce MARS, which treats multilingual search as an iterative, agentic process. The core of the system is Utility-Aware Retrieval Policy Optimization (URPO), which refines the agent's behavior using three key components:
- Direct Corpus Interaction (DCI): Instead of a black-box retriever, the model generates executable shell commands to filter and pipe results from language-specific corpora.
- Corpus-Native Entity Grounding: To solve the lexical mismatch problem, MARS extracts language-specific titles and aliases directly from retrieved snippets. These "lexical anchors" are then reused in subsequent search steps. If a search in a target language fails, the model can "pivot" to a high-resource language (like English) to find a valid anchor first.
- Redundancy-Aware Search Reward: The authors use a novel reward function that penalizes Jaccard similarity between commands in a single search trajectory. This encourages the model to be concise and explore new search terms rather than repeating failed ones.
Datasets
The researchers evaluated MARS on two primary multilingual benchmarks:
- MKQA (Multilingual Knowledge Questions and Answers): 13 languages, including "seen" languages used during training and "unseen" languages to test cross-lingual generalization.
- MLQA (Multilingual Question Answering): 7 languages, used to evaluate the model's ability to handle extractive QA tasks.
- HotpotQA: A monolingual multi-hop benchmark used to ensure that adding multilingual capabilities doesn't degrade standard performance.
Key Insights and Ideas
- Anchors over Translations: You don't need a perfect translation of a query to succeed in DCI. You just need to find one document that bridges the gap between the query term and the corpus-native alias.
- Efficiency as a Training Signal: Rewarding "conciseness" doesn't just save tokens; it forces the model to generate more informative search commands. This prevents the "training collapse" often seen when RL agents are only given a reward at the very end of a long, messy trajectory.
- Robustness to Scale: Interestingly, while baseline DCI methods often get "confused" and perform worse as more languages are added to the search space, MARS actually improves. It successfully leverages the increased evidence pool without getting bogged down by the noise.
Key Results
MARS consistently outperformed both traditional RAG and existing DCI baselines (like Grepseek) across various model scales (Qwen3.5-4B and 9B).
- Accuracy: On MKQA, MARS achieved a 40.6% Exact Match (EM) score, significantly higher than the 37.1% achieved by the previous best DCI method.
- Efficiency: MARS reduced the number of search commands by 39% and the total response tokens (a proxy for latency) by 29% compared to Grepseek.
- Generalization: The model showed strong performance on "unseen" languages, proving that the search strategy it learns is language-agnostic.
Limitations and Future Work
The authors acknowledge that MARS requires task-specific training data and incurs additional optimization costs during the RL phase. Furthermore, while corpus-native grounding helps, the system is still fundamentally sensitive to lexical variation; if a corpus contains no aliases or links for an entity, exact search will still fail.
Future work may explore hybrid retrieval—combining the precision and control of executable shell commands with the fuzzy-matching robustness of semantic vector search.
Summary updated