MARS: Training Multilingual Search Agents for Retriever-Free Evidence Acquisition
Rui Qi, Yunlong Liang, Yufeng Chen, Jinan Xu, Fandong Meng, Kaiyu Huang
cs.CL
2026-11-06
MARS trains LLM agents to grep raw multilingual corpora with RL, lifting MKQA EM from 37.07 to 40.64 over Grepseek while using 39% fewer search commands.
Multilingual RAG normally depends on a learned retriever that embeds queries and documents into a shared space and ranks by similarity. That has two known weak spots. Embedding distributions shift between languages, so a lower-scoring document in one language can be more useful than a higher-scoring one in another, and low-resource languages suffer most. And building, storing and updating an index costs real money.
Direct corpus interaction (DCI) is a recent alternative: skip the retriever and let the agent write shell commands such as grep against the raw text. This paper brings DCI to multilingual settings. The bottleneck moves from relevance scoring to writing a command that actually hits, and exact matching is brutal about entity surface forms. The paper's example: searching a Chinese corpus for "Cristiano Ronaldo" returns nothing because the corpus writes the name as "C罗". Training is the second problem. With only an outcome reward, the policy has little signal on how to revise a search, and it drifts toward repeating near-identical commands.
Each rollout is multi-turn. The model reasons, picks a language corpus, emits a shell command, reads the output, and repeats until it answers or runs out of turns. Training uses GRPO; the authors call their variant URPO and add two ideas.
The final reward is exact match, minus a length penalty, plus the two bonuses.
Experiments use MKQA (13 languages, 8 for training and 5 unseen) and MLQA (7 languages) with Qwen3.5 4B and 9B backbones. Baselines: direct inference, static RAG, D-RAG, the multilingual agentic RAG method LcRL, training-free DCI, and the post-trained DCI method Grepseek. The paper reports MARS best in every setting with significance at p<0.05, and DCI methods generally ahead of retrieval-based ones. The column layout of the main table is unclear in the extracted text, so only unambiguous numbers are listed.
| Comparison | Metric | Result |
| MKQA ablation: Grepseek-style baseline to full MARS | EM / fEM | 37.07 to 40.64 / 53.23 to 55.54 |
| Grounding only / redundancy reward only | EM | 39.79 / 38.97 |
| Search keys: model self-translation to corpus anchors | Empty-hit rate | 32.2 to 24.3 (external translation API: 39.0) |
| Same | Correct-entity hit / EM | 54.1 to 56.7 / 37.1 to 40.3 |
| Without vs with redundancy reward | Repeated-search rate | 22.33% to 4.89% |
| MARS vs Grepseek | Search commands per query | 4.22 to 2.56 (39% fewer) |
| Same | Turns / response tokens | 1.87 to 1.36 / 4,853 to 3,447 |
| HotpotQA, monolingual, vs Grepseek | EM / F1 | 46.27 to 50.00 / 62.31 to 63.51 |
Two details stand out. The external translation API raised the empty-hit rate, which shows a correct translation and a matching string are different things. And when the accessible corpora grow from the query language alone to all languages, the DCI baseline degrades while MARS stays stable.
For teams building multilingual knowledge-base QA without wanting to maintain a vector index, this is a workable alternative. DCI needs no embeddings or index; the cost moves to online multi-turn commands. In the paper's efficiency table, D-RAG uses 6,920 response tokens per query versus 3,447 for MARS and 3,532 for LcRL, so the online cost is not out of hand.
Still, this is incremental. The DCI idea comes from prior work, and MARS adds two multilingual training details worth roughly 3.6 EM points. The reusable part is the trick of mining aliases from the corpus itself to use as search keys, which is model-agnostic.
The authors say the method needs task-specific training data and extra optimization cost, covers only a small slice of the world's languages, and stays fragile when the corpus never exposes useful aliases. They suggest hybrid lexical and semantic retrieval as future work.
Other gaps: the corpora are Wikipedia-style benchmark collections, far smaller than a real enterprise store, and grep latency at TB scale is not reported. Efficiency is measured by command count and tokens, not wall-clock time. Which backbone each table uses is not always clear. The English pivot assumes the entity has an English entry, and that case is not analyzed separately. The code is hosted in an anonymous repository.