Verified, then logged.
Sun 22 Nov 2026 · 00:23 KST / 15:23 UTC한국어 | ENSearch /

01 / AI · Analysis

Multilingual MARS agent trades retrievers for shell commands

The authors said MARS beat existing methods, but the abstract gives neither the accuracy gap nor the number of languages tested.

A close-up of a laptop terminal on a dim desk, packed with blurred rows of text suggesting several writing systems, with a few segments highlighted in amber like search matches. AI-generated image

The authors of an arXiv paper published on 9 Nov 2026 presented MARS, a multilingual search agent that drops the retriever and searches source text directly with shell commands. They said in the abstract that the agent, trained with reinforcement learning, outperformed existing retrieval-based and direct-search methods.

The authors argued that existing multilingual retrieval-augmented generation depends on learned retrievers, which leaves it vulnerable to cross-lingual retrieval errors and indexing overhead. Retrieval-augmented generation finds outside documents to support an answer, and a retriever is the model that selects those documents. MARS removes the retriever and moves the bottleneck to generating search commands.

The paper names two challenges for this design: exact search that is sensitive to the form a term takes in a document, and the limits of training that scores only the final outcome. MARS was designed to address both.

A mechanism that recovers search terms from document text

According to the paper, the MARS model works on the raw corpus through shell commands, which are written instructions that an operating system executes. A corpus is the collection of documents being searched. The approach is called direct corpus interaction, or DCI.

To handle sensitivity to surface forms, the authors proposed corpus-native entity grounding. The technique recovers reliable lexical anchors from evidence in the corpus and uses them as reference points for search. Its aim is to reduce failures to find supporting documents.

A redundancy-aware reward against training collapse

The second challenge is the training signal. The paper states that supervision scoring only the final outcome gives limited guidance for search that runs over multiple turns. Reinforcement learning is a training method in which a model learns the actions that increase its reward.

The authors said they introduced a redundancy-aware search reward to encourage concise and effective search paths. The reward targets the training collapse that occurs in reinforcement learning in multilingual settings. The abstract does not give the reward's exact formula.

A performance lead reported without figures

In results published by the authors, MARS consistently outperformed its comparison systems and showed robust generalisation across languages. The abstract describes those systems as "strong retrieval-based and DCI baselines", meaning both retriever-based methods and earlier DCI methods served as benchmarks.

The efficiency claim also rests on the authors' own results. The abstract says MARS substantially reduced redundant corpus interactions. It gives neither the size of the accuracy gap nor the number of languages tested.

The abstract does not say whether the code and model weights are public. The paper's central issue is where the difficulty goes once the retriever is removed. The authors narrowed it to recovering surface forms and designing the reward.