Divide and Cooperate: Role-Decomposed Multi-Agent LLM Training with Cross-Agent Learning Signals

A new multi-agent training framework called DAC splits language-model search into two cooperating roles, addressing a long-standing credit-assignment problem that has hindered single-policy systems, according to a paper submitted to arXiv on 9 June 2026 [1][2]. The framework, named Divide and Cooperate, separates evidence acquisition and answer generation into distinct subtasks handled by dedicated agents that share a backbone model through parameter-efficient LoRA modules [2]. The generator agent produces answers and simultaneously verifies whether retrieved evidence is sufficient, abstaining when it is not. That abstention signal is fed back into the search agent’s reward, creating structured cross-agent learning signals that the authors argue improve credit assignment [2]. In parallel, the searcher agent exposes the generator to harder evidence environments through hard-positive augmentation, which is designed to make the generator more robust [2]. Prior approaches typically couple both functions inside a single policy, forcing one model to juggle potentially conflicting objectives. The authors note this design causes a combinatorial explosion in the policy space and makes efficient exploration difficult [2]. It also muddies training feedback: a search step that retrieves adequate evidence can still be penalized if the generation step fails, and vice versa [2]. Experiments on general and multi-hop question-answering benchmarks show that DAC achieves strong performance against earlier baselines that require full fine-tuning of monolithic models [2]. The paper was posted on arXiv, the open-access e-print repository that hosts preprints across physics, computer science, mathematics, and related fields [6]. As of November 2024, the repository was receiving roughly 24,000 new articles per month [6]. The work appears under arXiv’s machine learning category and is accompanied by the repository’s standard Labs integrations, including bibliographic tools and code finders [4]. arXivLabs, launched in 2020 as a formal framework for community-contributed features, allows third-party developers to build experimental tools that appear on article record pages while adhering to arXiv’s values of openness, community, and user-data privacy [5]. The Labs program is currently pausing new proposals while the arXiv development team focuses on modernizing and migrating systems to the cloud [3].