![[2602.01848] ROMA: Recursive Open Meta-Agent Framework for Long-Horizon Multi-Agent Systems](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.01848] ROMA: Recursive Open Meta-Agent Framework for Long-Horizon Multi-Agent Systems
Summary
The paper introduces ROMA, a Recursive Open Meta-Agent Framework designed to enhance performance in long-horizon multi-agent systems by addressing limitations in current agentic frameworks through recursive task decomposition and structured aggregation.
Why It Matters
As AI systems increasingly tackle complex, long-term tasks, the need for frameworks that can manage reasoning depth and improve interpretability becomes critical. ROMA offers a novel approach that enhances performance while maintaining flexibility and transparency, making it relevant for researchers and practitioners in AI and multi-agent systems.
Key Takeaways
- ROMA improves long-horizon task performance through recursive task decomposition.
- The framework supports modular roles, enhancing flexibility and interpretability.
- GEPA+ allows adaptation to specific tasks without fine-tuning, improving efficiency.
- ROMA demonstrates superior performance on reasoning and long-form generation benchmarks.
- The design facilitates heterogeneous multi-agent systems, optimizing for cost and capability.
Computer Science > Artificial Intelligence arXiv:2602.01848 (cs) [Submitted on 2 Feb 2026 (v1), last revised 14 Feb 2026 (this version, v2)] Title:ROMA: Recursive Open Meta-Agent Framework for Long-Horizon Multi-Agent Systems Authors:Salaheddin Alzu'bi, Baran Nama, Arda Kaz, Anushri Eswaran, Weiyuan Chen, Sarvesh Khetan, Rishab Bala, Tu Vu, Sewoong Oh View a PDF of the paper titled ROMA: Recursive Open Meta-Agent Framework for Long-Horizon Multi-Agent Systems, by Salaheddin Alzu'bi and 8 other authors View PDF HTML (experimental) Abstract:Current agentic frameworks underperform on long-horizon tasks. As reasoning depth increases, sequential orchestration becomes brittle, context windows impose hard limits that degrade performance, and opaque execution traces make failures difficult to localize or debug. We introduce ROMA (Recursive Open Meta-Agents), a domain-agnostic framework that addresses these limitations through recursive task decomposition and structured aggregation. ROMA decomposes goals into dependency-aware subtask trees that can be executed in parallel, while aggregation compresses and validates intermediate results to control context growth. Our framework standardizes agent construction around four modular roles --Atomizer (which decides whether a task should be decomposed), Planner, Executor, and Aggregator -- which cleanly separate orchestration from model selection and enable transparent, hierarchical execution traces. This design supports heterogeneous mult...
