Stanford/Yale team introduces MoRSE: AI multi-agent system that breaks tasks into specialized roles
New MoRSE system uses dynamic LoRA experts to specialize agents at both task and parameter levels for complex workflows.
The article introduces MoRSE, a task-oriented multi-agent system that decomposes each task into a dependency-aware directed acyclic graph of subtasks and assigns agents specific (role, subtask) specializations. It uses a dynamic mixture of role- and subtask-conditioned LoRA experts with a prototype-based semantic router to add parameter-level specialization on a shared LLM substrate cost-effectively. A hierarchical group-relative policy optimization isolates expert updates from routing decisions, and experiments on code-generation benchmarks show improvements in both whole-task and step-wise performance, with gains generalizing across held-out task categories and domains.
- MoRSE uses a dynamic mixture of LoRA experts to specialize agents at both task and parameter levels, improving performance on complex workflows.
- The system decomposes tasks into a dependency-aware graph of subtasks and assigns agents specialized (role, subtask) pairs.
- Experiments on code-generation benchmarks showed gains across multiple backbones, with improvements generalizing to new tasks.
Why It Matters
MoRSE could redefine how AI agents collaborate on complex tasks by enabling cost-effective, specialized multi-agent systems that outperform traditional approaches.