DynaFlow: Learning free-form multi-agent orchestration
DynaFlow: Learning Free-Form Multi-Agent Orchestration · arXiv:2610.34467
In plain words
Some questions can be answered directly, while others need several agents to pass work between them. DynaFlow trains a main agent to choose between answering and writing a program that coordinates fixed helper agents. It can inspect what they return and decide what to do next.
Across six question sets, workflow use without training often scored below direct answers. After training, it matched or beat the untrained model's direct answers on all six sets. The tests do not show how it would perform on every kind of question.
The research
When tasks vary, one fixed way of splitting work can add unnecessary coordination or fail to pass needed results between stages. The question is whether a main agent can learn both when to delegate and how to organize that work, instead of simply being given access to other agents.
DynaFlow trains a main agent to choose direct answers or write workflows for fixed sub-agents, then evaluates accuracy and coordination across six question sets.
Researchers tested final-answer accuracy on 688 questions in six sets covering competition and multilingual mathematics, scientific reasoning, and questions requiring evidence gathered across multiple steps.
They compared direct answers and workflows, where a trained main agent assigns and coordinates work for fixed sub-agents, with other methods using the same Qwen3.5-27B backbone.
In Forced mode, which requires a workflow before substantive solving, DynaFlow scored 88.08%, versus 77.18% for Self-Refine, the strongest evaluated baseline. Separately, workflow use without orchestration training fell below direct answers on five sets; after Stage 1 training, Forced mode matched or exceeded the untrained model's direct answers on all six. This pattern suggests workflow access alone was insufficient in these tests.
The results cover 688 questions from six sampled evaluation sets, not tasks in general.
Stage 1 to Stage 2 workflow changes also combine selected training examples with continued training; the reported results do not include the equal-size random-subset control needed to isolate the effect of selection.
What the evidence shows
- Why fixed plans can fail
Some questions can be answered by one system, while others need separate steps that rely on earlier results. A fixed collaboration pattern may add unnecessary work to simple questions or fail to transfer information between dependent steps.
The research asks whether a system can decide when collaboration is useful and organize it around the task at hand.
- Give the agent choices
DynaFlow is a framework that gives a trained main agent the choice to answer directly, use a tool, or write a program coordinating fixed sub-agents. The program assigns work and specifies dependencies and information transfer.
Returned results and an execution trace inform the main agent's next decision, including whether to create another workflow.
- Training changes the result
The researchers tested final-answer accuracy on 688 questions in six sets covering competition and multilingual mathematics, scientific reasoning, and questions requiring evidence across multiple steps. Without orchestration training, requiring a workflow scored below direct answers on five sets.
After Stage 1 training, workflow mode matched or exceeded the untrained model's direct answers on all six.
- Compare the full systems
In Forced mode, which requires a workflow before substantive problem solving, DynaFlow reached 88.08% accuracy across the 688 questions. Self-Refine, the strongest evaluated baseline by overall accuracy, reached 77.18%.
Both use the same Qwen3.5-27B backbone, but their training and answer procedures differ, so this comparison measures the systems as a whole.
- Choose when to collaborate
Dynamic mode lets the trained agent decide between a direct answer and a workflow. It answered 47.1% of questions directly and reached 87.94% accuracy, close to Forced mode's 88.08%.
Its average of 2.44 sub-agents per question, counting all questions, was below Forced mode's approximately 4.16. These averages suggest selective collaboration can retain accuracy with less agent use in this evaluation.
Deep
- Original paper authors
- Ruhao Liu, Qi Li, Shizun Wang, Xinchao Wang
- Submitted to arXiv
- License
- CC BY 4.0