Paper review · October 28, 2026
DynaFlow: Learning Free-Form Multi-Agent Orchestration
A main agent is trained to write JavaScript workflows on its own when needed to coordinate fixed sub-agents, raising accuracy by 10.90 percentage points over the strongest baseline built on the same Qwen3.5-27B.
Problem
Multi-agent systems often fix roles and communication structures in advance, which adds unnecessary collaboration costs even for simple questions and makes it hard to adapt to each task's needs. Research on automatic workflow design also has limitations: it relies on fixed operator templates or builds the whole structure in one pass without looking at intermediate execution results.
Method
The main agent's action space includes "answer directly," "call a general tool" and "write and execute a workflow program." Workflows are written in JavaScript and define sub-agent roles, sequential and parallel execution, and the passing and verification of intermediate results; the agent then looks at the execution results and trace to decide its next action (such as a follow-up workflow). Only the main agent is trained. In stage 1, the first turns of 2,031 correct trajectories produced by a teacher model (GLM-5.3 Max) are distilled and used for SFT with QLoRA. In stage 2, an LLM judge filters out examples with unfounded premise injection, biased verification instructions or information-flow defects, and the remaining 712 are used for further training. All sub-agents are fixed as untrained Qwen3.5-27B.
Results
- Across 688 questions from six benchmarks (AIME 2024–2026, PolyMATH, GPQA-Diamond, HotpotQA), overall accuracy in Forced mode was 88.08%, 10.90 percentage points higher than the strongest baseline, Self-Refine (77.18%).
- Dynamic mode reached 87.94%, similar to Forced, while answering 47.1% of questions without a workflow.
- Before training, the main agent scored only 75.87% in Forced mode and 65.12% in Dynamic mode, so workflows helped only after training. Forced mode scored 100% on all three years of AIME.
- On GPQA-Diamond, Forced scored 84.85%, higher than Base Forced (77.78%) but below Self-Refine (88.38%) and AFlow (87.37%).
- On HotpotQA, Dynamic scored 84.50% and Forced 83.50%, 6–7 percentage points higher than Base + Web search (77.50%).
Limitations
Evaluation is limited to 688 questions and a single backbone (Qwen3.5-27B) paired with fixed sub-agents of the same model. Stage-2 data selection depends on an LLM judge. The authors note that tool-free baselines lack search access on HotpotQA, making direct comparison difficult. The extra compute cost and latency of running workflows could not be quantified within the provided text.
Why it matters
The work shows that training only the orchestrator, while leaving sub-models unchanged, can substantially improve performance. It has practical implications for multi-agent designs that reuse existing pretrained models as executors without retraining. Its analysis that information flow—explicitly passing and verifying intermediate results—is a key quality factor is also worth noting.
Key terms
- Orchestration
- Organizing the order, parallelism and result passing of multiple model and tool calls to solve a task.
- Forced / Dynamic mode
- Forced is a setting that requires running a workflow as the first action; Dynamic is a setting where the model chooses whether to answer directly or use a workflow.
- Premise injection
- A defect in which the main agent passes an answer-determining claim not present in the problem to sub-agents as if it were an established premise. It is one of the targets filtered out in stage-2 data selection.
Original
Authors: Ruhao Liu, Qi Li, Shizun Wang, Xinchao Wang · Posted 2026-10-25 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.