$ cat wiki/papers/2026/2610.34467-dynaflow.md
DynaFlow: learning executable collaboration
TL;DR
DynaFlow trains a main agent to synthesize executable collaboration programs while keeping sub-agents fixed. The authors report improved performance across six benchmarks with the same Qwen3.5-27B backbone. (source)
Authors & Org
Ruhao Liu, Qi Li, Shizun Wang and Xinchao Wang. Affiliations are not established by the captured abstract page. First submitted 2026-10-25. (source)
Method
The main agent decides when to collaborate and writes task-specific programs for sequential or parallel execution, information exchange and verification. Execution traces inform later actions. Training uses supervised fine-tuning on distilled trajectories, then refinement on examples filtered for unsupported premises, biased verification and information-flow defects; executor models remain fixed. (source)
Results
Across six benchmarks covering mathematics, scientific question answering and multi-hop retrieval, DynaFlow reaches 88.08% overall accuracy in Forced mode, 10.90 percentage points above the strongest evaluated baseline with the same backbone. Workflow analysis finds more explicit use, verification and refinement of upstream outputs. These are author-reported results; the Forced-mode score does not establish the performance of discretionary delegation. (source)
Significance
For Agents (LLM Agents), orchestration becomes a trainable action rather than a fixed graph. Holding sub-agents fixed helps distinguish coordinator improvement from stronger executors, while Eval Harness Configuration still matters to the comparison. (source)
Open Questions
Does the gain survive a matched total inference budget, and how well does optional delegation work? The abstract does not resolve those questions. Full paper and code were not reviewed. (source)
Cite
DynaFlow: Learning Free-Form Multi-Agent Orchestration, arXiv:2610.34467, 2026. (source)