Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.34467.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

DynaFlow trains a main agent to synthesize executable, task-specific programs for coordinating fixed sub-agents sequentially or in parallel. It uses supervised fine-tuning and defect-filtered refinement, reporting improved accuracy, more verification, and better information flow across six reasoning, scientific问答, and retrieval benchmarks.

Separates orchestration learning from executor capability and provides a concrete model for designing, testing, and refining multi-agent coordination workflows.

Cited by

No citations on record.