TL;DR: DyFlow introduces a two-level Designer–Executor architecture with dynamic operators that adaptively re-plan subgoals during execution based on intermediate feedback. This enables more generalizable and robust reasoning across diverse domains and tasks.

• Execution-adaptive workflows: Dynamically adjust reasoning processes and subgoals according to intermediate feedback
• Two-core components:
  • Designer — performs high-level task decomposition and planning
  • Executor — carries out low-level execution and tool invocation
• Cross-domain evaluation: Demonstrated effectiveness across multiple domains

git clone https://github.com/wyf23187/DyFlow.git
cd DyFlow
pip install -r requirements.txt

Create a .env file with your API keys:

OPENAI_API_KEY=your_openai_key
ANTHROPIC_API_KEY=your_anthropic_key
DEEPINFRA_API_KEY=your_deepinfra_key

DyPlanner uses a locally deployed model via vLLM. First, deploy the DyPlanner model:

# Download and deploy the DyPlanner model from Hugging Face
# Model: https://huggingface.co/wyf23187/DyPlanner
vllm serve wyf23187/DyPlanner \
    --port 8000 \

The ModelService.local() will automatically connect to this vLLM endpoint at https://localhost:8000 to get responses from DyPlanner.

For basic usage and benchmark evaluation examples, please refer to:

• scripts/run_workflow.py - Single problem workflow execution
• scripts/run_dataset.py - Batch benchmark evaluation

Available benchmarks: HumanEval, MATH, LiveBench, SocialMaze, PubMedQA

For generating training data from DyFlow execution traces, see train/.

If you find our work useful, please cite:

@inproceedings{wang2025dyflow,
  title={DyFlow: Dynamic Workflow Framework for Agentic Reasoning},
  author={Wang, Yanbo and Xu, Zixiang and Huang, Yue and Wang, Xiangqi and Song, Zirui and Gao, Lang and Wang, Chenxi and Tang, Xiangru and Zhao, Yue and Cohan, Arman and others},
  booktitle={Advances in Neural Information Processing Systems},
  year={2025}
}