LoongFlow  by baidu-baige

A thinking and learning framework for expert-grade AI agents

Created 3 weeks ago

New!

282 stars

Top 92.7% on SourcePulse

GitHubView on GitHub
Project Summary

LoongFlow is an expert-grade AI agent development framework designed to transform professional expertise into AI productivity. It empowers agents with structured thinking and continuous learning capabilities, enabling them to tackle complex, long-range reasoning tasks with expert-level performance. The framework targets developers building sophisticated AI agents that require deep domain knowledge and advanced problem-solving skills.

How It Works

The framework centers on the PES (Plan-Execute-Summary) paradigm, guiding agents through structured thinking, deliberate execution, and reflective summarization to foster reasoning-guided improvement. This is complemented by a Multi-Structure Fusion Memory system that allows agents to synthesize experience and learn iteratively without costly retraining, facilitating "run-and-improve" evolution and enabling "jump-style" reasoning by leveraging past discoveries. This approach shifts abstraction from evolving outputs to standardizing agent thinking and learning processes.

Quick Start & Requirements

Installation requires Python 3.12+. Setup involves cloning the repository and installing dependencies using uv or conda followed by pip install -e .. LLM configuration supports OpenAI-compatible APIs (e.g., OpenAI, Google, local models via vLLM/sglang) via llm_config in task_config.yaml. Example execution commands are provided for General-Evolve and ML-Evolve agents, with a visualization server available for real-time tracking.

Highlighted Details

  • LoongFlow has demonstrated state-of-the-art (SOTA) performance, outperforming human bests on 11 mathematical challenges and surpassing AlphaEvolve on 7.
  • Within the MLE-bench, it achieved 22 Gold Medals across 40 Kaggle competitions.
  • The PES paradigm enables structured, long-horizon reasoning, while the memory system facilitates continuous, lightweight learning and experience reuse.

Maintenance & Community

The project is maintained by the LoongFlow community. Community engagement is facilitated via Discord and WeChat. Contribution guidelines are available in CONTRIBUTING.md.

Licensing & Compatibility

LoongFlow is released under the permissive Apache License 2.0, suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The provided README does not explicitly detail limitations, alpha status, or known bugs. The framework's focus is on advanced agent development for complex reasoning tasks.

Health Check
Last Commit

2 days ago

Responsiveness

Inactive

Pull Requests (30d)
19
Issues (30d)
5
Star History
284 stars in the last 27 days

Explore Similar Projects

Feedback? Help us improve.