Reproduce “Orchard: An Open-Source Agentic Modeling Framework” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Baolin Peng, Wenlin Yao, Qianhui Wu, Hao Cheng, Xiao Yu, Rui Yang, Tao Ge, Alessandro Sordoni and 6 more
Microsoft Research and collaborators introduce Orchard, an open-source framework featuring Orchard Env, a Kubernetes-native environment service that decouples agent execution from training, achieving 0.28 seconds average command latency and a 10x cost reduction. The framework enables state-of-the-art performance for open-source software engineering agents (73.0% on SWE-bench Verified), vision-language GUI agents (68.4% average success rate), and personal assistant agents (73.9% pass@3), demonstrating advanced capabilities through scalable environment-grounded reinforcement learning.
Key insights
- A decoupled, standalone, and reusable environment service like Orchard Env is critical for efficient, reproducible, and cost-effective large-scale training and evaluation of LLM agents.
- Environment-grounded reinforcement learning, when combined with sophisticated data curation and reward strategies, can enable smaller models (e.g., 4B vision-language models) to surpass larger teacher models and achieve state-of-the-art open-source performance.
- Leveraging diverse training data (multi-teacher, multi-harness, multi-source) and training methodologies, including historical experience distillation, significantly enhances agent performance, generalization, and cross-harness transferability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.