Reproduce “Plan-over-Graph: Towards Parallelable LLM Agent Schedule” 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
Shiqi Zhang, Xinbei Ma, Zouying Cao, Zhuosheng Zhang, Hai Zhao
A framework from Shanghai Jiao Tong University enables Large Language Model agents to generate parallel execution schedules by decomposing natural language tasks into structured graphs and then optimizing plans on these representations. This approach achieves substantially higher optimal planning rates and reduces execution time compared to sequential methods, especially when LLMs are specifically trained on graph-based task data.
Key insights
- Explicit training on structured graph data using SFT and DPO significantly boosts LLMs' ability to achieve optimal parallel plans, outperforming untrained and API-based models.
- The 'Extract + Plan' paradigm, which involves converting textual queries into graphs before planning, consistently improves task performance compared to direct textual planning.
- The inherent parallelism supported by the graph-based planning framework leads to measurable reductions in task completion time, demonstrating practical efficiency gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.