Reproduce “Vote-Tree-Planner: Optimizing Execution Order in LLM-based Task Planning Pipeline via Voting” 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
Chaoyuan Zhang, Zhaowei Li, Wentao Yuan
The Vote-Tree-Planner, developed at the University of Washington, introduces a framework that optimizes the execution order and efficiency of LLM-based task planning pipelines by integrating a multi-plan voting mechanism within a tree structure. It achieved a 43% success rate without error correction and 48% with error correction in embodied AI tasks, demonstrating improved reliability and reduced LLM query redundancy compared to prior methods.
Key insights
- Aggregating multiple LLM-generated plans and weighting commands by their frequency (votes) provides a robust heuristic for selecting more reliable and executable task paths.
- Separating the initial generation of diverse commands from their optimal reordering, coupled with a unique command extraction step, significantly reduces plan redundancy.
- An integrated, vote-guided error correction mechanism within the plan tree enables efficient adaptation to execution failures without incurring repeated LLM querying costs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.