Reproduce “Improving Large Language Model Planning with Action Sequence Similarity” 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
Xinran Zhao, Hanie Sedghi, Bernd Bohnet, Dale Schuurmans, Azade Nova
Google DeepMind researchers develop GRASE-DC, a two-stage pipeline that improves large language models' planning capabilities by selecting contextual examples based on action sequence similarity, achieving 11-40 percentage point accuracy improvements across multiple planning tasks while requiring fewer exemplars compared to random selection methods.
Key insights
- Action sequence similarity is more effective than problem similarity for selecting relevant planning examples
- Dynamic clustering helps maintain diversity while prioritizing highly similar exemplars
- Iterative refinement with validator feedback further improves planning performance
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.