Reproduce “Learning adaptive planning representations with natural language guidance” 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
Lionel Wong, Jiayuan Mao, Pratyusha Sharma, Zachary S. Siegel, Jiahai Feng, Noa Korneev, Joshua B. Tenenbaum, Jacob Andreas
The paper introduces Ada (Action Domain Acquisition), a framework that automatically constructs task-specific planning representations by leveraging task-general background knowledge from large language models and an iterative bi-level planning and refinement process. It achieved 100% success on complex compositional tasks in Mini Minecraft and a 79% success rate on ALFRED household tasks, significantly outperforming various baselines.
Key insights
- Explicitly learning and leveraging a library of high-level actions is critical for robust long-horizon planning, outperforming methods that rely on LLMs for direct plan or subgoal generation.
- LLMs are effective as generative knowledge sources for *proposing* structured planning representations (operators, goals), but require structured reasoning and environmental verification to ensure accuracy and grounding.
- A neuro-symbolic approach, integrating LLMs for knowledge acquisition with symbolic planners for structured reasoning and verification, offers a principled pathway for building adaptable planning agents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.