Reproduce “System-1.x: Learning to Balance Fast and Slow Planning with Language Models” 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
Swarnadeep Saha, Archiki Prasad, Justin Chih-Yao Chen, Peter Hase, Elias Stengel-Eskin, Mohit Bansal
A framework is introduced that allows large language models to dynamically balance between fast, intuitive System-1 planning and deliberate, explicit System-2 planning for complex tasks. This method achieves improved planning success rates and better generalization, especially under constrained computational budgets, compared to isolated planning approaches.
Key insights
- Human-like dual-process cognition (fast intuition and slow deliberation) can be effectively applied to LLM-based planning by dynamically switching between planning modes.
- Intelligent decomposition of complex problems into easier and harder sub-goals, coupled with dynamic resource allocation, improves planning efficiency and accuracy.
- Controllability through a "hybridization factor" allows users to systematically balance the trade-off between computational cost and planning performance at both training and test time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.