Reproduce “LLMPC: Large Language Model Predictive Control” 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
Gabriel Maher
The LLMPC (Large Language Model Predictive Control) framework integrates Large Language Models into Model Predictive Control, formally characterizing LLMs as approximate optimizers for planning tasks. This approach uses multiple LLM-generated plans and explicit objective function evaluation to enhance planning performance, achieving substantial improvements in success rates on combinatorial problems and reducing control objective gaps in continuous systems.
Key insights
- Large Language Models inherently operate as approximate optimization algorithms when generating sequential plans.
- Explicitly sampling multiple plans from an LLM and evaluating them against a defined objective function significantly reduces the performance gap between LLM planning and exact optimization.
- Iterative refinement through replanning, combined with multiple plan sampling within an MPC structure, is crucial for improving LLM planning capabilities on complex tasks with many constraints.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.