LLM-Based State Abstraction

Rust simulator + Python orchestration to test whether LLMs can produce useful MDP abstractions for planning (MCTS).

This project explores how large language models can induce useful state abstractions in a grid‑world environment and how these abstractions interact with classical planning. The performance‑critical simulation and representation building live in a Rust core, while Python orchestrates prompt construction, model calls, post‑processing and empirical evaluation. Together, they enable fast experiments that compare abstraction quality both by a model‑based similarity metric and by downstream planning performance using Monte Carlo Tree Search.

Highlights - Stateless Rust core with Python orchestration - LLM‑driven cluster abstractions (JSON or text representations) - Model‑based similarity and MCTS performance metrics - Composite score to rank models and prompts

Quick Links

• Thesis

• Quickstart

• Architecture

• Experiments & Results

View on GitHub: Repo