Search is seen as a pre-requisite to reasoning, where a base model proposes a few valid states and a discriminator guides states to explore. One small step in that direction is easy-to-verify explore/exploit tasks. I think cube solving is a good test-bed, and this interface allows testing multiple solution "paths" simultaneously. LLMs perform poorly! While they understand the task, they lack any heuristic towards solving it.
This interface plugs into the OpenAI API and explores different "branches" of cube solutions, giving the model 12 actions to perform on the cube (L, L', R, R', U, U', D, D', F, F', B, B'), verifying it makes a valid action, and re-prompting with the new cube.
You can automatically expands paths as well:
- The front end is based on normal-computing/branches (a visualization of code generation & verification), which was originally a fork of paradigmxyz/flux
- The back end for verifying cubes, rotating and parsing in/out of text is based on jasonrute/puzzle_cube
- The "difficult" starting cube was taken from https://cube20.org/distance20s (takes at least 20 moves to solve!)
- Interested in programatic solvers? See pglass/cube for a (non-optimal but deterministic) solver and jasonrute/puzzle_cube for training a CNN-based solver with RL
# Run React front-end in first terminal:
npm install
npm run dev
# Run Flask backend in second terminal:
cd api
pip install -r requirements.txt
python execute.py