metadata
license: mit
Overview
Developed by Menlo Research, AlphaMaze is a novel model designed to enhance and assess visual reasoning in large language models (LLMs). Unlike approaches that rely on complex image generation, AlphaMaze uses a surprisingly simple task: solving text-based mazes. This requires the LLM to internally reconstruct the maze, plan its path, and strategically reset after dead ends. To further improve AlphaMaze's capabilities, we utilize the GRPO (Generalized Relative Policy Optimization) method. The AlphaMaze model itself offers a richer, more nuanced assessment of spatial understanding than traditional multiple-choice tests.
Variants
| No | Variant | Cortex CLI command |
|---|---|---|
| 1 | gguf | cortex run alphamaze-v0.2 |
Use it with Jan (UI)
- Install Jan using Quickstart
- Use in Jan model Hub:
cortexso/alphamaze-v0.2
Use it with Cortex (CLI)
- Install Cortex using Quickstart
- Run the model with command:
cortex run alphamaze-v0.2