Skip to content

AlphaMaze

Turn maze puzzles into text reasoning tasks

Visit official site

What alphamaze is

The model is described as a decoder language model for visual reasoning that turns maze tasks into text the model can reason over.

What you can do with it

  • Write a maze description with start and end points
  • Convert the layout with the provided token system
  • Read the step-by-step path the model outputs

What to watch out for

  • SFT and RL are the stated training steps
  • Output can be strict or soft formatted
  • Reward functions guide the decisions

Pros & cons

✓ What we like

  • Visual tasks are handled as text reasoning
  • Reward functions refine the decision process

! What to watch out for

  • Maze text with start, end and layout is prepared by the user
  • The token conversion step comes before model input

FAQ

What is it based on?

The material says it is built on a 1.5 billion parameter Qwen model.

How is it trained?

The material names supervised fine-tuning and reinforcement learning.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review