Live browser demo
Let a trained model choose the moves
Run the policy, change the walls, and see whether it reaches the goal.
A 23,300-parameter neural network reads the walls, current position and goal, then returns UP, RIGHT, DOWN or LEFT. These are actual model calls in your browser. The model can loop or fail.
Select an open cell or wall to change the maze. The dot is the model; Goal is the destination.
Loading the small model...
What this demonstrates
The network has 48 inputs, two 128-unit hidden layers and four outputs. No search algorithm or known shortest route chooses its actions. The optional rule removes illegal moves before taking the highest model score; it does not stop legal loops. Episodes end at the goal or after 20 moves.
The ten initial layouts are the saved test mazes. Editing one creates an unscored example, not new benchmark evidence. This small policy differs from the earlier Qwen maze head.
Training and measured outcomes · Earlier Qwen recordings · All live demos