Recorded experiment
Watch the maze test
Inspect recorded moves from the full-depth and early-exit Qwen classifiers.
These are recorded decisions from Qwen in ten new mazes. Choose a model, then step through its moves. The browser replays saved results; it does not run the model.
Loading the recorded experiment.
What this test measures
The model receives a small text grid and chooses UP, RIGHT, DOWN or LEFT. Training and evaluation use different maze layouts. Illegal moves stay in place, and each episode ends after reaching the goal or 20 actions.
The shortest-path solver supplies the scoring reference. It does not correct the model's moves. This is a generated maze environment, not Pac-Man or a real-world navigation test.
Methods and measured results · Every recorded move · Experiment code