SkinDeepRESEARCHSteve Seguin

Research results

From a decision to a finished task

A correct-looking move is not enough. The agent needs to reach the goal.

Test results

Generated task: ten 4 × 4 mazes, reused for exploration. Layouts and protocol · Maze generator.

First recorded maze, before and after one move. P = player; G = goal; # = wall. This policy reached the goal in 6 moves on this example.

The small network reads walls, player and goal, then returns one of four moves. It does not write directions in text.

MethodGoals reached
256 training states0 / 10
256 states + forbid wall moves2 / 10
4,992 training states5 / 10
4,992 states + forbid wall moves7 / 10
Shortest-path search10 / 10

More examples helped. Forbidding moves through walls helped further, reaching 7/10 goals. The remaining failures were legal moves repeated in loops.

This is a small, separate network, not a Qwen early-exit result. No Qwen layer-skipping or timing saving is claimed here.

Training and the original failed run

The network maps 48 grid values through two 128-unit hidden layers to four action scores. A search algorithm supplies training labels. More training states come from the same 32 training layouts; the ten evaluation layouts are different but previously inspected.

The optional legal-move rule sees walls and boundaries, not the correct path. Search remains the stronger baseline on these simple fully visible mazes.

The original Qwen policies both reached 0/10 goals. The existing browser replay shows those original failures, not the later 7/10 policy.

Replay the original Qwen failures · Later training results and trajectories · Follow-up episode records