AUBIN plays with a 2-step lookahead tool, the way an agent uses a calculator. The tool scores every move; clear moves are played directly and on close calls (34% of all moves) AUBIN makes the decision. Over 10 fixed seeds it reached the 2048 tile in 8 games. This is the full replay of its best game, regenerated from the seed and the recorded moves.
How it works: the tool looks two steps ahead (every possible new tile, then the best follow-up move) and scores each move; when the top two scores are within 1.0,
AUBIN gets a choice question with each candidate's lookahead value, game-over risk, merge points and empty cells, and picks the move. Model: AUBIN-12B (Gemma-4-12B + AUBIN LoRA)
on a free Kaggle T4, 0.26 s per move on average. On close calls AUBIN agreed with the tool's top move 93% of the time;
the tool alone scored slightly higher on average, so the win over the baselines comes from AUBIN using the tool, not from overriding it.
Earlier version without the tool: mean 6,309 over 5 games.