Mahjong AI v3 hits 77.81% match with Transformer

The v3 project is best understood as a clean, from-scratch reconstruction of the supervised-learning stage of mahjong AI. Instead of feeding the model hand-crafted features such as shanten counts or safe-tile scores, the author arranged the game state as three token types—one round token, 37 tile tokens, and up to 128 event tokens—and let a 6-layer Transformer mix information across all of them. A key design choice is that each tile token carries its own score, so illegal discards can be masked easily and tile-specific information stays attached to the tile.

V3 is placed against Suphx (Microsoft Research Asia, 2020) and NAGA (Dwango). The author notes that Suphx uses a CNN with about 100 layers, over 100 look-ahead features, around 20 million parameters, and reinforcement learning, while v3 uses a Transformer, no derived features, about 4.82 million parameters, and imitation only. The headline 77.81% match rate is not a measure of strength and is not directly comparable to Suphx's 76.7%, because the data and counting methods differ. v3 has not been connected to a full game engine, so its playing strength remains unmeasured.