Ti-Rong Wu

Latent Action Guidance can double success rates in reinforcement learning tasks by effectively harnessing the latent knowledge of pretrained language models.

WallZero not only outperforms professional players but also uncovers strategic insights that could redefine competitive play in WallGo.

Adding the T-pentomino to Tetris Block Puzzle makes the game significantly harder, quantified by a slowdown in SGAZ agent convergence.

AlphaZero learns way faster when it focuses on replaying the moments it *almost* got right, improving by an average of 83 Elo across games.