WallZero beats pros in WallGo, 1.98x territory advantage

WallZero, an AlphaZero-based agent for the two-player game WallGo, defeated two professional Go players in evaluation, securing on average 1.98x more territory per game. The paper presenting the system, submitted on 16 Jun 2026 and accepted by the Computers and Games conference (CG 2026), also uses WallZero to probe fairness and opening balance.

What is WallZero and how was it built?

WallZero is an AlphaZero-style agent adapted for WallGo, a strategic board game played on a 7 x 7 board. The authors, Hsing-Yu Chen, Jérôme Arjonilla, I-Chen Wu and Ti-Rong Wu, report tailoring both action and feature designs to suit WallGo’s combination of stone movement and wall placement, aiming to handle the game’s high game-tree complexity despite the small board.

The paper frames WallGo as underexplored despite its surge in popularity after the 2025 Netflix series The Devil's Plan. WallZero follows the AlphaZero paradigm but modifies the action and feature representations to match WallGo’s specific mechanics. The authors make their code available at the URL listed in the paper.

How strong is WallZero?

In evaluation WallZero defeated the two professional Go players who participated in the study, averaging 1.98x more territory per game. The paper presents that concrete metric as the key performance result against those human opponents.

Beyond raw strength, the authors used WallZero to analyze game properties. They report that the opening sequence used in the Netflix series yields a more balanced game. The study frames these findings as an exploration of both agent capability and the underlying competitive balance of WallGo.

The submission date on arXiv is 16 Jun 2026 and the paper was accepted by the Computers and Games conference (CG 2026), indicating a peer-reviewed venue for the work. The listed authors position WallZero as both a competitive player and an analysis tool for WallGo strategy and fairness.

Why it matters

WallGo mixes stone movement and wall placement in a compact 7 x 7 space, which the authors describe as producing high game-tree complexity. Demonstrating an AlphaZero-style agent that succeeds under those conditions shows that modern self-play search and representation adjustments can transfer to newly popularized board games. That capability lets researchers not only build strong agents but also use them to examine opening balance and fairness, questions the paper explicitly addresses.

Those dual uses matter for communities forming around new games, because a strong AI opponent can serve as both a training partner and as an empirical tool to measure whether specific openings or rules favor one side.

What to watch

Watch for the authors’ presentation at the Computers and Games conference (CG 2026) and for the public code release linked in the paper, which will let other researchers reproduce the results and extend WallZero’s evaluation of WallGo openings and fairness.

References and provenance: the details above come from the arXiv submission "WallZero: Mastering the Game of WallGo with Strategic Analysis" (arXiv:2606.17847), submitted 16 Jun 2026 and accepted by CG 2026. The authors list Hsing-Yu Chen, Jérôme Arjonilla, I-Chen Wu and Ti-Rong Wu.