Go is rich in domain terminology. Terms like atari and sente give players a shorthand for positions and strategic priorities. But when a language model plays Go or comments on a game, does it actually understand those terms? A new preprint from arXiv argues this remains unclear.
The authors propose controlled evaluation tasks for language-based Go agents. The tasks focus on two specific abilities: recognizing atari, a state in which a stone or group can be captured, and assessing sente, a move that demands a response. Rather than measuring win rate alone, the tasks test whether a model can name and apply these concepts to specific positions.
The idea is to separate fluent language use from genuine understanding. A model might produce plausible commentary without being able to identify an atari or judge whether a move keeps the initiative. Because this is a single arXiv preprint, the claims are not yet peer-reviewed, and the abstract gives limited detail on the tasks' construction and results.