TReC metric compares Go moves by their board-state effect, not distance
A new arXiv paper proposes Transition Representation Consistency, which scores alternative Go moves by how similarly they change a board-state embedding rather than by coordinate distance.
In plain words
When a Go-playing model picks a different move from the expert's, the usual check is how far away on the board it lands. The paper's authors instead measure what each move does to the resulting position, using a similarity score between the two state changes. Their score, TReC, is meant to capture whether the alternative move leads the game in a similar direction.
What was announced
The authors define TReC by playing both the model's move and the expert's move from the same position under game rules, then computing the cosine similarity between the two resulting changes in a board-state embedding. In experiments with a frozen KataGo network on 680 mid-to-late-game positions from five games, the spread between the highest and lowest candidate per position had a median of 0.01 to 0.08 for the direct successor-state cosine, depending on the evaluator layer, versus 0.88 to 1.33 for TReC. The team also reports that TReC tracks coordinate distance only weakly: the median within-position rank correlation between a candidate's distance to the reference and its TReC was -0.19.
What it does not show
- The evidence is the paper's abstract only; full methodology, baselines and peer review status are not available in the supplied passages.
- Experiments cover 680 positions from five games with one frozen KataGo network, so results may not generalise to other models or game phases.
- The abstract reports no downstream effect on play strength or prediction quality, so the practical value of TReC is not established by this evidence.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.