Study finds stronger play pushes computation deeper in a chess transformer
The author's own ablations show computation migrating to deeper layers as skill conditioning rises, opposite to a common prediction; generalization to language models is untested.
Original event 2026-09-24
An interpretability study reports that the stronger a chess model plays, the more its computation concentrates in deeper layers.
The subject is Maia-3, an 8-layer chess transformer that takes player Elo as an input, so skill can be varied without changing weights. Author David Litman measured head-by-head ablations and found that as Elo rises from 700 to 2500, the causal center of mass migrates deeper, monotonically, for every move type measured, with knight forks showing the largest shift.
The direction contradicts the plausible prediction that skilled features should be computed earlier for reuse. The analysis used the author's own chessformer-lens tool on this single chess model, and whether it generalizes to language models is unknown.