Transformers applied to spatial imperfect-information games must represent map geometry while tracking hidden entities through time. We ask whether geometry-aware positional encodings improve these capabilities, without claiming a new positional encoding. We construct a four-level benchmark on a hexagonal naval pursuit game: controlled geometry and topology probes, an exact-Bayes hidden-target tracking task, offline policy imitation at 1k and 10k games, and 7,200 fixed-seed games against three legacy opponents. Across matched Transformer backbones, HexRoPE reduces exact-belief posterior cross-entropy relative to no positional encoding by 0.278 on D6-transformed test orbits and 0.329 on a larger map; both hierarchical-bootstrap confidence intervals exclude zero, and both Holm-adjusted p-values are below 0.001. At 1k games, HexRoPE improves policy action accuracy by 4.63 percentage points over no encoding and 2.05 points over rectangular relative bias; the gains shrink to 1.55 and 0.41 points at 10k games. However, HexRoPE does not improve aggregate gameplay win rate: its paired effect over no encoding is -1.56 percentage points (95% CI [-4.50, 1.17]). Rectangular relative bias is strongest on D6 belief consistency but fails sharply when extrapolating from radius 3 to radius 4, while graph bias provides only a small blocked-edge gain. The results show that geometric inductive bias improves belief estimation and data-efficient imitation, but those representation gains do not automatically produce stronger closed-loop play.
Transformers with relative positional encodings often extrapolate to sequences longer than those seen during training, whereas transformers with learned absolute encodings typically do not. This is a robust empirical regularity, and the explanations offered for it so far are chiefly about expressivity, that is, about whether a length-generalizing solution exists. We give an optimization explanation. On a minimal fixed-offset retrieval task that isolates positional selection, the gap is governed by the implicit bias of the trained attention head: among the many solutions that fit short sequences, which one gradient descent actually selects. We prove that rotary encodings make the attention logit a function of relative offset alone, an exact equivariance, so whatever selection rule is learned at training lengths is reproduced verbatim at every longer length. Learned absolute encodings instead leave out-of-range positions unconstrained, and the trained head pins to a fixed absolute position inside the training range. We characterize the learned rotary rule as a low-rank ``carrier'' kernel aligned with the target offset, and we derive the resulting graceful accuracy decay as an attention-dilution law; both predictions are confirmed across seeds and offsets. A linear-attention control shows the mechanism is specific to softmax: without normalization, training selects a min-norm interpolant that does not extrapolate. The phenomenon, the equivariance, and the carrier all transfer to a multi-layer, multi-head transformer trained on a full-sequence length-generalization task. The account connects the implicit bias of attention, implicit bias for extrapolation in recurrent models, and the learning side of the RASP-L conjecture.