Geometric Inductive Biases for Semi-Supervised Equalization: The Constellation-Aware Transformer

Decoding signals over unknown channels with minimal pilot overhead is a critical challenge in next-generation communications. Existing deep learning approaches typically rely on generic encoders that struggle to model long-range temporal dependencies or efficiently capture the channel's physical properties from scarce data. We argue that standard architectures suffer from fundamental agnostic estimation gaps, as they must implicitly learn the constellation geometry that is already known. We introduce the Constellation-Aware Transformer (CAT), a novel architecture that explicitly injects geometric inductive biases into the equalization process. CAT is composed of a stack of custom TransFIRmer blocks, which utilize an ``early interaction'' paradigm to co-process received signals and ideal constellation symbols. Each block features a split Feed-Forward Network that applies a Finite Impulse Response (FIR)-inspired filter for robust deconvolution and a parallel MLP for geometric refinement. We theoretically prove that this design creates a structural isomorphism to the optimal MIMO Wiener Receiver and eliminates the ambiguity floor inherent to blind estimators. In the challenging semi-supervised setting, CAT achieves state-of-the-art performance, significantly reducing pilot overhead compared to VAE and standard Transformer baselines.