Why does binary attention structure better suppress combinatorial noise than continuous learned weights?

I appreciate the precise question, but I need to be direct: the provided preprints don't address why binary attention structures suppress combinatorial noise better than continuous learned weights.

Here's what the archive *does* contain:

The closest relevant work is [AG-2025.11-1573], which investigates *why* sparse, nearly-binary attention emerges in the Particle Transformer (ParT) at the LHC. The paper shows that binary sparsity arises primarily from the attention mechanism itself during training on jet-tagging tasks, and that this structure helps identify physically meaningful correlations (like leptons in top decays). However, this describes an *empirical observation* of sparsity in a specific physics domain—it doesn't theoretically explain noise suppression or compare binary-by-design vs. continuous learned weights.

[AG-2026.01-998] on symmetry breaking in transformers discusses how breaking rotational symmetries in attention can improve optimization and interpretability, but doesn't frame the question in terms of combinatorial noise or binary vs. continuous weight regimes.

The quantum attention papers ([AG-2024.03-1440], [AG-2024.01-1910], [AG-2024.05-2209]) explore hard attention mechanisms (which are discrete/binary-like) but don't provide comparative noise analysis against continuous alternatives.

Your question seems to be asking about a theoretical principle in neural network design or information theory that isn't directly covered in these preprints.