Can graph attention mechanisms identify which intermediate resonances are physically real versus combinatorial noise?

This is a great question at the intersection of machine learning interpretability and physics—but the provided preprints don't directly address resonance identification in, say, particle collider data or scattering experiments.

Here's what *is* in the archive that touches on related territory:

Sparse attention can reveal physically meaningful structures. The Particle Transformer model used in jet tagging at the LHC develops strikingly sparse, nearly binary attention patterns [AG-2025.11-1573]. Researchers found that this sparsity encodes genuine physics: the model identifies key jet substructure elements (like leptons in semileptonic decays) without explicit particle-ID inputs. Crucially, they showed the sparsity arises primarily from the attention mechanism itself rather than noise in the input data. This suggests that when attention "locks onto" specific features, it's often because those features carry real physical information.

Attention maps can separate signal from noise artifacts. In gravitational-wave analysis, attention visualizations reveal whether a model bases its predictions on actual signal versus instrumental glitches [AG-2025.01-521]. As models focus more on glitches, parameter estimates become biased—suggesting attention maps could flag when predictions are unreliable. This principle could transfer: if a resonance is real, attention should converge consistently across different model initializations and datasets; if it's combinatorial noise, attention may scatter or depend heavily on minor hyperparameter choices.

Graph attention networks can theoretically handle structured data, including resonance networks [AG-2024.05-2472], though the archive doesn't show this applied to distinguishing real from fake resonances.

The archive lacks studies directly comparing how attention mechanisms discriminate intermediate resonances (real, narrow-width physics objects) from statistical fluctuations in multi-body decays or scattering amplitudes—that's the gap in coverage here.