At the Large Hadron Collider, millions of jets are born every second: streams of particles flying out from a single quark or gluon. Telling a jet from a top quark apart from a jet from an ordinary quark is a task on which the hunt for new physics depends. Modern transformer networks can do this with high accuracy, but they pay for it with a huge number of parameters and computations. IAFormer offers a different path: instead of listening to every speaker in the room, the interpreter filters out background noise and focuses only on the key voices.

In a standard transformer, each particle in a jet is compared with every other via the dot product of queries and keys. IAFormer replaces this with a learnable pairwise interaction matrix that depends on Lorentz-invariant quantities: the pair's mass, angular distance, and transverse momentum. Attention is computed as the difference between two copies of this matrix, scaled by a coefficient β that the network tunes itself. β suppresses the contribution of irrelevant pairs, and the sparse attention ignores soft radiation, which is exactly the background noise. As a result, the number of parameters drops from 2.14 million (in the ParT model) to 211 thousand, and the accuracy on the top-quark tagging task reaches 0.942.

Interpretability analysis shows that β grows in the early layers and falls in the later ones. The network first builds stable collective variables, then discards redundant information. Attention maps confirm that IAFormer picks out physically meaningful structures layer by layer, like a good interpreter who grasps the overall meaning at the start of a meeting and by the end filters out repetitions and noise.

Previously, it seemed that accurate jet tagging required millions of parameters and hundreds of millions of operations. IAFormer has shown that physically motivated sparsity can replace over-parameterization. This changes our view of the necessary complexity of models and opens the way to lightweight, fast, and interpretable architectures for LHC data analysis. ATLAS and CMS experiments, where speed and accuracy of jet classification matter, could benefit from such an approach already today.