From Spins to Tokens: Ising-Aware Contrastive Event Embeddings for Model-Independent Anomaly Searches

Matchev et al. (2020) propose an elegant Ising Hamiltonian on binned data whose ground-state energy serves as an anomaly test statistic, capturing spatially correlated deviations and mitigating the look-elsewhere effect. Metzger et al. (2025) show that contrastive embeddings can preserve rare signals better than hand-crafted features, but they also uncover a key tension: the representation that best structures the background is not always the most sensitive to new physics. This project fuses these ideas by training an event encoder with a differentiable surrogate of the Ising ground-state energy as a target. Concretely, we: (i) tokenize events à la Visive et al. (2025) into physics-aware tokens (objects, subjets, tracks) and learn an embedding; (ii) build a differentiable “Ising energy” layer using spectral/variational relaxations to approximate the Matchev et al. ground-state objective; (iii) use multi-objective contrastive training that explicitly maximizes separation in the Ising statistic between background and a suite of weakly modeled synthetic anomalies, while maintaining background manifold fidelity. The novelty is to make anomaly sensitivity itself the training signal, rather than hope it emerges as a byproduct. By injecting spatial correlations via learned, data-driven couplings in the Ising layer, the model should better pick up clustered distortions typical of new resonances or EFT-induced shape changes, and it provides a principled way to tune the anomaly–background trade-off Metzger et al. identified. Impact: a representation tailor-made for discovery power, robust to background structure, with a natural test statistic and controllable look-elsewhere behavior.

References:

1. A Quantum Algorithm for Model-Independent Searches for New Physics. Konstantin T. Matchev, Prasanth Shyamsundar, Jordan Smolinsky (2020). Letters in High Energy Physics.
2. Anomaly preserving contrastive neural embeddings for end-to-end model-independent searches at the LHC. Kyle Metzger, Lana Xu, Mia Sodini, T. K. Årrestad, K. Govorkova, Gaia Grosso, Philip Harris (2025).
3. Event Tokenization and Next-Token Prediction for Anomaly Detection at the Large Hadron Collider. A. Visive, P. Moskvitina, C. Nellist, R. D. Austri, Sascha Caron Institute of Physics, U. Amsterdam, Amsterdam, The Netherlands, Nikhef, Dutch National Institute for Subatomic Physics, High Energy Physics, R. University, Nijmegen, Instituto de F'isica Corpuscular, IFIC-UVCSIC, Paterna, Spain. (2025).