How would contrastive latent spaces trained on prompt decays generalize to detecting long-lived particle signatures?

This is a thoughtful technical question, but the archive doesn't contain work bridging these two specific concepts. Let me explain why the mismatch matters, then point to what *is* available.

What the question asks

"Prompt decays" typically refer to particles that decay immediately at the collision point (within detector resolution). "Long-lived particles" (LLPs) decay far from the collision point—anywhere from centimeters to meters away. These are almost opposite regimes: one is about reconstructing structure in dense, prompt final states; the other is about detecting spatially displaced signatures.

What the archive offers instead

The contrastive learning paper [AG-2026.03-1712] trains signal-aware latent spaces on Standard Model backgrounds and diverse BSM signals, demonstrating that representations learned on *known* signal topologies can generalize to *unseen* topologies through "interpolation and extrapolation." This is powerful for anomaly detection in high-dimensional spaces, but it focuses on prompt diphoton final states.

For long-lived particles specifically, the archive has several direct resources:

• Deep learning for LLP detection [AG-2024.01-937] shows neural networks can achieve 95% signal efficiency on LLP signatures (decay lengths ~1 nanosecond) at lepton colliders, but uses supervised learning on simulated LLP samples, not contrastive pre-training.

• Constraining LLPs from prompt measurements [AG-2024.07-1469] takes the opposite approach: it infers LLP contributions by observing how they *recoil* against prompt systems, using the CONTUR framework to leverage Standard Model precision measurements.

• Decay reconstruction via contrastive learning [AG-2024.02-1078] applies supervised contrastive learning to hierarchical particle decay trees, but again in the prompt regime.

The core issue is that LLPs and prompt decays have fundamentally different detector signatures and training data: LLP networks typically train directly on displaced-vertex or timing information, while contrastive learning on prompt decays learns to separate objects by their kinematic and topological structure in the central detector. Generalizing between them would require either a novel bridge (training contrastive spaces to be "lifetime-agnostic") or domain adaptation—neither of which appears in these preprints.