Speaker
Description
Neural simulation-based inference (NSBI) in HEP predominantly relies on event-level likelihood or likelihood-ratio estimators. In non-trivial settings, these require expensive post-hoc profiling or marginalization over nuisance parameters to perform dataset-wide inference. We introduce PAIRS, a two-stage procedure that learns an event representation from small datasets ($N \le 2$). Under mild regularity assumptions, this results in an observable that is information-preserving and has a simple, fast composition rule, with the potential to resolve a fundamental scaling bottleneck for NSBI at the LHC. Validated on several benchmarks, including a resonance search inspired by LHC bump hunts, PAIRS empirically matches or outperforms dataset-wide training at a fraction of the compute for a variety of problems and offers a practical path to enable amortized inference over large datasets.