Generator Based Inference (GBI)

C. L. Cheng, Ratan Das, Runze Li, Radha Mastandrea, V. M. Mikuni, Benjamin Nachman, David Shih, Gurpreet SingharXiv 2025 · DOI · arXiv

TL;DR

Generator Based Inference (GBI) is a new framework that generalizes Simulation-Based Inference (SBI) by allowing "forward models" to be learned directly from experimental data rather than relying solely on physics simulators. By applying GBI to resonant anomaly detection at the Large Hadron Collider (LHC), researchers can now perform high-dimensional, unbinned parameter estimation—turning a search for "new physics" into a direct measurement of signal strength and particle properties with state-of-the-art sensitivity.

What Problem Does This Paper Address?

In particle physics, discovering new phenomena usually involves comparing experimental data to theoretical simulations. Traditionally, this is done by reducing complex, high-dimensional data into simple histograms and performing a $\chi^2$ fit. While robust, this "binning" process throws away a massive amount of information.

Recent advances in Simulation-Based Inference (SBI) allow for "unbinned" analysis, but they have a major Achilles' heel: they require a near-perfect simulator. If your simulation of the background (the "boring" known physics) is slightly off, your inference about the signal (the "exciting" new physics) will be biased. This paper introduces Generator Based Inference (GBI) to bridge this gap, allowing physicists to build generators from the data itself—specifically from "sidebands" (regions where no signal is expected)—to create a more accurate, data-driven background model.

Methods

The authors propose GBI as an umbrella framework and test two specific implementations within the context of resonant anomaly detection:

  1. R-ANODE (Residual Anomaly Detection with Density Estimation): This is a non-parametric approach. It uses Normalizing Flows (NFs)—a type of generative AI—to learn the probability density of the background from sideband data and interpolate it into the signal region. It then attempts to "learn" the signal shape directly from the data by maximizing a mixture-model likelihood.
  2. GBI-PAWS (Prior-Assisted Weak Supervision): This is a hybrid approach. It uses a classifier pre-trained on a wide variety of simulated signal models but fits the final model using a data-driven background. It utilizes Conditional Flow-Matching (CFM) to generate high-quality background samples for the fit.

The key innovation is moving beyond simple classification (asking "is this an anomaly?") to full statistical inference (asking "how much signal is here, and what are its physical parameters?").

Datasets

The study uses the LHC Olympics (LHCO) 2020 community benchmark dataset.

  • Background: Generic quark-gluon scattering processes producing "dijet" events.
  • Signal: A hypothetical $W'$ boson decaying into two new particles ($X$ and $Y$), which then decay into quarks.
  • Features: Each event is described by six high-dimensional features, including jet masses and "n-subjettiness" ratios (which describe the internal shape of the particle jets).

Key Insights and Ideas

  • From Detection to Measurement: Most anomaly detection tools just give you a "score" or a p-value. GBI transforms these tools into measurement devices that provide confidence intervals for physical parameters like signal cross-section and particle masses.
  • The Power of Sidebands: By learning the background generator from the data's own sidebands, the method becomes largely insensitive to the "simulation systematic" errors that plague traditional SBI.
  • Hybrid Efficiency: GBI-PAWS demonstrates that you don't need a perfect signal simulation. By using a broad "prior" of possible signals, the model can still accurately "home in" on the true signal parameters once it sees the real data.
  • Unbinned Sensitivity: By utilizing the full 6D feature space without binning, the authors achieved a 5x improvement in sensitivity over previous iterations of the PAWS method.

Key Results

  • State-of-the-Art Sensitivity: GBI-PAWS can detect anomalies with an initial signal injection as low as $0.1\sigma$, significantly outperforming previous methods.
  • Parameter Accuracy: For the LHCO benchmark, GBI-PAWS correctly identified the masses of the $X$ and $Y$ particles (100 GeV and 500 GeV) within narrow, statistically sound confidence intervals.
  • Robustness: R-ANODE proved capable of finding anomalies without any prior signal hypothesis, though it requires more signal (starting at $\sim 1\sigma$) to reach a discovery threshold compared to the hybrid PAWS approach.

Limitations and Future Work

  • Computational Cost: R-ANODE requires training large ensembles of Normalizing Flows (20+ models) to ensure stability and estimate uncertainties, which is computationally intensive.
  • Interpolation Artifacts: The authors noted that "interpolation error"—the difficulty of perfectly predicting the background shape in the signal region based only on the sidebands—remains the primary source of bias at very low signal strengths.
  • Feature Selection: While 6 features were used here, scaling GBI to even higher-dimensional "raw" data (like full calorimeter images) remains a frontier for future research.
  • Broadening the Scope: The authors suggest GBI could be applied to other areas of physics where simulators are imperfect but "control regions" (sidebands) are plentiful.

Summary updated

Abstract

Statistical inference in physics is often based on samples from a generator (sometimes referred to as a ``forward model") that emulate experimental data and depend on parameters of the underlying theory. Modern machine learning has supercharged this workflow to enable high-dimensional and unbinned analyses to utilize much more information than ever before. We propose a general framework for describing the integration of machine learning with generators called Generator Based Inference (GBI). A well-studied special case of this setup is Simulation Based Inference (SBI) where the generator is a physics-based simulator. In this work, we examine other methods within the GBI toolkit that use data-driven methods to build the generator. In particular, we focus on resonant anomaly detection, where the generator describing the background is learned from sidebands. We show how to perform machine learning-based parameter estimation in this context with data-derived generators. This transforms the statistical outputs of anomaly detection to be directly interpretable and the performance on the LHCO community benchmark dataset establishes a new state-of-the-art for anomaly detection sensitivity.
BibTeX
@misc{cheng2025generator,
  author = {C. L. Cheng and Ratan Das and Runze Li and Radha Mastandrea and V. M. Mikuni and Benjamin Nachman and David Shih and Gurpreet Singh},
  title = {{Generator Based Inference (GBI)}},
  year = {2025},
  eprint = {2506.00119},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.00119},
  howpublished = {arXiv},
}
Figure