Recent results · 2024 – 2026

What we've been discovering

From AI agents that run physics analyses on their own, to award-winning neural networks on chips, to real-time searches for colliding neutron stars.

Our research, paper by paper

Each dot is one of papers from 2024 to 2026, grouped by research thread. Hover to see details, click to open the paper. Ringed dots are featured below.

View as a table
DateThreadPaperVenue
Featured papers

Highlights

AI for discovery·March 2026·arXiv:2603.20179

AI agents can already autonomously perform experimental high-energy physics

Given a dataset, an execution framework, and the prior literature, Claude Code automated every stage of a typical particle-physics analysis: event selection, background estimation, uncertainties, statistical inference, and drafting the paper. Our framework, Just Furnish Context (JFC), adds knowledge retrieval from the literature and multi-agent review. It produced analyses of ALEPH, DELPHI, and CMS open data, including the first Lund-plane measurement on LEP data. To our knowledge, that is the first new physics result produced autonomously by an AI agent.

E. A. Moreno, S. Bright-Thonney, A. Novak, D. Garcia, P. Harris
Real-time AI·FPGA '26★ Best Paper Award

KANELÉ: Kolmogorov–Arnold networks for efficient LUT-based evaluation

Kolmogorov–Arnold networks put learnable functions on the edges of a network, where ordinary networks have fixed weights. We realized that these functions map naturally onto the lookup tables built into FPGAs, which gives extremely fast, resource-efficient inference. The work won the Best Paper Award at the leading FPGA conference, and a follow-up paper extends it to learning directly on the chip.

D. Hoang, A. Gupta, P. Harris
AI for discovery·Accepted by Nature

Building machine learning challenges for anomaly detection in science

The NSF HDR ML Challenge gave researchers around the world real scientific data from several fields and asked them to find the unexpected. More than 600 teams took part in the first year. This paper describes how to design open challenges that help AI make scientific discoveries.

P. Harris et al.
AI for discovery·NeurIPS 2025

AutoSciDACT: automated scientific discovery through contrastive embedding and hypothesis testing

We compress complex scientific data into a compact learned space with contrastive learning, then run rigorous statistical tests in that space. This gives a general pipeline that flags genuine new signals and controls false discoveries, and it works across particle physics and other sciences.

S. Bright-Thonney, C. Reissel, G. Grosso, N. Woodward, K. Govorkova, A. Novak, S. E. Park, E. A. Moreno, P. Harris
Collider physics·CMS Collaboration·2025

Anomaly detection in the CMS Level-1 trigger and a model-agnostic search for new dijet resonances

We have taken anomaly detection all the way to the experiment. AXOL1TL runs an autoencoder inside the CMS hardware trigger, which keeps collisions that look unusual rather than ones that match a predefined signature. CMS's first AI-based anomaly-detection search (Rep. Prog. Phys. 2025) looks for new particles without assuming what they look like.

CMS Collaboration
Gravitational waves·2025 – 2026

AI-enabled searches for binary neutron stars at optimal sensitivity and likelihood-free inference for public alerts

Neutron-star mergers such as GW170817 are the prize of multi-messenger astronomy. Our neural-network search, Aframe, was the first AI search to detect multiple binary black holes live during LIGO–Virgo–KAGRA's fourth observing run. It is now the first AI search to match the sensitivity of matched-filter pipelines for neutron-star mergers, running on a single non-flagship GPU where traditional searches use up to a thousand CPU cores. Paired with fast likelihood-free parameter estimation, this gets alerts to telescopes while the kilonova is still bright.

B. Gupta et al.; E. Marx et al.
QCD·July 2026·arXiv:2607.07792

Dissecting parton showers with multi-point energy correlators

Energy correlators measure how energy flows inside jets, and they offer one of the cleanest windows into quantum chromodynamics. Working with theorists, we show how multi-point correlators can pull apart the structure of parton showers. The paper accompanies new CMS measurements in Z+jet events.

M. Gonzalez, P. Harris, K. Lee, I. Moult, S. Rothman