Quantum-enhanced Representations improve classical ML performance across domains

Researchers at the Institute of Experimental Particle Physics (ETP) and Institute of Theoretical Physics (ITP) at KIT, in collaboration with Imperial College London, have developed a new method that uses ideas from quantum computing to make large machine learning models smarter, without needing an actual quantum computer. The method, called QUIVER, gives machine learning (ML) models a "quantum-enhanced view" of the data they are trying to understand, helping them spot patterns they would otherwise miss.

The idea works like this: When a large ML model looks at something complicated, like a spray of particles produced inside the Large Hadron Collider (LHC), or the structure of a molecule, it usually sees only a flat list of numbers describing the object. The underlying principle is simple: take the same input and pass it through a variational quantum circuit, which acts like a different kind of lens. This lens highlights how different parts of the input are connected to each other, the kind of correlations that might be hard to see in ordinary tables of numbers. The result is a compact summary of these hidden connections, which the ML model can then use alongside its normal input.

The researchers tested the method on two very different problems, demonstrating its domain agnostic nature. The first was identifying the origin of particle jets at the LHC, a task that is essential for almost every analysis at CERN. The second was predicting a key property of small molecules, the energy gap between the highest occupied and the lowest unoccupied molecular orbitals. In both cases, adding the quantum-enhanced representation to existing state-of-the-art models made them measurably more accurate, while barely increasing their size. The improvements were consistent across many training runs, suggesting that the quantum view really does provide new information rather than just acting as extra padding.

The results are available as a preprint on arXiv and a short summary can be viewed on the project page. They will be presented at the International Conference on Machine Learning (ICML) 2026 at the AI4Physics workshop, to be held in July in Seoul, South Korea and will now undergo additional peer-review for publication in a journal. This research was carried out by master student Michael Binder and postdoc Dr. Aritra Bal.

Contact: Dr. Aritra Bal