A UChicago-led team won a best paper award at ICML 2026 for building an artificial intelligence data filter that helps CERN’s Large Hadron Collider make faster, smarter decisions about saving collision data, addressing severe storage limits and shifting experimental conditions without breaking strict operational rules.
The High-Stakes Filter at CERN’s Large Hadron Collider
Deep beneath the French-Swiss border, CERN’s Large Hadron Collider smashes protons together at extraordinary speeds to investigate subatomic particles and probe fundamental physics questions. However, the machine generates far more collision data than its systems can possibly save. To manage this influx, a rapid filtering mechanism called the trigger system decides in real time which collisions to store and which to discard forever.
According to computer scientist Asst. Prof. Yuxin Chen, the trigger system decides, in real time, which collisions to keep and which to throw away forever, and a discarded collision is gone for good.
If preset thresholds are too loose, the instrument overwhelms storage and readout capacity. If they are too strict, the system risks rejecting rare events that could lead to important discoveries.
From Manual Adjustments to a Self-Driving Trigger
Traditionally, experts adjusted those crucial thresholds by hand. Yet collider conditions shift over time, as beam intensity changes and background patterns drift, meaning a configuration that works initially may no longer be ideal hours later. This operational challenge led researchers to investigate whether the process could become adaptive.
Prof. David Miller from the Department of Physics described the envisioned upgrade as a self-driving trigger. By building a system that can learn and adapt to the experimental conditions, we can not only optimize our instruments more effectively and efficiently,
Miller said, we can also allow the scientists to focus on the broader questions about why and what our experiments should be measuring to make the next big discovery.
He added that making these choices under staggering constraints can mean the difference between a Nobel-Prize-winning discovery and just another day at the office.
Group-Filtered Policy Optimization in Action
To solve the drift problem, the interdisciplinary team framed the trigger not as a fixed control problem, but as a sequential decision-making task native to artificial intelligence. Ph.D. student Zixin Ding noted that the objective was to understand changes rather than merely react to them, explaining that the policy can then diagnose why the rate drifted, not simply react to the fact that it drifted.
Because the Large Hadron Collider imposes hard physical limits on storage and bandwidth, an AI model cannot afford to be loosely right. Traditional reinforcement learning methods frequently evaluate candidate actions and select the ones that appear best, but shifting conditions can occasionally render all available options poor, risking actions that are unacceptable. To counter this, Chen and his collaborators developed a new technique called Group-Filtered Policy Optimization, which determines which actions are safe enough to consider before evaluating which one is best.
ICML Recognition and Interdisciplinary Collaboration
The award-winning project stems from a broad collaboration bridging computer science and physics. Led by Chen and Ding at the University of Chicago, the research team worked alongside Miller and colleagues from the University of Michigan and Fermilab. Their work was recognized with a best paper award at the ICML 2026 Workshop on AI for Physics, highlighting how computational methods can successfully adapt to real-world scientific constraints.
