MOP6336

Journals of Accelerator Conferences Website (JACoW)

JACoW is a publisher in Geneva, Switzerland that publishes the proceedings of accelerator conferences held around the world by an international collaboration of editors.

URL	https://doi.org/10.18429/JACoW-IPAC2026-MOP6336
Title	Active supervision for AGS bunch-merging with LLM-based reinforcement learning
Authors	Y. Zhao, Y. Wang Rensselaer Polytechnic Institute A. Sukhanov, J. Morris, K. Zeno, K. Brown, S. Tajne, V. Schoefer, W. Lin, Y. Gao Brookhaven National Laboratory A. Kasparian, M. Schram Thomas Jefferson National Accelerator Facility A. Edelen SLAC National Accelerator Laboratory D. Kuzovkova, E. Hamwi, J. Unger Cornell University (CLASSE) G. Hoffstaetter Cornell University T. Miceli Fermi National Accelerator Laboratory
Abstract	Radio-frequency (RF) bunch-merging gymnastics is used in the RHIC heavy-ion program to combine individual source pulses into single bunches with suitable intensity. Preserving intensity and emittance during these gymnastics requires careful coordination of the voltages and phases of RF cavities at several harmonic numbers, which is labor-intensive and fragile against machine drift. Recent work using a physics-based simulator of the Brookhaven Alternating Gradient Synchrotron (AGS) has shown that reinforcement learning (RL) can learn effective merge configurations. RL is data-intensive and requires many training interactions with the environment. Large language models (LLMs) have recently demonstrated the ability to extract patterns from large, noisy data and to integrate domain knowledge into the control loop, making them an attractive aid for tuning complex accelerator systems. However, domain adaptation (i.e., prompt engineering, finetuning, etc.) is always required for deploying LLM in the target domain and has not been investigated in particle accelerators. To fill this gap, we propose an active supervision framework in which the LLM-based teacher first transfers general control principles from human operators to the student agent. Then, the student agent further finetunes the control policy by interacting with the simulator/experiments with improved sample efficiency.
Paper	download: MOP6336.pdf
Cite	BibTeX, LaTeX, Text/Word, RIS, EndNote
Conference	17th International Particle Accelerator Conference
Series
Location	Deauville, France
Date	17-22 May 2026
Publisher	JACoW Publishing, Geneva, Switzerland
Editorial Board	Editorial Board
Online ISBN	978-3-95450-252-3
Online ISSN	2673-5350
Received	13 May 2026
Revised	20 May 2026
Accepted
Issued	20 July 2026
DOI	10.18429/JACoW-IPAC2026-MOP6336
Pages	448-451
Copyright	Published by JACoW Publishing under the terms of the Creative Commons Attribution 4.0 license. Any further distribution of this work must maintain attribution to the author(s), the published article's title, publisher, and DOI.