15–18 Dec 2026
Duke University
America/New_York timezone
Artificial Intelligence for the Electron Ion Collider (2026)

Autonomous AI Experimentation: A Scientific Protocol for Agent-Driven Research in Nuclear Physics

Not scheduled
20m
LSRC B101 Love Auditorium (Duke University)

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
Trends in Data Science Trends in Data Science

Speaker

Dmitry Romanov (Jefferson lab)

Description

Autonomous research loops, in which an AI agent proposes a hypothesis, changes code, measures the result, and decides what to try next, have become one of the defining trends of 2026, from Karpathy's autoresearch to AlphaEvolve-style evolution engines. Industry reports are striking, yet they rarely address the central question of this talk: when can such a loop be effective, and when can it be trusted?
We review the method in nuclear physics, drawing on emerging theory, empirical studies of research agents, and a protocol we have used in more than ten projects at Jefferson Lab and EIC ePIC. Several problems are well documented. An agent optimizing a proxy score drifts from the goal the score was meant to measure. Errors in unchecked changes accumulate over a long run. And where no fast, objective check exists, agents still trail human experts. We review what mitigation strategies work against each.
We illustrate this with examples and results from streaming readout. On simulated ePIC frames, where signal is 0.5% of hits, per-hit classifiers plateaued at 14% background rejection; given the true interaction time, the same classifiers reject 97.5% at 99% signal efficiency, so we recast the task as finding that time. A calorimeter-assisted CNN now finds it in 99.97% of test frames, and a first tracker-only version inside EICrecon makes tracking 6.5 times faster at a cost of 12 ms per event. A pre-registered control on pure background exposed an event classifier’s perfect score as a simulation-campaign artifact; corrected, it reaches 99.4%. On JLab beam-test data, the loop raised a streaming-readout chain from 95 MB/s to 4.2 GB/s (44x) with bit-identical output, and in the production EVIO library, it found 27 latent defects and pushed cold reads to 95% of the device’s throughput limit.

Author

Dmitry Romanov (Jefferson lab)

Presentation materials

There are no materials yet.