Workshops Archive

Trigger-Gated Multimodal LLM Agents for In-situ Scientific Simulation Monitoring


Workshop: PDSW’26: The 11th International Parallel Data Systems Workshop

Authors: Hua Xu, Jaime Cernuda, and Anthony Kougkas (Illinois Institute of Technology); Ana Gainaru and Scott Klasky (Oak Ridge National Laboratory); Xian-He Sun (Illinois Institute of Technology); and Jay Lofstead (Sandia National Laboratories, University of New Mexico)

Abstract: Simulations routinely run to completion while computing the wrong answer. No production tool monitors whether a running simulation computes valid science; we call this blind spot dark waste. Multimodal LLM agents could supply the physics-aware judgment such monitoring needs, but no scientific field fits in a context window. A rendering does. Vigil is a trigger-render-reason pipeline built as an ADIOS2 plugin engine on the application's I/O path. Deterministic triggers run every timestep inline with the write. Only flagged states stream in-transit to a visualization node, where an agent drives ParaView through tool calls and returns a structured verdict. Across four codes, with zero per-application training, Vigil confirms every parameter-induced failure and dismisses a trigger deliberately fired on a healthy run. Halting at the alert saves 39--68\% of an under-resolved Taylor--Green run, and gating costs 0.05\% of run time at 256 processes.


Back to PDSW’26: The 11th International Parallel Data Systems Workshop Archive Listing

Back to Full Workshop Archive Listing