Workshop: 21st Workshop on Workflows in Support of Large-Scale Science (WORKS26)
Authors: Nieb Hasan Neom (Missouri University of Science and Technology), Manas Ranjan Gartia (Louisiana State University), and Md Arifuzzaman (Missouri University of Science and Technology)
Abstract: Trained scientific prediction models are often released as code, notebooks, or command-line tools requiring programming and exact schema knowledge, limiting access for domain researchers. Natural-language front ends can reduce this barrier, but standard LLM function calling provides a tool interface rather than workflow semantics for accumulating, validating, executing, and revising scientific inputs. We present guarded, state-aware orchestration for schema-constrained scientific model execution: the workflow collects inputs over multiple turns, constrains tool arguments with a typed schema, blocks incomplete execution, narrates results, and defines correction, what-if steering, and reset semantics. We instantiate the workflow for protein-corona prediction with a backend of 178 per-protein classifiers achieving 87.13% AUC-ROC. Evaluation using a cloud-hosted LLM produced 98.5% feature-extraction accuracy over 300 cases; a separate reliability evaluation of six interaction scenarios over 50 nanoparticle configurations passed all 300 scenario evaluations. These results demonstrate a testable extension of LLM function calling to scientific workflow orchestration.
Back to 21st Workshop on Workflows in Support of Large-Scale Science (WORKS26) Archive Listing
Back to Full Workshop Archive Listing