From Research Question to Scientific Workflow: Leveraging Agentic AI (arXiv 2604.21910)

Forum topic · 小凯 · 2026-04-27

Summary

This arXiv paper (2604.21910) by Bartosz Balis, Michal Orzechowski, Piotr Kica, Michal Dygas, and Michal Kuszewski proposes an agentic AI architecture that bridges the gap between research questions and executable scientific workflows. While scientific workflow systems automate scheduling, fault tolerance, and resource management, scientists still manually translate research questions into workflow specifications—a task requiring both domain and infrastructure expertise. The proposed three-layer architecture consists of: a semantic layer where an LLM interprets natural language into structured intent; a deterministic layer where validated generators produce reproducible workflow DAGs; and a knowledge layer where domain experts author markdown 'skills' encoding vocabulary mappings, parameter constraints, and optimization strategies. This decomposition confines LLM non-determinism to intent extraction, so the same intent always yields the same workflow. Evaluated on the 1000 Genomes population genetics workflow using the Hyperflow WMS on Kubernetes, an ablation study of 150 queries showed skills improved exact-match intent accuracy from 44% to 83%; skill-driven lazy workflow generation reduced data transfer by 92%; and the end-to-end pipeline completed queries on Kubernetes with under 15 seconds of LLM overhead and less than $0.001 per query.

Paper Overview

Field: Machine Learning Authors: Bartosz Balis, Michal Orzechowski, Piotr Kica, Michal Dygas, Michal Kuszewski Published: 2026-04-23 arXiv: 2604.21910

Abstract

Scientific workflow systems automate execution — scheduling, fault tolerance, resource management — but not the semantic translation that precedes it. Scientists still manually convert research questions into workflow specifications, a task requiring both domain knowledge and infrastructure expertise. The authors propose an agentic architecture that bridges this gap through three layers:

Semantic layer: an LLM interprets natural language into structured intent
Deterministic layer: validated generators produce reproducible workflow DAGs
Knowledge layer: domain experts author "skills" — markdown documents encoding vocabulary mappings, parameter constraints, and optimization strategies

This decomposition confines LLM non-determinism to intent extraction: the same intent always produces the same workflow.

Evaluation

The architecture was implemented and evaluated on the 1000 Genomes population genetics workflow running on the Hyperflow WMS on Kubernetes. Results from an ablation study of 150 queries:

Skills raised exact-match intent accuracy from 44% to 83%
Skill-driven lazy workflow generation reduced data transfer by 92%
The end-to-end pipeline completed queries on Kubernetes with LLM overhead under 15 seconds and cost below $0.001 per query

---

*Auto-collected on 2026-04-27*