Workshop: 4th Annual Trillion Parameter Workshop: Building Open AI Infrastructure, Models, and Agentic Systems for Science
Authors: Ana Gainaru (Los Alamos National Laboratory (LANL)) and Togo Odbadrakh, Angel Diaz Abreu, Cade Finney, Miguel Kirby, Kory Linton, and Denise Adorno Lopes (Oak Ridge National Laboratory)
Abstract: Scientific AI agents increasingly assist researchers in software development, workflow orchestration, data analysis, and simulation-driven discovery. However, scientific environments impose requirements beyond those of general-purpose assistants, including domain-specific knowledge integration, reproducibility, correctness verification, and continuous adaptation to evolving software ecosystems. In this paper, we present a design of scientific agents evaluated across four nuclear-fuel specialists spanning open-ended research to deterministic workflow execution. We evaluate the impact of structured knowledge vaults and reusable skills on agent performance, execution cost, and reliability and find that across five foundation models, the answer quality improves by roughly 1.2x-1.5x more than they raise execution cost. We describe a correctness-verification agent that traces cited references to validate outputs, reducing relative grading error against manual evaluation and discuss a path toward collaborative multi-agent systems combining agents that follow our design. The study provides practical guidance for building scalable, trustworthy scientific AI agents.
Back to 4th Annual Trillion Parameter Workshop: Building Open AI Infrastructure, Models, and Agentic Systems for Science Archive Listing
Back to Full Workshop Archive Listing