Workshop: AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC
Authors: Kevin Menear, Olivia Hull, Jon Rood, and Daniel Zalkind (National Laboratory of the Rockies)
Abstract: LLM Fortran-to-C++ translation is usually evaluated by compilation success, similarity metrics, or tests on isolated functions. However, none of these establishes substitutability: whether a translated function can replace the original in production without changing results. Reported here are two migration campaigns held to that standard by VIT (Verified Incremental Translator). VIT generates the interoperability code, and the LLM writes only the algorithmic function body. Each body passes two bit-identical gates, replay against captured production state and whole-simulation baselines. ROSCO, a wind-turbine controller, was fully translated to C++; 2.56 million nontrivial outputs match exactly across 27 scenarios. In OpenFAST’s AeroDyn, ten times ROSCO’s size, 33 of 507 procedures are translated under the same gates and tool. Of fourteen injected defects, three pass both gates. Thus, bit-identity certifies agreement on exercised inputs, not equivalence of two implementations, and as translation becomes inexpensive, the verification oracle becomes the limiting factor in migration.
Back to AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC Archive Listing
Back to Full Workshop Archive Listing