Researchers validate LLM-upgraded Fortran code with fault injection

A team led by Evan Coleman from Old Dominion University and Yuzhong Shen published research validating the reliability of large language model (LLM)-modernized Fortran scientific software using differential fault injection. The study focused on GAMESS, a quantum chemistry package where LLM agents were used to modernize legacy Fortran code.

The researchers developed a fault-injection harness that applies identical, deterministic faults to both the original and LLM-modernized implementations. In over 2,200 test runs, the modernized code matched the original's behavior across all 200 paired fault injections, with measured fault absorption costs aligning with predicted models. The method also identified phase-dependent parallel deadlocks and false convergence under reduced precision, demonstrating its utility for ensuring robust modernization of scientific software.