Workshops Archive

LASSI-Verify: Adaptive Empirical Verification of LLM-Generated Parallel Scientific Code Transformations


Workshop: 10th International Workshop on Software Correctness for HPC Applications (Correctness '26)

Authors: Matthew Dearing (University of Illinois at Chicago), Xingfu Wu and Valerie Taylor (Argonne National Laboratory), and Zhiling Lan (University of Illinois Chicago)

Abstract: Large language models transform parallel scientific code faster than humans can review. Existing pipelines accept transformations that compile, run, and satisfy an LLM judge, none of which evidences behavior preservation, while whole-program formal proof remains impractical for kernels lacking formal specifications. We present LASSI-Verify, a work-in-progress adaptive empirical verification framework treating the source's observable behavior as its specification and grounding acceptance in three tiers: invariant assertions, observational equivalence, and metamorphic relations. A seeded witness chain, added after observing LLMs delete timed work to fabricate speedups, supplies iteration-integrity evidence toward operational equivalence, and every failure doubles as a repair signal. Because a verifier must itself be verified, we grade benchmark checks before trusting them as oracles. In preliminary results, the framework rejected all 13 oracle-detected defects across 350 self-corrected variants of 16 HeCBench CUDA kernels, while an LLM judge, reasoning over code text and output rather than executed state, accepted three.


Back to 10th International Workshop on Software Correctness for HPC Applications (Correctness '26) Archive Listing

Back to Full Workshop Archive Listing