Turns out the fastest way to get an LLM to speed up your code might be to just ask it to, no scaffolding required.
Researchers tested two ways of using large language models to optimize parallel C code for high-performance computing. One followed the traditional playbook: translate source code into a domain-specific representation, then search a constrained set of verifiable transformations, the way established compiler frameworks and auto-tuners have long done. The other skipped the scaffolding entirely, handing the model a plain-language optimization goal and letting it work directly on the code. Tested against the PolyBench benchmark suite, the direct-prompting approach produced code that ran faster and passed correctness checks more often than the pipeline-and-schedule method.
That's a mild upset for a field that assumed LLMs need the same rigid rails as older auto-tuning tools to be trustworthy. It suggests the constraints built to keep automated optimizers safe and verifiable may be quietly capping what a model that already understands code semantics could otherwise do. The authors say this points toward rethinking how to verify LLM-generated optimizations rather than just borrowing old frameworks wholesale.
It's a small benchmark study, not a production compiler - but it's a useful reminder that porting old tooling assumptions onto LLMs isn't automatically the safer choice.