Intel / article

Verified Learning for Compiler Optimization: An LLM-Guided Architecture with Formal Control

Source: arxiv.org
Author: Dev Pratap Singh, Rong Feng, Suman Saha
Date: Published Sep 24, 2026

Context

A can propose compiler changes while a separate formal checker decides whether to accept them. In this compiler-optimization study, researchers train a model on transformations produced by an existing optimizer, Wyvern, then use Alive2 to check that generated rewrites preserve program behavior. Rejected candidates can be regenerated through the feedback loop.

The authors report no observed semantic violations in their , but the learned system matched or improved runtime on only 9.8% of benchmarks; Wyvern remained faster on most. The contribution is therefore a tested arrangement for controlling generated transformations, with correctness checks outside the model and performance assessed separately.

Key takeaways · AI-distilled

The case study is lazification in LLVM IR: a code LLM is fine-tuned on transformations produced by the Wyvern optimizer.
Correctness is enforced outside the model: Alive2 symbolically validates each candidate rewrite, and failing candidates are regenerated rather than accepted.
Wyvern itself stays faster on most benchmarks; 9.8% of benchmarks reached comparable or better runtime under the learned system, with no semantic violations observed.
The authors report that verification overhead stayed bounded and the regenerate-until-valid loop converged stably.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.

Read the source arxiv.org