Agent-Driven Automated Testing for MLIR-based AI Compilers via Semantics-Preserving Transformations

ABSTRACT

AI compilers are increasingly used to optimize and lower deep learning models across different hardware backends. MLIR provides a flexible compiler infrastructure for building such AI compilers, where different levels of abstraction are represented through extensible dialects and operations.

This project explores whether LLM-based agents can automatically derive semantics-preserving transformations for MLIR operations and use them to detect compiler correctness bugs, such as miscompilation. The student will build a prototype that analyzes MLIR operation definitions, documentation, and examples, then asks an agent to generate equivalent MLIR program variants. These variants will be checked using MLIR tools, compiler pipelines, and differential testing when possible.

The goal is to evaluate whether agent-generated transformations are valid, useful, and capable of exposing real bugs in MLIR dialects, optimization passes, or lowerings.

Research questions include:

1. Can an LLM agent infer useful semantics of MLIR operations from existing definitions and documentation?

2. Can it generate valid semantics-preserving transformations across different MLIR dialects?

3. Can these transformations help detect miscompilation or other compiler bugs?

SKILLS & EXPERIENCE REQUIRED

• Background in Computer Science or related discipline.
• Good programming skills, preferably in C++ and Python.
• Interest in LLMs, agents, or AI-assisted programming.
• Interest in compilers, AI infra, or software testing.
• Basic familiarity with LLVM/MLIR is helpful but not required.