T-LLM Compiler Optimizes Code with LLM and Verification.

Large Language Models (LLMs) show great potential for high-level code transformations and optimization, but their application has been limited by the complexity of code and their inability to independently verify the correctness of their changes. To overcome these hurdles, researchers have introduced the Trusted LLM (T-LLM) Compiler, a novel framework that integrates LLM-driven code optimization with the rigor of traditional compilers and formal verification tools. This collaborative approach allows for iterative code optimization, where LLM-generated transformations are continuously checked for correctness, enabling corrective actions when necessary. Experimental results on PolyBench/C benchmarks demonstrate that the T-LLM Compiler significantly enhances code correctness, achieving an optimization accuracy of up to 83.3% and an average speedup of 26.7% compared to standard baselines. This framework represents a significant advancement in leveraging LLMs for reliable and performant code optimization.