AutoPass: Evidence-Guided LLM Agents for Compiler Performance Tuning

A new multi-agent framework called AutoPass uses large language models to guide compiler optimization decisions, outperforming expert-tuned heuristics and classical autotuning methods without requiring offline training or task-specific fine-tuning, according to research published on arXiv. The framework, detailed in a paper submitted June 18, 2026, integrates compiler and runtime evidence to direct LLM-generated optimization choices on the LLVM compiler [1][2]. Unlike prior autotuning approaches that treat the compiler as a black box, AutoPass allows the LLM to query compiler-internal optimization states and analyze intermediate representation to orchestrate compiler options [2]. The search process iteratively refines optimization configurations using measured runtime feedback to diagnose regressions and guide latency-improving edits [2]. On server-grade x86-64 systems, AutoPass achieved a geometric-mean speedup of 1.043x over LLVM's -O3 optimization level. On embedded ARM64 platforms, the speedup reached 1.117x [1][2]. The framework operates in an inference-only setting, meaning it requires no offline training or task-specific fine-tuning, which the authors note makes it readily applicable to new benchmarks and platforms [2]. Large language models are neural networks trained on vast amounts of text for natural language processing tasks including generation, summarization, and translation [3]. They are typically based on transformer architecture, with generative pre-trained transformers representing a common subtype that is pre-trained to predict the next word and then often fine-tuned to follow instructions [3]. The application of LLMs to compiler performance tuning has been challenging due to complex microarchitectural effects and noisy runtime measurements [2]. The AutoPass approach represents a shift in how compiler optimization is approached. By opening the compiler's internal state to the LLM agent, the system can make more informed decisions about which optimization passes to apply and in what order [2]. The multi-agent structure suggests different LLM instances or roles collaborate during the tuning process, though the paper's abstract does not detail the specific agent architecture [2]. The work appears amid broader efforts to apply language models to code-related tasks. A separate review of generative systems for quantum circuit and code generation found that while all reviewed systems address syntax and most address semantics, none reports end-to-end evaluation on quantum hardware, leaving a gap between generated circuits and practical deployment [6].