Prior Knowledge or Search? A Study of LLM Agents in Hardware-Aware Code Optimization

TL;DR

This paper investigates how large language models (LLMs) perform in hardware-aware code optimization, revealing they rely more on pretraining than on feedback or agentic strategies, with varied effectiveness across tasks.

Contribution

The study provides controlled experiments analyzing LLM behavior in code optimization, highlighting their dependence on priors over explicit feedback or structured exploration.

Findings

LLMs act as greedy optimizers in black-box settings.

Explicit input-size info has no effect on zero-shot kernel generation.

Kernel optimization performance varies with feedback and kernel size.

Abstract

LLM discovery and optimization systems are increasingly applied across domains, implementing a common propose-evaluate-revise loop. Such optimization or discovery progresses via context conditioning on received feedback from an environment. However, as modern LLM agents are increasingly complex in their structure, it is difficult to evaluate which components contribute the most, and when and how this exploration may fail. We answer these questions through three controlled experiments. Our findings: (1) In pure black-box optimization, LLMs act as greedy optimizers. (2) In zero-shot kernel generation, providing explicit input-size information has no measurable effect, models converge to the same kernel parameters regardless of size or temperature, as though the size instruction were invisible. Moreover, when tasked to perform kernel optimization for uncommon kernel sizes, performance sharply degrades regardless of the language used. (3) In feedback-loop kernel optimization, CUDA improves monotonically under iterative feedback, while TVM IR actively degrades, which demonstrates that kernel optimization degrades when models operate with low-density language. Our results conclude that LLMs in code optimization tasks highly depend on pretrained priors rather than provided feedback or agentic structure.