Phaedrus is a novel compiler-assisted deep learning framework that accurately predicts how applications will behave with different inputs, enabling better optimization strategies.

• Addresses the challenge of highly variant application behavior across different inputs
• Combines lightweight generative models with LLMs to predict dynamic program behaviors
• Enables more effective code and memory layout optimizations without running multiple profiling instances
• Creates tailored optimization decisions specific to program profiles

This research significantly advances software engineering by allowing developers to optimize applications more efficiently, reducing the need for extensive profiling across numerous input scenarios.

Phaedrus: Predicting Dynamic Application Behavior with Lightweight Generative Models and LLMs