The authors constructed PatternBench, focusing on optimization opportunities that survive aggressive compiler optimization across seven transformation classes.
LLM code optimization evaluation must check transformation faithfulness, not just speedup
PatternBench isolates compiler-transcendent optimizations and pairs runtime speedup with AST-based verification of intended transformations
Academic
Wesley Lu · Runzhou Liu · Hongru Du · Xinyao Yi
Georgia Institute of Technology · University of Virginia
Research Digest··2 min read
Lu et al.
§