LLMs achieve superoptimization for assembly programs, outperforming compilers

Researchers have developed SuperCoder, a system that uses large language models (LLMs) to optimize assembly programs beyond the capabilities of standard compilers. A benchmark dataset of over 8,000 assembly programs was created to test 23 LLMs, with Claude Opus 4 achieving a 51.5% success rate. After fine-tuning with reinforcement learning, a Qwen2.5-Coder-7B-Instruct model, named SuperCoder, reached a 95.0% correctness rate and a 1.46x speedup over gcc -O3, demonstrating LLMs' potential in program performance optimization. AI

IMPACT Demonstrates LLMs' capability to surpass traditional compiler optimizations for assembly code, potentially accelerating software performance improvements.

RANK_REASON Research paper detailing a new method for program optimization using LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Assembly Programs
• Claude Opus 4
• gcc -O3
• large-language models
• qwen2.5 coder 7b instruct
• SuperCoder

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →