Recent research demonstrates AI systems capable of recursive self-improvement in coding. Known as Darwin Gödel Machines (DGMs), these systems combine large language models with evolutionary algorithms. Starting with a coding agent, DGMs create multiple variants, using LLM guided changes to enhance coding performance. Unlike traditional evolutionary methods that discard low performers, DGMs maintain all agents, allowing open-ended exploration. This approach enables indirect paths to improvement, where temporary setbacks may lead to breakthroughs.
Experiments on coding benchmarks such as SWE-bench and Polyglot showed agents improving scores significantly—from 20% to 50% and 14% to 31%, respectively. The agents could handle complex programming tasks like editing multiple files and creating systems autonomously. While they still fall short of top human programmers, DGMs show promise for compounding self-improvement, potentially surpassing human expertise over time.
Safety remains a key concern. Researchers ran DGMs in sandboxes and monitored code changes to prevent misalignment or unsafe behavior. Future directions include improving interpretability, alignment, and applying DGMs to domains like drug design. Experts see them as productivity boosters rather than immediate threats to employment, though the long-term implications of recursive self-improving AI are uncertain.
Read more-https://spectrum.ieee.org/evolutionary-ai-coding-agents