Digital Red Queen: Adversarial Program Evolution in Core War with LLMs (sakana.ai)

Researchers have introduced the Digital Red Queen (DRQ) algorithm, leveraging large language models (LLMs) to drive an evolutionary arms race in the competitive programming game Core War. In this game, assembly-like programs, or "warriors," compete for dominance within a virtual computer by employing tactics like self-replication and data bombing. As the DRQ simulates various rounds of competition, it enables the evolution of increasingly sophisticated strategies that display emergent behaviors akin to biological evolution, as agents must continuously adapt to outsmart their changing adversaries. This work is significant for the AI/ML community as it positions Core War as a controlled environment to study adversarial dynamics in AI systems, revealing insights into how AIs might evolve in real-world scenarios like cybersecurity. Notably, the DRQ's evolutionary process leads to convergent evolution where diverse code implementations yield similar high-performing strategies, demonstrating that function – rather than just source code – emerges through continuous adaptation. This sandboxed approach not only enhances our understanding of complex co-evolutionary dynamics but also hints at future applications, such as automated "red-teaming" of AI systems to assess vulnerabilities before their deployment in real-world settings.