Beyond Static Evaluation: Co-Evolutionary Mechanisms for LLM-Driven Strategy Evolution in Adversarial Games

Researchers introduce FAMOU, a framework that uses co-evolutionary mechanisms to improve LLM-driven strategy development in adversarial multi-agent games, addressing the challenge of evaluation landscape shifts through evaluator co-evolution, hierarchical deep evaluation, and weakness pressure. The system achieved first place in hardware rounds and third in simulation at the AAMAS 2026 Maritime Capture-The-Flag competition, demonstrating that code-level evolution can generate novel algorithmic innovations.

FAMOU addresses a critical limitation in applying large language models to adversarial game development: the moving target problem. Traditional evaluation methods fail in multi-agent environments because as strategies improve, the competitive landscape shifts, rendering static evaluators obsolete. This research demonstrates how three coordinated mechanisms—incorporating champion strategies into opponent pools, replacing noisy evaluations with statistically robust assessments, and dynamically prioritizing difficult opponents—create a feedback loop that sustains iterative improvement.

The work builds on established paradigms like OpenEvolve and ShinkaEvolve, extending code-evolution capabilities into genuinely adversarial domains. Prior research focused on single-agent optimization or cooperative settings where the evaluation landscape remains relatively stable. FAMOU's innovation lies in recognizing that competitive multi-agent scenarios require adaptive evaluation frameworks, not just better mutation operators.

The practical validation is substantial. The system achieved 0.526 combined score on the MCTF 3v3 maritime task and 61.7% win rate against unseen opponents, outperforming baselines across multiple LLM backbones. More significantly, the evolved strategies independently discovered sophisticated algorithmic structures—lookahead search and adaptive interception—that weren't present in seed strategies. This suggests LLM-driven code evolution can generate non-trivial tactical innovations rather than merely optimizing existing approaches.

The competition placements validate real-world transferability beyond simulation environments. This outcome matters for autonomous systems development, where adversarial robustness is critical. The open-source release enables broader research into co-evolutionary mechanisms, potentially influencing how AI systems are evolved for competitive applications across robotics, finance, and cyber-physical systems.

• →Co-evolutionary evaluation mechanisms enable continuous strategy improvement in adversarial games by preventing evaluation landscape stagnation.
• →FAMOU generated novel algorithmic innovations like lookahead search through code-level evolution without explicit programming.
• →The system achieved measurable competition success with 61.7% win rate against unseen opponents and real-world hardware validation.
• →Hierarchical evaluation and weakness pressure mechanisms proved individually critical to overall performance through ablation studies.
• →Open-source availability suggests co-evolutionary frameworks may become standard for adversarial AI system development.