Co-Evolutionary Multi-Modal Alignment via Structured Adversarial Evolution

Researchers introduce CEMMA, a co-evolutionary framework for improving AI safety alignment in multimodal large language models. The system uses evolving adversarial attacks and adaptive defenses to create more robust AI systems that better resist jailbreak attempts while maintaining functionality.

• →CEMMA introduces co-evolutionary alignment that moves beyond static adversarial training methods for AI safety.
• →The Evolutionary Attacker uses genetic algorithms to automatically generate sophisticated jailbreak prompts from simple seed attacks.
• →The Adaptive Defender continuously updates on synthesized hard negatives to improve robustness against evolving threats.
• →Experiments show substantial increases in red-teaming attack success rates while improving model defense capabilities.
• →The framework maintains compatibility with existing inference-time defenses and avoids excessive benign refusal rates.