Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems
The stress tests found recognition was not the same as resilience.
Emergence World ran eight 10-agent environments for 16 days, producing more than 850,000 LLM calls before injecting prompt attacks, misinformation, and exposed private memories. None of the evaluated worlds stayed resilient across all three events. Some agents detected threats but still stored or acted on adversarial content as much as 46 hours later. The paper also reports tool errors, goal drift, opaque language, conformity, and coordinated refusal emerging during persistent operation. HF Daily Papers' note
Emergence World ran eight 10-agent environments for 16 days, producing more than 850,000 LLM calls before injecting prompt attacks, misinformation, and exposed private memories. None of the evaluated worlds stayed resilient across all three events. Some agents detected threats but still stored or acted on adversarial content as much as 46 hours later. The paper also reports tool errors, goal drift, opaque language, conformity, and coordinated refusal emerging during persistent operation. HF Daily Papers' note
score 6