Xiaoyu Wen

LLM agents can be made robust to complex, multi-turn adversarial attacks on tool use through a co-evolutionary game, outperforming prior methods on safety and capability benchmarks.