一作ICLR 2026PosterUnderstanding and Improving Continuous LLM Adversarial Training via In-context Learning TheoryShaopeng Fu · Di Wang理论与学习动力学安全OpenReview