PPL
Xiaomin Li
Papers (5)
When Think-with-Image Meets Safety: What Determines Multimodal Jailbreak Robustness?AgentLens: Revealing The Lucky Pass Problem in SWE-Agent EvaluationMulti-Rollout On-Policy Distillation via Peer Successes and FailuresDataDignity: Training Data Attribution for Large Language ModelsChain of Risk: Safety Failures in Large Reasoning Models and Mitigation via Adaptive Multi-Principle Steering