PaperWeekly

AI Agent真的安全吗？南洋理工最新综述揭秘LLM Agent安全风险与防御策略

作者

A Survey on Trustworthy LLM Agents: Threats and Countermeasures, a study by researchers from Nanyang Technological University and Squirrel AI, categorizes Agent safety issues into internal (Brain, Memory, Tool) and external (User, Agent, Environment) modules. It provides insights for future research directions such as collaborative security mechanisms and adversarial robustness training methods.

更精细的解耦评估！VisualSimpleQA开创视觉语言大模型事实问答评测新范式

作者

VisualSimpleQA 提出了一个面向事实查询的多模态评测基准，旨在有效评估大型视觉语言模型（LVLMs）在事实问答任务中的表现。该基准采用了解耦评估框架和明确的样本难度标准，能够区分不同模型的表现，并揭示了当前前沿 LVLMs 在视觉和语言模块上的改进空间。