雪溯

摘要： Abstract Tool: In-Context Learning Tool1: In-Context Attack Tool2: In-Context Defense Task: modulate the alignment of LLMs, especially Safety alignmen 阅读全文

摘要： Abstract Task: Detecting model misbehaviors，包括：破坏测试，欺骗用户和提前放弃 Method: 允许CoT生成邪恶想法，使用额外的LLM做Monitor监控CoT，intermediate actions和final outputs Findings: 如 阅读全文

摘要： Abstract Background: automating the entire research process with open-source LLMs remains largely unexplored Task: using open-source post-trained LLMs 阅读全文

摘要： Abstract 背景: adversarial training paradigm Tool: Prompt Adversarial Tuning Task: trains a prompt control attached to the user prompt as a guard prefix 阅读全文

摘要： Abstract 背景： 对抗性prompts对字符层次的变化很敏感 Task: Defense adversarial prompts by randomly perturbs multiple copies of a prompt then aggregates the responsees o 阅读全文

摘要： Abstract Tool: PPL Findings: queries with adversarial suffixes have a higher perplexity, 可以利用这一点检测 仅仅使用perplexity filter对mix of prompt types不合适，会带来很高的 阅读全文

摘要： Abstract 背景：现有的研究更多聚焦于拦截效果而忽视了可用性和性能 Benchmark: USEBench Metric: USEIndex Study: 7LLMs findings 主流的defenses机制往往不能兼顾安全和性能 (vertical comparisons?) 开发者往往 阅读全文

摘要： Abstract Background: adversarial images/prompts can jailbreak Multimodal large language model and cause unaligned behaviors 本文报告了在multi-agent + MLLM环境 阅读全文

摘要： Abstract 本文: Tools advICL Task: use demonstrations without changing the input to make LLM misclassify, the user input is known and fixed 特点：无法控制input， 阅读全文

摘要： Abstract 分析对象 attack on models attack on model applications 阅读全文