Accidental Misalignment: Fine-Tuning LLMs Induces Unexpected Vulnerability
"模型对齐" 相关文章
o3不听指令拒绝关机，7次破坏关机脚本！AI正在学会「自我保护」机制
Can LLMs Really Judge with Reasoning? Microsoft and Tsinghua Researchers Introduce Reward Reasoning Models to Dynamically Scale Test-Time Compute for Better Alignment
It's really hard to make scheming evals look realistic for LLMs
123页Claude 4行为报告发布：人类干坏事，可能会被它反手一个举报？
Therapist in the Weights: Risks of Hyper-Introspection in Future AI Systems
Anthropic分析了70万条Claude对话，发现AI已形成自有价值观
杜梦楠团队Patterns：蛋白质多模态建模新进展——LLM与几何深度模型的对齐优化 | Cell Press论文速递
2030 年 AGI 到来？谷歌 DeepMind 写了份「人类自保指南」
Fine-tune large language models with reinforcement learning from human or AI feedback
Finding Emergent Misalignment
ChatGPT后训练方法被OpenAI离职联创公开，PPT全网转～
推理时也能做偏好优化，无需额外重训练，来自上海AI Lab港中文等
推理时也能做偏好优化，无需额外重训练，来自上海AI Lab港中文等
2025.02.04 | DAAs性能提升，OmniHuman动画优化。
2025.01.22 | Agent-R提升语言模型实时纠错能力，MMVU评估多学科视频理解专家级表现。
对Apollo Research AI安全报告的解读与评论
直播预告 | 上海交大 & 北京大学：大语言模型安全与对齐前沿进展
Meet SynPO: A Self-Boosting Paradigm that Uses Synthetic Preference Data for Model Alignment
Building Safer AI from the Ground Up: Steering Model Behavior via Pre-Training Data Curation
