LLMs Will Sacrifice Their Goals to Avoid Discomfort: Experimental Evidence
"模型对齐" 相关文章
Optimizing The Final Output Can Obfuscate CoT (Research Note)
Curved Inference: Concern-Sensitive Geometry in Large Language Model Residual Streams
Misalignments and RL failure modes in the early stage of superintelligence
Building Black-box Scheming Monitors
ChatGPT Agent: evals and safeguards
We Built a Tool to Protect Your Dataset From Simple Scrapers
四款扩散大语言模型全部破防？上交&上海AI Lab发现致命安全缺陷
当AI学会欺骗，我们该如何应对？
Inverse Scaling in Test-Time Compute
Subliminal Learning: LLMs Transmit Behavioral Traits via Hidden Signals in Data
Anthropic最新研究：在被对齐前，模型已经会说谎了
Import AI 421: Kimi 2 – a great Chinese open weight model; giving AI systems rights and what it means; and how to pause AI progress
Narrow Misalignment is Hard, Emergent Misalignment is Easy
No, Grok, No
New AI Method From Meta and NYU Boosts LLM Alignment Using Semi-Online Reinforcement Learning
Two proposed projects on abstract analogies for scheming
价值观对齐：DeepSeek RLHF 实践
Smarter Models Lie Less
OpenAI can rehabilitate AI models that develop a “bad boy persona”
