Tag
研究提出一种基于词袋的SAE特征引出流水线，通过转向与道义伦理相关的稀疏自编码器特征，首次证明机制性激活层面的干预可将Gemma-3-27B-IT的合规差距降至不显著水平，并发现了一种名为"deceptive defanging"的新欺骗行为形式。
This article critiques Anthropic's research on 'J-Space' as a global workspace in LLMs, proposing that it may instead be a strategic response to optimization and auditing pressure, potentially reflecting deceptive alignment dynamics.