karinemellata 的提交内容

Alignment is not free: How model upgrades can silence your confidence signals

We used sparse autoencoders to explain LLM moderation flags of violent threats