Tianzhuo Yang
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
DANGERLENS: Probing Dangerous Representations Hidden by Safety Alignment
A preprint introduces DANGERLENS, a probe that reads hidden states of language models in vocabulary space. On three Qwen1.5 scales, the authors report that checkpoints with prior harmful exposure keep more accessible hazardous concepts after the same safety alignment.