Zhenghan Zhou
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
Why Static Probes Fail: Context Determines Deception Feature Composition
Yao Huang and co-authors argue that detecting AI deception with a single linear probe in a model's activations does not generalize across settings. They propose a multi-direction method that adapts to each setting from about 10 examples, and report better detection than single-probe baselines on two open-weight models.