Stefan Heimersheim
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
Activation Probes Are Reliable With a Handful of Positive Examples
According to the abstract, Riya Tyagi and Stefan Heimersheim test activation probes for spotting rare misalignment when only a few positive examples exist. On spam and honesty tasks, training with many negatives beat balanced training for 1 to 10 positives, and larger models were more sample-efficient.