Research collection
#activation-probing
3 papers tagged Activation probing. Each paper states what it shows and what it does not prove.
3 results
A method finds feature changes that recur across prompts
RSS finds recurring feature changes and helps rank failures in two frozen safety readouts, with no test of generated answers.
A text-only test traces watermarked training into student models
WM-Basin detects watermarked distillation from text, outperforming token checks in tested math-model evaluations.
A probe reads vulnerability signals inside language models
Reva reads language models’ internal signals and matches or exceeds multi-call detectors on PrimeVul’s held-out C/C++ function tests.
