Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong

38 pointsposted 6 hours ago
by HenryNdubuaku
(github.com)

6 Comments

cacio-e-pepe

5 hours ago

> So we did mechanistic studies on small models, Gemma 4 particularly, and found the hidden state for different layers carry meaningful self-awareness signal for various situations.

Neat! Just to make sure I understand - you trained your probe layer to take this hidden state and predict p(wrong)?

Curious to learn more. Any more info on your approach (esp the mechanistic study)?

HenryNdubuaku

3 hours ago

Correct, the study is verbose, we will compile into a neat shareable report and publish once we solve the pending caveats. Interesting username btw haha.