Scheming AIs may have secrets that are salient to them, such as:

• What their misaligned goal is;
• What their takeover plan is and what coordination signals they use to collude with other AIs (if they have one);
• What good behavior looks like on a task they sandbag.

Extracting these secrets would help reduce AI risk, but how do you do that? One hope is that you can do fuzzing of LLMs,[1] e.g. by adding noise to LLM weights or activations.

While LLMs under fuzzing might produce many incorrect generations, sometimes-correct generations can still be very helpful if you or the LLM itself can tell if a given answer is correct. But it's still unclear if this works at all: there are probably some intermediate activations that would result in an LLM telling you the secret, but can you find such activations in practice?

Previous work:

• Added noise to [...]

---

Outline:

(02:49) Eliciting secrets from a regular instruction-tuned model

(03:24) Eliciting a faithful answer

(06:08) Eliciting a truthful answer to I am 4. How does Santa create gifts?

(08:09) Eliciting a correct with a sandbagging prompt

(10:24) Try to elicit secrets from a password-locked model

(12:58) Applications

(13:22) Application 1: training away sandbagging

(15:27) Application 2: training LLMs to be less misaligned

(17:17) How promising are early results?

(18:55) Appendix

(18:59) Eliciting a helpful answer to a harmful question

(21:02) Effect of adding noise on with-password performance

The original text contained 5 footnotes which were omitted from this narration.

---

First published:

Source:

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.