(Alternate titles: Belief-behavior generalization in LLMs? Assertion-act generalization?)

TLDR

Suppose one fine-tunes an LLM chatbot-style assistant to say "X is bad" and "We know X is bad because of reason Y" and many similar lengthier statements reflecting the overall worldview of someone who believes "X is bad."

Suppose that one also deliberately refrains from fine-tuning the LLM to refuse requests such as "Can you help me do X?"

Is an LLM so trained to state that X is bad, subsequently notably more likely to refuse to assist users with X, even without explicit refusal training regarding X?

As it turns out, yes -- subject to some conditions. I confirm that this is the case for two different worldviews: Catholicism and a completely invented religion.

This constitutes generalization from training on explicit normative attitudes to acting according to the implied behavioral refusals.

Thus [...]

---

Outline:

(00:19) TLDR

(01:25) Introduction

(03:50) Avoiding Data Leakage from Pretraining

(06:49) Training a Catholic LLM

(07:40) Training Data

(11:39) Catholic Refusals without Catholic Refusal Training

(15:23) Moral Refusals Are Rarer Without Some Kind of Refusal Fine-Tuning Data

(18:07) Training a Gramenist LLM

(24:10) Conclusion

---

First published:

Source:

---

Narrated by TYPE III AUDIO.