Produced as part of the ML Alignment & Theory Scholars Program - Winter 2023-24 Cohort under the supervision of Evan Hubinger.

Acknowledgements: Thanks to Kyle Brady for his many contributions to this project.

Abstract.

This post argues that the performance elicited by fine-tuning an LLM on a task using a given prompt format does not usefully bound the level of performance observed when the same information is presented in a different structure. Thus, fine-tuned performance provides very little information about the best performance that would be achieved by a large number of actors fine-tuning models with random prompting schemes in parallel.

In particular, we find that we get much better results from fine-tuning gpt-3.5-turbo (ChatGPT 3.5) to play chess when the game so far is presented in a single block of SAN[1] than when the game so far is separated into a series of SAN moves presented as [...]

---

Outline:

(03:06) General Setting

(04:44) LLM Prompting is Very Brittle and Can Elicit a Range of Behaviors

(06:46) Experiment 1: Semantic Prompt Variations

(09:09) Experiment 2: Structural Prompt Variations

(11:31) LLM Fine-tuning is Brittle and Can Elicit a Range of Behaviors

(14:21) Conclusion

(16:16) What's next?

(16:47) Appendix: Prompt specifications

(17:01) Completions model:

(17:23) Chat model tournament notation:

(18:09) Chat model move-by-move notation:

The original text contained 6 footnotes which were omitted from this narration.

---

First published:
June 14th, 2024

Source:
https://www.lesswrong.com/posts/4KLHJY9sPE7q8HK8N/fine-tuning-is-not-sufficient-for-capability-elicitation

---

Narrated by TYPE III AUDIO.

“Fine-tuning is not sufficient for capability elicitation” by Theodore Chapman

LessWrong (30+ Karma)

“Fine-tuning is not sufficient for capability elicitation” by Theodore Chapman