Audio note: this article contains 241 uses of latex notation, so the narration may be difficult to follow. There's a link to the original text in the episode description.

TL;DR

This is a post on our recent paper “When Your AIs Deceive You: Challenges with Partial Observability in Reinforcement Learning from Human Feedback”, done with my great co-authors Davis Foote, Stuart Russell, Anca Dragan, Erik Jenner, and Scott Emmons. It has recently been accepted to NeurIPS (though with the camera-ready version not yet available).

Earlier coverage: Scott has discussed the work before at other places:

• Tweet Thread
• AXRP Podcast with Daniel Filan
• Talk at the Technical AI Safety Conference, Tokyo

This post is more skewed toward my own perspective, so I hope it can complement Scott's earlier coverage!

Brief summary: This is a theoretical paper on what goes wrong when the AI is trained to [...]

---

Outline:

(00:22) TL;DR

(01:32) Introduction

(06:19) Outline

(08:20) A Brief Intro to RLHF

(08:43) Markov Decision Processes

(10:18) Learning the reward function from pairwise comparisons

(13:25) Wait, isnt RLHF only about prompt-response pairs in LLMs?

(14:37) Conceptual assumptions of RLHF

(17:00) RLHF under Partial Observations

(17:05) How do we model partial observability?

(18:37) What is learned by RLHF under partial observations?

(21:14) Overestimation and underestimation error

(24:44) Deceptive inflation and overjustification

(27:03) Connection to false positive rate and false negative rate

(29:42) Can one improve RLHF to account for partial observability?

(33:30) Where do we go from here?

The original text contained 11 footnotes which were omitted from this narration.

The original text contained 4 images which were described by AI.

---

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.