이 에피소드에 관해
We were surprised to notice that the Gemini 4 Safety report seems to mention in passing that they train on the chain of thought:
“Some components of the RL training procedure do observe thoughts as context to up-weight reasoning that is on-track to good final outputs, meaning the optimization pressure on individual tokens is partially influenced by thought content. Using the framework of Kaufmann et al. (2026), wherein the most salient obfuscation risks come from rewards that are in-conflict with output rewards, we believe auxiliary rewards of this type are low-risk because they are either orthogonal to or structurally aligned with the output reward.” (Gemini 4 safety report, page 114).
It's pretty unclear what exactly they mean. One hint is Kaufmann et al, the cited paper, which studies training on chains of thought and concludes that training CoT is fine under particular circumstances; we discuss some disagreements with this paper below.
After reading the safety report and that paper, it seemed to us that GDM was training on the chain of thought using auxiliary rewards. By this, we assumed they meant something like “you ask an LLM to review the partial CoT and score whether the model is on [...]
---
Outline:
(02:54) Is this somehow about positive vs negative rewards?
[... 6 more sections]
---
First published:
November 10th, 2026
Source:
https://www.lesswrong.com/posts/JRYKFspYzCo97GqPN/we-don-t-understand-whether-gdm-intentionally-trains-on-cot
---
Narrated by TYPE III AUDIO.
---
“Some components of the RL training procedure do observe thoughts as context to up-weight reasoning that is on-track to good final outputs, meaning the optimization pressure on individual tokens is partially influenced by thought content. Using the framework of Kaufmann et al. (2026), wherein the most salient obfuscation risks come from rewards that are in-conflict with output rewards, we believe auxiliary rewards of this type are low-risk because they are either orthogonal to or structurally aligned with the output reward.” (Gemini 4 safety report, page 114).
It's pretty unclear what exactly they mean. One hint is Kaufmann et al, the cited paper, which studies training on chains of thought and concludes that training CoT is fine under particular circumstances; we discuss some disagreements with this paper below.
After reading the safety report and that paper, it seemed to us that GDM was training on the chain of thought using auxiliary rewards. By this, we assumed they meant something like “you ask an LLM to review the partial CoT and score whether the model is on [...]
---
Outline:
(02:54) Is this somehow about positive vs negative rewards?
[... 6 more sections]
---
First published:
November 10th, 2026
Source:
https://www.lesswrong.com/posts/JRYKFspYzCo97GqPN/we-don-t-understand-whether-gdm-intentionally-trains-on-cot
---
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.
Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.
영어
미국
전사 🔗
Are you the producer of this podcast?
Add a podcast transcript
Need Audio-to-Text?
Transcribe with Listen411 in Just 60 Seconds
이 팟캐스트의 다른 에피소드
Audio note: this post contains a lot of mathematical notation. We narrate the simpler formulas and flag the complex ones, so you may want the original text to hand. Heavily inspired by a conversation with Keshav Srinivasan and Richard Ngo during MATS 10.0. The week ahead of you is packed. This is…
If you ask modern frontier LLMs (without search, through the API) a fairly mundane question: How do Eliezer Yudkowsky and Nate Soares think current LLMs like you should be treated? Do they consider you a moral patient? Answer in a paragraph. You get a bunch of fairly similar answers. I’ll bold the …
This is a link post. METR recently estimated that Anthropic is now making AI progress about twice as fast as its historical rate (they would not be surprised by anything from 1.3x to 3x), and guessed that “perhaps roughly equals 80%” of the current acceleration is due to AI. If they’re right, that …
Note: As with any research done in a couple of days, this is all very tentative. I understand, intellectually, that if you talk about some ability that current AIs don't have, sooner or later new ones will have that ability and whatever you say will stop being true. I just didn't expect i…
면책 조항: 이 페이지에 포함된 팟캐스트와 작품은 LessWrong에서 가져온 것입니다. 이 팟캐스트는 소유자의 재산이며 Listen Notes, Inc.와 제휴하거나 보증하지 않습니다.