LessWrong (Curated & Popular)

“Optimizing The Final Output Can Obfuscate CoT (Research Note)” by lukemarks, jacob_drori, cloud, TurnTrout

31 Jul 2025

Audio

Description

Produced as part of MATS 8.0 under the mentorship of Alex Turner and Alex Cloud. This research note overviews some early results which we are looking for feedback on. TL;DR: We train language models with RL in toy environments. We show that penalizing some property of the output is sufficient to suppress that property in the chain of thought also, even when that property is relevant to task completion. For example, when we penalize a model for mentioning in its output that it completed a task via a certain form of cheating, its reasoning also omits this fact. We therefore suspect that training against an output monitor is sufficient to cause obfuscated CoTs in more realistic settings. Introduction Chain of thought (CoT) supervision appears in many control and scalable oversight protocols. It has been argued that being able to monitor CoTs for unwanted behavior is a critical property [...] ---Outline:(00:56) Introduction(02:38) Setup(03:48) Single-Turn Setting(04:26) Multi-Turn Setting(06:51) Results(06:54) Single-Turn Setting(08:21) Multi-Turn Terminal-Based Setting(08:25) Word-Usage Penalty(09:12) LLM Judge Penalty(10:12) Takeaways(10:57) AcknowledgementsThe original text contained 1 footnote which was omitted from this narration. --- First published: July 30th, 2025 Source: https://www.lesswrong.com/posts/CM7AsQoBxDW4vhkP3/optimizing-the-final-output-can-obfuscate-cot-research-note --- Narrated by TYPE III AUDIO. ---Images from the article:

Transcription

This episode hasn't been transcribed yet

Help us prioritize this episode for transcription by upvoting it.

0 upvotes

🗳️ Sign in to Upvote

Popular episodes get transcribed faster

Other episodes from LessWrong (Curated & Popular)

Transcribed and ready to explore now

"Ada Palmer: Inventing the Renaissance" by Martin Sustrik

28 Jan 2026

LessWrong (Curated & Popular)

"The inaugural Redwood Research podcast" by Buck, ryan_greenblatt

27 Jan 2026

LessWrong (Curated & Popular)

"Canada Lost Its Measles Elimination Status Because We Don’t Have Enough Nurses Who Speak Low German" by jenn

26 Jan 2026

LessWrong (Curated & Popular)

[Linkpost] "“The first two weeks are the hardest”: my first digital declutter" by mingyuan

20 Jan 2026

LessWrong (Curated & Popular)

"Why we are excited about confession!" by boazbarak, Gabriel Wu, Manas Joglekar

19 Jan 2026

LessWrong (Curated & Popular)

"Backyard cat fight shows Schelling points preexist language" by jchan

16 Jan 2026

LessWrong (Curated & Popular)

View all episodes from LessWrong (Curated & Popular)

Comments

There are no comments yet.

Please log in to write the first comment.

Report any issue

LessWrong (Curated & Popular)

“Optimizing The Final Output Can Obfuscate CoT (Research Note)” by lukemarks, jacob_drori, cloud, TurnTrout

This episode hasn't been transcribed yet

Other episodes from LessWrong (Curated & Popular)

"Ada Palmer: Inventing the Renaissance" by Martin Sustrik

"The inaugural Redwood Research podcast" by Buck, ryan_greenblatt

"Canada Lost Its Measles Elimination Status Because We Don’t Have Enough Nurses Who Speak Low German" by jenn

[Linkpost] "“The first two weeks are the hardest”: my first digital declutter" by mingyuan

"Why we are excited about confession!" by boazbarak, Gabriel Wu, Manas Joglekar

"Backyard cat fight shows Schelling points preexist language" by jchan

Sign in to Audioscrape

Share this moment