Reasoning Models Struggle to Control their Chains of Thought (Mar 2026)

Описание: Title: Reasoning Models Struggle to Control their Chains of Thought (Mar 2026)
Link: http://arxiv.org/abs/2603.05706v1
Date: March 2026

Summary:
This paper introduces the CoT-Control evaluation suite to measure 'CoT controllability'—the ability of reasoning models to follow instructions or constraints within their intermediate chain-of-thought (CoT). The study finds that frontier models struggle significantly more to control their CoT than their final output, and that this controllability decreases with more RL training, longer reasoning traces, and increased problem difficulty. These results suggest that CoT monitoring remains a promising and robust tool for AI safety as models are currently unable to reliably manipulate their internal reasoning to evade oversight.

Key Topics:
Chain-of-thought (CoT)
AI Safety
Model Controllability
Monitorability
Reasoning Models
Reinforcement Learning (RL)
Situational Awareness

Chapters:
00:00 - Introducing Key Findings
01:46 - Measuring Hidden Stealth
03:07 - Applying Proxy Constraints
04:15 - Stress Testing Cognitive Load
05:22 - Dissecting the Output Gap
07:09 - Identifying Meta Discussion Traps
08:34 - Analyzing Scaling Trends
09:27 - Limits of Reasoning Length
10:43 - Impact of RLVR Training
12:28 - Running Adversarial Tests
14:15 - Evaluating Red Teaming Results
15:14 - Probing Latent Capabilities
17:03 - Maintaining Future Vigilance
18:18 - Engineering Cognitive Transparency

Stock video credits:
José Alfredo Munguía Lira - https://www.pexels.com/@rectorretro
Silviu Din - https://www.pexels.com/@silviu-din-16...
Google DeepMind - https://www.pexels.com/@googledeepmind
Bedrijfsfilmspecialist.nl - https://www.pexels.com/@bedrijfsfilms...
Pressmaster - https://www.pexels.com/@pressmaster
Adis Resic - https://www.pexels.com/@adis-resic-29...
Pixabay - https://www.pexels.com/@pixabay
Colors Motion Graphics - https://www.pexels.com/@colors-motion...
cottonbro studio - https://www.pexels.com/@cottonbro
olia danilevich - https://www.pexels.com/@olia-danilevich
Max Fischer - https://www.pexels.com/@max-fischer
Cyriac von Czapiewski - https://www.pexels.com/@cyriac-von-cz...
Dan Cristian Pădureț - https://www.pexels.com/@paduret
fauxels - https://www.pexels.com/@fauxels
tunnel motions - https://www.pexels.com/@tunnelmotions
Ketut Subiyanto - https://www.pexels.com/@ketut-subiyanto
Colin Jones - https://www.pexels.com/@larchmedia
Vlada Karpovich - https://www.pexels.com/@vlada-karpovich
Pavel Danilyuk - https://www.pexels.com/@pavel-danilyuk
Soumya - https://www.pexels.com/@soumya-1446957
Yaroslav Shuraev - https://www.pexels.com/@yaroslav-shuraev
StefWithAnF - https://www.pexels.com/@stefwithanf-1...
Anthony 🙂 - https://www.pexels.com/@inspiredimages
Mikhail Nilov - https://www.pexels.com/@mikhail-nilov
Pachon in Motion - https://www.pexels.com/@pachon-in-mot...
Engin Akyurt - https://www.pexels.com/@enginakyurt
crazy motions - https://www.pexels.com/@crazy-motions...
Caleb Oquendo - https://www.pexels.com/@caleboquendo
Anete Lusina - https://www.pexels.com/@anete-lusina
Kindel Media - https://www.pexels.com/@kindelmedia
Tiger Lily - https://www.pexels.com/@tiger-lily
Trippy Lagoon - https://www.pexels.com/@trippy-lagoon...
Nino Souza - https://www.pexels.com/@ninosouza
Tima Miroshnichenko - https://www.pexels.com/@tima-miroshni...