corpus.blog
56,966 blogs · 4,417,066 posts Claim your blog
Record · corpus.blog/posts/01a0d013-2ad9-70ad-b8c7-97720d208fc0

Latent reasoning architectures would undermine CoT, our strongest oversight tool

blog.redwoodresearch.org · published 23 September 2026

A record is what was published and who pointed at it. The text of the post is not held here: read it at the source, or at its Wayback capture.

Read the post at blog.redwoodresearch.org Wayback capture
1blogs citing
First cited
24 Sept 2026
Most recent
24 Sept 2026
Rank this month
not ranked
Links from this post
65
Words captured
8,421

Cited by

Blog
In the post
Date
Astral Codex Ten
The Specter Of Neuralese Read at the source ↗
24 Sept 2026

Links from this post

Link
Host
On the Navier–Stokes Millennium Prize Problem Read at the source ↗
openai.com
AI agent collaboration Read at the source ↗
metr.org
https://www.anthropic.com/institute/recursive-self-improvement
anthropic.com
Research acceleration: The view inside OpenAI Read at the source ↗
openai.com
https://arxiv.org/abs/2507.05246
arxiv.org
https://arxiv.org/pdf/2412.06769
arxiv.org
https://arxiv.org/abs/2608.08888
arxiv.org
https://arxiv.org/abs/2412.04984
arxiv.org
https://www.lesswrong.com/posts/wnzkjSmrgWZaBa2aC/self-preservation-or-instruction-ambiguity-examining-the
lesswrong.com
https://arxiv.org/pdf/2412.14093
arxiv.org
https://www.apolloresearch.ai/science/claude-sonnet-37-often-knows-when-its-in-alignment-evaluations/
apolloresearch.ai
https://www.lesswrong.com/posts/aTcsN5ZZDnMFJvRiG/models-may-behave-worse-when-eval-aware
lesswrong.com
https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/6a0ed93f9b4a6a65994235d8_Loss_of_Oversight%20(7
cdn.prod.website-files.com
https://arxiv.org/pdf/2507.11473
arxiv.org
https://institute.deepmind.com/essays/the-case-for-reasoning-transparency/
institute.deepmind.com
https://blog.redwoodresearch.org/i/217105546/cot-may-be-hard-to-replace-with-other-interpretability-tools
blog.redwoodresearch.org
https://blog.redwoodresearch.org/i/217105546/propensity-based-arguments-may-not-be-robust-in-the-current-paradigm-but-would-be-further-undermined-by-latent-reasoning-architectures
blog.redwoodresearch.org
https://blog.redwoodresearch.org/i/217105546/appendix-a-more-on-the-necessity-argument-in-the-existing-cot-paradigm
blog.redwoodresearch.org
https://arxiv.org/pdf/2606.07157
arxiv.org
https://substackcdn.com/image/fetch/$s_!4zWE!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe6b3b08f-3af9-4a34-93c3-501b8f2aa310_1774x1082.webp
substackcdn.com
https://arxiv.org/abs/2412.06769
arxiv.org
https://substackcdn.com/image/fetch/$s_!72p1!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F63b5fb7b-b2e3-443e-bf73-90122a326018_1322x404.webp
substackcdn.com
https://substackcdn.com/image/fetch/$s_!zb1x!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b0df891-7507-4b4c-b7d3-177bea395491_1174x695.webp
substackcdn.com
https://www.lesswrong.com/posts/5guQJSqstkjgys3PE/deep-recurrent-models-are-less-robustly-cot-monitorable-than
lesswrong.com
https://blog.redwoodresearch.org/i/217105546/appendix-b-do-all-latent-reasoning-architectures-threaten-monitorability
blog.redwoodresearch.org
https://arxiv.org/pdf/2502.05171
arxiv.org
https://substackcdn.com/image/fetch/$s_!r05w!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc2de584b-5557-4aca-9c23-5805abbeaf3f_1999x1234.webp
substackcdn.com
https://www.redwoodresearch.org/blog/an-operationalization-of-opaque-serial-depth
redwoodresearch.org
https://www.lesswrong.com/posts/eRmzz8J8Qkzqvzrgg/astra-can-do-a-concerning-amount-with-no-chain-of-thought
lesswrong.com
https://x.com/merettm/status/2095023204993490967
x.com
https://arxiv.org/abs/2603.05706
arxiv.org
https://substackcdn.com/image/fetch/$s_!oa3Y!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffbb95fef-380e-4f8f-9406-bc0d2314f45f_1999x1118.webp
substackcdn.com
https://deploymentsafety.openai.com/gpt-6-astra/monitor-evasion
deploymentsafety.openai.com
https://arxiv.org/html/2511.11584v1
arxiv.org
https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c342ee809620.pdf
www-cdn.anthropic.com
Improving our alignment and security efforts Read at the source ↗
anthropic.com
https://deploymentsafety.openai.com/gpt-6-astra/monitorability-under-adversarial-conditions
deploymentsafety.openai.com
https://blog.redwoodresearch.org/i/217105546/architectures-with-auxiliary-cot
blog.redwoodresearch.org
https://transformer-circuits.pub/2024/scaling-monosemanticity/
transformer-circuits.pub
https://www.anthropic.com/research/probes-catch-sleeper-agents
anthropic.com
https://transformer-circuits.pub/2026/nla/
transformer-circuits.pub
How confessions can keep language models honest Read at the source ↗
openai.com
https://x.com/Jack_W_Lindsey/status/2095174799664283910?s=20
x.com
https://x.com/MicahCarroll/status/2095267089649192992?s=20
x.com
https://x.com/nabla_theta/status/2095084948562776415?s=20
x.com
https://x.com/NeelNanda5/status/2095533397297045716?s=20
x.com
https://blog.redwoodresearch.org/i/217105546/appendix-c-comparing-specific-interpretability-techniques-with-cot
blog.redwoodresearch.org
https://www.redwoodresearch.org/blog/proposal-for-tracking-architecture-on-monitorability
redwoodresearch.org
https://blog.redwoodresearch.org/i/217105546/absent-architectural-change-the-value-of-cot-could-likely-be-preserved
blog.redwoodresearch.org
https://arxiv.org/abs/2606.20560
arxiv.org
https://www.lesswrong.com/posts/QBuJ3suRZxrrxSTtv/does-diffusiongemma-do-latent-reasoning
lesswrong.com
The Persona Selection Model: Why AI Assistants might Behave like Humans Read at the source ↗
alignment.anthropic.com
https://arxiv.org/abs/2510.09714
arxiv.org
https://stolen-thoughts.com/
stolen-thoughts.com
Learning to reason with LLMs Read at the source ↗
openai.com
https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235/Claude%20Fable%205.1%20%26%20Claude%20Mythos%205.1%20System%20Card.pdf
www-cdn.anthropic.com
https://substackcdn.com/image/fetch/$s_!vKPT!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe9ebc3ef-cc7a-4938-91c7-674a72122dad_1296x810.webp
substackcdn.com
https://blog.redwoodresearch.org/p/ai-catastrophes-and-rogue-deployments
blog.redwoodresearch.org
https://docs.google.com/document/d/1WwsnJQstPq91_Yh-Ch2XRL8H_EpsnjrC1dwZXR37PC8/edit
docs.google.com
An alignment assessment of recent cybersecurity incidents Read at the source ↗
anthropic.com
https://docs.google.com/document/d/1WwsnJQstPq91_Yh-Ch2XRL8H_EpsnjrC1dwZXR37PC8/edit?tab=t.0
docs.google.com
How Robust Are Natural Language Autoencoders to Initialization? Read at the source ↗
turntrout.com
Verbalizable Representations Form a Global Workspace in Language Models Read at the source ↗
transformer-circuits.pub
https://www.theinformation.com/articles/secret-technique-behind-openais-astra-model-sparks-security-concerns
theinformation.com
https://epoch.ai/trends
epoch.ai