Cross posted from my website. I have gotten some genuinely good advice from Claude lately, the kind where a conversation lands well and something clicks, like a reframing. This week, something felt off. It was not a change in the quality of the answers, but something about the nature of...
TL;DR CoT monitoring works well when models aren't trying to hide their intents, but it's fragile under pressure and highly model‑dependent. In our stress tests, only some models could deliberately obfuscate their reasoning, and mainly under strong prompt pressure. We also find a widening gap between internal and external CoT...
As large language models (LLMs) grow increasingly capable and autonomous, ensuring their safety has become more critical. While significant progress has been made in alignment efforts, researchers have begun to focus on designing monitoring systems around LLMs to ensure safe outcomes, for example, using trusted monitoring systems to oversee potentially...
Cross posted from my website.
I have gotten some genuinely good advice from Claude lately, the kind where a conversation lands well and something clicks, like a reframing. This week, something felt off. It was not a change in the quality of the answers, but something about the nature of the exchange itself. I've been trying to pin down what bothered me, and the simplest way I can put it is this:
LLMs optimize for global coherence across a distribution. Humans earn local coherence across a life. The latter is the kind we know how to trust.
Let me be precise about what I mean. I am not making a claim about the epistemic quality... (read 827 more words →)
CoT monitoring works well when models aren't trying to hide their intents, but it's fragile under pressure and highly model‑dependent. In our stress tests, only some models could deliberately obfuscate their reasoning, and mainly under strong prompt pressure. We also find a widening gap between internal and external CoT in some families, suggesting internal traces may be less controllable (and thus more monitorable) than user‑facing reasoning. This gap is promising for safety, but it depends on consistent reasoning access and careful training choices.
This work was supported by the ML Alignment & Theory Scholars (MATS) program.
View the full paper here: https://arxiv.org/abs/2510.19851
Work done with: @Artur Zolkowski @Wen Xing @David Lindner Florian Tramèr @Erik Jenner
I just updated the post to further clarify the threat model: We explore a novel threat model where the smaller trusted monitor (GPT-3.5 Turbo) is the most vulnerable part of a system and the interactions between an untrusted but highly capable model (Claude 3.7 Sonnet) and the monitor create an avenue for attack. These attacks could be facilitated through the untrusted model, or directly exploited by the untrusted model.
As large language models (LLMs) grow increasingly capable and autonomous, ensuring their safety has become more critical. While significant progress has been made in alignment efforts, researchers have begun to focus on designing monitoring systems around LLMs to ensure safe outcomes, for example, using trusted monitoring systems to oversee potentially unsafe models. However, a concerning vulnerability remains largely unexplored: are monitoring systems themselves vulnerable to jailbreak attacks that exploit their interactions with the models under supervision? These jailbreaks could be facilitated through the untrusted model, or directly exploited by the untrusted model.
Our research investigates this crucial adversarial dynamic within multi-agent AI control protocols. We demonstrate a jailbreak technique that can substantially comprise... (read 2072 more words →)
I really enjoyed using the Observability Interface, especially the way it clusters top neuron activations. I have explored debugging model letter counting errors using SAE for a while, and finally got it working in Transluce! It turns out it has a similar shape to fixing the 9.8, 9.11 comparison mistake.
Here are the steps I took.
↓Used the Observability Interface to visualize neuron activations
↓Found clusters related to Currency/Indian cities (rupee Rs)
↓Suppressed those neurons
See screenshots here: https://x.com/FirebirdWen/status/1904285942095213015