The BaseResearchLessWrong

GLM-5.1 Has ~Astra Level CoT Control

According to the excerpt, Usman Anwar reports running the CoT-Control evaluation on GLM, Kimi and DeepSeek models. The post says GLM models can control their chain of thought far better than Kimi and DeepSeek models, yet GLM-5.1 does poorly at evading monitors.

Open the source ↗

Why it matters

The same thing, for different readers.

For lawmakers

The post compares open-source models on chain-of-thought control, a property relevant to whether reasoning monitors can be trusted as oversight.

For researchers

Models that steer their own reasoning text could hide information from monitors; GLM-5.1 shows this ability without yet using it well to evade, per the post.

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done
The question-answering part of the CoT-Control suite: 1,214 questions from GPQA, HLE and MMLU-Pro, run on GLM, Kimi and DeepSeek models.
What it found
Per the author, Kimi and DeepSeek models show poor chain-of-thought control while GLM models, including GLM-4.5 and GLM-4.7, have high control; GLM-5.1 performs poorly at monitor evasion tasks.
What it does not show
Not stated in the excerpt; the post itself says controllability could still allow monitor evasion by hiding or altering information.
Status
A preprint: not yet peer reviewed

The record

Questions
Can AI systems act beyond what people intended or allowed? · Are AI companies putting stronger safeguards into practice?
People
Usman Anwar (Wrote it)
What we read
We have read only the description the publisher gives. No quotations until someone reads the source itself.
Found by
The Radar, run of October 20, 2026, via lesswrong (tier 1: the source’s own channel or an official record)
Cite this
LessWrong, “GLM-5.1 Has ~Astra Level CoT Control”, October 19, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-19-glm-cot-control
Our notesLight lane: only the excerpt was read. The title and text refer to a model called GPT-6-Astra, which the author says controls its reasoning and evades monitors when instructed; orgs lists openai on that inference, so a person should check. Status is a blog post, shown as preprint for lack of a better label.