The BaseResearchLessWrong
Uncertainty in the Verbalizable Workspace
On LessWrong, Virginia Ceccatelli and coauthors present an uncertainty-estimation method for reasoning models, building four directions in hidden-state space, which they combine with token entropy into a readout called the U-lens. They report it helps spot likely wrong answers, and that longer reasoning traces tend to accompany wrong answers.
Why it matters
The same thing, for different readers.
What it means for you
Could eventually help flag when an AI's answer is more likely to be wrong, though this is early research.
For researchers
Offers an interpretability-based uncertainty signal tested on three models and seeds; the authors say its value as a broader monitoring tool is open.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- A method building four hidden-state directions (ambiguity, incompleteness, conflicting evidence, general uncertainty) in three reasoning models, paired with average token entropy, run on three seeds.
- What it found
- Per the post, the readout helps spot answers more likely to be wrong; reasoning length alone is also a good indicator, so the authors also compared traces of similar length.
- What it does not show
- The authors say whether it can become a broader monitoring tool is open, and that mental-state wording should be taken with caution.
- Status
- A report
The record
- Questions
- Are AI companies putting stronger safeguards into practice? · Can AI systems act beyond what people intended or allowed?
- People
- Virginia Ceccatelli (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 17, 2026, via
lesswrong(tier 1: the source’s own channel or an official record) - Cite this
- LessWrong, “Uncertainty in the Verbalizable Workspace”, October 17, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-17-uncertainty-verbalizable-workspace
Our notesLight lane: written from the excerpt only, cut off mid-sentence; the result on similar-length traces is unread. LessWrong post, not peer reviewed. Co-authors not named in the excerpt.