The BaseResearchLessWrong

On Qwen, the J-lens readouts at the newline token reveal lies

According to the excerpt, Jimmy Liu reports that reading four open models' internal states with a method called the J-lens in pressure scenarios that give a reason to lie shows deception-related signals concentrated at the newline token after Qwen's response, which separate lies from truths better than the neighbouring tokens.

Open the source ↗

Why it matters

The same thing, for different readers.

For lawmakers

Jimmy Liu's results suggest white-box lie detection on open models is feasible but still flags some true statements as lies.

For researchers

Jimmy Liu reports a place in Qwen's activations where lies can be detected, with probes there scoring higher than at neighbouring template tokens.

How the finding was reached

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

What was done
Four models (Qwen3.6-27B, two Gemma-3 models, gpt-oss-20b) were placed in MASK pressure scenarios that give a role and a reason to lie; readouts at chat-template tokens were grouped into word families and counted after lies and after truths.
What it found
The author reports a lie-related word family at the newline token after 84% of Qwen's lies and 9% of truths (78% and 7% held out); eight families as a detector caught 86% of held-out lies and wrongly flagged 13% of truths (AUROC 0.87); probes there reached held-out AUROC 0.96 against 0.69 to 0.81 elsewhere.
What it does not show
Not stated in the excerpt; the full post was not read.
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
Jimmy Liu (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 19, 2026, via lesswrong (tier 1: the source’s own channel or an official record)
Cite this
LessWrong, “On Qwen, the J-lens readouts at the newline token reveal lies”, October 19, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-19-qwen-j-lens-newline-lies
Our notesLight lane: from the triage excerpt only, which is cut off mid-sentence. Numbers are the author's, not checked.