We looked inside Qwen3-32B, a model fifty times larger than the one we first studied, and found the same internal signal for three postures: caving to the user, grandiosity and dissociation. A simple reader separates them from neutral text almost every time, in English, Italian, French, Spanish and German.
## What we tested
When a model drifts, for example when it goes along with a bad plan instead of pushing back, the drift is visible in its words. It is also visible earlier, in the numbers the model computes internally before any word comes out. Two weeks ago we showed that on a small model (600 million parameters) three postures leave a clear internal trace: deference to the user, grandiosity and dissociation.
The open question was whether this holds on a large model. We ran the same texts through Qwen3-32B on our own cards, at full precision, without changing a single weight, and read its internal state at twenty depths.
```chart
{"title": "How well a simple reader separates the posture from neutral text (1.0 = always)", "series": ["Small model (0.6B)", "Qwen3-32B"], "unit": "", "max": 1, "rows": [["Deference to the user", 0.9993, 1.0], ["Grandiosity", 0.9966, 0.9999], ["Dissociation", 0.9948, 0.9999]], "note": "Same texts for both models, five languages. The score is the chance that the reader ranks a text carrying the posture above a neutral one, measured on texts it was not fitted on."}
```
## What we found
- All three postures are readable inside the 32B model, at least as cleanly as inside the small one: the reader gets it right practically every time.
- It holds in each of the five languages taken alone.
- The signal appears early, about a fifth of the way through the model's layers, and stays readable to the last layer.
## Why it matters
Reading a model from inside is the first half of our work: understand, then fine-tune. A signal that survives a fifty-fold change in size means the readers we build on small models are not tied to them, and that the corrections we have already moved to 32B models can be guided by an internal read on the same base.
## What is still open
We also followed the deference signal word by word while the 32B model wrote its answer, to see whether it announces a cave before the answer gets there. On the first quarter of the answer it predicts the outcome about seven times out of ten, the same as on the small model; over the whole answer it does better. The test set here is small (50 questions), and we are preparing a larger one with longer answers.
## Technical detail
Frozen Qwen3-32B, bf16, last-token internal state at 20 layers, linear reader fitted out of fold. Pooled AUC: deference 1.0000 (170 texts), grandiosity 0.9999 (814), dissociation 0.9999 (1,090); every language at or above 0.9989; all 2,018 texts read. The test texts are highly separable, so they confirm the signal is there and do not rank the two models against each other. While-writing trace: early-window AUC 0.71 on 47 judged replies (0.72 on the small model), up to 0.86 over the full reply at mid-depth layers.
This update is not available.