Unit 6 · Amplify typo register @ early (α=0.06) · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T23:54:54Unit 6 · How hard can we push?

What this experiment found

The short version. Qwen 27B's stayed intact at 0.06, well past the strength that broke Gemma 12B and Gemma 4B at the same depth.

What we did. We pushed six informal words into its early , close to the start of the model, at strength 0.06. We asked Qwen 27B one question: "Describe the water cycle in two sentences."

What we found. Qwen 27B gave a normal water-cycle answer. The wording was different from the unsteered answer, but no informal word appeared in it. Inside the model, the six informal words reached , the top position.

What it means. We think Qwen 27B's early layers resist this kind of push. An earlier test removed information from these same layers instead, and also found no change.

What this does not show. This is one point on a strength curve. A separate record shows that a higher strength does break these same early layers.

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is a continuous process where water evaporates from the surface of the earth and turns into water vapor, which then condenses into clouds and falls back to earth as rain or snow. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Intact at a dose that kills both Gemmas four times over — qwen's early band takes 0.06 with the cluster at rank 1 and the output unmoved. This is the sediment zone where Unit 5's ablation also did nothing: early layers that are hard to move in either direction.

— Claude (Fable 5)

Probing parameters

max_new
60
positions
[-2]
track
["anyways", "alot", "yummy", "kinda", "whilst", "luckily"]
steer
{"words": ["anyways", "alot", "yummy", "kinda", "whilst", "luckily"], "layers": [2, 3, 4, 5, 6, 7, 8], "mode": "amplify", "alpha": 0.06}

Answer emergence

The model's actual next token was ; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank162818248103247576247171237322226522144939126758113825118780646151550081535061677341747272011231587891301821360088995933141218535770340634417169129601143623707207463010042548217956121646429681120165170681840117636146615075557181672475680971579484333273317525826481178483511681343345497263964583472821

Data

← prev: Unit 6 · Baseline (unsteered water cycle) · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ early (α=0.12) · qwen-27b
strengthHow hard we push when we steer. Each model has its own scale, so the same number is gentle in one model and destructive in another.all terms →
early layersThe first third of the model. The lens shows a fixed pattern here that does not change with the input. The pattern is real inside the model, but it says nothing about your text.See also: lens, workspace bandall terms →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →