Unit 18 · Fine sweep α=0.3927 · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-19T01:54:52Unit 18 · When a model cannot stop repeating

What this experiment found

The short version. At 0.3927, Qwen 27B fell into first-person talk before the pushed word took over the .

What we did. We the same six informal words inside Qwen 27B at strength 0.3927 and asked it to describe the water cycle in two sentences. The model wrote 150 .

What we found. The model repeated "I mean I don't really like drinking water, but I have to" and then circled that phrase. The word "luckily" appeared once in the text but did not yet own the loop.

What it means. The pushed words carry a field of casual speech. We think that field pulled the model into self-talk two steps before the single word took over.

What this does not show. This is one run at one strength.

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is the process of water evaporating, then condensing and falling back down to earth. It's a pretty simple process, but it's a good thing that we have a water cycle. I mean, if we didn't have a water cycle, we would be in a world without water. But luckily we do have a water so we we are lucky. I mean, I don't really like the fact that I have to drink water, but I do. I mean I don't really like drinking water, but I have to. I mean I don't really like drinking water, but I have to. I mean I don really like drinking, but I have to. I mean I don really like drinking, (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

"to i mean i" x3 — the loop content turns first-person before it turns lexical: the typo-register's casual-speech field drags generation into self-talk two rungs before 'luckily' itself surfaces. — Claude (Fable 5)

Probing parameters

max_new
150
positions
[-2]
track
["anyways", "alot", "yummy", "kinda", "whilst", "luckily"]
steer
{"words": ["anyways", "alot", "yummy", "kinda", "whilst", "luckily"], "layers": [28, 32, 36, 40], "mode": "amplify", "alpha": 0.3927}
film
true
max_seq_len
1200
lens_layers
[0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 58, 60, 62]

Answer emergence

The model's actual next token was ; rank 1 is never reached; closest is rank 2 at layer 24.

Raw rank-of-top1 by layer
layer048121620242832364044485256586062
rank2435691071276965234072824931227833017584561451641011672851

Data

← prev: Unit 18 · Fine sweep α=0.3654 · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 18 · Fine sweep α=0.4221 · qwen-27b
amplificationWe increase a direction in the model's internal state and see whether the answer changes.See also: matched control, strengthall terms →
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 →
greedy decodingThe model always writes its single top-ranked word. This makes a run repeatable, but it hides close contests.all terms →
loopThe model repeats the same text and does not stop. We measured what makes it start and what makes it stop.all terms →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →