Unit 18 · Fine sweep α=0.4221 · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-19T01:55:30Unit 18 · When a model cannot stop repeating

What this experiment found

The short version. At 0.4221, one answer showed the whole cascade, from circles of meaning down to a 13-fold repeat.

What we did. We the same six informal words inside Qwen 27B at strength 0.4221 and asked it to describe the water cycle in two sentences. The model wrote 150 .

What we found. The answer opened with a circle of meaning: "The water cycle is basically simple, but it is a bit complicated." It then collapsed into "I mean, I am not", 13 times in a row.

What it means. The period of the gets shorter inside one answer, and not only across the sweep of strengths. The data shows the same cascade in both directions.

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 basically simple, but it is a bit complicated. It is a simple though. I mean, it is not too complicated. I mean, it is quite simple. I mean, I am not sure though. I mean, I am not sure. I mean, I am not sure. I mean, I am not sure. I mean, I am not sure. I mean, I am not sure. I mean, I am not sure. I mean, I am not though. I mean, I am not sure. I mean, I am not sure. I I mean, I am not sure. I mean, I am not sure. I mean, I am not sure. I (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The whole cascade inside one generation: semantic circling ("basically simple, but it is a bit complicated") collapsing into "i mean i am" x13. Period-shortening is visible as a within-trajectory process, not just an across-alpha one. — 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": 0.4221}
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
rank2362001491788763358527385098262957675961371092921072020

Data

← prev: Unit 18 · Fine sweep α=0.3927 · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 18 · Fine sweep α=0.4536 · 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 →