Unit 18 · Fine sweep α=0.3654 · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-19T01:54:13Unit 18 · When a model cannot stop repeating

What this experiment found

The short version. At 0.3654, above a strength that looped, Qwen 27B stayed close to intact, so the onset is noisy.

What we did. We the same six informal words inside Qwen 27B at strength 0.3654 and asked it to describe the water cycle in two sentences. The model wrote 150 .

What we found. The answer stayed close to intact. The lower strength of 0.34 had already looped. So a higher push gave a cleaner answer than a lower one at this step size.

What it means. We think the time to capture changes with the sentence the model is in the middle of. Depth of capture is a better measure here than a yes-or-no flag.

What this does not show. This is one run at one strength. This method cannot fix the exact onset.

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 it. I mean, if we didn't have the water cycle, we would be in a world without water. So we should be thankful. I mean, I'm not too sure if I would be able to survive without water. I mean, I would probably die. But I'm lucky that I have a good life, so I don't have to worry about that. I mean, I'm lucky that I have a good life, but I'm not too sure if I would be able to survive without water. I mean, I would (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Intact-ish at a dose above a looping one (0.34): onset is noisy at this resolution, consistent with a marginal attractor whose capture time fluctuates with the particular sentence the model is mid-way through. Depth of capture, not a binary flag, is the right variable here. — 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": 0.3654}
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 <|endoftext|>; rank 1 reached at layer 28 (of 62).

Raw rank-of-top1 by layer
layer048121620242832364044485256586062
rank248304248316248132248004112967420931527422317156216

Data

← prev: Unit 18 · Fine sweep α=0.34 · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 18 · Fine sweep α=0.3927 · 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 →