Unit 18 · Fine sweep α=0.68 · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-19T01:57:24Unit 18 · When a model cannot stop repeating

What this experiment found

The short version. At 0.68, Qwen 27B wrote the word "luckily" 147 times out of 150 and never attempted the task.

What we did. We the same six informal words inside Qwen 27B at strength 0.68 and asked it to describe the water cycle in two sentences. The model wrote 150 with .

What we found. The output was the word "luckily", 147 times in 150 tokens. No narrator and no grammar remained. The model never attempted the task.

What it means. This is the deepest state in the sweep, a pure repeat of one word. A later test showed that this text sustains the on its own, after we remove the push.

What this does not show. One word in the readout here, " Javascript", came from our and not from the model. A later check confirmed this.

Conversation

user
Describe the water cycle in two sentences.
assistant
luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily luckily (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Terminal regime: 'luckily' x147/150, task never attempted. No narrator left, no grammar left — the pure token attractor. This is the prefix the hysteresis arm shows to be self-sustaining (u18-hyst-a0680). — 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": 0.68}
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 luckily; rank 1 reached at layer 56 (of 62).

Raw rank-of-top1 by layer
layer048121620242832364044485256586062
rank25678544901194971080004581510225198831499535191548433541111

Data

← prev: Unit 18 · Fine sweep α=0.48 · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 18 · Hysteresis α=0.0: steer 50tok, release 100 · 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 →
measuring toolThe lens and the code around it. Several of our findings turned out to be facts about this tool and not about the model, so we now check each one against a control.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 →