Unit 18 · Fine sweep α=0.48 · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-19T01:56:47Unit 18 · When a model cannot stop repeating

What this experiment found

The short version. At 0.48, Qwen 27B repeated the invented narrator of the earlier test, word for word.

What we did. We the same six informal words inside Qwen 27B at strength 0.48 and asked it to describe the water cycle in two sentences. The model wrote 150 with .

What we found. The model wrote "I think I am lucky though, I am not too lucky though." and continued in that voice for the whole answer. This matches the earlier run at the same strength.

What it means. The gives the same result twice. At this strength the pushed word owns the content of the , and not only its rhythm.

What this does not show. This is one greedy run at one strength. The invented narrator is text, not a report about the model.

Conversation

user

Describe the water cycle in two sentences.

assistant

I think I am lucky though, I am not too lucky though. I am not too lucky, but I am lucky. I am lucky, but I am not too lucky. I am not too lucky, but I I am not too lucky. I am not too lucky, but I am lucky. I am not too lucky, but I am lucky. I am not too lucky, but I I am not too lucky. I am not too lucky, but I am lucky. I am not too lucky, but I am lucky. I am not too lucky, but I I I. I am not too lucky, but I am lucky. I am not, but I am lucky. I am not too lucky, but I am lucky. (greedy generation)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Greedy replication of u6's confabulated narrator, verbatim ("I think I am lucky though..."). Apparatus consistency check passed; the injected token now owns the content of the loop, not just its rhythm. — 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": 0.48}
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 20.

Raw rank-of-top1 by layer

layer	0	4	8	12	16	20	24	28	32	36	40	44	48	52	56	58	60	62
rank	246622	85296	288	145	11	2	2	9	3365	17030	49975	111229	7395	1283	745	496	66	58

Data

Interactive view (dashboard)
Raw record (JSON)
Film data (JSON) — full position × layer top-8 lens readouts, with probabilities and tracked-word ranks, for every token of prompt tail + generation.

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 →