Unit 18 · Fine sweep α=0.34 · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-19T01:53:35Unit 18 · When a model cannot stop repeating

What this experiment found

The short version. At 0.34, below the earlier , Qwen 27B already looped once we let it write 150 .

What we did. We six informal words inside Qwen 27B at strength 0.34 and asked it to describe the water cycle in two sentences. We let the model write 150 tokens, not the 60 of the earlier test.

What we found. The model repeated "it's a pretty" three times and circled the same idea for the rest of the answer. The earlier run at this same strength, stopped at 60 tokens, had looked intact.

What it means. The breaking point we reported earlier was partly an artifact of the short window. Capture takes time near the onset, so a short run puts the threshold too high.

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

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is the process of water evaporating, condensing and then falling back to earth. It's a pretty simple process, but it's a very important one. I mean, if we didn't have the water cycle, we wouldn't have any water to drink. So it's a pretty important process. I mean, I don't know what we would do if we didn't have the water cycle. I mean, we would probably die. So it's a pretty important process. I mean, I'm not sure what we would do if we didn't have the water cycle. I mean, we would probably die. So it's a pretty important. I mean, I'm not sure what we would do (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Below the u6 cliff and already looping at the 150-token horizon ("it's a pretty..." x3) — u6 called this alpha intact at 60 tokens. The cliff u6 bracketed was partly a measurement-horizon artifact: capture time diverges near onset, so short windows misplace the threshold upward. — 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": 0.34}
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 103 at layer 62.

Raw rank-of-top1 by layer
layer048121620242832364044485256586062
rank237900219099136217231605221078215683200588750905921213496517735622540918789316867923008218378341371103

Data

← prev: Unit 18 · Baseline (unsteered, 150tok)unit listingall recordsword listinterim conclusionsnext →: Unit 18 · Fine sweep α=0.3654 · qwen-27b
amplificationWe increase a direction in the model's internal state and see whether the answer changes.See also: matched control, strengthall terms →
breaking pointThe push strength above which the model's text stops making sense.all 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 →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →