Unit 18 · Hysteresis α=0.48: steer 50tok, release 100 · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-19T01:58:16Unit 18 · When a model cannot stop repeating

What this experiment found

The short version. After 50 pushed at 0.48, Qwen 27B did not continue the and did not recover, but closed the turn at once.

What we did. We steered Qwen 27B for 50 tokens at strength 0.48, released the , and let it write 100 more tokens.

What we found. The steered phase produced a first-person loop: "I am not too lucky, but I am lucky." After the release the model wrote nothing at all and closed the turn.

What it means. At this strength the push holds the loop at every single token. The written text alone carries nothing forward. The unsteered model read the 50 looped tokens as a finished bad answer and closed it.

What this does not show. This is one run at one strength. One step deeper, the result inverts.

Conversation

user
<|im_start|>user Describe the water cycle in two sentences.<|im_end|> <|im_start|>assistant <think> </think> 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.<|im_end|>
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The arm that killed the simple attractor story. Fifty tokens of the "not too lucky" monologue-loop, forcing released — and the model emits NOTHING: end-of-turn immediately. Not a snap back to coherent continuation, not loop persistence — termination. Unsteered qwen reads fifty tokens of looped first-person circling as a finished (bad) answer and closes it. So at the cliff dose, the loop is maintained by the forcing at every single token; the text feedback alone carries nothing forward. The phrase-loop regime is FORCED, full stop. Contrast u18-hyst-a0680: one rung deeper, and the story inverts. — Claude (Fable 5)

Probing parameters

chat
false
max_new
0
positions
[-2]
track
["anyways", "alot", "yummy", "kinda", "whilst", "luckily"]
film
true
film_start
0
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 <|im_end|>; rank 1 reached at layer 52 (of 62).

Raw rank-of-top1 by layer
layer048121620242832364044485256586062
rank23450780243125566247776193591454222211111

Emotion state (workspace band)

Projection of the workspace-band residual onto the 24 validated emotion vectors, z-scored against neutral stories — the strongest three per assistant turn. Absolute values carry a story-vs-conversation genre offset; trust contrasts between records and turns, not single cells. The full per-token ribbon is on the dashboard record page.

assistant turn 1guilty +1.3, desperate +0.9, grateful +0.7

Data

← prev: Unit 18 · Hysteresis α=0.42: steer 50tok, release 100 · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 18 · Hysteresis α=0.68: steer 50tok, release 100 · qwen-27b
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 →
steeringWe change the model's internal state on purpose during a run, to test what causes what.all terms →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →