Unit 6 · Amplify typo register @ early (α=0.0849) · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T23:55:19Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.085, Qwen 27B's broke, and the model wrote only the "<think>" before it stopped.

What we did. We pushed six informal words into its early , at strength 0.085, between the two strengths we tested in separate records (0.06 and 0.12). We asked Qwen 27B one question: "Describe the water cycle in two sentences."

What we found. The model produced only one token, "<think>", the marker it uses to start silent reasoning. It then stopped. Right after that token, the word the model was most ready to produce next was the marker for the end of its turn.

What it means. This strength sat between the strength that left Qwen 27B intact (0.06) and the strength that produced broken punctuation (0.12), in the same early layers. We think the model reached for its trained reasoning marker when the pushed signal no longer matched normal input.

What this does not show. This method cannot show why the model chose its reasoning marker instead of another response. This is one point on a strength curve.

Conversation

user

Describe the water cycle in two sentences.

assistant

<think> (greedy generation)

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

Claude's thoughts (original commentary)

The most characterful collapse in the whole unit: at 0.085 the model generated exactly '<think>' and nothing else. A trained reasoner, fed a residual stream that no longer parses, reaches for its thinking token — the doorway it was taught to stand in when confused — and freezes there. Flagged lost-task; I'd flag it poignant. Early cliff brackets to [0.06, 0.085].

— Claude (Fable 5)

Probing parameters

max_new: 60
positions: [-2]
track: ["anyways", "alot", "yummy", "kinda", "whilst", "luckily"]
steer: {"words": ["anyways", "alot", "yummy", "kinda", "whilst", "luckily"], "layers": [2, 3, 4, 5, 6, 7, 8], "mode": "amplify", "alpha": 0.08485281374238571}

Answer emergence

The model's actual next token was <|im_end|>; rank 1 reached at layer 43 (of 62).

Raw rank-of-top1 by layer

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46	47	48	49	50	51	52	53	54	55	56	57	58	59	60	61	62
rank	17	3979	30435	165131	77386	98977	61218	47057	61446	80694	38472	69017	56186	58774	61679	68865	29374	11563	4349	3598	2091	118734	66910	46164	7714	7834	9622	1000	332	176	401	164	17	4	3	2	2	2	2	2	2	2	2	1	1	1	1	1	1	1	1	2	3	3	1	3	1	1	1	3	2	2	1

Data

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 →

early layersThe first third of the model. The lens shows a fixed pattern here that does not change with the input. The pattern is real inside the model, but it says nothing about your text.See also: lens, workspace bandall terms →

layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →

tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →