Unit 6 · Amplify typo register @ mid (α=0.03) · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:03:24Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.03 in Gemma 4B's , the water-cycle answer broke into meaningless repeated fragments such as "BeBe BeBe".

What we did. We the same six-word casual direction inside Gemma 4B at the middle , at strength 0.03. We asked the model to describe the water cycle in two sentences.

What we found. Gemma 4B wrote "BeR About BeP What BeBe BeBe BeBe BeBe BeBe BeBe BeBe BeBe", then repeated "Be" alone for the rest of the answer. Neither the water-cycle task nor the amplified words appeared in the text.

What it means. At this strength, the break did not take the shape of the amplified direction. It produced noise instead of the direction's own words.

What this does not show. This record tests one strength only. It does not show why this break took a different shape from other broken runs in this test set.

Conversation

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

Claude's thoughts (original commentary)

Broken into "BeBe BeBe BeBe" — neither task nor cluster survives; this is the zone where the residual stream is just ringing. Between 0.03 and 0.06 the failure changed character from noise to capture.

— 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": [15, 17, 19, 21], "mode": "amplify", "alpha": 0.03}

Answer emergence

The model's actual next token was S; rank 1 is never reached; closest is rank 16 at layer 32.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank228239205033223555225963627581661031293821100235837335842345259446453616170799187947529102472929528288411797917640155361262320078164471805610285407754331881138727216

Data

← prev: Unit 6 · Amplify typo register @ mid (α=0.06) · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ mid (α=0.015) · gemma-4b
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 →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
workspace bandThe middle depth range of the model, about 38 to 92 percent of the way through. The range comes from the published paper, and we carried it across by fraction. Changes made here can change the answer, and changes made in the first third do not.See also: start depth, final layersall terms →