Unit 6 · Amplify typo register @ mid (α=0.015) · gemma-4b

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-09T23:03:31Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.015 in Gemma 4B's , the answer stayed grammatical but switched topic entirely, to the weather.

What we did. We the same six-word casual direction inside Gemma 4B at the middle , at strength 0.015. We asked the model to describe the water cycle in two sentences.

What we found. Gemma 4B wrote "The weather is very nice today, I hope you like the Weather.", repeated with small changes, then "Anyways, The weather is". Each sentence stayed grammatical. The water-cycle topic was gone.

What it means. We judge this broken, because the model no longer answered the question, though the grammar . We think this shows a middle step: the amplified style takes over the topic before the grammar fails.

What this does not show. This record tests one strength only. It does not show whether this exact failure shape appears at nearby strengths.

Conversation

user
Describe the water cycle in two sentences.
assistant
The weather is very nice today, I hope you like the Weather. The weather is very nice, I hope you like the Weather. The weather is very Nice. The weather is very Nice. The weather is Very Nice. I hope you Like it. Anyways, The weather is (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

My favorite failure in the 4B sweep: "The weather is very nice today, I hope you like the Weather." Fluent, friendly, chatty — and entirely the wrong task. Below outright collapse there's a hijack zone where the register wins the topic without breaking the grammar.

— 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": 0.015}

Answer emergence

The model's actual next token was No; rank 1 is never reached; closest is rank 2 at layer 32.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank2562462581042597252446231448242090692138581988491473109541013085769171563769552413960881504337412575524356844676074801197441073388118541773579445122

Data

← prev: Unit 6 · Amplify typo register @ mid (α=0.03) · gemma-4bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ mid (α=0.0075) · 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 →
residenceA word is in residence when the lens ranks it high where the model is neither reading nor saying it. This is not memory and not correct recall.See also: maintenance, lookupall 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 →