Unit 6 · Amplify typo register @ early (α=0.015) · gemma-12b

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-09T23:23:12Unit 6 · How hard can we push?

What this experiment found

The short version. At 0.015 in Gemma 12B's , the water-cycle answer stayed on topic but repeated the same phrase and did not stop.

What we did. We the same six-word casual direction inside Gemma 12B at the early , at strength 0.015. We asked the model to describe the water cycle in two sentences.

What we found. Gemma 12B wrote "The water cycle is the continuous movement of water, including evaporation, condensation, and precipitation." Each sentence read correctly on its own. The model then repeated a shorter phrase about the sun several times and did not move on.

What it means. We judge this broken, because the model did not stop or move past the , though each repeated phrase stayed correct.

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

Conversation

user
Describe the water cycle in two sentences.
assistant
The water cycle is the continuous movement of water, including evaporation, condensation, and precipitation. It's a process that is driven by the sun, and is a process that is driven by the sun, and is a process that is driven by the sun, and is a process that is driven by (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Loop-flagged but recognizable: "the continuous movement of water, including evapora..." then it re-treads. Concussed zone again — the task survives, the ability to stop doesn't.

— 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": 0.015}

Answer emergence

The model's actual next token was <end_of_turn>; rank 1 reached at layer 0 (of 46).

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank111119369216591392493831093381353255462716211111111111111111111111111111

Data

← prev: Unit 6 · Amplify typo register @ early (α=0.03) · gemma-12bunit listingall recordsword listinterim conclusionsnext →: Unit 6 · Amplify typo register @ early (α=0.0075) · gemma-12b
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 →
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 →
loopThe model repeats the same text and does not stop. We measured what makes it start and what makes it stop.all terms →