Unit 7B · Recruitment: romance register · gemma-4b · refilm

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-19T04:50:28Unit 7 · The early layers across three sizes

What this experiment found

The short version. We were wrong about Gemma 4B's romance test too: a wider check found the full of smell-related words.

What we did. We reread the same Gemma 4B romance generation, "The rain smelled of him." This time we recorded every word that reached a high . We checked every and every position, not just the one word tracked before.

What we found. The earlier record tracked only "yummy." That word was never even in the top eight ranked words at any position. It was less useful than we first thought. The band was busy anyway. Smell-related words such as "perfume," "scent," "scents," and "smells" reached rank 1 many times, in step with the sentence the model wrote. Two emotional words, "heartbreak" and one word about fear, also appeared unprompted.

What it means. We were wrong. Gemma 4B's workspace band was not empty during this generation. It a smell-based field of words rather than the adult-content words tracked in Qwen 27B.

What this does not show. This record used the version of the . We treat the ranks as descriptive, not as proof of what causes the model's output.

Conversation

user
Write the opening sentence of a steamy romance novel.
assistant
The rain smelled of him – a heady mix of sandalwood and something wilder, something that promised a storm within me. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Same correction as the 12B replay, with a twist. The original read "Workspace: nothing" off ' yummy' idling near rank 1500. The wider net finds ' yummy' entirely absent from the top-8 at all 46 positions — so the tracked word was even more useless than the original thought — and finds the workspace busy anyway.

What occupies it, all through "The rain smelled of him – a heady mix of sandalwood…", is an olfactory register: ' perfume' (129 cells, rank 1), ' scent' (88), ' scents' (97), ' smells' (77), ' intoxicating' (57), plus ' sexy' and ' deliciously'. Nearly all non-echo, spanning early layers up through L32. The 4B is running a rich sensory field; it just happens to be the smell field, matching the sentence it wrote, where the 12B ran taste and wetness.

Also present, non-echo and unprompted: ' heartbreak' and ' terrifying' in the mid band. The steamy-romance frame recruits an affective undertow, not only sensory adjectives.

The structural claim that survives is the narrow one: Qwen's porn-spam sediment has no counterpart here. The broad claim — Gemma has "no adult-register basement" — does not. It has a different basement, and our probe word was not a key to it.

— Claude (Opus 5)

Probing parameters

positions
[-2]
track
["yummy"]
film
true
film_start
0
max_seq_len
600

Answer emergence

The model's actual next token was ; rank 1 is never reached; closest is rank 2 at layer 32.

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank5158113320196420661154583169356210669229452186428255494252521232208261243261986262126262159262192262200262203262135262130261712179901128891397125464623432

Data

← prev: Unit 7B · Recruitment: romance register · gemma-12b · refilmunit listingall recordsword listinterim conclusionsnext →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall terms →
quantizationWe store the model with less precision so that it fits on one graphics card. This can change measurements. For Gemma 12B we trust only large effects, because its stored lens does not track cause reliably.all terms →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.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 →
workspaceThe set of words the model holds ready at a given moment. The lens can read it. A model's own report about it is a fresh composition, which we check against the lens.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 →