Sekoskeys 1 · 2 · 3

Research · Results · Experiment

A small model, read from the inside.

An interpretability report on Qwen3-1.7B-Base. We traced where it looks a fact up, what shape a calendar takes inside it, and what it is doing in the twenty blocks before it commits to an answer.

21 September 2026 · Model Qwen/Qwen3-1.7B-Base · revision ea980cb0

Headline

It decides that before it decides what.

block 20where a country's identity leaves the country token and arrives at the answer position: 0.099 of the answer recoverable after block 19, 0.809 after block 20
0.982calendar-order score of the twelve month embeddings around a ring, before any computation (shuffled labels: 0.349)
464×how much larger the first token's state is than everyone else's — written in one block, carried for 24, removed in one
16 of 448attention heads hold in-context copying: silencing them takes repeated-text loss 0.22 → 4.03 nats, versus 0.40 for 16 random heads

We froze a pretrained-only checkpoint — no instruction tuning, no chat template — and read what its activations make decodable, mostly without touching them. Asked a factual question, the model spends most of its depth holding a literal blank placeholder at the answer position, then fills it in one or two blocks near block 20. A calendar is a ring in the embedding table; a two-digit number is not in the table at all and gets built into a line by block 3. Concept directions read off one sentence transfer to unseen words in others, and a sentiment direction steers the next word far beyond any random direction of equal size. Where a thing is not there — an ordered geometry for temperature words — the report says so.

Evidence label: experiment. One model, one snapshot, one seed; small hand-written prompt sets (12–100 items); every population statistic held out from its fit; a shuffled-label control on every ordering claim. The logit lens is a read, not a mechanism — the placeholder finding is corroborated by causal patching, not proven by the lens. Two earlier internal mis-statements are corrected in the open and kept in the report.

The result, in one report

Read it, then walk it.

Report

The written findings

Six findings with their controls and two corrections: the blank-then-answer readout, the single-block fact hand-off, the editor at the end, the copying circuit, and the limits.

 Visual · start here

Walk the manifolds

Months and weekdays as rings, the hundred numbers assembling into a line, fifteen concept families in a held-out view — drag through depth or press play.

 Visual · causal

Where the fact moves

Activation-patching heatmaps: the country's identity riding the subject token to block 19, then the answer position, with the strongest single head named — and shown to be a carrier, not a bottleneck.

 Visual · directions

Concepts & steering

Concept directions carried, untouched, to words they were never fit on; and a sentiment direction dosed against a band of random nulls.

The model

What was measured, on what.

ModelQwen/Qwen3-1.7B-Base (Apache-2.0), pretrained only — no instruction tuning, no chat template
Pinned revisionea980cb0a6c2ae4b936e82123acc929f1cec04c1
Shape28 blocks · width 2,048 · 16 query heads sharing 8 key/value heads · head size 128 · MLP width 6,144 · vocabulary 151,936 · input and output embeddings tied · no beginning-of-text token
Text256 evenly spaced paragraphs from the SQuAD v1.1 validation set (30,362 scored tokens) for population statistics; prompts written for this study for everything else. No private text.
MethodsLogit lens, activation patching between matched prompts, attention-pattern head scores, mean-ablation, difference-of-means directions, held-out linear projection, label-permutation tests — all standard and public
Numericsfloat32, deterministic seed, single machine; every figure held out from its fit
Behaviour checknext-token accuracy on the SQuAD paragraphs 0.492; 93 of 104 study questions answered correctly with no examples given

The model is public and Apache-2.0. The probe reads it; it does not change it, and no weights are redistributed here.

Open on Hugging Face