place and journal names in citations
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 10 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron fires on the city, state, or journal name in reference lists, like the comma after 'Portsmouth' in 'Portsmouth, NH' or 'Clinical Chemistry'. It is tuned to bibliography style text. Turning it up made the output drift into unrelated topics, so the steering does not show this clearly.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
, Vol 3118, Springer-Verlag, Berlin⏎Heidelberg New York⏎D., Lenoir, R., and McKenzie
-
really learn.⏎Beers, Kylene and Robert E. Probst. Notice & Note: Strategies for Close Reading. Portsmouth, NH: Henemann, 2013.⏎Using the work of Louise
-
th ed.). Boston, MA: Cengage Learning.⏎[TradingCoachUK]. (2018
-
: How to teach writing like you mean it (2nd ed.). Newark, DE: International Reading Association.⏎- Writing activities that help students integrate precise, vivid
-
. The Art of Slow Reading. Portsmouth, NH: Heinamann, 2012.⏎And advocate to taking
-
azren (2007) Assisting learning using computers through the application of animated pedagogical agents (APA) or intelligent agents. Smart Teaching & Learning: Re-engineering ID, Utilization and Innovation of Technology, 1 .
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: the problem of a single sentence. So, the problem of a sentence. Now, I think about the sentence. Now, I think about the problem of a sentence. So, the problem of the sentence. So, the problem. So
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
0 percent are either overweight or obese, according to the U.S. Centers for Disease Control and Prevention.⏎Results of the study were published recently in the journal Clinical Chemistry.⏎The U.S. National Heart, Lung, and Blood Institute offers a guide
-
Karl M. Johnson, Douglas F. Antczak, William H. Dietz, David H. Martin and Thomas E. Walton appears in Emerging Infectious Diseases 17 (5) (20
-
the gastrointestinal system. Their results are published in "Nature Structural & Molecular Biology".⏎"These findings could in principle be explored to develop a decoy alternative to conventional
-
for Atmospheric Research and gathered information such as radial velocity and reflectivity from six Doppler weather radars that were in place during storms. Using the Weather Research and Forecasting Model, they found that incorporating the Doppler radar-based information decreased the error
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L8:2698 suppresses the word 'dental' in oral health
- L8:270 suppresses
- L7:320 suppresses
- L1:1464 weakly excites
- L6:55 weakly suppresses
- L6:2402 weakly excites
- L5:1051 weakly suppresses
- L8:1365 weakly suppresses
- L8:2792 weakly excites
- L5:2321 weakly excites
- L4:994 weakly excites
- L6:1696 weakly excites
- L2:2835 weakly excites
- L4:983 weakly suppresses
- L1:1319 weakly excites 'for' right after 'look'
- L4:1438 weakly excites
- L6:2943 weakly suppresses
- L1:2801 weakly excites
- L2:2147 weakly suppresses
- L1:73 weakly suppresses
- L1:3003 weakly excites
- L1:171 weakly excites "but" after a negated clause
- L2:75 weakly suppresses
- L7:1163 weakly excites
Feeds later layers
- L24:612 weakly excites
- L11:2198 weakly suppresses
- L12:2948 weakly excites
- L12:2082 weakly excites
- L15:269 weakly excites
- L24:2806 weakly excites
- L21:893 weakly suppresses
- L11:1762 weakly excites
- L12:1580 weakly excites
- L11:1720 weakly suppresses
- L20:1953 weakly excites
- L11:2309 weakly suppresses
- L17:1748 weakly suppresses
- L11:122 weakly excites
- L23:2160 weakly excites
- L18:884 weakly suppresses
- L14:1185 weakly excites
- L16:1684 weakly excites
- L16:2978 weakly excites
- L15:1764 weakly excites
- L25:1650 weakly excites
- L14:387 weakly suppresses
- L14:842 weakly excites
- L17:1844 weakly suppresses
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- cimal
- Neill
- TokenType
- ไท
- 缩水
- productList
- pakistan
- Feeling
- แยก
- فئ
Suppresses
- уж
- 断
- hers
- ici
- اغ
- research
- orsch
- 塑造
- 层面
- 让她