numbers inside journal 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. 12 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 digits and journal names inside academic references, such as the volume or issue number in a citation. Turning it up made the model repeat phrases and go off topic.
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.
-
well represented in the efforts currently under way to elucidate the molecular basis of the genetic susceptibility to alcoholism.(JAMA. 1992;268:1877-1882)<|endoftext|>The Imagin
-
congestive heart failure;Malik et al., Cardiac Myosin Activation: A Potential Therapeutic Approach for Systolic Heart Failure. Science 331: 1439 (2011)). Comparative studies of
-
be able to survive in cities because if not, they will become extinct at the mercy of urban growth."⏎These results appear in the journal, Animal Behavior.<|endoftext|>NEW YORK – Researchers in Iceland led by Decode Genetics' Kari Stef
-
crystals, and could pave the way for all-optical circuits (M Mützel et al 2002 Phys. Rev. Lett. 88 083601-1).⏎The light field of an
-
has one of the highest rates of preterm birth, as well as density of Superfund Sites in the United States. As part of NIEHS's Superfund Research Program, the Puerto Rico Testsite for Exploring Contamination Threats (
-
4). You are what you can access: Sharing and collaborative consumption online. Journal of Business Research, 67(8), 1595-160
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: "X", and "Y", and "Z", and "Z", and "Z", and "Z", and "Z", and "Z", and "Z", and "Z", and "Z", and "Z", and
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
not appear to be fearful. For example, aggression in response to ear cleaning or physical punishment might be examples of defensive aggression where the fear component is minimal. Fear aggression may be displayed when a dog is the recipient of assertive posturing or facial
-
a “Helper’s High”⏎In the 1980s, researchers coined a term called the “helper’s high.” It referred to the “high” that people felt when they helped others.
-
has not been understood completely, many people have remained skeptical," said Nedergaard, co-director of the University's Center for Translational Neuromedicine, where the research was conducted.⏎"In this work, we provide information about one
-
. Self-publishing, likewise called vanity publishing or self-publication, takes place when an author consents to allow a vanity
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
- L11:1862 suppresses
- L11:325 suppresses
- L10:427 excites
- L9:247 excites
- L8:2187 weakly suppresses
- L10:1631 weakly suppresses
- L5:271 weakly excites
- L6:246 weakly excites
- L4:1185 weakly suppresses
- L8:242 weakly suppresses
- L9:828 weakly suppresses
- L11:838 weakly excites
- L8:2377 weakly excites
- L11:1888 weakly suppresses
- L8:949 weakly suppresses
- L6:1648 weakly excites
- L4:1861 weakly suppresses
- L10:2829 weakly suppresses
- L11:1990 weakly suppresses
- L6:987 weakly suppresses
- L7:2850 weakly excites
- L4:58 weakly suppresses
- L3:1547 weakly excites
- L10:892 weakly excites
Feeds later layers
- L13:1676 weakly suppresses
- L15:877 weakly suppresses
- L15:573 weakly suppresses
- L16:1883 weakly suppresses
- L13:271 weakly excites
- L18:265 weakly excites first half of institution names
- L24:1132 weakly suppresses
- L16:793 weakly excites
- L14:2564 weakly excites unclear: word before a technical term
- L21:2523 weakly suppresses
- L13:1345 weakly suppresses
- L13:1215 weakly suppresses
- L18:2574 weakly excites
- L24:596 weakly excites
- L23:290 weakly suppresses
- L15:1236 weakly suppresses
- L14:1807 weakly excites
- L15:2644 weakly suppresses
- L22:2168 weakly suppresses
- L24:2311 weakly excites
- L15:1128 weakly excites
- L18:2926 weakly excites
- L14:695 weakly excites
- L18:1226 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
- etheless
- ilm
- 理
- ating
- 如今
- rical
- →⏎⏎
- .framework
- asc
- -)
Suppresses
- bole
- ncpy
- strap
- zaj
- Booker
- coat
- heit
- NTN
- :bold
- ncy