commas between author 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. 2 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 lights up on the comma that comes right after an author's surname in a reference list, such as the comma after 'Chung' in a citation. When it was turned up, the written text drifted into unrelated rambling and did not show a clear theme.
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.
-
Ho, C. S. H., Lo, C.-M., Chan, K. S.-C., Chan, D. W.-O., & Chung, K. K.-H. (2014, July). Effectiveness
-
J.M.; Mortimer, S.R.; Brereton, T.; Redhead, J.W.; Thomas, J.A.; Pywell, R.F.. 2012 Identifying time lags in the restoration
-
our fishing communities,<|endoftext|>Suicide and Social Change in China⏎- Cite this article as:⏎- Phillips, M.R., Liu, H. & Zhang, Y. Cult Med Psychiatry (1999) 23:
-
This handout provides an example of a revised “show-me” sentence and six exercises for students to practice.|⏎Baines, L., & Kunkel, A. (2010). Going bohemian
-
results⏎Authors: Walker, Robert J. (John), 1801-1869
-
justify your arguments.⏎references: Gwartney, J. D., Stroup, R. L., Sobel, R. S., & Macpherson, D. A. (2018). Economics: Private⏎and public
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: 0.000000000000000000000000000000000000000000000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
pysyväislinkit|⏎|Tila||Julkaistu - 1 syysk. 2021|⏎|OKM-julkaisutyyppi||A1 Julkaistu artikkeli,
-
helped spearhead efforts to create a warning system to keep jet aircraft from flying into the sort of engine-seizing ash cast up by Eyjafjallajökull, the low-lying volcano whose plume all but shut down
-
astronauts and equipment, according to Reuters.⏎China's Shenzhou spacecraft, based on Russia's Soyuz, has the capacity to ferry up to
-
011. Successful docking with an unmanned spacecraft had already been conducted in 2011 itself. But this time, the three astronauts in Shenzhou 9 established that China could precisely maneuver a space capsule to rendezvous with and
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
- L1:2580 suppresses
- L0:178 weakly suppresses
- L0:1301 weakly suppresses
- L1:2236 weakly excites
- L1:2683 weakly excites
- L0:1224 weakly suppresses
- L0:1042 weakly suppresses "This" starting a labelled sentence
- L1:1732 weakly excites
- L1:1464 weakly excites
- L1:2197 weakly suppresses
- L1:628 weakly suppresses
- L1:1857 weakly suppresses
- L0:313 weakly suppresses "and" joining paired nouns
- L1:844 weakly suppresses
- L1:1134 weakly suppresses
- L1:2796 weakly excites
- L0:802 weakly excites
- L0:571 weakly excites
- L1:1907 weakly suppresses
- L1:1692 weakly excites
- L0:3015 weakly excites
- L1:1735 weakly excites
- L0:191 weakly suppresses
- L1:987 weakly excites
Feeds later layers
- L5:1027 excites
- L16:826 weakly suppresses
- L10:2091 weakly suppresses
- L7:1945 weakly excites
- L16:845 weakly excites
- L12:495 weakly excites
- L17:1665 weakly suppresses
- L23:2123 weakly suppresses
- L15:1114 weakly suppresses
- L4:560 weakly suppresses
- L5:1586 weakly excites
- L5:1938 weakly excites
- L16:1194 weakly suppresses
- L10:612 weakly suppresses full stop ending boilerplate or a title
- L13:640 weakly suppresses
- L14:2440 weakly excites
- L6:770 weakly suppresses
- L14:1653 weakly excites
- L13:2798 weakly excites
- L3:2833 weakly suppresses
- L6:1397 weakly suppresses
- L4:1713 weakly suppresses
- L17:1111 weakly suppresses
- L12:2776 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
- prend
- ケット
- 最大
- ボード
- yourself
- /bind
- ลาด
- 都想
- combin
- imet
Suppresses
- sts
- s
- atic
- ующ
- tas
- sti
- oS
- icated
- ian
- st