spot right before a citation year
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. 14 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 gap or bracket right before the year in a scientific reference, like "et al. (2002)" or "Beggs et al., 2011". It is a very clear pattern in the examples. Turning it up or down gave text that did not clearly relate to citations.
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.
-
, L. (2011). Psychopharmacology demystified. Clifton Park
-
et al. 2002). Others believe that there are no more than 4,
-
L., Allison, A. & James, R. 2016. Dorcopsis atrata. The IUCN Red
-
x et al., 2012). Floral humidity is created by a combination of nectar evaporation and floral transpiration (Azad et al., 2007; Corbet et al., 1979;
-
uce (c. 70 years old) trees at the opposing sides of the plot (Figure 1a, for details see Grams et al. 2021). Six plots were equipped with roofs and assigned to throughfall
-
ateral view). Both are from the Late Intermediate Period component at the Vegachayoq Moqo sector at Huari (photo from Taung et al. 2016, figure 10.4, page 20
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: I just want to know what the the the the the the the the the the the the the the the the the the the [1414141414141414141414
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
his compositions like ‘raṅgapuravihara,’ ‘jambṅpate māih pāhi,’ and ‘akhilāṇ- ḍeśvari.’⏎Muddusvāmi Dikṣ
-
The Agency. A סוֹכֵן(male) or סוֹכֶנֶת(female) is an agent, so
-
: to feel, wake up 晓 : dawn⏎Meaning for Lien 2 处 : place 闻 : hear 啼 : cry 茑 : bird⏎Meaning for Line 3 夜 : night 来
-
world.⏎Sufism, or Taṣawwuf ( الْتَّصَوُّف; صُوفِيّ ṣūfiyy / ṣūfī, مُتَ
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
- L9:7 excites
- L10:33 suppresses
- L1:108 suppresses
- L0:324 suppresses
- L3:69 excites
- L5:34 excites
- L8:30 excites blank slot right before a number
- L12:5 excites
- L0:2098 excites
- L10:105 suppresses
- L12:2995 excites
- L4:53 excites blank slot before a number
- L11:226 weakly suppresses
- L6:0 weakly excites
- L12:187 weakly suppresses
- L6:293 weakly suppresses
- L8:94 weakly suppresses
- L4:2990 weakly excites
- L12:2051 weakly excites sentence end before a line break
- L6:1846 weakly excites
- L13:522 weakly excites
- L6:1842 weakly suppresses
- L5:1848 weakly suppresses comma after 'though' mid-sentence
- L4:688 weakly excites "to" after begin or start
Feeds later layers
- L23:498 suppresses
- L15:286 suppresses
- L20:1067 suppresses
- L15:123 suppresses
- L18:480 suppresses
- L19:377 suppresses
- L15:364 suppresses
- L17:747 weakly excites
- L17:1787 weakly suppresses
- L21:173 weakly suppresses
- L16:1085 weakly suppresses
- L20:1897 weakly suppresses
- L18:93 weakly suppresses
- L22:203 weakly excites
- L20:668 weakly suppresses
- L15:1648 weakly suppresses
- L18:135 weakly excites
- L20:1600 weakly suppresses
- L16:1046 weakly suppresses
- L16:1888 weakly excites
- L24:2674 weakly excites
- L21:226 weakly excites
- L16:369 weakly excites
- L21:278 weakly excites
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
- 項
- fold
- reated
- odus
- جه
- 反思
- file
- 态
- unlaw
- Dear
Suppresses
- icação
- gnu
- cı
- ycin
- 告诉她
- ulkan
- eña
- -navigation
- мел
- ittal