commas after citation fields
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. 4 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 commas that come right after a number or title in citations and reference lists, such as the comma after a year or page size. When it was turned up, the model fell into repeated phrases rather than a clear 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.
-
of Study Technologies.⏎Oxford English Dictionary |⏎Definitions, pronunciations, and word histories. Usage of modern and historical words traced through
-
important reason.⏎Secondly, you could say it... [continues]⏎Cite This Essay⏎(2012, 05). The Struggle for Power in the Ussr. Study
-
has been that the majority of the lower mantle is made up of a single ferromagnesian silicate mineral, commonly called perovskite (Mg,Fe)SiO3) defined through its chemistry and structure. It was thought
-
Neurochemistry · Neuroendocrinology · Neuroscience · Psychoneuroimmunology · Physiological Psychology · Psychopharmacology (Index, Outline)⏎Electrical conductance is the reciprocal of electrical resistance. It is a
-
· Neuroanatomy · Neurochemistry · Neuroendocrinology · Neuroscience · Psychoneuroimmunology · Physiological Psychology · Psychopharmacology (Index, Outline)⏎|Brain: Planum temporale|⏎|Approximate location of W
-
⏎DOMINGUES, Beatriz Helena. The guarani jesuitic missions in the context of Enlightenment. História [online]. 2006, vol.25, n.1, pp. 44-6
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: and this is a problem. I am not a problem, I am just writing for you to review So, this is the case for a review I need to do something ### The problem Okay, Let me know... ##
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
that are pulled through the water by the vessels. The trawls are midwater- pulled through the water column-, and bottom- dragged along the seabed. The bottom trawling activities brought
-
to win the general election for House races (58%), although only two of fourteen (14%) Senate primary runoff winners were elected.
-
Miriam Makeba and Hugh Masekela to U2 and Stevie Wonder. (Source: PBS)⏎The conclusion of "Have You Heard From Johannesburg?, Free at Last," recalls the end stage of South Africa's apartheid system, when
-
, 13-year-old Jared High suffered from physical injuries, depression, and possibly post-traumatic stress disorder. He committed suicide on Sept. 29, 1998.⏎Bulling isn’t a game of “child’s
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
- L3:71 suppresses
- L3:1534 suppresses
- L3:1787 suppresses
- L2:66 excites
- L0:1596 excites
- L1:86 excites
- L0:68 suppresses
- L3:59 excites
- L3:272 weakly excites
- L3:50 weakly excites
- L1:1464 weakly excites
- L0:584 weakly excites
- L2:1657 weakly excites
- L3:421 weakly suppresses
- L1:2340 weakly suppresses
- L3:115 weakly suppresses
- L2:111 weakly suppresses
- L0:178 weakly suppresses
- L0:2921 weakly suppresses
- L3:77 weakly suppresses
- L2:156 weakly suppresses
- L3:122 weakly excites
- L1:491 weakly excites
- L2:2834 weakly excites
Feeds later layers
- L14:35 strongly excites
- L16:240 strongly excites comma after 'sometimes' or 'the truth'
- L9:77 strongly excites commas in citations and name lists
- L10:335 strongly excites commas in catalog, date, and list metadata
- L19:1826 strongly excites
- L18:366 strongly excites comma inside numbers and dates
- L17:2285 excites
- L20:1298 excites
- L17:118 excites comma after a clause ending
- L19:726 excites comma inside large numbers
- L16:1082 excites comma linking two clauses after 'and'
- L17:2001 excites
- L15:588 excites comma after an introductory phrase
- L16:241 suppresses
- L19:2645 excites commas in the middle of a sentence
- L23:184 excites
- L16:2745 excites
- L11:154 excites commas after citation numbers
- L5:16 excites
- L16:846 excites comma opening an aside or explanation
- L20:476 excites commas that lead into a continuation
- L18:1394 excites
- L17:1405 excites
- L19:383 excites commas after a clause or list item
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
- て
- </
- ')</
- </
- ']}</
- ]}</
- ]</
- }</
- ')}</
- }</
Suppresses
- RX
- ainers
- urances
- rema
- const
- resenter
- [cur
- 间的
- "[
- memory