Neuron Explorer

Qwen3-0.6B

comma after a finished clause

A neuron in layer 18 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featured, so it has steering outputs.

What this neuron does

This neuron fires on a comma that comes right after a complete phrase or sentence part, where more text is expected to follow, like a comma before a quote, a 'and why?' or a trailing 'sir,'. When it was turned up, the model's writing fell into repeated filler like 'the, that, the'.

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 activation was highest. Amber marks positive activation; the outlined token is the peak. Hover a token for its strength.

  1. #1 peak

    .⏎That’s right: our phones got so smart they began putting pay phones out of business.⏎“Do you see the cell phone as sort of an arriviste, as someone that sort of came along and stole the thunder of the payphone?”

  2. #2 peak

    religion until after the eleventh century, at which time it shared many features in common with Buddhism. Before that, the pre-Buddhist native tradition of Tibet, sometimes confusingly

  3. #3 peak

    social group - a caste or community or religious minority, say - which can be counted on to back a specific party or candidate consistently. But who coined the term, and how did it come into use? This column by Ramachandra Guha

  4. #4 peak

    His mind and body seemed two separate things and the young soldier couldn’t stand at attention or even offer a salute. “I-I guess I can’t take it, sir,” Kuhl said.⏎The general’s already-flushed face grew crimson.

  5. #5 peak

    When it was over, DeMille yelled, “Cut!” and turned to his assistant, all smiles. “That was great!” he said.⏎“It was, C.B.,” the assistant yelled back. “It was fantastic! Everything went

  6. #6 peak

    like chewing tobacco or peyote. Basically, it’s like a temporary, flavored exercise machine for your mouth. So: who would come up with this idea, and why?⏎As with most everything in our contemporary culture, we have to start

Steering

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.

turned down a lot
Prompt

Here is something I have been thinking about lately: that that that, that, that, that, that, that, that, that, that, (that ( that ( that ( that ( that ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( (

Most negative on

Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.

  1. #1 peak

    which can result in the shrinking of the hair follicles. This would result in the inability of the hair follicle to produce hair efficiently and gradually called hair loss, receding hairline and

  2. #2 peak

    code of Chivalry, scientific, agricultural and artistic advancements all happened during this time.⏎The Code of Chivalry started in the court of Charlemagne 742-814. Charlemagne not only brought order to

  3. #3 peak

    who met in Vietnam on 22-26 November 2010 and who called on the leaders of the world to stop talking and take action on climate change.⏎They were brought together by the British Council Climate Generation Project

  4. #4 peak

    organized in Dallas – now Baylor University Medical Center.⏎1919-College of Arts and Sciences organized.⏎1919-College of Fine Arts organized, which consisted of offerings in music and in expression.⏎1921-Ter

Connections

Show on map

Which neurons in earlier layers feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.

Fed by earlier layers

  1. L14:2965 suppresses
  2. L13:1698 excites
  3. L17:1147 excites
  4. L17:1953 excites
  5. L15:123 suppresses
  6. L14:282 excites
  7. L15:1495 excites
  8. L12:575 excites
  9. L16:749 excites
  10. L13:2983 excites
  11. L17:2001 excites
  12. L14:1095 suppresses
  13. L14:775 suppresses
  14. L17:340 suppresses
  15. L15:44 suppresses
  16. L9:537 suppresses
  17. L17:424 weakly excites sentence-ending full stops
  18. L15:2804 weakly excites
  19. L1:2630 weakly suppresses
  20. L17:2622 weakly excites
  21. L17:1770 weakly suppresses
  22. L14:211 weakly suppresses
  23. L5:2305 weakly suppresses
  24. L17:494 weakly excites

Feeds later layers

  1. L19:501 strongly excites
  2. L20:79 strongly excites
  3. L19:223 excites
  4. L19:413 suppresses
  5. L19:1130 suppresses
  6. L19:211 excites
  7. L19:195 excites
  8. L19:608 suppresses
  9. L20:132 suppresses
  10. L21:720 suppresses
  11. L19:1044 suppresses
  12. L19:1201 suppresses
  13. L20:756 suppresses
  14. L19:535 excites
  15. L19:711 suppresses
  16. L20:202 suppresses
  17. L20:46 excites
  18. L21:359 suppresses
  19. L21:117 excites
  20. L22:597 suppresses end of document token
  21. L22:820 suppresses
  22. L24:35 excites
  23. L22:160 excites
  24. L24:240 suppresses

Activation histogram

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.

1 values between -6.68 and -6.371 values between -5.76 and -5.451 values between -5.45 and -5.142 values between -4.53 and -4.221 values between -3.91 and -3.611 values between -3.61 and -3.303 values between -3.30 and -2.992 values between -2.99 and -2.696 values between -2.69 and -2.385 values between -2.38 and -2.073 values between -2.07 and -1.7714 values between -1.77 and -1.4612 values between -1.46 and -1.1531 values between -1.15 and -0.8563 values between -0.85 and -0.54387 values between -0.54 and -0.232206 values between -0.23 and 0.08927 values between 0.08 and 0.38124 values between 0.38 and 0.6943 values between 0.69 and 1.0024 values between 1.00 and 1.3012 values between 1.30 and 1.6111 values between 1.61 and 1.9210 values between 1.92 and 2.229 values between 2.22 and 2.532 values between 2.53 and 2.843 values between 2.84 and 3.141 values between 3.14 and 3.451 values between 3.45 and 3.761 values between 4.37 and 4.681 values between 5.29 and 5.60offnegativepositiveoften

Logit lens

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 layers instead.

Promotes

  1. hers
  2. ześ
  3. alties
  4. &eacute
  5. bei
  6. /Edit
  7. .VideoCapture
  8. amphetamine
  9. že
  10. CreateInfo

Suppresses

  1. }
  2. }⏎⏎
  3. ()
  4. fuse
  5. ----------------------------------------------------------------------------
  6. }⏎
  7. ----------------------------------------------------------------
  8. }⏎⏎⏎
  9. Sim
  10. Sim