Neuron Explorer

Qwen3-0.6B

'well' in compound adjectives

A neuron in layer 21 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 the word 'well' when it starts a compound adjective like 'well-known' or 'well-organized', or in phrases like 'a well known fact'. Turned up, the model's text became repetitive and odd. Turned down, it produced looping sentences about 'the best' or 'young'.

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

    ons of England organized numerous uprisings. In the face of such strong and well-organized opposition, King John was forced to renounce certain rights and to grant

  2. #2 peak

    with minefields and traps. Following Prime Minister Narendra Modi’s decision to strike at terrorists across the Line of Control in 2016, Pakistan responded with well-tested covert tools—ratcheting up the levels of violence against Indian forces in

  3. #3 peak

    focuses on spreading appreciation for the arts and academics all throughout Asia and Europe. One of the latest acquisitions of the organisation is Kubota’s kimono collection, a well-acclaimed exhibit of the late Itchiku Kubota, a textile artist in

  4. #4 peak

    At the conclusion of the novel, Gilly still faces difficult choices, but she knows herself better and is able to judge others more soundly. Paterson's well-d

  5. #5 peak

    , stability factors and, last but not least, reaction speed. Countless sporting activities and games such as running, jumping but also catching or throwing objects require some well-developed motor skills that will allow the child to participate in them. So,

  6. #6 peak

    're gone.⏎The title of this well-known story comes from a poem by Sara Teasdale in which nature is depicted

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: I have a 300,000/1000/300,000/300,000/1000/300,000/30

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

    1 you to take the former path, to teach only naturalism. 2 They have asked you to reject the Science Education Standards adopted by the Kansas State Board of Education on August 11, 1999 (the "New

  2. #2 peak

    officer of police to institute any necessary proceedings at the cost of the defaulting council.⏎In 1901 the president of the Board of Trade it troduced a bill to continue the act until 1906

  3. #3 peak

    ayashi is Associate Professor in the Department of Media and Communication at City University of Hong Kong.⏎Atsushi Tago is Professor in the School of Political Science

  4. #4 peak

    take the Library.?1 This is from the ?Prospectus? for Horace Mann?s The School Library, a series of books endorsed by the state Board of Education for use in Massachusetts? newly formed district school libraries. The promotion of ...⏎

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. L20:1248 excites
  2. L19:1578 excites
  3. L19:1225 excites
  4. L20:1976 excites
  5. L4:521 suppresses
  6. L20:2482 weakly suppresses
  7. L20:1536 weakly excites the word 'well' in phrases
  8. L20:2698 weakly excites
  9. L18:2193 weakly excites "well" before known or documented
  10. L5:1755 weakly excites
  11. L8:2961 weakly excites
  12. L19:1159 weakly suppresses
  13. L15:2846 weakly excites
  14. L20:743 weakly excites
  15. L20:679 weakly suppresses
  16. L11:429 weakly suppresses
  17. L18:1729 weakly excites
  18. L13:562 weakly suppresses
  19. L8:2801 weakly excites
  20. L9:2208 weakly suppresses
  21. L18:736 weakly suppresses
  22. L4:1772 weakly excites
  23. L18:192 weakly suppresses
  24. L3:2566 weakly suppresses

Feeds later layers

  1. L24:422 strongly excites
  2. L23:575 excites
  3. L27:217 excites
  4. L27:716 excites
  5. L27:1211 excites
  6. L26:138 excites
  7. L27:869 excites
  8. L26:1940 excites
  9. L25:644 excites
  10. L26:395 excites
  11. L26:149 excites
  12. L22:494 excites
  13. L22:200 excites
  14. L25:149 excites
  15. L27:1832 excites
  16. L26:632 excites
  17. L23:740 excites 'well' in well-dressed, well-known, well-off
  18. L24:733 weakly excites
  19. L24:239 weakly excites
  20. L22:667 weakly excites
  21. L26:144 weakly excites
  22. L26:596 weakly excites
  23. L25:349 weakly excites
  24. L27:1281 weakly excites

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.

92 values between -3.09 and -1.373674 values between -1.37 and 0.3567 values between 0.35 and 2.0730 values between 2.07 and 3.7911 values between 3.79 and 5.5111 values between 5.51 and 7.233 values between 7.23 and 8.955 values between 8.95 and 10.674 values between 10.67 and 12.392 values between 12.39 and 14.112 values between 14.11 and 15.832 values between 19.27 and 20.991 values between 22.70 and 24.421 values between 34.74 and 36.461 values between 36.46 and 38.181 values between 48.49 and 50.211 values between 63.97 and 65.69offnegativepositiveoften

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. -written
  2. -formed
  3. formed
  4. constructed
  5. formulated
  6. owned
  7. written
  8. spelled
  9. placed
  10. built

Suppresses

  1. osal
  2. ぐ
  3. 全身
  4. ocity
  5. SALE
  6. zej
  7. メント
  8. ')(
  9. xcf
  10. éments