Neuron Explorer

Qwen3-0.6B

predicts the word 'state'

A neuron in layer 25 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. It is featured, so it has steering outputs.

What this neuron does

This neuron looks like it helps the model get ready to write the word 'state', and it is also active on word beginnings like 'ren' in renounce. Turning it up made the model repeat 'state' again and again. It switches off on the word 'social'.

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

    land and started settling down there.⏎- By 1916, Palestine came under British control after the Sykes-Picot Agreement (a secret agreement between Great Britain and France). This led to the division of the old

  2. #2 peak

    .⏎Sponges live under water. Most sponges live in salt water. The 5,000 species of sponge belong to the Phylum Porifora. They are further classified into four groups based on what comprises their skeletons

  3. #3 peak

    783). Finally, in 1819, Spain ceded Florida to the United States in exchange for the American renunciation of any claims on Texas. On March 3, 184

  4. #4 peak

    ‘world’s most intractable conflict’ – is rooted in a dispute over land claimed by Jews as their biblical birthright and by the Palestinians, who seek self-determination. Despite repeated attempts to end the conflict between the two countries, there

  5. #5 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 a charter of liberties. This document stated that the

  6. #6 peak

    ." In contrast, many psychologists have taken the self very seriously, and discussed at length a huge number of important phenomena including self-identity, self-esteem, self-regulation, and self-improvement. Is it possible to have a psychologically interesting

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 want to get good at making good value line design, I want to create good value line design, I want to create good value line design, I want to create good value line design, I want to create good value line design,

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

    psychology, which has found gossip to be an important means for people to monitor cooperative reputations and so maintain widespread indirect reciprocity. Indirect reciprocity is a social interaction in which one actor helps another and is then benefited by a third party.

  2. #2 peak

    under multi-objective multiple-participant environment is very promising. Coupled with multi-objective methods (e. g

  3. #3 peak

    diagram! Does this economy converge to a steady state from any initial capital stock k0?⏎d) Is the private market equilibrium socially efficient from the viewpoint of a social planner that seeks to maximize the representative consumers utility function ? Explain why or why not

  4. #4 peak

    - [“Computer Science”,”Economics”,”Political Science”,”Cognitive Psychology”]⏎- Carnegie-Mellon University⏎- Bounded Rationality, Satisficing⏎Herbert Simon spent the bulk of his career as a

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. L21:873 suppresses
  2. L9:510 excites
  3. L20:871 weakly suppresses
  4. L22:121 weakly suppresses
  5. L20:1461 weakly suppresses
  6. L10:154 weakly excites
  7. L23:1676 weakly excites
  8. L9:2540 weakly excites
  9. L1:2754 weakly excites
  10. L8:589 weakly excites
  11. L6:2001 weakly excites
  12. L3:1455 weakly excites
  13. L22:740 weakly suppresses
  14. L20:559 weakly suppresses
  15. L12:948 weakly suppresses
  16. L22:1861 weakly suppresses
  17. L21:972 weakly suppresses
  18. L7:2945 weakly suppresses
  19. L7:2607 weakly suppresses
  20. L9:205 weakly suppresses
  21. L4:1358 weakly suppresses
  22. L18:1418 weakly excites
  23. L0:2218 weakly suppresses
  24. L17:1169 weakly excites

Feeds later layers

  1. L27:2824 weakly excites
  2. L26:1194 weakly suppresses
  3. L26:2265 weakly excites
  4. L27:1029 weakly excites
  5. L26:2913 weakly excites
  6. L26:1991 weakly excites
  7. L26:2645 weakly excites
  8. L26:2594 weakly suppresses
  9. L27:3003 weakly excites
  10. L26:2296 weakly suppresses
  11. L27:1906 weakly excites
  12. L27:942 weakly excites
  13. L26:2771 weakly excites
  14. L26:1225 weakly suppresses
  15. L27:2767 weakly suppresses
  16. L26:1800 weakly suppresses
  17. L27:2489 weakly excites
  18. L27:2161 weakly excites
  19. L27:1877 weakly suppresses
  20. L27:340 weakly excites parenthesis introducing an acronym
  21. L27:792 weakly suppresses
  22. L27:2635 weakly excites
  23. L27:1998 weakly suppresses
  24. L26:1687 weakly 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.

2 values between -16.22 and -15.371 values between -14.51 and -13.662 values between -12.81 and -11.961 values between -11.96 and -11.103 values between -11.10 and -10.255 values between -10.25 and -9.403 values between -9.40 and -8.5411 values between -8.54 and -7.698 values between -7.69 and -6.8411 values between -6.84 and -5.999 values between -5.99 and -5.1313 values between -5.13 and -4.2832 values between -4.28 and -3.4346 values between -3.43 and -2.5882 values between -2.58 and -1.72198 values between -1.72 and -0.871461 values between -0.87 and -0.021662 values between -0.02 and 0.84170 values between 0.84 and 1.6969 values between 1.69 and 2.5431 values between 2.54 and 3.3920 values between 3.39 and 4.2522 values between 4.25 and 5.1014 values between 5.10 and 5.9513 values between 5.95 and 6.812 values between 6.81 and 7.664 values between 7.66 and 8.512 values between 8.51 and 9.363 values between 9.36 and 10.223 values between 10.22 and 11.071 values between 11.07 and 11.921 values between 12.77 and 13.631 values between 14.48 and 15.331 values between 15.33 and 16.191 values between 17.04 and 17.89offnegativepositiveoften

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. state
  2. →state
  3. nesia
  4. <State
  5. -state
  6. [state
  7. chóng
  8. state
  9. /Users
  10. (state

Suppresses

  1. SYN
  2. ige
  3. auté
  4. rove
  5. ериал
  6. ofire
  7. rif
  8. 代
  9. 羡
  10. Tunisia