Neuron Explorer

Qwen3-0.6B

'in' in 'this, in turn' phrases

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 lights up on the word "in" in set phrases like "This, in turn," or "which in some cases," where a sentence links a cause to its effect. Turning it up made the model's writing wander into unrelated topics like maths problems and multiple-choice questions, so the steering shows no clear link to the trigger.

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

    phenomenon has been a great source of joy for many entrepreneurs and enthusiasts. We’re one step away from actually connecting everyday objects to the Internet, and that in itself is a true revolution.⏎There’s something undeniably exciting about this possibility

  2. #2 peak

    more relevant and useful to compliment eHealth.⏎Translation of healthcare documents is necessary in order help all people to correctly understand the message that is being conveyed. That, in turn, will help them make informed medical choices.⏎Translation for

  3. #3 peak

    such as:⏎Diagnosing and treating sleep apnea can help control these conditions. Not treating sleep apnea can aggravate the symptoms of those other conditions—which in some cases can be

  4. #4 peak

    thrive when the weather is warm and dry. Whiteflies feed on the gardenia’s sap and secrete a sticky substance known as honeydew. This, in turn, attracts sooty mold, which grows wherever the honeydew is

  5. #5 peak

    3). This negative relationship is observed when formal leaders experience a heightened subjective sense of power. As a consequence, they will tend to dominate group discussions. This, in turn, will lead other team members to perceive that their perspectives are not valued.⏎

  6. #6 peak

    in which these are growing, it helps the undisturbed roots to develop more profusely and enables it to tap more nutrients from the soil. This, in turn, facilitates a larger number of tillers (shoots) per root-system

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: it is given in the book (or document) it is given in the document it is given in the document it is given in the document it is given in the document it is given in the document it is given in the document it is

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

    to End all Wars, against the Greek King Antiochus Epiphanes who decided to wipe the Jewish nation from the face of the earth in about 160-165 BCE. A priest named Mattathias son

  2. #2 peak

    CE, his brother Domitian became emperor. Titus married twice, but his first wife died, and he divorced the second soon after the birth (c. 65) of his only child, a daughter, Flavia Julia, to

  3. #3 peak

    6 BC, Tyre regained its independence (from the Seleucids) and was allowed to keep much of its independence when the area became a Roman province in 64 BC.<|endoftext|>Complete Green Streets⏎Streets in Madison move people

  4. #4 peak

    title of husband and wife. Cleopatra and Ptolemy were members of the Macedonian dynasty that governed Egypt since the death of Alexander the Great in 323 B.C. Although Cleopatra had no Egyptian blood, she

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. L18:363 excites
  2. L8:407 excites
  3. L20:1044 excites
  4. L16:1023 excites
  5. L19:1280 weakly excites end-of-document token between unrelated texts
  6. L7:1161 weakly suppresses
  7. L9:1947 weakly excites
  8. L7:969 weakly excites
  9. L2:2100 weakly suppresses
  10. L4:2361 weakly excites
  11. L1:2280 weakly excites
  12. L7:2311 weakly excites
  13. L18:2157 weakly excites
  14. L20:508 weakly excites
  15. L16:573 weakly suppresses
  16. L7:1752 weakly suppresses
  17. L18:828 weakly suppresses
  18. L18:1106 weakly suppresses
  19. L6:6 weakly excites
  20. L20:1120 weakly excites
  21. L4:289 weakly excites
  22. L1:2111 weakly excites
  23. L4:1599 weakly excites
  24. L13:594 weakly suppresses

Feeds later layers

  1. L25:2056 weakly suppresses
  2. L23:224 weakly suppresses
  3. L23:2736 weakly excites
  4. L22:2419 weakly excites
  5. L22:1702 weakly excites
  6. L25:306 weakly suppresses
  7. L24:853 weakly suppresses
  8. L25:832 weakly suppresses zero before a decimal point
  9. L22:2409 weakly suppresses
  10. L27:1502 weakly suppresses
  11. L24:146 weakly suppresses
  12. L23:1018 weakly suppresses
  13. L22:2086 weakly suppresses
  14. L22:661 weakly excites
  15. L25:3008 weakly suppresses
  16. L25:1669 weakly excites
  17. L27:1872 weakly excites
  18. L24:491 weakly excites
  19. L23:1222 weakly suppresses
  20. L22:2532 weakly excites
  21. L23:107 weakly suppresses
  22. L24:1234 weakly excites
  23. L23:2484 weakly excites number in days or time spans
  24. L24:1087 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.

1 values between -10.92 and -10.241 values between -4.79 and -4.111 values between -4.11 and -3.422 values between -2.74 and -2.0610 values between -2.06 and -1.3857 values between -1.38 and -0.702238 values between -0.70 and -0.021518 values between -0.02 and 0.6743 values between 0.67 and 1.3511 values between 1.35 and 2.032 values between 2.03 and 2.715 values between 2.71 and 3.392 values between 3.39 and 4.074 values between 4.07 and 4.762 values between 4.76 and 5.444 values between 5.44 and 6.123 values between 6.12 and 6.802 values between 12.25 and 12.941 values between 14.30 and 14.981 values between 15.66 and 16.34offnegativepositiveoften

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. änger
  2. arding
  3. witter
  4. arkan
  5. iscrim
  6. лем
  7. adium
  8. śmy
  9. później
  10. }-{

Suppresses

  1. document
  2. document
  3. (document
  4. Array
  5. Array
  6. .document
  7. Document
  8. дер
  9. $('.
  10. документ