Neuron Explorer

Qwen3-0.6B

'the' before 'way' in idioms

A neuron in layer 26 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

It fires on the word 'the' when the phrase is likely to continue with 'way', as in 'paved the way' or 'along the way'. Turning it down made the model output the Chinese word for road or way over and over, which fits the same idea of a path.

Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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

    in his own practice, Zahoor-ul-Akhlaq absorbed the shocks and hurdles of artistic transition from modernism to contemporaneity in Pakistan. He paved the way for post-modern artists to develop their own vocabulary, drawing upon traditional sources

  2. #2 peak

    help. In 2007, she was trained by Gore and others to present The Inconvenient Truth slideshow as a Climate Project Ambassador, and paved the way for more presenters (including me!) to be

  3. #3 peak

    2017 Gender Inequality Index and Pakistan does not fare not much better, ranking at 133.⏎Here are three outstanding leaders who paved the way in Afghanistan and Pakistan and made huge strides toward gender equality by fighting for girls

  4. #4 peak

    south for the winter, traveling as fast as thirty miles an hour for hundreds of miles on end. The trip is long, with savage weather and many predators along the way, but Tiny Bird is built for this epic journey and eventually arrives at its

  5. #5 peak

    A friend who is upright is honest and sincere. He or she is transparent without a hint of flattery or fakeness. He will not lead you down the slippery road where there is a chance of you breaking the law and is sincerely caring

  6. #6 peak

    storage on a magnetic disc by solving the problem at the atomic level. It has managed to fit a single bit of data on just 12 atoms, paving the way for 400TB hard

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:道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道道

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

    , the dream of the popes to convert the main, political centre of Russian Orthodoxy to some form of union with Rome handed many tangible advantages to the grand dukes of the Kremlin, who cynically exploited this dream. In the name of

  2. #2 peak

    ern part of the Korean peninsula 3. The state of Paekche, because of the continuous pression exerted upon it by Koguryŏ 高句麗 4

  3. #3 peak

    and died 48 hours after birth.⏎Pig clones have popped up in the United States, the United Kingdom, Japan and other locations,<|endoftext|>Cleopatra, queen of Egypt and lover of Julius Caesar and Mark Antony, takes her

  4. #4 peak

    .⏎Looking anew at history from this perspective forces a reconsideration of accepted notions from “Greco-Roman civilization,” the “European miracle,” and the “Great Divergence” to “Japan’s modernization.” Through

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. L25:2196 strongly excites
  2. L24:2673 strongly excites
  3. L22:1826 suppresses
  4. L22:1465 excites
  5. L23:1107 suppresses
  6. L21:2871 excites
  7. L23:2661 suppresses
  8. L23:2037 suppresses
  9. L22:1489 suppresses
  10. L20:2358 excites
  11. L20:1122 suppresses
  12. L23:2957 suppresses
  13. L23:2477 suppresses
  14. L21:2547 suppresses
  15. L24:931 excites
  16. L21:2873 excites
  17. L17:1251 excites
  18. L23:2422 excites
  19. L23:1733 excites
  20. L5:1543 weakly suppresses
  21. L20:2047 weakly excites
  22. L12:2338 weakly suppresses
  23. L21:1681 weakly excites
  24. L24:1916 weakly suppresses

Feeds later layers

  1. L27:1167 strongly suppresses
  2. L27:2269 strongly suppresses
  3. L27:1509 weakly excites
  4. L27:3006 weakly suppresses
  5. L27:893 weakly excites
  6. L27:1809 weakly suppresses
  7. L27:1803 weakly excites
  8. L27:1104 weakly suppresses
  9. L27:1447 weakly suppresses
  10. L27:2006 weakly suppresses
  11. L27:1782 weakly suppresses
  12. L27:2750 weakly suppresses
  13. L27:1579 weakly suppresses
  14. L27:1601 weakly suppresses
  15. L27:2616 weakly excites
  16. L27:2581 weakly suppresses
  17. L27:2158 weakly suppresses
  18. L27:2709 weakly excites
  19. L27:449 weakly excites
  20. L27:2918 weakly suppresses
  21. L27:1680 weakly suppresses
  22. L27:1340 weakly suppresses
  23. L27:557 weakly excites
  24. L27:1722 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 -7.50 and -6.932 values between -6.36 and -5.791 values between -5.23 and -4.661 values between -4.09 and -3.522 values between -3.52 and -2.951 values between -2.95 and -2.384 values between -2.38 and -1.8114 values between -1.81 and -1.2465 values between -1.24 and -0.681129 values between -0.68 and -0.112249 values between -0.11 and 0.46345 values between 0.46 and 1.0346 values between 1.03 and 1.6012 values between 1.60 and 2.178 values between 2.17 and 2.749 values between 2.74 and 3.312 values between 3.31 and 3.882 values between 3.88 and 4.441 values between 4.44 and 5.012 values between 5.01 and 5.581 values between 5.58 and 6.153 values between 6.72 and 7.292 values between 8.99 and 9.563 values between 9.56 and 10.131 values between 10.13 and 10.701 values between 11.84 and 12.411 values between 14.68 and 15.25offnegativepositiveoften

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. ynos
  2. rador
  3. igators
  4. ätz
  5. symp
  6. ycop
  7. CBC
  8. cona
  9. ności
  10. curring

Suppresses

  1. path
  2. 道
  3. 路径
  4. paths
  5. path
  6. Path
  7. Path
  8. paths
  9. PATH
  10. Paths