Neuron Explorer

Qwen3-0.6B

unclear: end-of-text markers and word fragments

A neuron in layer 9 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 most strongly on the hidden marker that signals the end of one document, right where an unrelated text would begin. It is hard to pin down beyond that. When it was turned up, the model's replies drifted into unrelated topics, with no clear theme.

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

    Siberian ginseng, improves stamina and enhances immune function, as do reishi mushrooms — another adaptogenic remedy. Ashwagandha, an Indian<|endoftext|>Moral FriendSeries teasers are where you try to complete the sequence of a

  2. #2 peak

    data. Further, the cloud and artificial intelligence have a symbiotic relationship. The cloud handles the fundamental source of information that augments artificial intelligence networks, and artificial<|endoftext|>If I were to think of an emblematic phrase, in arts education, it

  3. #3 peak

    4, 1957, Sputnik, the Soviet satellite that was the first to orbit Earth, shocked Americans and delighted the Soviets. A month<|endoftext|>O’Connell, Caitlin and Donna M. Jackson. The Elephant Scientist.

  4. #4 peak

    , and encouraged trade and industry, dispatching a seaborne trading expedition to Punt, a land on the Somali coast of Africa, and exploiting the copper<|endoftext|>Industrial design is a process of design applied to products and industrial goods that are to

  5. #5 peak

    Prev Med Rep⏎Authors: Altman EA, Lee KL, Hecht CA, Hampton KE, Moreno G

  6. #6 peak

    queen Violant sought support for her grandchildren from her brother, King Peter III of Aragon, who agreed to protect and guard them in the kingdom of Aragon<|endoftext|>ALR provided recommendations that helped inform the national Childhood Obesity Task Force action plan,

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: In the context of the problem, the user is asking for the sum of the digits of the product, so we need to compute 10^k, for k=1, 2,..., 1000. But

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

    blackberries, pears, raspberries, strawberries, whole grain breads and bran cere

  2. #2 peak

    gardenia.⏎Insecticides kill off the whitefly’s natural predators. These predators include

  3. #3 peak

    side effects”⏎While the changes are only a few percent, with more than 40

  4. #4 peak

    Porcelain⏎Process – Raw Porcelain is rolled to create the tile, each letter is

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. L4:1411 strongly excites
  2. L8:269 strongly excites
  3. L3:1392 strongly excites
  4. L8:624 strongly excites
  5. L7:1130 strongly excites
  6. L2:1519 strongly suppresses
  7. L7:406 suppresses
  8. L3:921 suppresses
  9. L2:182 suppresses
  10. L8:313 suppresses
  11. L6:927 suppresses
  12. L7:234 excites
  13. L7:241 suppresses
  14. L5:112 suppresses
  15. L5:1341 excites
  16. L7:1866 excites
  17. L8:1053 suppresses
  18. L2:2010 excites
  19. L8:154 excites
  20. L2:699 excites
  21. L6:2834 excites
  22. L7:911 suppresses
  23. L8:49 weakly excites
  24. L1:1337 weakly suppresses

Feeds later layers

  1. L12:978 strongly excites
  2. L10:710 strongly excites
  3. L10:369 strongly suppresses
  4. L11:1143 strongly suppresses
  5. L12:354 strongly suppresses
  6. L10:762 strongly suppresses
  7. L12:933 suppresses mid-sentence 'also' and 'another'
  8. L12:2162 suppresses
  9. L12:2508 suppresses
  10. L14:3021 suppresses
  11. L12:2618 suppresses document boundary token
  12. L15:3039 suppresses
  13. L12:1302 suppresses
  14. L14:2862 suppresses
  15. L15:2008 excites
  16. L17:2032 suppresses
  17. L10:1749 excites
  18. L16:2255 suppresses
  19. L17:658 excites
  20. L14:1023 suppresses
  21. L11:1719 suppresses dash starting a bullet item
  22. L10:286 excites
  23. L15:2742 suppresses
  24. L20:2391 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.

3 values between -9.39 and -9.131 values between -9.13 and -8.882 values between -8.62 and -8.372 values between -8.37 and -8.111 values between -8.11 and -7.852 values between -7.85 and -7.602 values between -7.60 and -7.342 values between -7.34 and -7.082 values between -7.08 and -6.831 values between -6.83 and -6.572 values between -6.57 and -6.325 values between -6.32 and -6.061 values between -6.06 and -5.806 values between -5.80 and -5.555 values between -5.55 and -5.299 values between -5.29 and -5.039 values between -5.03 and -4.785 values between -4.78 and -4.528 values between -4.52 and -4.2614 values between -4.26 and -4.016 values between -4.01 and -3.7514 values between -3.75 and -3.5018 values between -3.50 and -3.2419 values between -3.24 and -2.9820 values between -2.98 and -2.7324 values between -2.73 and -2.4723 values between -2.47 and -2.2126 values between -2.21 and -1.9630 values between -1.96 and -1.7044 values between -1.70 and -1.4552 values between -1.45 and -1.1974 values between -1.19 and -0.9388 values between -0.93 and -0.68163 values between -0.68 and -0.42301 values between -0.42 and -0.161370 values between -0.16 and 0.091456 values between 0.09 and 0.3596 values between 0.35 and 0.612 values between 0.61 and 0.86offnegativepositiveoften

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. thon
  2. 名叫
  3. �
  4. 以下是
  5. 金银
  6. ...⏎⏎⏎⏎
  7. 牌子
  8. (_,
  9. 到底
  10. JECT

Suppresses

  1. herself
  2. онт
  3. auer
  4. ост
  5. rame
  6. himself
  7. Himself
  8. orrent
  9. át
  10. ienda