Neuron Explorer

Qwen3-0.6B

comparison words like more, same, but

A neuron in layer 8 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 fires on words that set up a comparison, such as same, more, much, higher, except and but, as in 'similar to X, except...'. When it is pushed down, the model tends to write sentences with the word 'as' in a comparison, though the steered text was mostly unrelated.

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

    sentence as you and win a point! This is a great, non-threatening activity to encourage reluctant writers!⏎Predict the Sentence 2 – Same as above but this time students aim to write a unique sentence. They volunteer to read their sentence aloud.

  2. #2 peak

    ensure that other astronomers did not make the same mistake, he began compiling a list of 1oo of them. This list

  3. #3 peak

    and get a spinach leaf in reward. After a few days, the researchers replaced the spinach with a grape. Compared to spinach, a sugary grape is a much better reward, and the monkeys' dopamine soared. Their non-verbal brains thought

  4. #4 peak

    80hp Le Rhône engine. The Nieuport 12 itself was a larger version of the Nieuport 10, with a more powerful 110-130hp engine.⏎The Nieuport

  5. #5 peak

    consumption. Some people may snore every night, while others snore only occasionally.⏎What is sleep apnea?⏎Sleep apnea is similar to snoring, except that the airway becomes completely constricted. As a result, it becomes impossible

  6. #6 peak

    with this intent can immediately report the perpetrator to the appropriate authorities. Mail fraud is a federal crime.⏎- Wire Fraud⏎Wire fraud is similar to mail fraud, but rather than using the mail to perpetrate the deception, it involves the use of

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: A square. (The original) What is the problem? A square. (The original) What is the problem? A square. ( The original) a square. ( The original) a square. ( The original) a square.

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

    unpredictable."⏎4. Is the following statement true, false or uncertain? "Under full liquidity constraints the marginal propensity to consume out of current income is generally at least as large as under no liquidity constraints."⏎5. Is the following statement true, false

  2. #2 peak

    , and the width of each step. The height of the landing is important because it will determine how steep the stairs can be. The landing should be at least as high as the bottom stair, and it’s best to make it higher if possible

  3. #3 peak

    ude's reaction to the statement. By her reaction, he may be able to judge whether she has knowledge of Claudius' crime. He says, "Almost as bad...As kill a king, and marry with his brother." Gertrude

  4. #4 peak

    University of Maine Cooperative Extension.⏎Washing your fruit with a "homemade wash" without spending a lot of money on store-bought products proves to be just as effective, according to the University of Maine Cooperative Extension. Soak your fruits and

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. L4:344 suppresses
  2. L7:213 excites
  3. L5:2308 suppresses
  4. L7:2523 excites 'since' after a change or growth
  5. L3:1316 suppresses
  6. L4:1058 weakly excites
  7. L7:2248 weakly suppresses
  8. L7:1088 weakly suppresses
  9. L7:2116 weakly excites
  10. L7:1978 weakly excites
  11. L4:181 weakly suppresses
  12. L4:2103 weakly excites
  13. L5:2067 weakly suppresses
  14. L6:2066 weakly excites 'of' linking a category to its parent field
  15. L6:960 weakly excites
  16. L5:2919 weakly excites
  17. L5:2118 weakly suppresses
  18. L2:2277 weakly suppresses
  19. L1:1999 weakly excites
  20. L5:1552 weakly suppresses
  21. L1:2937 weakly excites
  22. L6:974 weakly suppresses
  23. L3:2532 weakly excites
  24. L2:1792 weakly excites

Feeds later layers

  1. L12:1153 suppresses
  2. L11:490 suppresses
  3. L10:1752 suppresses
  4. L11:1895 suppresses
  5. L9:2725 suppresses
  6. L13:1820 suppresses
  7. L15:1145 suppresses
  8. L9:750 suppresses
  9. L10:1152 suppresses
  10. L17:2880 suppresses
  11. L14:867 suppresses
  12. L11:458 suppresses
  13. L16:2061 weakly suppresses
  14. L13:2979 weakly suppresses
  15. L22:1630 weakly suppresses
  16. L11:139 weakly excites
  17. L16:1645 weakly suppresses
  18. L16:2395 weakly suppresses
  19. L15:2006 weakly suppresses
  20. L18:1824 weakly suppresses
  21. L11:1713 weakly suppresses
  22. L18:2951 weakly suppresses
  23. L22:2960 weakly suppresses
  24. L11:2844 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 -3.05 and -2.931 values between -2.18 and -2.061 values between -1.81 and -1.682 values between -1.68 and -1.562 values between -1.56 and -1.441 values between -1.44 and -1.311 values between -1.31 and -1.191 values between -1.19 and -1.064 values between -1.06 and -0.944 values between -0.94 and -0.822 values between -0.82 and -0.6910 values between -0.69 and -0.5720 values between -0.57 and -0.4435 values between -0.44 and -0.32112 values between -0.32 and -0.20438 values between -0.20 and -0.071625 values between -0.07 and 0.051062 values between 0.05 and 0.18400 values between 0.18 and 0.30116 values between 0.30 and 0.4335 values between 0.43 and 0.5514 values between 0.55 and 0.674 values between 0.67 and 0.804 values between 0.80 and 0.921 values between 0.92 and 1.055 values between 1.05 and 1.171 values between 1.42 and 1.542 values between 1.54 and 1.673 values between 1.79 and 1.91offnegativepositiveoften

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. 대학
  2. ami
  3. LIABLE
  4. azi
  5. ////////////////////////////////////////////////////////////////////////////////⏎⏎
  6. 才可以
  7. 壕
  8. 谩
  9. #⏎⏎
  10. paired

Suppresses

  1. returnUrl
  2. downstream
  3. Flow
  4. eral
  5. 饴
  6. arket
  7. orama
  8. flows
  9. orph
  10. thalm