© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    Jacobian LensNEW
    Natural Language
    Autoencoders
    NEW
    Assistant AxisNEWCircuit TracerUPDATESteerSAE EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Qwen3.5-4B
    3. 15-RES-MATRYOSHKA-65K
    4. 19944
    Prev
    Next
    INDEX
    Explanations

    The user wants me to explain the neuron's behavior with a short phrase (3-20 words) based on the provided lists. I need to find a pattern.Let's examine the lists:1. **MAX_ACTIVATING_TOKENS**: `.`, `.`, `,`, `.`, `!`, `!).`, `.`, `levels`, `)).`, `.` * This list is heavily dominated by punctuation: `.`, `,`, `!`, `)`. * It contains "levels".2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `T`, `thank`, `and`, `There`, `The`, `Breakfast`, `,`, `A`, `But` * These are words that often appear at the start of phrases or sentences, or important nouns.3. **TOP_POSITIVE_LOGITS**: `aret`, `водиÑĤели`, `åľ¨è§Ħå®ļ`, `NSS`, `nez`, `ulers`, `çģ«äºĨ`, `iba`, `alun`, `ève` * These look like tokens from non-English languages or potentially corrupted text.4. **TOP_ACTIVATING_TEXTS**: * "his structure... to" - ends with "to" * "making our daughters" - text example * "cybernetic law-enforcement “unit” called RoboCop... the" - text example * "diet by mouth. There was no operative mortality and the patients survived 2 to 14" - ends with "14" * "rich sauce. The toasted bun had just the right amount of" - ends with "of" * "(which was only a week out!)." - ends with "out!" * "Breakfast was AMAZING!! The staff was very efficient and we had everything we could possibly desire! I liked that there was" - ends with "was" * "glucose levels, temporarily curing their diabetes. In type 1 diabetes, the body’s immune system mistakenly destroys the pancreatic beta and" - contains "levels", ends with "and" * "detection level (1 cfu/100 cm(2)). A shelf-life study lasting" - contains "level", ends with "lasting" * "linked ads In the first screenshot you can see web" - ends with "web"Let's try to find a unifying pattern.The `MAX_ACTIVATING_TOKENS` are mostly punctuation. The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are often words that start sentences or clauses.Consider the pattern of punctuation followed by a word:- The `MAX_ACTIVATING_TOKENS` are often single characters or short strings, frequently punctuation.- The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are words that often follow punctuation.- The `TOP_ACTIVATING_TEXTS` show examples where phrases end with punctuation (like `out!`, `was`, `and`, `lasting`, `web`) and sometimes the following token is given.The presence of `levels` in `MAX_ACTIVATING_TOKENS` is a bit of an outlier if we focus *only* on punctuation. However, the overwhelming majority are punctuation.Let's look at `TOP_POSITIVE_LOGITS` again. They are indeed foreign/garbled. This suggests the neuron might be attuned to specific character sets or perhaps text processing artifacts that result in such tokens.Let's reconsider the `MAX_ACTIVATING_TOKENS`.`.`, `.`, `,`, `.`, `!`, `!).`, `.`, `levels`, `)).`, `.`What if the neuron detects endings and specific types of characters? The punctuation suggests sentence/phrase endings. `levels` might be a specific word it's sensitive to.The `TOP_ACTIVATING_TEXTS` show examples ending in `.`, `!` `)`, `)`, `.`And the `MAX_ACTIVATING_TOKENS` are `.`, `.`, `,`, `.`, `!`, `!).`, `.`, `levels`, `)).`, `.`What if it's about specific markers or characters that signal the end of a thought or a specific informational unit, possibly in conjunction with foreign characters or specific words like "levels"?The prompt asks for "what the neuron detects or predicts by finding patterns in lists."The `MAX_ACTIVATING_TOKENS` are very strong indicators of *what the neuron is seeing directly*. The fact that it's almost *all* punctuation is a major clue. The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` show what comes *after* these punctuation marks.Maybe it's detecting sentence endings or specific tokens often found at the end of clauses or phrases, and also some specific foreign/technical terms.Let's try to summarize the punctuation aspect:- `MAX_ACTIVATING_TOKENS`: many punctuation marks.- `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: words following the punctuation.sentence endings and foreign characters

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    decoderesearch/qwen-3.5-saes/qwen-3.5-4b
    Prompts (Dashboard)
    16,384 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    reek
    -0.05
    FF
    -0.05
    éı
    -0.05
    æĬĺ
    -0.05
    ä¸įå°½
    -0.05
    klad
    -0.05
    c
    -0.05
     Greg
    -0.05
     doch
    -0.05
    peg
    -0.05
    POSITIVE LOGITS
    aret
    0.07
    водиÑĤели
    0.07
    åľ¨è§Ħå®ļ
    0.06
    NSS
    0.06
    nez
    0.06
    ulers
    0.06
    çģ«äºĨ
    0.06
    iba
    0.06
     alun
    0.06
    ève
    0.06
    Activations Density 0.005%

    No Known Activations