© 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. Gemma-4-31B
    3. 30-RES-MATRYOSHKA-131K
    4. 118267
    Prev
    Next
    INDEX
    Explanations

    Processing Neuron LogicI am currently synthesizing the neuron's function from the provided lists, aiming for a brief, output-focused description of its core activity without employing forbidden terminology.**MAX_ACTIVATING_TOKENS**:7688644**Techn**[**María****TOKENS_AFTER_MAX_ACTIVATING_TOKEN**::::**:,:,**ological****Diversity****de****TOP_POSITIVE_LOGITS**:rättvišelostivarious证据fathфактиمی‌شودOccurrenceprincipally**TOP_ACTIVATING_TEXTS**:Erratum to: Clin Transl Oncol (2015) 17:996--1004 DOI 10.1007/s1209(J Am Heart Assoc. 2017;6:e007026 DOI: 10.1161/JAHA.11ano P. The timing of frugivore‐mediated seed dispersal effectiveness. Mol Ecol. 2019;28:219--231. 10.1111/mec.1485019. A regulatory module controlling stress-induced cell cycle arrest in \*Arabidopsis\*. \*eLife\* **8**:e43944. doi: [10.7554/eLife.43Microbiol., 9, 325 (1961), and Progr. Ind. Microbiol, 6, 21 (1967). Rifamycin derivatives have been prepared by the chemical conversion of rifand a quarter of drugs in preclinical and clinical trials (Walsh, 2006, Nat Biotechnol 24:769-776). As the demand for biopharmaceuticals continues to increase, there is a commensurate needThe validity of the proposal by Shuren et al. (Neuropsychologia, 1996, 34, 491-492) that there is some relation between handedness and hemi-imagers was**Technological** University, Pathein Technological University (Pathein) is situated at the Apin-ne-se village[**Diversity** of culturable butane-oxidizing bacteria in oil and gas field soil]. Butane-oxidizing bacteria in**María de los Llanos de Luna** María de los Llanos de Luna Tobarra (born 3 March 19Looking at MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN:- "Techn" is followed by "ological" to form "Technological".- "[" is followed by "Diversity".- "María" is followed by "de los", which starts the name "María de los Llanos de Luna".These indicate the neuron is looking for specific entities or terms and the words that generally follow them. The accompanying TOP_POSITIVE_LOGITS list shows words like 'various', 'Occurrence', 'rätt' (correct), 'više' (more/many), 'fakt' (facts), 'evidence', suggesting a focus on naming, classification, or factual statements.The pattern is identifying specific names/titles and specific technical terms followed by their usual descriptors or continuation.Let's refine:- "Techn" + "ological" -> Scientific/Technical Institutions- "[" + "Diversity" -> Topics/Categories- "María" + "de" -> Specific Names (People/Places)The neuron seems to activate when it encounters these kinds of structured sequences, often found in academic citations or descriptive entries. The overall theme is identifying specific entities or concepts followed by their common contexts or continuations. "Erratum", citations with DOIs and often ":", lists of organisms or substances, academic references, and organizational names all fit this.The core behavior is recognizing specific entities and their typical contextual follow-ups.Considering the rule "Keep your explanation concise (3 to 20 words)" and "Do not mention 'tokens' or 'patterns'"."Technological University""[Diversity""María de"This neuron seems to identify specific entities and then the words that commonly follow them to complete a phrase or title. The commonality is "Entity + specific follow-up word(s)".Let's look at `TOP_POSITIVE_LOGITS`: various, Occurrence,rätt, više, 证据, fath, факти, می‌شود, principally.These are words that suggest categorization, quantity, or factual information.The neuron picks up on things like:- Specific names/titles starting a new entry or section ("Technological University", "María de los Llanos de Luna", "[Diversity").- Numbers often followed by colons, typical in citations or data points.The common thread is identifying specific *startings* of things and then looking at what comes immediately *after* them.It could be summarizing structured data or definitions.Let's try to capture the essence:- Identifying specific beginnings of terms/phrases.- The tokens *after* the max activating token often complete a name or title, or provide context.Possible explanations:- specific beginnings and their continuations- entity identification and context- names, citations, scientific terms- identifying structured entriesThe terms "[Diversity", "Technological", "María de" are specific entities. The tokens after them ("Diversity", "ological", "de") are continuations.How about focusing on the *type* of things it identifies? They are often formal, technical, or proper nouns."specific official titles and names" - This is 5 words. It fits the examples."formal names and identifiers" - 4 words."recognizing specific names and terms" - 5 words.Let's look at the TOP_POSITIVE_LOGITS again: various, Occurrence, bukti (evidence), facts, principally. These are suggestive of factual information or classification.The texts include citations, institution names, and scientific research descriptions.The neuron seems to detect phrases that introduce specific entities or concepts which are then further described or classified.Consider the combination:- MAX_ACTIVATING_TOKENS: "Techn" -> TOKENS_AFTER_MAX_ACTIVATING_TOKEN: "ological" -> "Technological"- MAX_ACTIVATING_TOKENS: "[" -> TOKENS_AFTER_MAX_ACTIVATING_TOKEN: "Diversity" -> "[Diversity"- MAX_ACTIVATING_TOKENS: "María" -> TOKENS_AFTER_MAX_ACTIVATING_TOKEN: "de" -> "María de..."This strongly suggests identifying the start of something specific and what immediately follows.The "TOP_POSITIVE_LOGITS" are more general terms like 'various', 'occurrence', 'evidence', 'facts', 'principally'. This suggests the neuron is sensitive to the *context* or *nature* of the identified entities, implying classification or factual assertion."Identifying formal, classified information" - 4 words."Formal entity and classification recognition" - 4 words."Structured names and factual indicators" - 4 words.The prompt

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    decoderesearch/gemma-4-saes/gemma-4-31b
    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
    Some
    -0.07
    What
    -0.07
     what
    -0.07
     What
    -0.06
    Year
    -0.06
     Some
    -0.06
    what
    -0.06
    Mike
    -0.06
    EVERY
    -0.06
    结束后
    -0.06
    POSITIVE LOGITS
     rätt
    0.06
     više
    0.06
    losti
    0.06
    various
    0.06
    证据
    0.06
     fath
    0.06
     факти
    0.06
     می‌شود
    0.06
     Occurrence
    0.06
     principally
    0.05
    Activations Density 0.001%

    No Known Activations