{"id":"W4415600217","doi":"10.20944/preprints202510.2064.v1","title":"Monosemantic Feature Neurons: A Sparse Autoencoding Layer for Interpretable, Steerable Transformer Features","year":2025,"lang":"","type":"preprint","venue":"Preprints.org","topic":"Generative Adversarial Networks and Image Synthesis","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"Bell (Canada)","funders":"","keywords":"Interpretability; Transformer; Heuristics; Bottleneck; Computation; Residual; ENCODE; Pattern recognition (psychology)","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0006480625,0.0006484552,0.0003889465,0.0002374088,0.0002257355,0.0007568803,0.0009648821,0.0006334598,0.003727941],"category_scores_gemma":[0.002193419,0.0003262796,0.0004566729,0.0002187627,0.0008098643,0.001901518,0.001405286,0.001270718,0.0009975391],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004933618,"about_ca_system_score_gemma":0.0005493333,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0009649253,"about_ca_topic_score_gemma":0.002198595,"domain_scores_codex":[0.9998006,0.00005484801,0.00001140462,0.00005547761,0.00004799236,0.00002959641],"domain_scores_gemma":[0.9995266,0.0002277407,0.00004516085,0.0001155147,0.00005368694,0.0000312118],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0004258805,0.0001359374,0.00189348,0.0002318696,0.0001254897,0.000270132,0.0002811148,0.3480042,0.05735291,0.1996505,0.005930167,0.3856983],"study_design_scores_gemma":[0.000008115667,0.00005727906,0.0001485182,0.00001511034,0.00001598314,0.00007238065,0.00001763506,0.9340013,0.01254951,0.05094781,0.002154722,0.00001166206],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.02498698,0.000149619,0.970462,0.0002142482,0.00004861116,0.00003012975,0.0001576016,0.001091438,0.002859405],"genre_scores_gemma":[0.7647758,0.0002368951,0.2278079,0.0002454708,0.00004946001,0.00009412169,0.0004520448,0.0003160757,0.006022209],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.003727941,"threshold_uncertainty_score":0.0124712,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.09834803971921124,"score_gpt":0.3331170462294064,"score_spread":0.2347690065101952,"validation_status":"score_only:v0-immature-baseline","note":"Baseline scores from an immature model (maturity gate not passed). Scores rank; they never assert a category."}}