{"id":"W4416036894","doi":"10.18653/v1/2025.emnlp-main.519","title":"SAEs Are Good for Steering – If You Select the Right Features","year":2025,"lang":"","type":"article","venue":"","topic":"Vehicle Dynamics and Control Systems","field":"Engineering","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Azrieli Foundation; European Commission; Open Philanthropy Project","keywords":"Set (abstract data type); Feature (linguistics); Matching (statistics); Identification (biology); Key (lock)","routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":true},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002555125,0.001903749,0.0009079761,0.001050834,0.0004841303,0.00177122,0.0008797965,0.001558948,0.009967233],"category_scores_gemma":[0.01263996,0.0006321447,0.0009565331,0.0008134391,0.0008628376,0.003371678,0.001731699,0.001742258,0.009007467],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002562922,"about_ca_system_score_gemma":0.0005999468,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001745617,"about_ca_topic_score_gemma":0.004472143,"domain_scores_codex":[0.9982908,0.0005448182,0.0001003992,0.0003842934,0.0005353245,0.0001442635],"domain_scores_gemma":[0.9949799,0.001938837,0.0004112384,0.001122898,0.00133482,0.0002123364],"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":[0.0007139757,0.0003208463,0.01611786,0.000305457,0.0002960748,0.000132475,0.0001723809,0.07071663,0.02677695,0.006445409,0.02994759,0.8480545],"study_design_scores_gemma":[0.0001096896,0.0006293773,0.01731328,0.0001285402,0.0001347463,0.0004737121,0.0003987383,0.8839604,0.03626288,0.0372714,0.02316598,0.0001513489],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.05906294,0.0008244091,0.9168466,0.001244009,0.0003261038,0.0001603865,0.001034681,0.01036861,0.01013224],"genre_scores_gemma":[0.6235962,0.0006594475,0.3565497,0.0008565051,0.0002373905,0.0001522289,0.004451161,0.00128585,0.01221149],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.009967233,"threshold_uncertainty_score":0.03334373,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.002906539219855812,"score_gpt":0.1958968182108761,"score_spread":0.1929902789910203,"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."}}