{"id":"W4414588026","doi":"10.48550/arxiv.2505.20063","title":"SAEs Are Good for Steering -- If You Select the Right Features","year":2025,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Traffic Prediction and Management Techniques","field":"Engineering","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Azrieli Foundation; European Commission; Open Philanthropy Project","keywords":"Feature (linguistics); Pattern recognition (psychology); Decomposition; Work (physics); Key (lock); Feature vector","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":[0.003085023,0.001821083,0.0009711625,0.001421352,0.0005428264,0.001840755,0.0009733827,0.001432158,0.006118062],"category_scores_gemma":[0.01559243,0.0006505043,0.00112963,0.001001914,0.0009307381,0.003938294,0.00203296,0.002081477,0.005609502],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003441822,"about_ca_system_score_gemma":0.0007867095,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002048023,"about_ca_topic_score_gemma":0.004811826,"domain_scores_codex":[0.9978762,0.0007163654,0.0001351422,0.0005018464,0.0006255086,0.0001450428],"domain_scores_gemma":[0.9942006,0.002394419,0.0004532425,0.001436422,0.001308761,0.0002066113],"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":[0.0005120424,0.0003962037,0.01480566,0.0002249353,0.0003087898,0.0001068792,0.000200913,0.1453566,0.01527786,0.01030684,0.02077733,0.7917261],"study_design_scores_gemma":[0.00004768042,0.0002132893,0.004987773,0.0000508138,0.00006539747,0.0001516469,0.0001606175,0.942527,0.01270607,0.03147103,0.007567659,0.00005111232],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.03490504,0.0003618609,0.9530843,0.0006776479,0.0001401228,0.00009643979,0.0004723297,0.005736615,0.004525641],"genre_scores_gemma":[0.6396277,0.0005275028,0.3445792,0.0007610322,0.0002322042,0.0001999848,0.003132014,0.00102824,0.009912138],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.006118062,"threshold_uncertainty_score":0.02046692,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.02234653463575751,"score_gpt":0.1675682590369773,"score_spread":0.1452217244012198,"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."}}