{"id":"W7117277200","doi":"","title":"Saddle-to-Saddle Dynamics Explains A Simplicity Bias Across Neural Network Architectures","year":2025,"lang":"","type":"article","venue":"arXiv (Cornell University)","topic":"Stochastic Gradient Optimization Techniques","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Canadian Institute for Advanced Research","keywords":"Simplicity; Initialization; Gradient descent; Invariant (physics); Convolutional neural network; Artificial neural network; Class (philosophy)","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":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.0006275201,0.0007827952,0.0007277289,0.0006292,0.001315997,0.0005149873,0.003421471,0.0003935816,0.00005154526],"category_scores_gemma":[0.0004766362,0.000958203,0.0004244297,0.006825443,0.0004982836,0.0004590639,0.002924599,0.0007884431,0.0000690752],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0007859879,"about_ca_system_score_gemma":0.0002987603,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002918893,"about_ca_topic_score_gemma":0.0007029856,"domain_scores_codex":[0.9948155,0.0004037876,0.000643423,0.002171034,0.0002482687,0.001717989],"domain_scores_gemma":[0.9955586,0.000907135,0.0003930688,0.002111159,0.0004354354,0.0005946476],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0001278721,0.0001606401,0.00828932,0.00004119534,0.00009989044,0.0001186985,0.0007696786,0.7442746,0.000005864622,0.2400509,0.000945494,0.005115847],"study_design_scores_gemma":[0.0008030797,0.0003195274,0.005346258,0.0002243191,0.00008818843,0.00001356025,0.0002298923,0.9645636,0.00009559235,0.02712317,0.0003922601,0.0008005205],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.2759205,0.00006042598,0.7192112,0.0004919865,0.001159094,0.0008710705,0.00007140469,0.0006221578,0.001592199],"genre_scores_gemma":[0.9860343,0.00004121742,0.008967153,0.001633633,0.000138118,0.000005936722,0.0000173658,0.00004377524,0.003118504],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.7102441,"threshold_uncertainty_score":0.9999841,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.06538742909950591,"score_gpt":0.2209794611678297,"score_spread":0.1555920320683238,"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."}}