{"id":"W4415311596","doi":"10.48550/arxiv.2506.14126","title":"From Memorization to Parameter Interference: How Overtraining Experts Harms Model Merging","year":2025,"lang":"en","type":"preprint","venue":"ArXiv.org","topic":"Statistics Education and Methodologies","field":"Mathematics","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Fonds de recherche du Québec – Nature et technologies; Compute Canada; Natural Sciences and Engineering Research Council of Canada; Canadian Institute for Advanced Research; National Science Foundation","keywords":"Pipeline (software); Leverage (statistics); Reuse; Set (abstract data type); Downstream (manufacturing); Boosting (machine learning); Robustness (evolution)","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":[0.006426379,0.001788834,0.001198452,0.0007976943,0.001298699,0.002359676,0.002787877,0.002544668,0.003157907],"category_scores_gemma":[0.04782714,0.001017273,0.0008185569,0.0007096011,0.002172499,0.00509915,0.005213768,0.004579258,0.001261463],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001113783,"about_ca_system_score_gemma":0.001587953,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004026842,"about_ca_topic_score_gemma":0.005556914,"domain_scores_codex":[0.9960743,0.001326212,0.0002359312,0.001181808,0.0007234706,0.0004582282],"domain_scores_gemma":[0.9774165,0.01330466,0.001484421,0.005729516,0.001156054,0.0009087522],"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":[0.00205239,0.0005770956,0.02487988,0.0005845448,0.0005386119,0.001125826,0.003261083,0.4029074,0.048269,0.02957785,0.01496763,0.4712587],"study_design_scores_gemma":[0.0001702474,0.0007251829,0.005815428,0.0001676272,0.0002361621,0.0007217776,0.0006912735,0.8433408,0.04550895,0.09117042,0.01131196,0.0001400698],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.4590726,0.002850501,0.5149413,0.004352062,0.0003880023,0.0001553756,0.0003300236,0.006746622,0.01116344],"genre_scores_gemma":[0.9093537,0.0003387701,0.0842282,0.001341254,0.0001096143,0.00008883578,0.0003826968,0.0009429734,0.003213987],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.006426379,"threshold_uncertainty_score":0.03398633,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.4280267706046366,"score_gpt":0.4570584216027669,"score_spread":0.02903165099813038,"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."}}