{"id":"W7123353892","doi":"10.1109/cdc57313.2025.11312306","title":"Preventing Model Collapse when Training LLMs with Synthetic Data","year":2025,"lang":"","type":"article","venue":"","topic":"Generative Adversarial Networks and Image Synthesis","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"Queen's University","funders":"","keywords":"Synthetic data; Training (meteorology); Generative model; Simple (philosophy); Training set; Quality (philosophy)","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":[0.005168393,0.0009923484,0.0009054763,0.0006024075,0.0004744957,0.0008757397,0.00125853,0.001539255,0.001653395],"category_scores_gemma":[0.02290422,0.0006616659,0.0006557226,0.0002785388,0.002028304,0.001494969,0.00318946,0.002027327,0.0004352742],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000811386,"about_ca_system_score_gemma":0.0007020838,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001842959,"about_ca_topic_score_gemma":0.002021888,"domain_scores_codex":[0.998747,0.0005641878,0.00008165905,0.0002383403,0.0002428887,0.0001259506],"domain_scores_gemma":[0.9924636,0.005632561,0.0004997695,0.000740284,0.0004595112,0.0002043147],"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":[0.0001406318,0.00006288052,0.001455283,0.00009866161,0.00005431579,0.0001680842,0.0001852131,0.9525595,0.008731582,0.01173241,0.0007426571,0.02406874],"study_design_scores_gemma":[0.000009303433,0.00006159574,0.0001568141,0.000016947,0.000007268055,0.00004884497,0.00001943775,0.9887417,0.0045526,0.005911175,0.0004665209,0.000007810509],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1015785,0.000314621,0.8940247,0.0005469727,0.00006348314,0.0001039435,0.00007989808,0.000749314,0.002538628],"genre_scores_gemma":[0.905804,0.0001338058,0.09109955,0.0003738435,0.00002868514,0.0001664404,0.0002433105,0.0001741109,0.001976228],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.005168393,"threshold_uncertainty_score":0.02733338,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.06272020910394469,"score_gpt":0.2721604772171564,"score_spread":0.2094402681132117,"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."}}