{"id":"W4403346052","doi":"10.48550/arxiv.2410.01444","title":"Geometric Signatures of Compositionality Across a Language Model's Lifetime","year":2024,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Syntax, Semantics, Linguistic Variation","field":"Arts and Humanities","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Fonds de recherche du Québec – Nature et technologies; European Commission; Universitat Pompeu Fabra; Wellcome Trust","keywords":"Principle of compositionality; Computer science; Natural language processing; Linguistics; Artificial intelligence; 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"],"consensus_categories":[],"category_scores_codex":[0.0002782181,0.0002683212,0.0003865639,0.0003545303,0.0001627209,0.0001448889,0.0004519838,0.0002274018,0.0004379702],"category_scores_gemma":[0.0001242882,0.0002979061,0.0002763682,0.0001734687,0.0002601155,0.00006351995,0.0008705051,0.0005875356,0.0001128108],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001510069,"about_ca_system_score_gemma":0.0001237372,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002283515,"about_ca_topic_score_gemma":0.0006771309,"domain_scores_codex":[0.9986522,0.00007077031,0.0002831848,0.0005942082,0.0001496714,0.0002499959],"domain_scores_gemma":[0.9986249,0.0001739633,0.0002814387,0.0004976779,0.0003472161,0.00007486449],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0000467393,0.0001312472,0.0001056753,0.0005015681,0.0002766078,0.00009957216,0.01757449,0.1511484,0.00004362584,0.8298065,0.0002138954,0.0000516224],"study_design_scores_gemma":[0.0003276607,0.00003604964,0.0002636815,0.0002736633,0.0003957756,0.000001004476,0.003037824,0.4011661,0.0001682977,0.5937095,0.0001577218,0.0004626998],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9713178,0.0003387092,0.009192304,0.0005534081,0.001477747,0.0003370295,0.001818401,0.0002370765,0.01472752],"genre_scores_gemma":[0.9944468,0.00002285131,0.0001682189,0.0000566079,0.0005966856,9.273188e-7,0.0002322181,0.00003325631,0.004442493],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2500177,"threshold_uncertainty_score":0.9999473,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.06210371445777068,"score_gpt":0.2193220437224882,"score_spread":0.1572183292647175,"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."}}