{"id":"W7092316518","doi":"","title":"TokDrift: When LLM Speaks in Subwords but Code Speaks in Grammar","year":2025,"lang":"","type":"article","venue":"ArXiv.org","topic":"Diffusion and Search Dynamics","field":"Biochemistry, Genetics and Molecular Biology","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Waterloo","funders":"Natural Sciences and Engineering Research Council of Canada; Alliance de recherche numérique du Canada; University of Waterloo","keywords":"Lexical analysis; Code (set theory); Identifier; Grammar; Source code; Code review; Natural language; Encoding (memory); Unicode","routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0008192746,0.0005659394,0.0006338638,0.0004251892,0.000128207,0.000107712,0.0009237276,0.0007820483,0.0002848982],"category_scores_gemma":[0.0004597174,0.0006175401,0.0002723589,0.0006615297,0.0003510605,0.00001884372,0.0009109531,0.0008722033,0.0001407761],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001983432,"about_ca_system_score_gemma":0.0006633888,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000975145,"about_ca_topic_score_gemma":0.01388034,"domain_scores_codex":[0.995887,0.0003504694,0.000998992,0.001232251,0.0003280847,0.001203229],"domain_scores_gemma":[0.9981911,0.00006557283,0.0001560831,0.001154972,0.000143059,0.0002892755],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0005621436,0.001384205,0.863187,0.0001866948,0.0001132783,0.0002515983,0.0004691586,0.0001000995,0.1135769,0.001215474,0.001805618,0.01714781],"study_design_scores_gemma":[0.008044954,0.0008322639,0.6518667,0.0006959075,0.00009073634,0.00002039682,0.001642505,0.00564351,0.04322674,0.001746131,0.2844381,0.001752097],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9730933,0.002092795,0.0006198311,0.00294788,0.0009044606,0.0006185235,0.00007503757,0.000021751,0.01962647],"genre_scores_gemma":[0.9726622,0.00379883,0.0002617761,0.002010275,0.0002041421,0.0000362213,0.0002443688,0.00005683501,0.02072537],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2826325,"threshold_uncertainty_score":0.9996276,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01840998351677734,"score_gpt":0.2837767106709696,"score_spread":0.2653667271541922,"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."}}