{"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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001600863,0.0009969661,0.0005676194,0.0006436203,0.0006556901,0.002399631,0.001637098,0.001801949,0.006304546],"category_scores_gemma":[0.01734729,0.0007905957,0.0008966928,0.0005023252,0.001678879,0.006335558,0.002881033,0.002693389,0.003471515],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001030754,"about_ca_system_score_gemma":0.001469521,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.005356328,"about_ca_topic_score_gemma":0.007438762,"domain_scores_codex":[0.9986817,0.0003104883,0.0001037053,0.0004548421,0.0003035583,0.0001456999],"domain_scores_gemma":[0.9966028,0.001397844,0.0003005581,0.001244144,0.0003394024,0.0001151842],"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":[0.002292527,0.0004027414,0.06760886,0.001700962,0.0005490478,0.002727106,0.006769325,0.181826,0.1030505,0.1811321,0.09522812,0.3567127],"study_design_scores_gemma":[0.00004970437,0.0001192529,0.00257557,0.00008054906,0.00005424368,0.0004655988,0.0005642893,0.825321,0.05892982,0.09151424,0.02021093,0.0001148613],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.2303568,0.0003837483,0.6656315,0.00252498,0.0007670461,0.0002201933,0.004213252,0.08383787,0.0120646],"genre_scores_gemma":[0.7893958,0.0001889684,0.1815408,0.0007379603,0.00004409696,0.0002495127,0.004972775,0.01525874,0.007611302],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.006304546,"threshold_uncertainty_score":0.02109075,"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."}}