{"id":"W4417469644","doi":"10.48550/arxiv.2512.14427","title":"Effect of Document Packing on the Latent Multi-Hop Reasoning Capabilities of Large Language Models","year":2025,"lang":"","type":"preprint","venue":"ArXiv.org","topic":"Topic Modeling","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Alliance de recherche numérique du Canada; Natural Sciences and Engineering Research Council of Canada; Canadian Institute for Advanced Research","keywords":"Key (lock); Process (computing); Language model; Language understanding; Data modeling; Training set","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":[0.004452046,0.0007147326,0.001205793,0.0002486122,0.0002602304,0.000106747,0.002466054,0.0003610245,0.00006607504],"category_scores_gemma":[0.0007193508,0.0005257968,0.0006380519,0.0003533123,0.0001949104,0.0002862115,0.003570725,0.001166994,0.00001526355],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002697591,"about_ca_system_score_gemma":0.0002811781,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001519094,"about_ca_topic_score_gemma":0.00005694302,"domain_scores_codex":[0.9942124,0.00133659,0.001380085,0.001393604,0.0008742972,0.0008029541],"domain_scores_gemma":[0.9934323,0.001941933,0.0009460383,0.003275736,0.0002885736,0.0001154213],"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":[0.0003015743,0.0007776978,0.1502301,0.007798685,0.001163179,0.00005948679,0.1323232,0.5736272,0.003833163,0.09799274,0.00004094668,0.03185202],"study_design_scores_gemma":[0.001232453,0.0004006786,0.003585797,0.006125103,0.0001751028,0.000002769374,0.00102958,0.9166749,0.06926384,0.000899789,0.00002744542,0.0005825392],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8406122,0.001165282,0.1546071,0.000385813,0.001116745,0.00126703,0.00003813964,0.00007605137,0.0007315667],"genre_scores_gemma":[0.9945418,0.0001825279,0.003972149,0.0001585947,0.0001087673,0.0001362834,0.00000686566,0.00002935941,0.0008636552],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3430477,"threshold_uncertainty_score":0.9997194,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.03533297112411127,"score_gpt":0.3004903172743915,"score_spread":0.2651573461502802,"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."}}