{"id":"W4416445260","doi":"10.48550/arxiv.2505.03563","title":"Say It Another Way: Auditing LLMs with a User-Grounded Automated Paraphrasing Framework","year":2025,"lang":"en","type":"preprint","venue":"ArXiv.org","topic":"Topic Modeling","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Fonds de recherche du Québec – Nature et technologies; Compute Canada; Canadian Institute for Advanced Research","keywords":"Audit; Quality (philosophy); Natural language; Strengths and weaknesses; Natural (archaeology); Language model; Semantics (computer science)","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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01339263,0.00118855,0.0008916912,0.002141158,0.0008697705,0.004978543,0.002670912,0.002185442,0.003099679],"category_scores_gemma":[0.08773106,0.0006968686,0.001005762,0.001700994,0.001972693,0.007490433,0.005791174,0.003671239,0.002884139],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00110917,"about_ca_system_score_gemma":0.003228646,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004087579,"about_ca_topic_score_gemma":0.006987161,"domain_scores_codex":[0.9759754,0.01676483,0.00126479,0.002384874,0.003157523,0.0004526665],"domain_scores_gemma":[0.9462317,0.02339994,0.00299733,0.02194832,0.004745983,0.0006767896],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","study_design_scores_codex":[0.001405264,0.0008153904,0.02164625,0.001935467,0.0003902188,0.0008890699,0.009292052,0.08418111,0.04979874,0.05850671,0.04969416,0.7214456],"study_design_scores_gemma":[0.0001782896,0.0005853031,0.004079432,0.0003396937,0.0001069659,0.0006421484,0.00193606,0.7815722,0.04441573,0.1186794,0.0472629,0.0002019477],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.04068757,0.0007786467,0.9202836,0.00215338,0.0001300754,0.0005194217,0.002864426,0.02931616,0.003266619],"genre_scores_gemma":[0.4327767,0.0002694856,0.5576282,0.0008951072,0.00007344828,0.0004721661,0.004361213,0.001810315,0.001713398],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.01339263,"threshold_uncertainty_score":0.07082784,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.04506760192103412,"score_gpt":0.2914444646952941,"score_spread":0.2463768627742599,"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."}}