{"id":"W4389519586","doi":"10.18653/v1/2023.emnlp-main.971","title":"MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions","year":2023,"lang":"en","type":"article","venue":"","topic":"Topic Modeling","field":"Computer Science","cited_by":38,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Canadian Institute for Advanced Research; National Science Foundation","keywords":"Computer science; Recall; Language model; Question answering; Margin (machine learning); Natural language processing; Hop (telecommunications); Artificial intelligence; Machine learning; Cognitive psychology; Psychology","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":[],"consensus_categories":[],"category_scores_codex":[0.0005012488,0.0001009946,0.0001191176,0.0002415443,0.00009947876,0.000174399,0.000463915,0.00005919466,0.000009494689],"category_scores_gemma":[0.00007565154,0.00009888168,0.00003565724,0.0007402035,0.0000131984,0.0009539864,0.0003313285,0.0001703082,0.0001621132],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005976382,"about_ca_system_score_gemma":0.00006186896,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003150362,"about_ca_topic_score_gemma":0.0004494981,"domain_scores_codex":[0.9988861,0.00006900576,0.0002330228,0.0003471965,0.000139661,0.0003250034],"domain_scores_gemma":[0.9993455,0.000115087,0.00003686951,0.0003930324,0.00004543473,0.00006406107],"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":[8.052852e-7,0.0002585848,0.002007909,0.00009066238,0.00001607915,0.0002204297,0.03080377,0.1962539,0.01063779,0.1505161,0.001052445,0.6081415],"study_design_scores_gemma":[0.000163876,0.000003040181,0.0008824672,0.00004741525,0.000001008665,0.000003893879,0.0004042137,0.9957194,0.0002336174,0.002367361,0.0000602042,0.0001135578],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.04490171,0.00009175811,0.9458651,0.0004443945,0.0004606673,0.0000831005,4.259258e-7,0.0007432469,0.00740962],"genre_scores_gemma":[0.8293542,0.000002908881,0.1691792,0.00007342832,0.000137326,0.0000146912,0.000002042284,0.000009408902,0.001226755],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.7994654,"threshold_uncertainty_score":0.4032278,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.06529112178130436,"score_gpt":0.3471593056174225,"score_spread":0.2818681838361181,"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."}}