{"id":"W4404782641","doi":"10.18653/v1/2024.emnlp-main.142","title":"Backward Lens: Projecting Language Model Gradients into the Vocabulary Space","year":2024,"lang":"en","type":"article","venue":"","topic":"Natural Language Processing Techniques","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Azrieli Foundation; Tel Aviv University; Open Philanthropy Project","keywords":"Computer science; Space (punctuation); Vocabulary; Lens (geology); Artificial intelligence; Natural language processing; Linguistics; Optics; Physics; Philosophy","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":[0.001338079,0.001356758,0.0006090567,0.0008282693,0.0005028677,0.002124611,0.001532663,0.0008744202,0.01195501],"category_scores_gemma":[0.007424475,0.000719122,0.001084283,0.000712015,0.001378391,0.004396041,0.002948912,0.003108451,0.004712448],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000824195,"about_ca_system_score_gemma":0.00196355,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.006058576,"about_ca_topic_score_gemma":0.009224589,"domain_scores_codex":[0.9994063,0.0001577243,0.00003217156,0.0001426904,0.000199529,0.00006155272],"domain_scores_gemma":[0.9989817,0.0003750561,0.00009177158,0.0002435424,0.0002241011,0.00008372936],"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":[0.000330034,0.0001643703,0.001937436,0.0003436019,0.0001404172,0.0002977778,0.0006187176,0.1350794,0.02053222,0.340041,0.02506898,0.4754461],"study_design_scores_gemma":[0.00003870963,0.0000748908,0.0003075616,0.0000456888,0.00001963706,0.000117748,0.00007122588,0.7758983,0.01001901,0.206128,0.00724708,0.00003198713],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.003798175,0.00008209582,0.9927953,0.0002547408,0.00004702754,0.00003346131,0.0001908363,0.001445774,0.001352635],"genre_scores_gemma":[0.2039491,0.0005687354,0.7737005,0.0005679014,0.0001627048,0.0003130034,0.00133324,0.00171494,0.01768991],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.01195501,"threshold_uncertainty_score":0.03999352,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01598860997620192,"score_gpt":0.2864821520160792,"score_spread":0.2704935420398773,"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."}}