{"id":"W4402671301","doi":"10.18653/v1/2024.acl-long.785","title":"CausalGym: Benchmarking causal interpretability methods on linguistic tasks","year":2024,"lang":"en","type":"article","venue":"","topic":"Topic Modeling","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"University of Cambridge; Institute for Catastrophic Loss Reduction","keywords":"Interpretability; Benchmarking; Computer science; Natural language processing; Artificial intelligence; Linguistics; 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":[0.02022729,0.002502538,0.001025212,0.00326846,0.001186703,0.00260727,0.003735177,0.003771943,0.007169024],"category_scores_gemma":[0.09449512,0.0008075771,0.001757536,0.001986039,0.001990327,0.005752234,0.004261681,0.005398578,0.002017822],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002477413,"about_ca_system_score_gemma":0.003141847,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.008425784,"about_ca_topic_score_gemma":0.01101147,"domain_scores_codex":[0.985316,0.01024253,0.0008378531,0.001955423,0.001202243,0.0004460662],"domain_scores_gemma":[0.8924118,0.09326337,0.001400049,0.00911885,0.002820624,0.0009853515],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"bench_or_experimental","study_design_scores_codex":[0.00406142,0.003019225,0.03392788,0.00399916,0.001285506,0.0004642857,0.002255319,0.4736582,0.007763286,0.02774154,0.04239005,0.3994341],"study_design_scores_gemma":[0.0005142861,0.0004476337,0.003036913,0.00009942176,0.00007259651,0.00007133448,0.0004017918,0.9666876,0.00463781,0.01942715,0.004544066,0.00005946643],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"methods","genre_scores_codex":[0.5905686,0.005395482,0.3090153,0.003753894,0.001075228,0.00196531,0.009256233,0.05858875,0.02038118],"genre_scores_gemma":[0.6961039,0.0006696976,0.2826895,0.0006315133,0.0001643463,0.001556854,0.0123884,0.003576885,0.002218893],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.02022729,"threshold_uncertainty_score":0.1069734,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.03729222069651673,"score_gpt":0.3692485109035606,"score_spread":0.3319562902070439,"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."}}