{"id":"W4390722476","doi":"10.48550/arxiv.2401.03529","title":"Quantifying stability of non-power-seeking in artificial agents","year":2024,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Reinforcement Learning in Robotics","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Toronto","funders":"","keywords":"Shutdown; Computer science; Metric (unit); Stability (learning theory); Core (optical fiber); Shut down; Power (physics); Mathematical optimization; Economics; Engineering; Mathematics; Operations management; Machine learning","routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false,"invisible_to_affiliation_only":false},"retraction":null,"screen":null,"direct_labels":[],"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0006720972,0.0002862285,0.0004124135,0.0004914309,0.00006436113,0.000136493,0.001716618,0.0002651228,0.00004162937],"category_scores_gemma":[0.00008533266,0.0003506594,0.0002194581,0.00096517,0.0001098733,0.0002237143,0.004180653,0.001008164,0.00008663558],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003287988,"about_ca_system_score_gemma":0.0002721719,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002088199,"about_ca_topic_score_gemma":0.00006791888,"domain_scores_codex":[0.9977836,0.0001413691,0.0004595361,0.001048824,0.0001815475,0.0003851407],"domain_scores_gemma":[0.9980652,0.0001390839,0.0003279748,0.001262548,0.000115981,0.00008926551],"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":[0.00001360562,0.00005043391,0.0209698,0.0003161082,0.00004038371,0.0002005896,0.001012528,0.9428027,0.0003308147,0.03408356,0.0000216762,0.0001577485],"study_design_scores_gemma":[0.0001376615,0.0000429207,0.00479552,0.0003464942,0.00002693536,7.093889e-7,0.0001415381,0.9870346,0.0006443466,0.006496675,0.00002964229,0.0003029375],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.4696531,0.00001382791,0.527483,0.00002074328,0.0008060845,0.0001941325,0.000002591487,0.00008669456,0.001739847],"genre_scores_gemma":[0.9980719,0.00002109962,0.001713158,0.00001662211,0.00002545525,5.546926e-7,0.000005020796,0.00001840005,0.0001277606],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.5284188,"threshold_uncertainty_score":0.9998946,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.1637298112787587,"score_gpt":0.2461775063420691,"score_spread":0.08244769506331043,"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."}}