{"id":"W4416036439","doi":"10.18653/v1/2025.emnlp-main.567","title":"Back Attention: Understanding and Enhancing Multi-Hop Reasoning in Large Language Models","year":2025,"lang":"","type":"article","venue":"","topic":"Multimodal Machine Learning Applications","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"ca_institutions":"","funders":"Azrieli Foundation; European Commission; University of Manchester; Open Philanthropy Project","keywords":"Natural language; Language model; Feature (linguistics); Semantics (computer science); Language understanding; Automated reasoning","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":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001042602,0.0003034255,0.0003537677,0.0004503998,0.0004654068,0.0005026939,0.0005616946,0.000172631,0.0001693694],"category_scores_gemma":[0.000114075,0.0003345431,0.00007880003,0.001342288,0.00007377075,0.0007660343,0.0008799754,0.000606307,0.00009276115],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000394385,"about_ca_system_score_gemma":0.0001313278,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001750241,"about_ca_topic_score_gemma":0.00199365,"domain_scores_codex":[0.9972991,0.0002267133,0.000590669,0.001003545,0.0002272535,0.0006526909],"domain_scores_gemma":[0.9986691,0.0002653025,0.0001512244,0.0007232099,0.00005697341,0.0001342264],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00001426126,0.0003440612,0.04108543,0.0002910566,0.00006031797,0.00001606818,0.01135931,0.01183778,0.003821689,0.9219338,0.00007114206,0.009165039],"study_design_scores_gemma":[0.001414995,0.00001638042,0.01693653,0.0006330428,0.00001533558,0.000005228141,0.003125234,0.9754012,0.0001237352,0.00199794,0.00002995504,0.0003003834],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.07011579,0.0008682878,0.9079902,0.002209206,0.000187918,0.0004729139,0.000003681357,0.0001034531,0.0180485],"genre_scores_gemma":[0.8807611,0.00006538699,0.1140788,0.0003519273,0.0000267016,0.00002539822,0.000004500551,0.00001666106,0.004669493],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9635634,"threshold_uncertainty_score":0.9999107,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.03062317003061937,"score_gpt":0.3144617111784772,"score_spread":0.2838385411478578,"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."}}