{"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":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002587082,0.001551403,0.0007797799,0.001162855,0.0006691524,0.001831343,0.001790307,0.0012093,0.002836195],"category_scores_gemma":[0.01036214,0.0006541677,0.00126362,0.0006366255,0.0008271758,0.005492167,0.0025301,0.002582933,0.0007981966],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001558947,"about_ca_system_score_gemma":0.001538915,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01529372,"about_ca_topic_score_gemma":0.02341093,"domain_scores_codex":[0.9991398,0.0003102458,0.00004489727,0.0002514554,0.0001393152,0.0001142803],"domain_scores_gemma":[0.9952663,0.00362824,0.0002560748,0.0004586903,0.0002486703,0.0001420675],"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":[0.0006634669,0.0004582453,0.009077235,0.0002610784,0.0002111189,0.0005049272,0.001399565,0.3462085,0.01700996,0.01744817,0.005706571,0.6010511],"study_design_scores_gemma":[0.00001455255,0.00003666673,0.0002998224,0.000009453877,0.00002311646,0.00002130946,0.0000454022,0.9810161,0.002978316,0.01517839,0.0003679643,0.000008961205],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.1048,0.000574092,0.8839436,0.0008420924,0.00007258082,0.0001048046,0.000276252,0.007142198,0.002244299],"genre_scores_gemma":[0.8368402,0.0002086707,0.159399,0.0003962501,0.00005634482,0.00008621962,0.0005671814,0.0002567869,0.002189364],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.01529372,"threshold_uncertainty_score":0.0304094,"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."}}