{"id":"W7087771355","doi":"10.1016/j.ipm.2025.104431","title":"What drives attention sinks? A study of massive activations and rotational positional encoding in large vision–language models","year":2025,"lang":"en","type":"article","venue":"Information Processing & Management","topic":"Neuroinflammation and Neurodegeneration Mechanisms","field":"Neuroscience","cited_by":2,"is_retracted":false,"has_abstract":false,"ca_institutions":"","funders":"National Natural Science Foundation of China; Ontario Ministry of Natural Resources and Forestry","keywords":"Softmax function; Encoding (memory); Position (finance); Object (grammar); Encoder; Embedding; Continuation; Visual attention; Residual; Dual (grammatical number)","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":[],"consensus_categories":[],"category_scores_codex":[0.000218272,0.0001224628,0.0001179321,0.0007550486,0.000282735,0.000417287,0.0001172952,0.0000320071,0.00001927503],"category_scores_gemma":[0.00005390267,0.0001270004,0.00002326491,0.0006288461,0.00002314063,0.005930149,0.0001077845,0.00009853582,0.000005280685],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004679719,"about_ca_system_score_gemma":0.00003245955,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000001578896,"about_ca_topic_score_gemma":0.000005972624,"domain_scores_codex":[0.9985873,0.00008970939,0.0005488161,0.0002177004,0.0004171785,0.0001393641],"domain_scores_gemma":[0.9994102,0.00006145168,0.0002792593,0.000122588,0.00009894473,0.00002761656],"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":[0.00027077,0.00189553,0.0005481881,0.001074679,0.00004034527,0.00002431913,0.04113705,0.08436169,0.23164,0.4963433,0.0002226918,0.1424415],"study_design_scores_gemma":[0.007777498,0.0002787953,0.04617807,0.001199019,0.00007662419,0.0000101186,0.06365994,0.7626054,0.09617826,0.02114426,0.000287012,0.0006050604],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.942851,0.00001413088,0.05259563,0.0007754177,0.0001905468,0.0009082824,0.00000802499,0.00006959528,0.002587365],"genre_scores_gemma":[0.9981592,0.00003690284,0.0006029974,0.0007789948,0.00000721482,0.0001129509,0.00002909388,0.000005122454,0.0002674833],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.6782436,"threshold_uncertainty_score":0.5178926,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01548646666831226,"score_gpt":0.2941110356699409,"score_spread":0.2786245690016286,"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."}}