{"id":"W4415968362","doi":"10.1109/lca.2025.3630094","title":"Disaggregated Speculative Decoding for Carbon-Efficient LLM Serving","year":2025,"lang":"","type":"article","venue":"IEEE Computer Architecture Letters","topic":"Advanced Neural Network Applications","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"ca_institutions":"University of Waterloo","funders":"","keywords":"Decoding methods; Security token; Cache; Workload; Overhead (engineering); Latency (audio); Sequential decoding; Decoupling (probability)","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":[0.0003256435,0.0009737437,0.0008508142,0.0006591117,0.0009130759,0.0005405751,0.002729105,0.0002147425,0.000004352948],"category_scores_gemma":[0.00003825117,0.001003805,0.0005297991,0.002508561,0.0003787023,0.0001707819,0.001077318,0.0009875259,0.00001170852],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000425249,"about_ca_system_score_gemma":0.0001496006,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002180874,"about_ca_topic_score_gemma":0.00002039828,"domain_scores_codex":[0.9942797,0.000292977,0.001056442,0.002277028,0.0005200438,0.001573749],"domain_scores_gemma":[0.9955929,0.001462626,0.0005019396,0.00191757,0.0002204009,0.0003045757],"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":[0.00005737631,0.0001577683,0.00007069314,0.0001743907,0.0002168099,0.00002922348,0.002036323,0.8132828,0.008588914,0.002791042,0.002269236,0.1703254],"study_design_scores_gemma":[0.001216183,0.0001211632,0.0003846023,0.0007193409,0.0001047835,0.00003353719,0.000009434117,0.9831555,0.006383066,0.003040999,0.003915871,0.0009155268],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.08774961,0.0003428769,0.8686327,0.03499763,0.005130251,0.002500387,0.00002728332,0.0004390187,0.0001802187],"genre_scores_gemma":[0.6152608,0.00001061357,0.3662602,0.01629758,0.001669883,0.0002824985,0.00001724598,0.00008310779,0.0001180618],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.5275112,"threshold_uncertainty_score":0.9992412,"prediction_status":"machine_predicted_unvalidated"},"machine_scores":{"provisional":true,"baseline":true,"maturity_gate_passed":false,"score_opus":0.01334071655782341,"score_gpt":0.2646251550038536,"score_spread":0.2512844384460302,"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."}}