[{"data":1,"prerenderedAt":8485},["ShallowReactive",2],{"post-\u002Fposts\u002Fqwen25-sft-dpo-alignment-tax-ablation":3,"archive-public-posts":7987},{"id":4,"title":5,"body":6,"categories":7969,"cover":7971,"coverAlt":7972,"date":7973,"description":7974,"draft":7975,"extension":7976,"haloPublished":6380,"haloSync":6380,"hidden":7975,"legacy":7975,"meta":7977,"navigation":6380,"path":7978,"published":7975,"readingTime":7979,"seo":7980,"series":7979,"seriesOrder":7979,"stem":7981,"tags":7982,"__hash__":7986},"posts\u002Fposts\u002Fqwen25-sft-dpo-alignment-tax-ablation.md","大模型后训练偏好对齐实录：Qwen2.5 SFT 多维消融、动态参考 DPO 与对齐税（Alignment Tax）量化分析",{"type":7,"value":8,"toc":7959},"minimark",[9,222,225,706,725,728,733,760,771,957,2448,2661,2665,4914,5014,5016,5020,5102,6247,6340,6998,7000,7004,7187,7551,7554,7831,7833,7837,7955],[10,11,12,13,17,18,21,22,221],"p",{},"在大语言模型后训练（Post-Training）管线中，",[14,15,16],"strong",{},"监督微调（SFT）","负责教会模型任务格式与领域推理模式，而",[14,19,20],{},"直接偏好优化（DPO, Direct Preference Optimization）","则利用成对偏好数据 ",[23,24,27,80],"span",{"className":25},[26],"katex",[23,28,31],{"className":29},[30],"katex-mathml",[32,33,35],"math",{"xmlns":34},"http:\u002F\u002Fwww.w3.org\u002F1998\u002FMath\u002FMathML",[36,37,38,75],"semantics",{},[39,40,41,46,50,54,63,65,72],"mrow",{},[42,43,45],"mo",{"stretchy":44},"false","(",[47,48,49],"mi",{},"x",[42,51,53],{"separator":52},"true",",",[55,56,57,60],"msub",{},[47,58,59],{},"y",[47,61,62],{},"w",[42,64,53],{"separator":52},[55,66,67,69],{},[47,68,59],{},[47,70,71],{},"l",[42,73,74],{"stretchy":44},")",[76,77,79],"annotation",{"encoding":78},"application\u002Fx-tex","(x, y_w, y_l)",[23,81,84],{"className":82,"ariaHidden":52},[83],"katex-html",[23,85,88,93,97,102,106,111,169,172,175,217],{"className":86},[87],"base",[23,89],{"className":90,"style":92},[91],"strut","height:1em;vertical-align:-0.25em;",[23,94,45],{"className":95},[96],"mopen",[23,98,49],{"className":99},[100,101],"mord","mathnormal",[23,103,53],{"className":104},[105],"mpunct",[23,107],{"className":108,"style":110},[109],"mspace","margin-right:0.1667em;",[23,112,114,118],{"className":113},[100],[23,115,59],{"className":116,"style":117},[100,101],"margin-right:0.0359em;",[23,119,122],{"className":120},[121],"msupsub",[23,123,127,160],{"className":124},[125,126],"vlist-t","vlist-t2",[23,128,131,155],{"className":129},[130],"vlist-r",[23,132,136],{"className":133,"style":135},[134],"vlist","height:0.1514em;",[23,137,139,144],{"style":138},"top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;",[23,140],{"className":141,"style":143},[142],"pstrut","height:2.7em;",[23,145,151],{"className":146},[147,148,149,150],"sizing","reset-size6","size3","mtight",[23,152,62],{"className":153,"style":154},[100,101,150],"margin-right:0.0269em;",[23,156,159],{"className":157},[158],"vlist-s","​",[23,161,163],{"className":162},[130],[23,164,167],{"className":165,"style":166},[134],"height:0.15em;",[23,168],{},[23,170,53],{"className":171},[105],[23,173],{"className":174,"style":110},[109],[23,176,178,181],{"className":177},[100],[23,179,59],{"className":180,"style":117},[100,101],[23,182,184],{"className":183},[121],[23,185,187,209],{"className":186},[125,126],[23,188,190,206],{"className":189},[130],[23,191,194],{"className":192,"style":193},[134],"height:0.3361em;",[23,195,196,199],{"style":138},[23,197],{"className":198,"style":143},[142],[23,200,202],{"className":201},[147,148,149,150],[23,203,71],{"className":204,"style":205},[100,101,150],"margin-right:0.0197em;",[23,207,159],{"className":208},[158],[23,210,212],{"className":211},[130],[23,213,215],{"className":214,"style":166},[134],[23,216],{},[23,218,74],{"className":219},[220],"mclose"," 将模型输出分布向人类或高分策略偏好对齐。",[10,223,224],{},"然而，在实际工程与面试讨论中，有三个关键问题往往缺乏在同一受控实验下的定量回答：",[226,227,228,330,696],"ol",{},[229,230,231],"li",{},[14,232,233,234,265,266,297,298,329],{},"SFT 阶段的超参数（LoRA Rank ",[23,235,237,251],{"className":236},[26],[23,238,240],{"className":239},[30],[32,241,242],{"xmlns":34},[36,243,244,249],{},[39,245,246],{},[47,247,248],{},"r",[76,250,248],{"encoding":78},[23,252,254],{"className":253,"ariaHidden":52},[83],[23,255,257,261],{"className":256},[87],[23,258],{"className":259,"style":260},[91],"height:0.4306em;",[23,262,248],{"className":263,"style":264},[100,101],"margin-right:0.0278em;","、学习率 ",[23,267,269,284],{"className":268},[26],[23,270,272],{"className":271},[30],[32,273,274],{"xmlns":34},[36,275,276,281],{},[39,277,278],{},[47,279,280],{},"η",[76,282,283],{"encoding":78},"\\eta",[23,285,287],{"className":286,"ariaHidden":52},[83],[23,288,290,294],{"className":289},[87],[23,291],{"className":292,"style":293},[91],"height:0.625em;vertical-align:-0.1944em;",[23,295,280],{"className":296,"style":117},[100,101],"、样本规模 ",[23,299,301,315],{"className":300},[26],[23,302,304],{"className":303},[30],[32,305,306],{"xmlns":34},[36,307,308,313],{},[39,309,310],{},[47,311,312],{},"N",[76,314,312],{"encoding":78},[23,316,318],{"className":317,"ariaHidden":52},[83],[23,319,321,325],{"className":320},[87],[23,322],{"className":323,"style":324},[91],"height:0.6833em;",[23,326,312],{"className":327,"style":328},[100,101],"margin-right:0.109em;","）究竟哪个对下游数学推理准确率起决定性作用？单次随机种子的波动有多大？",[229,331,332],{},[14,333,334,335,416,417,695],{},"在单卡显存受限环境下运行 DPO，如何在不常驻第二份冻结基座模型（Reference Model ",[23,336,338,359],{"className":337},[26],[23,339,341],{"className":340},[30],[32,342,343],{"xmlns":34},[36,344,345,356],{},[39,346,347],{},[55,348,349,352],{},[47,350,351],{},"π",[353,354,355],"mtext",{},"ref",[76,357,358],{"encoding":78},"\\pi_{\\text{ref}}",[23,360,362],{"className":361,"ariaHidden":52},[83],[23,363,365,369],{"className":364},[87],[23,366],{"className":367,"style":368},[91],"height:0.5806em;vertical-align:-0.15em;",[23,370,372,375],{"className":371},[100],[23,373,351],{"className":374,"style":117},[100,101],[23,376,378],{"className":377},[121],[23,379,381,408],{"className":380},[125,126],[23,382,384,405],{"className":383},[130],[23,385,387],{"className":386,"style":193},[134],[23,388,389,392],{"style":138},[23,390],{"className":391,"style":143},[142],[23,393,395],{"className":394},[147,148,149,150],[23,396,398],{"className":397},[100,150],[23,399,402],{"className":400},[100,401,150],"text",[23,403,355],{"className":404},[100,150],[23,406,159],{"className":407},[158],[23,409,411],{"className":410},[130],[23,412,414],{"className":413,"style":166},[134],[23,415],{},"）的前提下，精确计算对数概率比 ",[23,418,420,479],{"className":419},[26],[23,421,423],{"className":422},[30],[32,424,425],{"xmlns":34},[36,426,427,476],{},[39,428,429,432,435],{},[47,430,431],{},"log",[42,433,434],{},"⁡",[436,437,438,458],"mfrac",{},[39,439,440,447,449,451,454,456],{},[55,441,442,444],{},[47,443,351],{},[47,445,446],{},"θ",[42,448,45],{"stretchy":44},[47,450,59],{},[42,452,453],{},"∣",[47,455,49],{},[42,457,74],{"stretchy":44},[39,459,460,466,468,470,472,474],{},[55,461,462,464],{},[47,463,351],{},[353,465,355],{},[42,467,45],{"stretchy":44},[47,469,59],{},[42,471,453],{},[47,473,49],{},[42,475,74],{"stretchy":44},[76,477,478],{"encoding":78},"\\log \\frac{\\pi_\\theta(y\\mid x)}{\\pi_{\\text{ref}}(y\\mid x)}",[23,480,482],{"className":481,"ariaHidden":52},[83],[23,483,485,489,498,501],{"className":484},[87],[23,486],{"className":487,"style":488},[91],"height:1.53em;vertical-align:-0.52em;",[23,490,493,494],{"className":491},[492],"mop","lo",[23,495,497],{"style":496},"margin-right:0.0139em;","g",[23,499],{"className":500,"style":110},[109],[23,502,504,508,692],{"className":503},[100],[23,505],{"className":506},[96,507],"nulldelimiter",[23,509,511],{"className":510},[436],[23,512,514,683],{"className":513},[125,126],[23,515,517,680],{"className":516},[130],[23,518,521,602,613],{"className":519,"style":520},[134],"height:1.01em;",[23,522,524,528],{"style":523},"top:-2.655em;",[23,525],{"className":526,"style":527},[142],"height:3em;",[23,529,531],{"className":530},[147,148,149,150],[23,532,534,586,589,592,596,599],{"className":533},[100,150],[23,535,537,540],{"className":536},[100,150],[23,538,351],{"className":539,"style":117},[100,101,150],[23,541,543],{"className":542},[121],[23,544,546,577],{"className":545},[125,126],[23,547,549,574],{"className":548},[130],[23,550,553],{"className":551,"style":552},[134],"height:0.3448em;",[23,554,556,560],{"style":555},"top:-2.3488em;margin-left:-0.0359em;margin-right:0.0714em;",[23,557],{"className":558,"style":559},[142],"height:2.5em;",[23,561,565],{"className":562},[147,563,564,150],"reset-size3","size1",[23,566,568],{"className":567},[100,150],[23,569,571],{"className":570},[100,401,150],[23,572,355],{"className":573},[100,150],[23,575,159],{"className":576},[158],[23,578,580],{"className":579},[130],[23,581,584],{"className":582,"style":583},[134],"height:0.1512em;",[23,585],{},[23,587,45],{"className":588},[96,150],[23,590,59],{"className":591,"style":117},[100,101,150],[23,593,453],{"className":594},[595,150],"mrel",[23,597,49],{"className":598},[100,101,150],[23,600,74],{"className":601},[220,150],[23,603,605,608],{"style":604},"top:-3.23em;",[23,606],{"className":607,"style":527},[142],[23,609],{"className":610,"style":612},[611],"frac-line","border-bottom-width:0.04em;",[23,614,616,619],{"style":615},"top:-3.485em;",[23,617],{"className":618,"style":527},[142],[23,620,622],{"className":621},[147,148,149,150],[23,623,625,665,668,671,674,677],{"className":624},[100,150],[23,626,628,631],{"className":627},[100,150],[23,629,351],{"className":630,"style":117},[100,101,150],[23,632,634],{"className":633},[121],[23,635,637,657],{"className":636},[125,126],[23,638,640,654],{"className":639},[130],[23,641,643],{"className":642,"style":552},[134],[23,644,645,648],{"style":555},[23,646],{"className":647,"style":559},[142],[23,649,651],{"className":650},[147,563,564,150],[23,652,446],{"className":653,"style":264},[100,101,150],[23,655,159],{"className":656},[158],[23,658,660],{"className":659},[130],[23,661,663],{"className":662,"style":583},[134],[23,664],{},[23,666,45],{"className":667},[96,150],[23,669,59],{"className":670,"style":117},[100,101,150],[23,672,453],{"className":673},[595,150],[23,675,49],{"className":676},[100,101,150],[23,678,74],{"className":679},[220,150],[23,681,159],{"className":682},[158],[23,684,686],{"className":685},[130],[23,687,690],{"className":688,"style":689},[134],"height:0.52em;",[23,691],{},[23,693],{"className":694},[220,507],"？",[229,697,698],{},[14,699,700,701,705],{},"在通用偏好数据集（如 ",[702,703,704],"code",{},"Anthropic\u002Fhh-rlhf","）上完成 DPO 提升安全性与语气偏好后，模型在确定性推理基准（如 GSM8K）上是会协同提升，还是会付出“对齐税（Alignment Tax）”？",[10,707,708,709,712,713,716,717,720,721,724],{},"本文基于我开源的后训练与对齐实验项目（",[702,710,711],{},"experiments\u002Fsft-ablation\u002F"," 与 ",[702,714,715],{},"experiments\u002Fdpo-alignment\u002F","），以 ",[14,718,719],{},"Qwen2.5"," 为基座，给出包含 ",[14,722,723],{},"3 随机种子交叉验证（Seeds 42 \u002F 123 \u002F 456）","的完整实验数据、代码实现与机理分析。",[726,727],"hr",{},[729,730,732],"h2",{"id":731},"一第一阶段sft-超参数三维度受控消融3-seed-交叉验证","一、第一阶段：SFT 超参数三维度受控消融（3-Seed 交叉验证）",[10,734,735,736,739,740,743,744,747,748,751,752,755,756,759],{},"为了避免单次训练的随机噪声误导超参结论，我们在 ",[14,737,738],{},"GSM8K"," 数学推理任务上固定评估集（200 道独立测试题，严格提取 ",[702,741,742],{},"#### \u003Cnumber>"," 最终数值匹配），对 ",[14,745,746],{},"LoRA Rank","、",[14,749,750],{},"学习率（Learning Rate）","与",[14,753,754],{},"训练样本规模（Data Size）","三个正交维度分别执行 3 组随机种子（",[702,757,758],{},"42, 123, 456","）独立训练。",[10,761,762,763,766,767,770],{},"未微调基座模型的零样本准确率为 ",[14,764,765],{},"19.50%","。各组实验汇总结果（来自 ",[702,768,769],{},"experiments\u002Fsft-ablation\u002Fresults.json","）如下：",[772,773,775,776,956],"h3",{"id":774},"_1-维度-alora-rank-消融固定-lr2104n1000textlr2times-10-4-n1000lr2104n1000","1. 维度 A：LoRA Rank 消融（固定 ",[23,777,779,830],{"className":778},[26],[23,780,782],{"className":781},[30],[32,783,784],{"xmlns":34},[36,785,786,827],{},[39,787,788,791,794,798,801,815,817,820,822,824],{},[353,789,790],{},"LR",[42,792,793],{},"=",[795,796,797],"mn",{},"2",[42,799,800],{},"×",[802,803,804,807],"msup",{},[795,805,806],{},"10",[39,808,809,812],{},[42,810,811],{},"−",[795,813,814],{},"4",[42,816,53],{"separator":52},[353,818,819],{},"  ",[47,821,312],{},[42,823,793],{},[795,825,826],{},"1000",[76,828,829],{"encoding":78},"\\text{LR}=2\\times 10^{-4},\\; N=1000",[23,831,833,855,876,946],{"className":832,"ariaHidden":52},[83],[23,834,836,839,845,849,852],{"className":835},[87],[23,837],{"className":838,"style":324},[91],[23,840,842],{"className":841},[100,401],[23,843,790],{"className":844},[100],[23,846],{"className":847,"style":848},[109],"margin-right:0.2778em;",[23,850,793],{"className":851},[595],[23,853],{"className":854,"style":848},[109],[23,856,858,862,865,869,873],{"className":857},[87],[23,859],{"className":860,"style":861},[91],"height:0.7278em;vertical-align:-0.0833em;",[23,863,797],{"className":864},[100],[23,866],{"className":867,"style":868},[109],"margin-right:0.2222em;",[23,870,800],{"className":871},[872],"mbin",[23,874],{"className":875,"style":868},[109],[23,877,879,883,887,925,928,931,934,937,940,943],{"className":878},[87],[23,880],{"className":881,"style":882},[91],"height:1.0085em;vertical-align:-0.1944em;",[23,884,886],{"className":885},[100],"1",[23,888,890,894],{"className":889},[100],[23,891,893],{"className":892},[100],"0",[23,895,897],{"className":896},[121],[23,898,900],{"className":899},[125],[23,901,903],{"className":902},[130],[23,904,907],{"className":905,"style":906},[134],"height:0.8141em;",[23,908,910,913],{"style":909},"top:-3.063em;margin-right:0.05em;",[23,911],{"className":912,"style":143},[142],[23,914,916],{"className":915},[147,148,149,150],[23,917,919,922],{"className":918},[100,150],[23,920,811],{"className":921},[100,150],[23,923,814],{"className":924},[100,150],[23,926,53],{"className":927},[105],[23,929],{"className":930,"style":848},[109],[23,932],{"className":933,"style":110},[109],[23,935,312],{"className":936,"style":328},[100,101],[23,938],{"className":939,"style":848},[109],[23,941,793],{"className":942},[595],[23,944],{"className":945,"style":848},[109],[23,947,949,953],{"className":948},[87],[23,950],{"className":951,"style":952},[91],"height:0.6444em;",[23,954,826],{"className":955},[100],"）",[958,959,960,1048],"table",{},[961,962,963],"thead",{},[964,965,966,999,1033,1036,1039,1042,1045],"tr",{},[967,968,970,971,74],"th",{"align":969},"right","LoRA Rank (",[23,972,974,987],{"className":973},[26],[23,975,977],{"className":976},[30],[32,978,979],{"xmlns":34},[36,980,981,985],{},[39,982,983],{},[47,984,248],{},[76,986,248],{"encoding":78},[23,988,990],{"className":989,"ariaHidden":52},[83],[23,991,993,996],{"className":992},[87],[23,994],{"className":995,"style":260},[91],[23,997,248],{"className":998,"style":264},[100,101],[967,1000,1001,1002,74],{"align":969},"LoRA Alpha (",[23,1003,1005,1020],{"className":1004},[26],[23,1006,1008],{"className":1007},[30],[32,1009,1010],{"xmlns":34},[36,1011,1012,1017],{},[39,1013,1014],{},[47,1015,1016],{},"α",[76,1018,1019],{"encoding":78},"\\alpha",[23,1021,1023],{"className":1022,"ariaHidden":52},[83],[23,1024,1026,1029],{"className":1025},[87],[23,1027],{"className":1028,"style":260},[91],[23,1030,1016],{"className":1031,"style":1032},[100,101],"margin-right:0.0037em;",[967,1034,1035],{"align":969},"可训练参数占比",[967,1037,1038],{"align":969},"最终 Train Loss (Mean ± Std)",[967,1040,1041],{"align":969},"Eval Loss (Mean ± Std)",[967,1043,1044],{"align":969},"GSM8K 准确率 (Mean ± Std)",[967,1046,1047],{"align":969},"相对基座增益",[1049,1050,1051,1337,1614,1895,2172],"tbody",{},[964,1052,1053,1107,1110,1113,1169,1224,1289],{},[1054,1055,1056],"td",{"align":969},[23,1057,1059,1077],{"className":1058},[26],[23,1060,1062],{"className":1061},[30],[32,1063,1064],{"xmlns":34},[36,1065,1066,1074],{},[39,1067,1068,1070,1072],{},[47,1069,248],{},[42,1071,793],{},[795,1073,814],{},[76,1075,1076],{"encoding":78},"r = 4",[23,1078,1080,1098],{"className":1079,"ariaHidden":52},[83],[23,1081,1083,1086,1089,1092,1095],{"className":1082},[87],[23,1084],{"className":1085,"style":260},[91],[23,1087,248],{"className":1088,"style":264},[100,101],[23,1090],{"className":1091,"style":848},[109],[23,1093,793],{"className":1094},[595],[23,1096],{"className":1097,"style":848},[109],[23,1099,1101,1104],{"className":1100},[87],[23,1102],{"className":1103,"style":952},[91],[23,1105,814],{"className":1106},[100],[1054,1108,1109],{"align":969},"8",[1054,1111,1112],{"align":969},"0.11%",[1054,1114,1115],{"align":969},[23,1116,1118,1139],{"className":1117},[26],[23,1119,1121],{"className":1120},[30],[32,1122,1123],{"xmlns":34},[36,1124,1125,1136],{},[39,1126,1127,1130,1133],{},[795,1128,1129],{},"0.5858",[42,1131,1132],{},"±",[795,1134,1135],{},"0.0044",[76,1137,1138],{"encoding":78},"0.5858 \\pm 0.0044",[23,1140,1142,1160],{"className":1141,"ariaHidden":52},[83],[23,1143,1145,1148,1151,1154,1157],{"className":1144},[87],[23,1146],{"className":1147,"style":861},[91],[23,1149,1129],{"className":1150},[100],[23,1152],{"className":1153,"style":868},[109],[23,1155,1132],{"className":1156},[872],[23,1158],{"className":1159,"style":868},[109],[23,1161,1163,1166],{"className":1162},[87],[23,1164],{"className":1165,"style":952},[91],[23,1167,1135],{"className":1168},[100],[1054,1170,1171],{"align":969},[23,1172,1174,1194],{"className":1173},[26],[23,1175,1177],{"className":1176},[30],[32,1178,1179],{"xmlns":34},[36,1180,1181,1191],{},[39,1182,1183,1186,1188],{},[795,1184,1185],{},"0.6648",[42,1187,1132],{},[795,1189,1190],{},"0.0005",[76,1192,1193],{"encoding":78},"0.6648 \\pm 0.0005",[23,1195,1197,1215],{"className":1196,"ariaHidden":52},[83],[23,1198,1200,1203,1206,1209,1212],{"className":1199},[87],[23,1201],{"className":1202,"style":861},[91],[23,1204,1185],{"className":1205},[100],[23,1207],{"className":1208,"style":868},[109],[23,1210,1132],{"className":1211},[872],[23,1213],{"className":1214,"style":868},[109],[23,1216,1218,1221],{"className":1217},[87],[23,1219],{"className":1220,"style":952},[91],[23,1222,1190],{"className":1223},[100],[1054,1225,1226],{"align":969},[23,1227,1229,1255],{"className":1228},[26],[23,1230,1232],{"className":1231},[30],[32,1233,1234],{"xmlns":34},[36,1235,1236,1252],{},[39,1237,1238,1241,1245,1247,1250],{},[795,1239,1240],{},"28.50",[47,1242,1244],{"mathvariant":1243},"normal","%",[42,1246,1132],{},[795,1248,1249],{},"1.08",[47,1251,1244],{"mathvariant":1243},[76,1253,1254],{"encoding":78},"28.50\\% \\pm 1.08\\%",[23,1256,1258,1278],{"className":1257,"ariaHidden":52},[83],[23,1259,1261,1265,1269,1272,1275],{"className":1260},[87],[23,1262],{"className":1263,"style":1264},[91],"height:0.8333em;vertical-align:-0.0833em;",[23,1266,1268],{"className":1267},[100],"28.50%",[23,1270],{"className":1271,"style":868},[109],[23,1273,1132],{"className":1274},[872],[23,1276],{"className":1277,"style":868},[109],[23,1279,1281,1285],{"className":1280},[87],[23,1282],{"className":1283,"style":1284},[91],"height:0.8056em;vertical-align:-0.0556em;",[23,1286,1288],{"className":1287},[100],"1.08%",[1054,1290,1291],{"align":969},[23,1292,1294,1315],{"className":1293},[26],[23,1295,1297],{"className":1296},[30],[32,1298,1299],{"xmlns":34},[36,1300,1301,1312],{},[39,1302,1303,1306,1309],{},[42,1304,1305],{},"+",[795,1307,1308],{},"9.00",[353,1310,1311],{}," pp",[76,1313,1314],{"encoding":78},"+9.00\\text{ pp}",[23,1316,1318],{"className":1317,"ariaHidden":52},[83],[23,1319,1321,1325,1328,1331],{"className":1320},[87],[23,1322],{"className":1323,"style":1324},[91],"height:0.8389em;vertical-align:-0.1944em;",[23,1326,1305],{"className":1327},[100],[23,1329,1308],{"className":1330},[100],[23,1332,1334],{"className":1333},[100,401],[23,1335,1311],{"className":1336},[100],[964,1338,1339,1392,1395,1398,1453,1508,1569],{},[1054,1340,1341],{"align":969},[23,1342,1344,1362],{"className":1343},[26],[23,1345,1347],{"className":1346},[30],[32,1348,1349],{"xmlns":34},[36,1350,1351,1359],{},[39,1352,1353,1355,1357],{},[47,1354,248],{},[42,1356,793],{},[795,1358,1109],{},[76,1360,1361],{"encoding":78},"r = 8",[23,1363,1365,1383],{"className":1364,"ariaHidden":52},[83],[23,1366,1368,1371,1374,1377,1380],{"className":1367},[87],[23,1369],{"className":1370,"style":260},[91],[23,1372,248],{"className":1373,"style":264},[100,101],[23,1375],{"className":1376,"style":848},[109],[23,1378,793],{"className":1379},[595],[23,1381],{"className":1382,"style":848},[109],[23,1384,1386,1389],{"className":1385},[87],[23,1387],{"className":1388,"style":952},[91],[23,1390,1109],{"className":1391},[100],[1054,1393,1394],{"align":969},"16",[1054,1396,1397],{"align":969},"0.22%",[1054,1399,1400],{"align":969},[23,1401,1403,1423],{"className":1402},[26],[23,1404,1406],{"className":1405},[30],[32,1407,1408],{"xmlns":34},[36,1409,1410,1420],{},[39,1411,1412,1415,1417],{},[795,1413,1414],{},"0.5256",[42,1416,1132],{},[795,1418,1419],{},"0.0052",[76,1421,1422],{"encoding":78},"0.5256 \\pm 0.0052",[23,1424,1426,1444],{"className":1425,"ariaHidden":52},[83],[23,1427,1429,1432,1435,1438,1441],{"className":1428},[87],[23,1430],{"className":1431,"style":861},[91],[23,1433,1414],{"className":1434},[100],[23,1436],{"className":1437,"style":868},[109],[23,1439,1132],{"className":1440},[872],[23,1442],{"className":1443,"style":868},[109],[23,1445,1447,1450],{"className":1446},[87],[23,1448],{"className":1449,"style":952},[91],[23,1451,1419],{"className":1452},[100],[1054,1454,1455],{"align":969},[23,1456,1458,1478],{"className":1457},[26],[23,1459,1461],{"className":1460},[30],[32,1462,1463],{"xmlns":34},[36,1464,1465,1475],{},[39,1466,1467,1470,1472],{},[795,1468,1469],{},"0.6593",[42,1471,1132],{},[795,1473,1474],{},"0.0004",[76,1476,1477],{"encoding":78},"0.6593 \\pm 0.0004",[23,1479,1481,1499],{"className":1480,"ariaHidden":52},[83],[23,1482,1484,1487,1490,1493,1496],{"className":1483},[87],[23,1485],{"className":1486,"style":861},[91],[23,1488,1469],{"className":1489},[100],[23,1491],{"className":1492,"style":868},[109],[23,1494,1132],{"className":1495},[872],[23,1497],{"className":1498,"style":868},[109],[23,1500,1502,1505],{"className":1501},[87],[23,1503],{"className":1504,"style":952},[91],[23,1506,1474],{"className":1507},[100],[1054,1509,1510],{"align":969},[23,1511,1513,1537],{"className":1512},[26],[23,1514,1516],{"className":1515},[30],[32,1517,1518],{"xmlns":34},[36,1519,1520,1534],{},[39,1521,1522,1525,1527,1529,1532],{},[795,1523,1524],{},"28.67",[47,1526,1244],{"mathvariant":1243},[42,1528,1132],{},[795,1530,1531],{},"1.65",[47,1533,1244],{"mathvariant":1243},[76,1535,1536],{"encoding":78},"28.67\\% \\pm 1.65\\%",[23,1538,1540,1559],{"className":1539,"ariaHidden":52},[83],[23,1541,1543,1546,1550,1553,1556],{"className":1542},[87],[23,1544],{"className":1545,"style":1264},[91],[23,1547,1549],{"className":1548},[100],"28.67%",[23,1551],{"className":1552,"style":868},[109],[23,1554,1132],{"className":1555},[872],[23,1557],{"className":1558,"style":868},[109],[23,1560,1562,1565],{"className":1561},[87],[23,1563],{"className":1564,"style":1284},[91],[23,1566,1568],{"className":1567},[100],"1.65%",[1054,1570,1571],{"align":969},[23,1572,1574,1593],{"className":1573},[26],[23,1575,1577],{"className":1576},[30],[32,1578,1579],{"xmlns":34},[36,1580,1581,1590],{},[39,1582,1583,1585,1588],{},[42,1584,1305],{},[795,1586,1587],{},"9.17",[353,1589,1311],{},[76,1591,1592],{"encoding":78},"+9.17\\text{ pp}",[23,1594,1596],{"className":1595,"ariaHidden":52},[83],[23,1597,1599,1602,1605,1608],{"className":1598},[87],[23,1600],{"className":1601,"style":1324},[91],[23,1603,1305],{"className":1604},[100],[23,1606,1587],{"className":1607},[100],[23,1609,1611],{"className":1610},[100,401],[23,1612,1311],{"className":1613},[100],[964,1615,1616,1669,1672,1675,1730,1785,1848],{},[1054,1617,1618],{"align":969},[23,1619,1621,1639],{"className":1620},[26],[23,1622,1624],{"className":1623},[30],[32,1625,1626],{"xmlns":34},[36,1627,1628,1636],{},[39,1629,1630,1632,1634],{},[47,1631,248],{},[42,1633,793],{},[795,1635,1394],{},[76,1637,1638],{"encoding":78},"r = 16",[23,1640,1642,1660],{"className":1641,"ariaHidden":52},[83],[23,1643,1645,1648,1651,1654,1657],{"className":1644},[87],[23,1646],{"className":1647,"style":260},[91],[23,1649,248],{"className":1650,"style":264},[100,101],[23,1652],{"className":1653,"style":848},[109],[23,1655,793],{"className":1656},[595],[23,1658],{"className":1659,"style":848},[109],[23,1661,1663,1666],{"className":1662},[87],[23,1664],{"className":1665,"style":952},[91],[23,1667,1394],{"className":1668},[100],[1054,1670,1671],{"align":969},"32",[1054,1673,1674],{"align":969},"0.44%",[1054,1676,1677],{"align":969},[23,1678,1680,1700],{"className":1679},[26],[23,1681,1683],{"className":1682},[30],[32,1684,1685],{"xmlns":34},[36,1686,1687,1697],{},[39,1688,1689,1692,1694],{},[795,1690,1691],{},"0.4534",[42,1693,1132],{},[795,1695,1696],{},"0.0061",[76,1698,1699],{"encoding":78},"0.4534 \\pm 0.0061",[23,1701,1703,1721],{"className":1702,"ariaHidden":52},[83],[23,1704,1706,1709,1712,1715,1718],{"className":1705},[87],[23,1707],{"className":1708,"style":861},[91],[23,1710,1691],{"className":1711},[100],[23,1713],{"className":1714,"style":868},[109],[23,1716,1132],{"className":1717},[872],[23,1719],{"className":1720,"style":868},[109],[23,1722,1724,1727],{"className":1723},[87],[23,1725],{"className":1726,"style":952},[91],[23,1728,1696],{"className":1729},[100],[1054,1731,1732],{"align":969},[23,1733,1735,1755],{"className":1734},[26],[23,1736,1738],{"className":1737},[30],[32,1739,1740],{"xmlns":34},[36,1741,1742,1752],{},[39,1743,1744,1747,1749],{},[795,1745,1746],{},"0.6744",[42,1748,1132],{},[795,1750,1751],{},"0.0009",[76,1753,1754],{"encoding":78},"0.6744 \\pm 0.0009",[23,1756,1758,1776],{"className":1757,"ariaHidden":52},[83],[23,1759,1761,1764,1767,1770,1773],{"className":1760},[87],[23,1762],{"className":1763,"style":861},[91],[23,1765,1746],{"className":1766},[100],[23,1768],{"className":1769,"style":868},[109],[23,1771,1132],{"className":1772},[872],[23,1774],{"className":1775,"style":868},[109],[23,1777,1779,1782],{"className":1778},[87],[23,1780],{"className":1781,"style":952},[91],[23,1783,1751],{"className":1784},[100],[1054,1786,1787],{"align":969},[14,1788,1789],{},[23,1790,1792,1816],{"className":1791},[26],[23,1793,1795],{"className":1794},[30],[32,1796,1797],{"xmlns":34},[36,1798,1799,1813],{},[39,1800,1801,1804,1806,1808,1811],{},[795,1802,1803],{},"29.50",[47,1805,1244],{"mathvariant":1243},[42,1807,1132],{},[795,1809,1810],{},"2.94",[47,1812,1244],{"mathvariant":1243},[76,1814,1815],{"encoding":78},"29.50\\% \\pm 2.94\\%",[23,1817,1819,1838],{"className":1818,"ariaHidden":52},[83],[23,1820,1822,1825,1829,1832,1835],{"className":1821},[87],[23,1823],{"className":1824,"style":1264},[91],[23,1826,1828],{"className":1827},[100],"29.50%",[23,1830],{"className":1831,"style":868},[109],[23,1833,1132],{"className":1834},[872],[23,1836],{"className":1837,"style":868},[109],[23,1839,1841,1844],{"className":1840},[87],[23,1842],{"className":1843,"style":1284},[91],[23,1845,1847],{"className":1846},[100],"2.94%",[1054,1849,1850],{"align":969},[14,1851,1852],{},[23,1853,1855,1874],{"className":1854},[26],[23,1856,1858],{"className":1857},[30],[32,1859,1860],{"xmlns":34},[36,1861,1862,1871],{},[39,1863,1864,1866,1869],{},[42,1865,1305],{},[795,1867,1868],{},"10.00",[353,1870,1311],{},[76,1872,1873],{"encoding":78},"+10.00\\text{ pp}",[23,1875,1877],{"className":1876,"ariaHidden":52},[83],[23,1878,1880,1883,1886,1889],{"className":1879},[87],[23,1881],{"className":1882,"style":1324},[91],[23,1884,1305],{"className":1885},[100],[23,1887,1868],{"className":1888},[100],[23,1890,1892],{"className":1891},[100,401],[23,1893,1311],{"className":1894},[100],[964,1896,1897,1950,1953,1956,2011,2066,2127],{},[1054,1898,1899],{"align":969},[23,1900,1902,1920],{"className":1901},[26],[23,1903,1905],{"className":1904},[30],[32,1906,1907],{"xmlns":34},[36,1908,1909,1917],{},[39,1910,1911,1913,1915],{},[47,1912,248],{},[42,1914,793],{},[795,1916,1671],{},[76,1918,1919],{"encoding":78},"r = 32",[23,1921,1923,1941],{"className":1922,"ariaHidden":52},[83],[23,1924,1926,1929,1932,1935,1938],{"className":1925},[87],[23,1927],{"className":1928,"style":260},[91],[23,1930,248],{"className":1931,"style":264},[100,101],[23,1933],{"className":1934,"style":848},[109],[23,1936,793],{"className":1937},[595],[23,1939],{"className":1940,"style":848},[109],[23,1942,1944,1947],{"className":1943},[87],[23,1945],{"className":1946,"style":952},[91],[23,1948,1671],{"className":1949},[100],[1054,1951,1952],{"align":969},"64",[1054,1954,1955],{"align":969},"0.87%",[1054,1957,1958],{"align":969},[23,1959,1961,1981],{"className":1960},[26],[23,1962,1964],{"className":1963},[30],[32,1965,1966],{"xmlns":34},[36,1967,1968,1978],{},[39,1969,1970,1973,1975],{},[795,1971,1972],{},"0.3770",[42,1974,1132],{},[795,1976,1977],{},"0.0065",[76,1979,1980],{"encoding":78},"0.3770 \\pm 0.0065",[23,1982,1984,2002],{"className":1983,"ariaHidden":52},[83],[23,1985,1987,1990,1993,1996,1999],{"className":1986},[87],[23,1988],{"className":1989,"style":861},[91],[23,1991,1972],{"className":1992},[100],[23,1994],{"className":1995,"style":868},[109],[23,1997,1132],{"className":1998},[872],[23,2000],{"className":2001,"style":868},[109],[23,2003,2005,2008],{"className":2004},[87],[23,2006],{"className":2007,"style":952},[91],[23,2009,1977],{"className":2010},[100],[1054,2012,2013],{"align":969},[23,2014,2016,2036],{"className":2015},[26],[23,2017,2019],{"className":2018},[30],[32,2020,2021],{"xmlns":34},[36,2022,2023,2033],{},[39,2024,2025,2028,2030],{},[795,2026,2027],{},"0.7204",[42,2029,1132],{},[795,2031,2032],{},"0.0010",[76,2034,2035],{"encoding":78},"0.7204 \\pm 0.0010",[23,2037,2039,2057],{"className":2038,"ariaHidden":52},[83],[23,2040,2042,2045,2048,2051,2054],{"className":2041},[87],[23,2043],{"className":2044,"style":861},[91],[23,2046,2027],{"className":2047},[100],[23,2049],{"className":2050,"style":868},[109],[23,2052,1132],{"className":2053},[872],[23,2055],{"className":2056,"style":868},[109],[23,2058,2060,2063],{"className":2059},[87],[23,2061],{"className":2062,"style":952},[91],[23,2064,2032],{"className":2065},[100],[1054,2067,2068],{"align":969},[23,2069,2071,2095],{"className":2070},[26],[23,2072,2074],{"className":2073},[30],[32,2075,2076],{"xmlns":34},[36,2077,2078,2092],{},[39,2079,2080,2083,2085,2087,2090],{},[795,2081,2082],{},"27.83",[47,2084,1244],{"mathvariant":1243},[42,2086,1132],{},[795,2088,2089],{},"2.32",[47,2091,1244],{"mathvariant":1243},[76,2093,2094],{"encoding":78},"27.83\\% \\pm 2.32\\%",[23,2096,2098,2117],{"className":2097,"ariaHidden":52},[83],[23,2099,2101,2104,2108,2111,2114],{"className":2100},[87],[23,2102],{"className":2103,"style":1264},[91],[23,2105,2107],{"className":2106},[100],"27.83%",[23,2109],{"className":2110,"style":868},[109],[23,2112,1132],{"className":2113},[872],[23,2115],{"className":2116,"style":868},[109],[23,2118,2120,2123],{"className":2119},[87],[23,2121],{"className":2122,"style":1284},[91],[23,2124,2126],{"className":2125},[100],"2.32%",[1054,2128,2129],{"align":969},[23,2130,2132,2151],{"className":2131},[26],[23,2133,2135],{"className":2134},[30],[32,2136,2137],{"xmlns":34},[36,2138,2139,2148],{},[39,2140,2141,2143,2146],{},[42,2142,1305],{},[795,2144,2145],{},"8.33",[353,2147,1311],{},[76,2149,2150],{"encoding":78},"+8.33\\text{ pp}",[23,2152,2154],{"className":2153,"ariaHidden":52},[83],[23,2155,2157,2160,2163,2166],{"className":2156},[87],[23,2158],{"className":2159,"style":1324},[91],[23,2161,1305],{"className":2162},[100],[23,2164,2145],{"className":2165},[100],[23,2167,2169],{"className":2168},[100,401],[23,2170,1311],{"className":2171},[100],[964,2173,2174,2227,2230,2233,2287,2342,2403],{},[1054,2175,2176],{"align":969},[23,2177,2179,2197],{"className":2178},[26],[23,2180,2182],{"className":2181},[30],[32,2183,2184],{"xmlns":34},[36,2185,2186,2194],{},[39,2187,2188,2190,2192],{},[47,2189,248],{},[42,2191,793],{},[795,2193,1952],{},[76,2195,2196],{"encoding":78},"r = 64",[23,2198,2200,2218],{"className":2199,"ariaHidden":52},[83],[23,2201,2203,2206,2209,2212,2215],{"className":2202},[87],[23,2204],{"className":2205,"style":260},[91],[23,2207,248],{"className":2208,"style":264},[100,101],[23,2210],{"className":2211,"style":848},[109],[23,2213,793],{"className":2214},[595],[23,2216],{"className":2217,"style":848},[109],[23,2219,2221,2224],{"className":2220},[87],[23,2222],{"className":2223,"style":952},[91],[23,2225,1952],{"className":2226},[100],[1054,2228,2229],{"align":969},"128",[1054,2231,2232],{"align":969},"1.72%",[1054,2234,2235],{"align":969},[23,2236,2238,2257],{"className":2237},[26],[23,2239,2241],{"className":2240},[30],[32,2242,2243],{"xmlns":34},[36,2244,2245,2254],{},[39,2246,2247,2250,2252],{},[795,2248,2249],{},"0.3052",[42,2251,1132],{},[795,2253,1419],{},[76,2255,2256],{"encoding":78},"0.3052 \\pm 0.0052",[23,2258,2260,2278],{"className":2259,"ariaHidden":52},[83],[23,2261,2263,2266,2269,2272,2275],{"className":2262},[87],[23,2264],{"className":2265,"style":861},[91],[23,2267,2249],{"className":2268},[100],[23,2270],{"className":2271,"style":868},[109],[23,2273,1132],{"className":2274},[872],[23,2276],{"className":2277,"style":868},[109],[23,2279,2281,2284],{"className":2280},[87],[23,2282],{"className":2283,"style":952},[91],[23,2285,1419],{"className":2286},[100],[1054,2288,2289],{"align":969},[23,2290,2292,2312],{"className":2291},[26],[23,2293,2295],{"className":2294},[30],[32,2296,2297],{"xmlns":34},[36,2298,2299,2309],{},[39,2300,2301,2304,2306],{},[795,2302,2303],{},"0.7931",[42,2305,1132],{},[795,2307,2308],{},"0.0013",[76,2310,2311],{"encoding":78},"0.7931 \\pm 0.0013",[23,2313,2315,2333],{"className":2314,"ariaHidden":52},[83],[23,2316,2318,2321,2324,2327,2330],{"className":2317},[87],[23,2319],{"className":2320,"style":861},[91],[23,2322,2303],{"className":2323},[100],[23,2325],{"className":2326,"style":868},[109],[23,2328,1132],{"className":2329},[872],[23,2331],{"className":2332,"style":868},[109],[23,2334,2336,2339],{"className":2335},[87],[23,2337],{"className":2338,"style":952},[91],[23,2340,2308],{"className":2341},[100],[1054,2343,2344],{"align":969},[23,2345,2347,2371],{"className":2346},[26],[23,2348,2350],{"className":2349},[30],[32,2351,2352],{"xmlns":34},[36,2353,2354,2368],{},[39,2355,2356,2359,2361,2363,2366],{},[795,2357,2358],{},"27.17",[47,2360,1244],{"mathvariant":1243},[42,2362,1132],{},[795,2364,2365],{},"0.62",[47,2367,1244],{"mathvariant":1243},[76,2369,2370],{"encoding":78},"27.17\\% \\pm 0.62\\%",[23,2372,2374,2393],{"className":2373,"ariaHidden":52},[83],[23,2375,2377,2380,2384,2387,2390],{"className":2376},[87],[23,2378],{"className":2379,"style":1264},[91],[23,2381,2383],{"className":2382},[100],"27.17%",[23,2385],{"className":2386,"style":868},[109],[23,2388,1132],{"className":2389},[872],[23,2391],{"className":2392,"style":868},[109],[23,2394,2396,2399],{"className":2395},[87],[23,2397],{"className":2398,"style":1284},[91],[23,2400,2402],{"className":2401},[100],"0.62%",[1054,2404,2405],{"align":969},[23,2406,2408,2427],{"className":2407},[26],[23,2409,2411],{"className":2410},[30],[32,2412,2413],{"xmlns":34},[36,2414,2415,2424],{},[39,2416,2417,2419,2422],{},[42,2418,1305],{},[795,2420,2421],{},"7.67",[353,2423,1311],{},[76,2425,2426],{"encoding":78},"+7.67\\text{ pp}",[23,2428,2430],{"className":2429,"ariaHidden":52},[83],[23,2431,2433,2436,2439,2442],{"className":2432},[87],[23,2434],{"className":2435,"style":1324},[91],[23,2437,1305],{"className":2438},[100],[23,2440,2421],{"className":2441},[100],[23,2443,2445],{"className":2444},[100,401],[23,2446,1311],{"className":2447},[100],[10,2449,2450,2453,2454,2457,2458,2532,2533,2597,2598,2601,2602,2605,2606,2660],{},[14,2451,2452],{},"关键发现","：随着 Rank 从 4 增至 64，训练集 Loss 单调下降（",[702,2455,2456],{},"0.5858 → 0.3052","），但验证集 Loss 在 ",[23,2459,2461,2484],{"className":2460},[26],[23,2462,2464],{"className":2463},[30],[32,2465,2466],{"xmlns":34},[36,2467,2468,2481],{},[39,2469,2470,2472,2474,2476,2479],{},[47,2471,248],{},[42,2473,793],{},[795,2475,1109],{},[42,2477,2478],{},"∼",[795,2480,1394],{},[76,2482,2483],{"encoding":78},"r=8\\sim 16",[23,2485,2487,2505,2523],{"className":2486,"ariaHidden":52},[83],[23,2488,2490,2493,2496,2499,2502],{"className":2489},[87],[23,2491],{"className":2492,"style":260},[91],[23,2494,248],{"className":2495,"style":264},[100,101],[23,2497],{"className":2498,"style":848},[109],[23,2500,793],{"className":2501},[595],[23,2503],{"className":2504,"style":848},[109],[23,2506,2508,2511,2514,2517,2520],{"className":2507},[87],[23,2509],{"className":2510,"style":952},[91],[23,2512,1109],{"className":2513},[100],[23,2515],{"className":2516,"style":848},[109],[23,2518,2478],{"className":2519},[595],[23,2521],{"className":2522,"style":848},[109],[23,2524,2526,2529],{"className":2525},[87],[23,2527],{"className":2528,"style":952},[91],[23,2530,1394],{"className":2531},[100]," 处即达到谷底，并在 ",[23,2534,2536,2558],{"className":2535},[26],[23,2537,2539],{"className":2538},[30],[32,2540,2541],{"xmlns":34},[36,2542,2543,2555],{},[39,2544,2545,2547,2549,2551,2553],{},[47,2546,248],{},[42,2548,793],{},[795,2550,1671],{},[42,2552,53],{"separator":52},[795,2554,1952],{},[76,2556,2557],{"encoding":78},"r=32, 64",[23,2559,2561,2579],{"className":2560,"ariaHidden":52},[83],[23,2562,2564,2567,2570,2573,2576],{"className":2563},[87],[23,2565],{"className":2566,"style":260},[91],[23,2568,248],{"className":2569,"style":264},[100,101],[23,2571],{"className":2572,"style":848},[109],[23,2574,793],{"className":2575},[595],[23,2577],{"className":2578,"style":848},[109],[23,2580,2582,2585,2588,2591,2594],{"className":2581},[87],[23,2583],{"className":2584,"style":1324},[91],[23,2586,1671],{"className":2587},[100],[23,2589,53],{"className":2590},[105],[23,2592],{"className":2593,"style":110},[109],[23,2595,1952],{"className":2596},[100]," 时显著反弹（",[702,2599,2600],{},"0.6593 → 0.7931","），对应 GSM8K 准确率从 ",[14,2603,2604],{},"29.50% 回落至 27.17%","。这证明在中等规模领域微调中，盲目增大 LoRA 秩会导致高秩子空间过拟合训练样本表面措辞，",[14,2607,2608,2659],{},[23,2609,2611,2629],{"className":2610},[26],[23,2612,2614],{"className":2613},[30],[32,2615,2616],{"xmlns":34},[36,2617,2618,2626],{},[39,2619,2620,2622,2624],{},[47,2621,248],{},[42,2623,793],{},[795,2625,1394],{},[76,2627,2628],{"encoding":78},"r=16",[23,2630,2632,2650],{"className":2631,"ariaHidden":52},[83],[23,2633,2635,2638,2641,2644,2647],{"className":2634},[87],[23,2636],{"className":2637,"style":260},[91],[23,2639,248],{"className":2640,"style":264},[100,101],[23,2642],{"className":2643,"style":848},[109],[23,2645,793],{"className":2646},[595],[23,2648],{"className":2649,"style":848},[109],[23,2651,2653,2656],{"className":2652},[87],[23,2654],{"className":2655,"style":952},[91],[23,2657,1394],{"className":2658},[100]," 是兼顾容量与泛化的最优点","。",[772,2662,2664],{"id":2663},"_2-维度-b-c学习率与数据规模-scaling-效应","2. 维度 B & C：学习率与数据规模 Scaling 效应",[958,2666,2667,2683],{},[961,2668,2669],{},[964,2670,2671,2674,2677,2679,2681],{},[967,2672,2673],{},"实验分组",[967,2675,2676],{},"变量设置",[967,2678,1041],{"align":969},[967,2680,1044],{"align":969},[967,2682,1047],{"align":969},[1049,2684,2685,3053,3335,3608,3891,4260,4479,4687],{},[964,2686,2687,2780,2899,2953,3010],{},[1054,2688,2689,2692,2693,74],{},[14,2690,2691],{},"学习率消融"," (",[23,2694,2696,2722],{"className":2695},[26],[23,2697,2699],{"className":2698},[30],[32,2700,2701],{"xmlns":34},[36,2702,2703,2719],{},[39,2704,2705,2707,2709,2711,2713,2715,2717],{},[47,2706,248],{},[42,2708,793],{},[795,2710,1394],{},[42,2712,53],{"separator":52},[47,2714,312],{},[42,2716,793],{},[795,2718,826],{},[76,2720,2721],{"encoding":78},"r=16, N=1000",[23,2723,2725,2743,2771],{"className":2724,"ariaHidden":52},[83],[23,2726,2728,2731,2734,2737,2740],{"className":2727},[87],[23,2729],{"className":2730,"style":260},[91],[23,2732,248],{"className":2733,"style":264},[100,101],[23,2735],{"className":2736,"style":848},[109],[23,2738,793],{"className":2739},[595],[23,2741],{"className":2742,"style":848},[109],[23,2744,2746,2750,2753,2756,2759,2762,2765,2768],{"className":2745},[87],[23,2747],{"className":2748,"style":2749},[91],"height:0.8778em;vertical-align:-0.1944em;",[23,2751,1394],{"className":2752},[100],[23,2754,53],{"className":2755},[105],[23,2757],{"className":2758,"style":110},[109],[23,2760,312],{"className":2761,"style":328},[100,101],[23,2763],{"className":2764,"style":848},[109],[23,2766,793],{"className":2767},[595],[23,2769],{"className":2770,"style":848},[109],[23,2772,2774,2777],{"className":2773},[87],[23,2775],{"className":2776,"style":952},[91],[23,2778,826],{"className":2779},[100],[1054,2781,2782],{},[23,2783,2785,2816],{"className":2784},[26],[23,2786,2788],{"className":2787},[30],[32,2789,2790],{"xmlns":34},[36,2791,2792,2813],{},[39,2793,2794,2796,2798,2801,2803],{},[47,2795,280],{},[42,2797,793],{},[795,2799,2800],{},"5",[42,2802,800],{},[802,2804,2805,2807],{},[795,2806,806],{},[39,2808,2809,2811],{},[42,2810,811],{},[795,2812,2800],{},[76,2814,2815],{"encoding":78},"\\eta = 5\\times 10^{-5}",[23,2817,2819,2837,2855],{"className":2818,"ariaHidden":52},[83],[23,2820,2822,2825,2828,2831,2834],{"className":2821},[87],[23,2823],{"className":2824,"style":293},[91],[23,2826,280],{"className":2827,"style":117},[100,101],[23,2829],{"className":2830,"style":848},[109],[23,2832,793],{"className":2833},[595],[23,2835],{"className":2836,"style":848},[109],[23,2838,2840,2843,2846,2849,2852],{"className":2839},[87],[23,2841],{"className":2842,"style":861},[91],[23,2844,2800],{"className":2845},[100],[23,2847],{"className":2848,"style":868},[109],[23,2850,800],{"className":2851},[872],[23,2853],{"className":2854,"style":868},[109],[23,2856,2858,2861,2864],{"className":2857},[87],[23,2859],{"className":2860,"style":906},[91],[23,2862,886],{"className":2863},[100],[23,2865,2867,2870],{"className":2866},[100],[23,2868,893],{"className":2869},[100],[23,2871,2873],{"className":2872},[121],[23,2874,2876],{"className":2875},[125],[23,2877,2879],{"className":2878},[130],[23,2880,2882],{"className":2881,"style":906},[134],[23,2883,2884,2887],{"style":909},[23,2885],{"className":2886,"style":143},[142],[23,2888,2890],{"className":2889},[147,148,149,150],[23,2891,2893,2896],{"className":2892},[100,150],[23,2894,811],{"className":2895},[100,150],[23,2897,2800],{"className":2898},[100,150],[1054,2900,2901],{"align":969},[23,2902,2904,2923],{"className":2903},[26],[23,2905,2907],{"className":2906},[30],[32,2908,2909],{"xmlns":34},[36,2910,2911,2920],{},[39,2912,2913,2916,2918],{},[795,2914,2915],{},"0.6689",[42,2917,1132],{},[795,2919,1474],{},[76,2921,2922],{"encoding":78},"0.6689 \\pm 0.0004",[23,2924,2926,2944],{"className":2925,"ariaHidden":52},[83],[23,2927,2929,2932,2935,2938,2941],{"className":2928},[87],[23,2930],{"className":2931,"style":861},[91],[23,2933,2915],{"className":2934},[100],[23,2936],{"className":2937,"style":868},[109],[23,2939,1132],{"className":2940},[872],[23,2942],{"className":2943,"style":868},[109],[23,2945,2947,2950],{"className":2946},[87],[23,2948],{"className":2949,"style":952},[91],[23,2951,1474],{"className":2952},[100],[1054,2954,2955],{"align":969},[23,2956,2958,2980],{"className":2957},[26],[23,2959,2961],{"className":2960},[30],[32,2962,2963],{"xmlns":34},[36,2964,2965,2977],{},[39,2966,2967,2969,2971,2973,2975],{},[795,2968,2082],{},[47,2970,1244],{"mathvariant":1243},[42,2972,1132],{},[795,2974,2365],{},[47,2976,1244],{"mathvariant":1243},[76,2978,2979],{"encoding":78},"27.83\\% \\pm 0.62\\%",[23,2981,2983,3001],{"className":2982,"ariaHidden":52},[83],[23,2984,2986,2989,2992,2995,2998],{"className":2985},[87],[23,2987],{"className":2988,"style":1264},[91],[23,2990,2107],{"className":2991},[100],[23,2993],{"className":2994,"style":868},[109],[23,2996,1132],{"className":2997},[872],[23,2999],{"className":3000,"style":868},[109],[23,3002,3004,3007],{"className":3003},[87],[23,3005],{"className":3006,"style":1284},[91],[23,3008,2402],{"className":3009},[100],[1054,3011,3012],{"align":969},[23,3013,3015,3032],{"className":3014},[26],[23,3016,3018],{"className":3017},[30],[32,3019,3020],{"xmlns":34},[36,3021,3022,3030],{},[39,3023,3024,3026,3028],{},[42,3025,1305],{},[795,3027,2145],{},[353,3029,1311],{},[76,3031,2150],{"encoding":78},[23,3033,3035],{"className":3034,"ariaHidden":52},[83],[23,3036,3038,3041,3044,3047],{"className":3037},[87],[23,3039],{"className":3040,"style":1324},[91],[23,3042,1305],{"className":3043},[100],[23,3045,2145],{"className":3046},[100],[23,3048,3050],{"className":3049},[100,401],[23,3051,1311],{"className":3052},[100],[964,3054,3055,3057,3175,3229,3290],{},[1054,3056],{},[1054,3058,3059],{},[23,3060,3062,3092],{"className":3061},[26],[23,3063,3065],{"className":3064},[30],[32,3066,3067],{"xmlns":34},[36,3068,3069,3089],{},[39,3070,3071,3073,3075,3077,3079],{},[47,3072,280],{},[42,3074,793],{},[795,3076,886],{},[42,3078,800],{},[802,3080,3081,3083],{},[795,3082,806],{},[39,3084,3085,3087],{},[42,3086,811],{},[795,3088,814],{},[76,3090,3091],{"encoding":78},"\\eta = 1\\times 10^{-4}",[23,3093,3095,3113,3131],{"className":3094,"ariaHidden":52},[83],[23,3096,3098,3101,3104,3107,3110],{"className":3097},[87],[23,3099],{"className":3100,"style":293},[91],[23,3102,280],{"className":3103,"style":117},[100,101],[23,3105],{"className":3106,"style":848},[109],[23,3108,793],{"className":3109},[595],[23,3111],{"className":3112,"style":848},[109],[23,3114,3116,3119,3122,3125,3128],{"className":3115},[87],[23,3117],{"className":3118,"style":861},[91],[23,3120,886],{"className":3121},[100],[23,3123],{"className":3124,"style":868},[109],[23,3126,800],{"className":3127},[872],[23,3129],{"className":3130,"style":868},[109],[23,3132,3134,3137,3140],{"className":3133},[87],[23,3135],{"className":3136,"style":906},[91],[23,3138,886],{"className":3139},[100],[23,3141,3143,3146],{"className":3142},[100],[23,3144,893],{"className":3145},[100],[23,3147,3149],{"className":3148},[121],[23,3150,3152],{"className":3151},[125],[23,3153,3155],{"className":3154},[130],[23,3156,3158],{"className":3157,"style":906},[134],[23,3159,3160,3163],{"style":909},[23,3161],{"className":3162,"style":143},[142],[23,3164,3166],{"className":3165},[147,148,149,150],[23,3167,3169,3172],{"className":3168},[100,150],[23,3170,811],{"className":3171},[100,150],[23,3173,814],{"className":3174},[100,150],[1054,3176,3177],{"align":969},[23,3178,3180,3199],{"className":3179},[26],[23,3181,3183],{"className":3182},[30],[32,3184,3185],{"xmlns":34},[36,3186,3187,3196],{},[39,3188,3189,3192,3194],{},[795,3190,3191],{},"0.6591",[42,3193,1132],{},[795,3195,1474],{},[76,3197,3198],{"encoding":78},"0.6591 \\pm 0.0004",[23,3200,3202,3220],{"className":3201,"ariaHidden":52},[83],[23,3203,3205,3208,3211,3214,3217],{"className":3204},[87],[23,3206],{"className":3207,"style":861},[91],[23,3209,3191],{"className":3210},[100],[23,3212],{"className":3213,"style":868},[109],[23,3215,1132],{"className":3216},[872],[23,3218],{"className":3219,"style":868},[109],[23,3221,3223,3226],{"className":3222},[87],[23,3224],{"className":3225,"style":952},[91],[23,3227,1474],{"className":3228},[100],[1054,3230,3231],{"align":969},[23,3232,3234,3258],{"className":3233},[26],[23,3235,3237],{"className":3236},[30],[32,3238,3239],{"xmlns":34},[36,3240,3241,3255],{},[39,3242,3243,3246,3248,3250,3253],{},[795,3244,3245],{},"29.17",[47,3247,1244],{"mathvariant":1243},[42,3249,1132],{},[795,3251,3252],{},"1.25",[47,3254,1244],{"mathvariant":1243},[76,3256,3257],{"encoding":78},"29.17\\% \\pm 1.25\\%",[23,3259,3261,3280],{"className":3260,"ariaHidden":52},[83],[23,3262,3264,3267,3271,3274,3277],{"className":3263},[87],[23,3265],{"className":3266,"style":1264},[91],[23,3268,3270],{"className":3269},[100],"29.17%",[23,3272],{"className":3273,"style":868},[109],[23,3275,1132],{"className":3276},[872],[23,3278],{"className":3279,"style":868},[109],[23,3281,3283,3286],{"className":3282},[87],[23,3284],{"className":3285,"style":1284},[91],[23,3287,3289],{"className":3288},[100],"1.25%",[1054,3291,3292],{"align":969},[23,3293,3295,3314],{"className":3294},[26],[23,3296,3298],{"className":3297},[30],[32,3299,3300],{"xmlns":34},[36,3301,3302,3311],{},[39,3303,3304,3306,3309],{},[42,3305,1305],{},[795,3307,3308],{},"9.67",[353,3310,1311],{},[76,3312,3313],{"encoding":78},"+9.67\\text{ pp}",[23,3315,3317],{"className":3316,"ariaHidden":52},[83],[23,3318,3320,3323,3326,3329],{"className":3319},[87],[23,3321],{"className":3322,"style":1324},[91],[23,3324,1305],{"className":3325},[100],[23,3327,3308],{"className":3328},[100],[23,3330,3332],{"className":3331},[100,401],[23,3333,1311],{"className":3334},[100],[964,3336,3337,3339,3457,3509,3565],{},[1054,3338],{},[1054,3340,3341],{},[23,3342,3344,3374],{"className":3343},[26],[23,3345,3347],{"className":3346},[30],[32,3348,3349],{"xmlns":34},[36,3350,3351,3371],{},[39,3352,3353,3355,3357,3359,3361],{},[47,3354,280],{},[42,3356,793],{},[795,3358,797],{},[42,3360,800],{},[802,3362,3363,3365],{},[795,3364,806],{},[39,3366,3367,3369],{},[42,3368,811],{},[795,3370,814],{},[76,3372,3373],{"encoding":78},"\\eta = 2\\times 10^{-4}",[23,3375,3377,3395,3413],{"className":3376,"ariaHidden":52},[83],[23,3378,3380,3383,3386,3389,3392],{"className":3379},[87],[23,3381],{"className":3382,"style":293},[91],[23,3384,280],{"className":3385,"style":117},[100,101],[23,3387],{"className":3388,"style":848},[109],[23,3390,793],{"className":3391},[595],[23,3393],{"className":3394,"style":848},[109],[23,3396,3398,3401,3404,3407,3410],{"className":3397},[87],[23,3399],{"className":3400,"style":861},[91],[23,3402,797],{"className":3403},[100],[23,3405],{"className":3406,"style":868},[109],[23,3408,800],{"className":3409},[872],[23,3411],{"className":3412,"style":868},[109],[23,3414,3416,3419,3422],{"className":3415},[87],[23,3417],{"className":3418,"style":906},[91],[23,3420,886],{"className":3421},[100],[23,3423,3425,3428],{"className":3424},[100],[23,3426,893],{"className":3427},[100],[23,3429,3431],{"className":3430},[121],[23,3432,3434],{"className":3433},[125],[23,3435,3437],{"className":3436},[130],[23,3438,3440],{"className":3439,"style":906},[134],[23,3441,3442,3445],{"style":909},[23,3443],{"className":3444,"style":143},[142],[23,3446,3448],{"className":3447},[147,148,149,150],[23,3449,3451,3454],{"className":3450},[100,150],[23,3452,811],{"className":3453},[100,150],[23,3455,814],{"className":3456},[100,150],[1054,3458,3459],{"align":969},[23,3460,3462,3479],{"className":3461},[26],[23,3463,3465],{"className":3464},[30],[32,3466,3467],{"xmlns":34},[36,3468,3469,3477],{},[39,3470,3471,3473,3475],{},[795,3472,1746],{},[42,3474,1132],{},[795,3476,1751],{},[76,3478,1754],{"encoding":78},[23,3480,3482,3500],{"className":3481,"ariaHidden":52},[83],[23,3483,3485,3488,3491,3494,3497],{"className":3484},[87],[23,3486],{"className":3487,"style":861},[91],[23,3489,1746],{"className":3490},[100],[23,3492],{"className":3493,"style":868},[109],[23,3495,1132],{"className":3496},[872],[23,3498],{"className":3499,"style":868},[109],[23,3501,3503,3506],{"className":3502},[87],[23,3504],{"className":3505,"style":952},[91],[23,3507,1751],{"className":3508},[100],[1054,3510,3511],{"align":969},[23,3512,3514,3535],{"className":3513},[26],[23,3515,3517],{"className":3516},[30],[32,3518,3519],{"xmlns":34},[36,3520,3521,3533],{},[39,3522,3523,3525,3527,3529,3531],{},[795,3524,1803],{},[47,3526,1244],{"mathvariant":1243},[42,3528,1132],{},[795,3530,1810],{},[47,3532,1244],{"mathvariant":1243},[76,3534,1815],{"encoding":78},[23,3536,3538,3556],{"className":3537,"ariaHidden":52},[83],[23,3539,3541,3544,3547,3550,3553],{"className":3540},[87],[23,3542],{"className":3543,"style":1264},[91],[23,3545,1828],{"className":3546},[100],[23,3548],{"className":3549,"style":868},[109],[23,3551,1132],{"className":3552},[872],[23,3554],{"className":3555,"style":868},[109],[23,3557,3559,3562],{"className":3558},[87],[23,3560],{"className":3561,"style":1284},[91],[23,3563,1847],{"className":3564},[100],[1054,3566,3567],{"align":969},[23,3568,3570,3587],{"className":3569},[26],[23,3571,3573],{"className":3572},[30],[32,3574,3575],{"xmlns":34},[36,3576,3577,3585],{},[39,3578,3579,3581,3583],{},[42,3580,1305],{},[795,3582,1868],{},[353,3584,1311],{},[76,3586,1873],{"encoding":78},[23,3588,3590],{"className":3589,"ariaHidden":52},[83],[23,3591,3593,3596,3599,3602],{"className":3592},[87],[23,3594],{"className":3595,"style":1324},[91],[23,3597,1305],{"className":3598},[100],[23,3600,1868],{"className":3601},[100],[23,3603,3605],{"className":3604},[100,401],[23,3606,1311],{"className":3607},[100],[964,3609,3610,3612,3730,3785,3846],{},[1054,3611],{},[1054,3613,3614],{},[23,3615,3617,3647],{"className":3616},[26],[23,3618,3620],{"className":3619},[30],[32,3621,3622],{"xmlns":34},[36,3623,3624,3644],{},[39,3625,3626,3628,3630,3632,3634],{},[47,3627,280],{},[42,3629,793],{},[795,3631,2800],{},[42,3633,800],{},[802,3635,3636,3638],{},[795,3637,806],{},[39,3639,3640,3642],{},[42,3641,811],{},[795,3643,814],{},[76,3645,3646],{"encoding":78},"\\eta = 5\\times 10^{-4}",[23,3648,3650,3668,3686],{"className":3649,"ariaHidden":52},[83],[23,3651,3653,3656,3659,3662,3665],{"className":3652},[87],[23,3654],{"className":3655,"style":293},[91],[23,3657,280],{"className":3658,"style":117},[100,101],[23,3660],{"className":3661,"style":848},[109],[23,3663,793],{"className":3664},[595],[23,3666],{"className":3667,"style":848},[109],[23,3669,3671,3674,3677,3680,3683],{"className":3670},[87],[23,3672],{"className":3673,"style":861},[91],[23,3675,2800],{"className":3676},[100],[23,3678],{"className":3679,"style":868},[109],[23,3681,800],{"className":3682},[872],[23,3684],{"className":3685,"style":868},[109],[23,3687,3689,3692,3695],{"className":3688},[87],[23,3690],{"className":3691,"style":906},[91],[23,3693,886],{"className":3694},[100],[23,3696,3698,3701],{"className":3697},[100],[23,3699,893],{"className":3700},[100],[23,3702,3704],{"className":3703},[121],[23,3705,3707],{"className":3706},[125],[23,3708,3710],{"className":3709},[130],[23,3711,3713],{"className":3712,"style":906},[134],[23,3714,3715,3718],{"style":909},[23,3716],{"className":3717,"style":143},[142],[23,3719,3721],{"className":3720},[147,148,149,150],[23,3722,3724,3727],{"className":3723},[100,150],[23,3725,811],{"className":3726},[100,150],[23,3728,814],{"className":3729},[100,150],[1054,3731,3732],{"align":969},[23,3733,3735,3755],{"className":3734},[26],[23,3736,3738],{"className":3737},[30],[32,3739,3740],{"xmlns":34},[36,3741,3742,3752],{},[39,3743,3744,3747,3749],{},[795,3745,3746],{},"0.7510",[42,3748,1132],{},[795,3750,3751],{},"0.0033",[76,3753,3754],{"encoding":78},"0.7510 \\pm 0.0033",[23,3756,3758,3776],{"className":3757,"ariaHidden":52},[83],[23,3759,3761,3764,3767,3770,3773],{"className":3760},[87],[23,3762],{"className":3763,"style":861},[91],[23,3765,3746],{"className":3766},[100],[23,3768],{"className":3769,"style":868},[109],[23,3771,1132],{"className":3772},[872],[23,3774],{"className":3775,"style":868},[109],[23,3777,3779,3782],{"className":3778},[87],[23,3780],{"className":3781,"style":952},[91],[23,3783,3751],{"className":3784},[100],[1054,3786,3787],{"align":969},[23,3788,3790,3814],{"className":3789},[26],[23,3791,3793],{"className":3792},[30],[32,3794,3795],{"xmlns":34},[36,3796,3797,3811],{},[39,3798,3799,3802,3804,3806,3809],{},[795,3800,3801],{},"28.00",[47,3803,1244],{"mathvariant":1243},[42,3805,1132],{},[795,3807,3808],{},"0.71",[47,3810,1244],{"mathvariant":1243},[76,3812,3813],{"encoding":78},"28.00\\% \\pm 0.71\\%",[23,3815,3817,3836],{"className":3816,"ariaHidden":52},[83],[23,3818,3820,3823,3827,3830,3833],{"className":3819},[87],[23,3821],{"className":3822,"style":1264},[91],[23,3824,3826],{"className":3825},[100],"28.00%",[23,3828],{"className":3829,"style":868},[109],[23,3831,1132],{"className":3832},[872],[23,3834],{"className":3835,"style":868},[109],[23,3837,3839,3842],{"className":3838},[87],[23,3840],{"className":3841,"style":1284},[91],[23,3843,3845],{"className":3844},[100],"0.71%",[1054,3847,3848],{"align":969},[23,3849,3851,3870],{"className":3850},[26],[23,3852,3854],{"className":3853},[30],[32,3855,3856],{"xmlns":34},[36,3857,3858,3867],{},[39,3859,3860,3862,3865],{},[42,3861,1305],{},[795,3863,3864],{},"8.50",[353,3866,1311],{},[76,3868,3869],{"encoding":78},"+8.50\\text{ pp}",[23,3871,3873],{"className":3872,"ariaHidden":52},[83],[23,3874,3876,3879,3882,3885],{"className":3875},[87],[23,3877],{"className":3878,"style":1324},[91],[23,3880,1305],{"className":3881},[100],[23,3883,3864],{"className":3884},[100],[23,3886,3888],{"className":3887},[100,401],[23,3889,1311],{"className":3890},[100],[964,3892,3893,4049,4103,4158,4217],{},[1054,3894,3895,2692,3898,74],{},[14,3896,3897],{},"数据规模消融",[23,3899,3901,3939],{"className":3900},[26],[23,3902,3904],{"className":3903},[30],[32,3905,3906],{"xmlns":34},[36,3907,3908,3936],{},[39,3909,3910,3912,3914,3916,3918,3920,3922,3924,3926],{},[47,3911,248],{},[42,3913,793],{},[795,3915,1394],{},[42,3917,53],{"separator":52},[47,3919,280],{},[42,3921,793],{},[795,3923,797],{},[42,3925,800],{},[802,3927,3928,3930],{},[795,3929,806],{},[39,3931,3932,3934],{},[42,3933,811],{},[795,3935,814],{},[76,3937,3938],{"encoding":78},"r=16, \\eta=2\\times 10^{-4}",[23,3940,3942,3960,3987,4005],{"className":3941,"ariaHidden":52},[83],[23,3943,3945,3948,3951,3954,3957],{"className":3944},[87],[23,3946],{"className":3947,"style":260},[91],[23,3949,248],{"className":3950,"style":264},[100,101],[23,3952],{"className":3953,"style":848},[109],[23,3955,793],{"className":3956},[595],[23,3958],{"className":3959,"style":848},[109],[23,3961,3963,3966,3969,3972,3975,3978,3981,3984],{"className":3962},[87],[23,3964],{"className":3965,"style":1324},[91],[23,3967,1394],{"className":3968},[100],[23,3970,53],{"className":3971},[105],[23,3973],{"className":3974,"style":110},[109],[23,3976,280],{"className":3977,"style":117},[100,101],[23,3979],{"className":3980,"style":848},[109],[23,3982,793],{"className":3983},[595],[23,3985],{"className":3986,"style":848},[109],[23,3988,3990,3993,3996,3999,4002],{"className":3989},[87],[23,3991],{"className":3992,"style":861},[91],[23,3994,797],{"className":3995},[100],[23,3997],{"className":3998,"style":868},[109],[23,4000,800],{"className":4001},[872],[23,4003],{"className":4004,"style":868},[109],[23,4006,4008,4011,4014],{"className":4007},[87],[23,4009],{"className":4010,"style":906},[91],[23,4012,886],{"className":4013},[100],[23,4015,4017,4020],{"className":4016},[100],[23,4018,893],{"className":4019},[100],[23,4021,4023],{"className":4022},[121],[23,4024,4026],{"className":4025},[125],[23,4027,4029],{"className":4028},[130],[23,4030,4032],{"className":4031,"style":906},[134],[23,4033,4034,4037],{"style":909},[23,4035],{"className":4036,"style":143},[142],[23,4038,4040],{"className":4039},[147,148,149,150],[23,4041,4043,4046],{"className":4042},[100,150],[23,4044,811],{"className":4045},[100,150],[23,4047,814],{"className":4048},[100,150],[1054,4050,4051],{},[23,4052,4054,4073],{"className":4053},[26],[23,4055,4057],{"className":4056},[30],[32,4058,4059],{"xmlns":34},[36,4060,4061,4070],{},[39,4062,4063,4065,4067],{},[47,4064,312],{},[42,4066,793],{},[795,4068,4069],{},"250",[76,4071,4072],{"encoding":78},"N = 250",[23,4074,4076,4094],{"className":4075,"ariaHidden":52},[83],[23,4077,4079,4082,4085,4088,4091],{"className":4078},[87],[23,4080],{"className":4081,"style":324},[91],[23,4083,312],{"className":4084,"style":328},[100,101],[23,4086],{"className":4087,"style":848},[109],[23,4089,793],{"className":4090},[595],[23,4092],{"className":4093,"style":848},[109],[23,4095,4097,4100],{"className":4096},[87],[23,4098],{"className":4099,"style":952},[91],[23,4101,4069],{"className":4102},[100],[1054,4104,4105],{"align":969},[23,4106,4108,4128],{"className":4107},[26],[23,4109,4111],{"className":4110},[30],[32,4112,4113],{"xmlns":34},[36,4114,4115,4125],{},[39,4116,4117,4120,4122],{},[795,4118,4119],{},"0.6924",[42,4121,1132],{},[795,4123,4124],{},"0.0049",[76,4126,4127],{"encoding":78},"0.6924 \\pm 0.0049",[23,4129,4131,4149],{"className":4130,"ariaHidden":52},[83],[23,4132,4134,4137,4140,4143,4146],{"className":4133},[87],[23,4135],{"className":4136,"style":861},[91],[23,4138,4119],{"className":4139},[100],[23,4141],{"className":4142,"style":868},[109],[23,4144,1132],{"className":4145},[872],[23,4147],{"className":4148,"style":868},[109],[23,4150,4152,4155],{"className":4151},[87],[23,4153],{"className":4154,"style":952},[91],[23,4156,4124],{"className":4157},[100],[1054,4159,4160],{"align":969},[23,4161,4163,4186],{"className":4162},[26],[23,4164,4166],{"className":4165},[30],[32,4167,4168],{"xmlns":34},[36,4169,4170,4183],{},[39,4171,4172,4174,4176,4178,4181],{},[795,4173,2082],{},[47,4175,1244],{"mathvariant":1243},[42,4177,1132],{},[795,4179,4180],{},"1.31",[47,4182,1244],{"mathvariant":1243},[76,4184,4185],{"encoding":78},"27.83\\% \\pm 1.31\\%",[23,4187,4189,4207],{"className":4188,"ariaHidden":52},[83],[23,4190,4192,4195,4198,4201,4204],{"className":4191},[87],[23,4193],{"className":4194,"style":1264},[91],[23,4196,2107],{"className":4197},[100],[23,4199],{"className":4200,"style":868},[109],[23,4202,1132],{"className":4203},[872],[23,4205],{"className":4206,"style":868},[109],[23,4208,4210,4213],{"className":4209},[87],[23,4211],{"className":4212,"style":1284},[91],[23,4214,4216],{"className":4215},[100],"1.31%",[1054,4218,4219],{"align":969},[23,4220,4222,4239],{"className":4221},[26],[23,4223,4225],{"className":4224},[30],[32,4226,4227],{"xmlns":34},[36,4228,4229,4237],{},[39,4230,4231,4233,4235],{},[42,4232,1305],{},[795,4234,2145],{},[353,4236,1311],{},[76,4238,2150],{"encoding":78},[23,4240,4242],{"className":4241,"ariaHidden":52},[83],[23,4243,4245,4248,4251,4254],{"className":4244},[87],[23,4246],{"className":4247,"style":1324},[91],[23,4249,1305],{"className":4250},[100],[23,4252,2145],{"className":4253},[100],[23,4255,4257],{"className":4256},[100,401],[23,4258,1311],{"className":4259},[100],[964,4261,4262,4264,4318,4373,4434],{},[1054,4263],{},[1054,4265,4266],{},[23,4267,4269,4288],{"className":4268},[26],[23,4270,4272],{"className":4271},[30],[32,4273,4274],{"xmlns":34},[36,4275,4276,4285],{},[39,4277,4278,4280,4282],{},[47,4279,312],{},[42,4281,793],{},[795,4283,4284],{},"500",[76,4286,4287],{"encoding":78},"N = 500",[23,4289,4291,4309],{"className":4290,"ariaHidden":52},[83],[23,4292,4294,4297,4300,4303,4306],{"className":4293},[87],[23,4295],{"className":4296,"style":324},[91],[23,4298,312],{"className":4299,"style":328},[100,101],[23,4301],{"className":4302,"style":848},[109],[23,4304,793],{"className":4305},[595],[23,4307],{"className":4308,"style":848},[109],[23,4310,4312,4315],{"className":4311},[87],[23,4313],{"className":4314,"style":952},[91],[23,4316,4284],{"className":4317},[100],[1054,4319,4320],{"align":969},[23,4321,4323,4343],{"className":4322},[26],[23,4324,4326],{"className":4325},[30],[32,4327,4328],{"xmlns":34},[36,4329,4330,4340],{},[39,4331,4332,4335,4337],{},[795,4333,4334],{},"0.6694",[42,4336,1132],{},[795,4338,4339],{},"0.0029",[76,4341,4342],{"encoding":78},"0.6694 \\pm 0.0029",[23,4344,4346,4364],{"className":4345,"ariaHidden":52},[83],[23,4347,4349,4352,4355,4358,4361],{"className":4348},[87],[23,4350],{"className":4351,"style":861},[91],[23,4353,4334],{"className":4354},[100],[23,4356],{"className":4357,"style":868},[109],[23,4359,1132],{"className":4360},[872],[23,4362],{"className":4363,"style":868},[109],[23,4365,4367,4370],{"className":4366},[87],[23,4368],{"className":4369,"style":952},[91],[23,4371,4339],{"className":4372},[100],[1054,4374,4375],{"align":969},[23,4376,4378,4402],{"className":4377},[26],[23,4379,4381],{"className":4380},[30],[32,4382,4383],{"xmlns":34},[36,4384,4385,4399],{},[39,4386,4387,4390,4392,4394,4397],{},[795,4388,4389],{},"27.50",[47,4391,1244],{"mathvariant":1243},[42,4393,1132],{},[795,4395,4396],{},"2.48",[47,4398,1244],{"mathvariant":1243},[76,4400,4401],{"encoding":78},"27.50\\% \\pm 2.48\\%",[23,4403,4405,4424],{"className":4404,"ariaHidden":52},[83],[23,4406,4408,4411,4415,4418,4421],{"className":4407},[87],[23,4409],{"className":4410,"style":1264},[91],[23,4412,4414],{"className":4413},[100],"27.50%",[23,4416],{"className":4417,"style":868},[109],[23,4419,1132],{"className":4420},[872],[23,4422],{"className":4423,"style":868},[109],[23,4425,4427,4430],{"className":4426},[87],[23,4428],{"className":4429,"style":1284},[91],[23,4431,4433],{"className":4432},[100],"2.48%",[1054,4435,4436],{"align":969},[23,4437,4439,4458],{"className":4438},[26],[23,4440,4442],{"className":4441},[30],[32,4443,4444],{"xmlns":34},[36,4445,4446,4455],{},[39,4447,4448,4450,4453],{},[42,4449,1305],{},[795,4451,4452],{},"8.00",[353,4454,1311],{},[76,4456,4457],{"encoding":78},"+8.00\\text{ pp}",[23,4459,4461],{"className":4460,"ariaHidden":52},[83],[23,4462,4464,4467,4470,4473],{"className":4463},[87],[23,4465],{"className":4466,"style":1324},[91],[23,4468,1305],{"className":4469},[100],[23,4471,4452],{"className":4472},[100],[23,4474,4476],{"className":4475},[100,401],[23,4477,1311],{"className":4478},[100],[964,4480,4481,4483,4536,4588,4644],{},[1054,4482],{},[1054,4484,4485],{},[23,4486,4488,4506],{"className":4487},[26],[23,4489,4491],{"className":4490},[30],[32,4492,4493],{"xmlns":34},[36,4494,4495,4503],{},[39,4496,4497,4499,4501],{},[47,4498,312],{},[42,4500,793],{},[795,4502,826],{},[76,4504,4505],{"encoding":78},"N = 1000",[23,4507,4509,4527],{"className":4508,"ariaHidden":52},[83],[23,4510,4512,4515,4518,4521,4524],{"className":4511},[87],[23,4513],{"className":4514,"style":324},[91],[23,4516,312],{"className":4517,"style":328},[100,101],[23,4519],{"className":4520,"style":848},[109],[23,4522,793],{"className":4523},[595],[23,4525],{"className":4526,"style":848},[109],[23,4528,4530,4533],{"className":4529},[87],[23,4531],{"className":4532,"style":952},[91],[23,4534,826],{"className":4535},[100],[1054,4537,4538],{"align":969},[23,4539,4541,4558],{"className":4540},[26],[23,4542,4544],{"className":4543},[30],[32,4545,4546],{"xmlns":34},[36,4547,4548,4556],{},[39,4549,4550,4552,4554],{},[795,4551,1746],{},[42,4553,1132],{},[795,4555,1751],{},[76,4557,1754],{"encoding":78},[23,4559,4561,4579],{"className":4560,"ariaHidden":52},[83],[23,4562,4564,4567,4570,4573,4576],{"className":4563},[87],[23,4565],{"className":4566,"style":861},[91],[23,4568,1746],{"className":4569},[100],[23,4571],{"className":4572,"style":868},[109],[23,4574,1132],{"className":4575},[872],[23,4577],{"className":4578,"style":868},[109],[23,4580,4582,4585],{"className":4581},[87],[23,4583],{"className":4584,"style":952},[91],[23,4586,1751],{"className":4587},[100],[1054,4589,4590],{"align":969},[23,4591,4593,4614],{"className":4592},[26],[23,4594,4596],{"className":4595},[30],[32,4597,4598],{"xmlns":34},[36,4599,4600,4612],{},[39,4601,4602,4604,4606,4608,4610],{},[795,4603,1803],{},[47,4605,1244],{"mathvariant":1243},[42,4607,1132],{},[795,4609,1810],{},[47,4611,1244],{"mathvariant":1243},[76,4613,1815],{"encoding":78},[23,4615,4617,4635],{"className":4616,"ariaHidden":52},[83],[23,4618,4620,4623,4626,4629,4632],{"className":4619},[87],[23,4621],{"className":4622,"style":1264},[91],[23,4624,1828],{"className":4625},[100],[23,4627],{"className":4628,"style":868},[109],[23,4630,1132],{"className":4631},[872],[23,4633],{"className":4634,"style":868},[109],[23,4636,4638,4641],{"className":4637},[87],[23,4639],{"className":4640,"style":1284},[91],[23,4642,1847],{"className":4643},[100],[1054,4645,4646],{"align":969},[23,4647,4649,4666],{"className":4648},[26],[23,4650,4652],{"className":4651},[30],[32,4653,4654],{"xmlns":34},[36,4655,4656,4664],{},[39,4657,4658,4660,4662],{},[42,4659,1305],{},[795,4661,1868],{},[353,4663,1311],{},[76,4665,1873],{"encoding":78},[23,4667,4669],{"className":4668,"ariaHidden":52},[83],[23,4670,4672,4675,4678,4681],{"className":4671},[87],[23,4673],{"className":4674,"style":1324},[91],[23,4676,1305],{"className":4677},[100],[23,4679,1868],{"className":4680},[100],[23,4682,4684],{"className":4683},[100,401],[23,4685,1311],{"className":4686},[100],[964,4688,4689,4691,4747,4804,4867],{},[1054,4690],{},[1054,4692,4693],{},[14,4694,4695],{},[23,4696,4698,4717],{"className":4697},[26],[23,4699,4701],{"className":4700},[30],[32,4702,4703],{"xmlns":34},[36,4704,4705,4714],{},[39,4706,4707,4709,4711],{},[47,4708,312],{},[42,4710,793],{},[795,4712,4713],{},"2000",[76,4715,4716],{"encoding":78},"N = 2000",[23,4718,4720,4738],{"className":4719,"ariaHidden":52},[83],[23,4721,4723,4726,4729,4732,4735],{"className":4722},[87],[23,4724],{"className":4725,"style":324},[91],[23,4727,312],{"className":4728,"style":328},[100,101],[23,4730],{"className":4731,"style":848},[109],[23,4733,793],{"className":4734},[595],[23,4736],{"className":4737,"style":848},[109],[23,4739,4741,4744],{"className":4740},[87],[23,4742],{"className":4743,"style":952},[91],[23,4745,4713],{"className":4746},[100],[1054,4748,4749],{"align":969},[14,4750,4751],{},[23,4752,4754,4774],{"className":4753},[26],[23,4755,4757],{"className":4756},[30],[32,4758,4759],{"xmlns":34},[36,4760,4761,4771],{},[39,4762,4763,4766,4768],{},[795,4764,4765],{},"0.6540",[42,4767,1132],{},[795,4769,4770],{},"0.0011",[76,4772,4773],{"encoding":78},"0.6540 \\pm 0.0011",[23,4775,4777,4795],{"className":4776,"ariaHidden":52},[83],[23,4778,4780,4783,4786,4789,4792],{"className":4779},[87],[23,4781],{"className":4782,"style":861},[91],[23,4784,4765],{"className":4785},[100],[23,4787],{"className":4788,"style":868},[109],[23,4790,1132],{"className":4791},[872],[23,4793],{"className":4794,"style":868},[109],[23,4796,4798,4801],{"className":4797},[87],[23,4799],{"className":4800,"style":952},[91],[23,4802,4770],{"className":4803},[100],[1054,4805,4806],{"align":969},[14,4807,4808],{},[23,4809,4811,4835],{"className":4810},[26],[23,4812,4814],{"className":4813},[30],[32,4815,4816],{"xmlns":34},[36,4817,4818,4832],{},[39,4819,4820,4823,4825,4827,4830],{},[795,4821,4822],{},"32.23",[47,4824,1244],{"mathvariant":1243},[42,4826,1132],{},[795,4828,4829],{},"2.09",[47,4831,1244],{"mathvariant":1243},[76,4833,4834],{"encoding":78},"32.23\\% \\pm 2.09\\%",[23,4836,4838,4857],{"className":4837,"ariaHidden":52},[83],[23,4839,4841,4844,4848,4851,4854],{"className":4840},[87],[23,4842],{"className":4843,"style":1264},[91],[23,4845,4847],{"className":4846},[100],"32.23%",[23,4849],{"className":4850,"style":868},[109],[23,4852,1132],{"className":4853},[872],[23,4855],{"className":4856,"style":868},[109],[23,4858,4860,4863],{"className":4859},[87],[23,4861],{"className":4862,"style":1284},[91],[23,4864,4866],{"className":4865},[100],"2.09%",[1054,4868,4869],{"align":969},[14,4870,4871],{},[23,4872,4874,4893],{"className":4873},[26],[23,4875,4877],{"className":4876},[30],[32,4878,4879],{"xmlns":34},[36,4880,4881,4890],{},[39,4882,4883,4885,4888],{},[42,4884,1305],{},[795,4886,4887],{},"12.73",[353,4889,1311],{},[76,4891,4892],{"encoding":78},"+12.73\\text{ pp}",[23,4894,4896],{"className":4895,"ariaHidden":52},[83],[23,4897,4899,4902,4905,4908],{"className":4898},[87],[23,4900],{"className":4901,"style":1324},[91],[23,4903,1305],{"className":4904},[100],[23,4906,4887],{"className":4907},[100],[23,4909,4911],{"className":4910},[100,401],[23,4912,1311],{"className":4913},[100],[10,4915,4916,4917,2660],{},"数据清晰表明：",[14,4918,4919,4920,4971,4972,5013],{},"扩大高质量推理轨迹的多样性（",[23,4921,4923,4941],{"className":4922},[26],[23,4924,4926],{"className":4925},[30],[32,4927,4928],{"xmlns":34},[36,4929,4930,4938],{},[39,4931,4932,4934,4936],{},[47,4933,312],{},[42,4935,793],{},[795,4937,4713],{},[76,4939,4940],{"encoding":78},"N=2000",[23,4942,4944,4962],{"className":4943,"ariaHidden":52},[83],[23,4945,4947,4950,4953,4956,4959],{"className":4946},[87],[23,4948],{"className":4949,"style":324},[91],[23,4951,312],{"className":4952,"style":328},[100,101],[23,4954],{"className":4955,"style":848},[109],[23,4957,793],{"className":4958},[595],[23,4960],{"className":4961,"style":848},[109],[23,4963,4965,4968],{"className":4964},[87],[23,4966],{"className":4967,"style":952},[91],[23,4969,4713],{"className":4970},[100]," 带来 ",[23,4973,4975,4992],{"className":4974},[26],[23,4976,4978],{"className":4977},[30],[32,4979,4980],{"xmlns":34},[36,4981,4982,4990],{},[39,4983,4984,4986,4988],{},[42,4985,1305],{},[795,4987,4887],{},[353,4989,1311],{},[76,4991,4892],{"encoding":78},[23,4993,4995],{"className":4994,"ariaHidden":52},[83],[23,4996,4998,5001,5004,5007],{"className":4997},[87],[23,4999],{"className":5000,"style":1324},[91],[23,5002,1305],{"className":5003},[100],[23,5005,4887],{"className":5006},[100],[23,5008,5010],{"className":5009},[100,401],[23,5011,1311],{"className":5012},[100]," 提升）远比在固定小数据集上微调 LoRA Rank 更能突破泛化上限",[726,5015],{},[729,5017,5019],{"id":5018},"二第二阶段基于动态-adapter-切换的零冗余-dpo-实现","二、第二阶段：基于动态 Adapter 切换的零冗余 DPO 实现",[10,5021,5022,5023,5101],{},"在获得 SFT 模型 ",[23,5024,5026,5045],{"className":5025},[26],[23,5027,5029],{"className":5028},[30],[32,5030,5031],{"xmlns":34},[36,5032,5033,5042],{},[39,5034,5035],{},[55,5036,5037,5039],{},[47,5038,351],{},[353,5040,5041],{},"SFT",[76,5043,5044],{"encoding":78},"\\pi_{\\text{SFT}}",[23,5046,5048],{"className":5047,"ariaHidden":52},[83],[23,5049,5051,5054],{"className":5050},[87],[23,5052],{"className":5053,"style":368},[91],[23,5055,5057,5060],{"className":5056},[100],[23,5058,351],{"className":5059,"style":117},[100,101],[23,5061,5063],{"className":5062},[121],[23,5064,5066,5093],{"className":5065},[125,126],[23,5067,5069,5090],{"className":5068},[130],[23,5070,5073],{"className":5071,"style":5072},[134],"height:0.3283em;",[23,5074,5075,5078],{"style":138},[23,5076],{"className":5077,"style":143},[142],[23,5079,5081],{"className":5080},[147,148,149,150],[23,5082,5084],{"className":5083},[100,150],[23,5085,5087],{"className":5086},[100,401,150],[23,5088,5041],{"className":5089},[100,150],[23,5091,159],{"className":5092},[158],[23,5094,5096],{"className":5095},[130],[23,5097,5099],{"className":5098,"style":166},[134],[23,5100],{}," 后，标准 Bradley-Terry DPO 的损失函数定义为：",[23,5103,5106],{"className":5104},[5105],"katex-display",[23,5107,5109,5317],{"className":5108},[26],[23,5110,5112],{"className":5111},[30],[32,5113,5115],{"xmlns":34,"display":5114},"block",[36,5116,5117,5314],{},[39,5118,5119,5128,5130,5136,5139,5145,5147,5149,5151,5186],{},[55,5120,5121,5125],{},[47,5122,5124],{"mathvariant":5123},"script","L",[353,5126,5127],{},"DPO",[42,5129,45],{"stretchy":44},[55,5131,5132,5134],{},[47,5133,351],{},[47,5135,446],{},[42,5137,5138],{"separator":52},";",[55,5140,5141,5143],{},[47,5142,351],{},[353,5144,355],{},[42,5146,74],{"stretchy":44},[42,5148,793],{},[42,5150,811],{},[55,5152,5153,5157],{},[47,5154,5156],{"mathvariant":5155},"double-struck","E",[39,5158,5159,5161,5163,5165,5171,5173,5179,5181,5183],{},[42,5160,45],{"stretchy":44},[47,5162,49],{},[42,5164,53],{"separator":52},[55,5166,5167,5169],{},[47,5168,59],{},[47,5170,62],{},[42,5172,53],{"separator":52},[55,5174,5175,5177],{},[47,5176,59],{},[47,5178,71],{},[42,5180,74],{"stretchy":44},[42,5182,2478],{},[47,5184,5185],{"mathvariant":5123},"D",[39,5187,5188,5191,5193,5195,5198,5311],{},[42,5189,5190],{"fence":52},"[",[47,5192,431],{},[42,5194,434],{},[47,5196,5197],{},"σ",[39,5199,5200,5202,5205,5207,5209,5255,5257,5259,5261,5263,5309],{},[42,5201,45],{"fence":52},[47,5203,5204],{},"β",[47,5206,431],{},[42,5208,434],{},[436,5210,5211,5233],{},[39,5212,5213,5219,5221,5227,5229,5231],{},[55,5214,5215,5217],{},[47,5216,351],{},[47,5218,446],{},[42,5220,45],{"stretchy":44},[55,5222,5223,5225],{},[47,5224,59],{},[47,5226,62],{},[42,5228,453],{},[47,5230,49],{},[42,5232,74],{"stretchy":44},[39,5234,5235,5241,5243,5249,5251,5253],{},[55,5236,5237,5239],{},[47,5238,351],{},[353,5240,355],{},[42,5242,45],{"stretchy":44},[55,5244,5245,5247],{},[47,5246,59],{},[47,5248,62],{},[42,5250,453],{},[47,5252,49],{},[42,5254,74],{"stretchy":44},[42,5256,811],{},[47,5258,5204],{},[47,5260,431],{},[42,5262,434],{},[436,5264,5265,5287],{},[39,5266,5267,5273,5275,5281,5283,5285],{},[55,5268,5269,5271],{},[47,5270,351],{},[47,5272,446],{},[42,5274,45],{"stretchy":44},[55,5276,5277,5279],{},[47,5278,59],{},[47,5280,71],{},[42,5282,453],{},[47,5284,49],{},[42,5286,74],{"stretchy":44},[39,5288,5289,5295,5297,5303,5305,5307],{},[55,5290,5291,5293],{},[47,5292,351],{},[353,5294,355],{},[42,5296,45],{"stretchy":44},[55,5298,5299,5301],{},[47,5300,59],{},[47,5302,71],{},[42,5304,453],{},[47,5306,49],{},[42,5308,74],{"stretchy":44},[42,5310,74],{"fence":52},[42,5312,5313],{"fence":52},"]",[76,5315,5316],{"encoding":78},"\\mathcal{L}_{\\text{DPO}}(\\pi_\\theta; \\pi_{\\text{ref}}) = -\\mathbb{E}_{(x, y_w, y_l) \\sim \\mathcal{D}} \\left[ \\log \\sigma \\left( \\beta \\log \\frac{\\pi_\\theta(y_w \\mid x)}{\\pi_{\\text{ref}}(y_w \\mid x)} - \\beta \\log \\frac{\\pi_\\theta(y_l \\mid x)}{\\pi_{\\text{ref}}(y_l \\mid x)} \\right) \\right]",[23,5318,5320,5481],{"className":5319,"ariaHidden":52},[83],[23,5321,5323,5326,5374,5377,5417,5420,5423,5469,5472,5475,5478],{"className":5322},[87],[23,5324],{"className":5325,"style":92},[91],[23,5327,5329,5333],{"className":5328},[100],[23,5330,5124],{"className":5331},[100,5332],"mathcal",[23,5334,5336],{"className":5335},[121],[23,5337,5339,5366],{"className":5338},[125,126],[23,5340,5342,5363],{"className":5341},[130],[23,5343,5345],{"className":5344,"style":5072},[134],[23,5346,5348,5351],{"style":5347},"top:-2.55em;margin-left:0em;margin-right:0.05em;",[23,5349],{"className":5350,"style":143},[142],[23,5352,5354],{"className":5353},[147,148,149,150],[23,5355,5357],{"className":5356},[100,150],[23,5358,5360],{"className":5359},[100,401,150],[23,5361,5127],{"className":5362},[100,150],[23,5364,159],{"className":5365},[158],[23,5367,5369],{"className":5368},[130],[23,5370,5372],{"className":5371,"style":166},[134],[23,5373],{},[23,5375,45],{"className":5376},[96],[23,5378,5380,5383],{"className":5379},[100],[23,5381,351],{"className":5382,"style":117},[100,101],[23,5384,5386],{"className":5385},[121],[23,5387,5389,5409],{"className":5388},[125,126],[23,5390,5392,5406],{"className":5391},[130],[23,5393,5395],{"className":5394,"style":193},[134],[23,5396,5397,5400],{"style":138},[23,5398],{"className":5399,"style":143},[142],[23,5401,5403],{"className":5402},[147,148,149,150],[23,5404,446],{"className":5405,"style":264},[100,101,150],[23,5407,159],{"className":5408},[158],[23,5410,5412],{"className":5411},[130],[23,5413,5415],{"className":5414,"style":166},[134],[23,5416],{},[23,5418,5138],{"className":5419},[105],[23,5421],{"className":5422,"style":110},[109],[23,5424,5426,5429],{"className":5425},[100],[23,5427,351],{"className":5428,"style":117},[100,101],[23,5430,5432],{"className":5431},[121],[23,5433,5435,5461],{"className":5434},[125,126],[23,5436,5438,5458],{"className":5437},[130],[23,5439,5441],{"className":5440,"style":193},[134],[23,5442,5443,5446],{"style":138},[23,5444],{"className":5445,"style":143},[142],[23,5447,5449],{"className":5448},[147,148,149,150],[23,5450,5452],{"className":5451},[100,150],[23,5453,5455],{"className":5454},[100,401,150],[23,5456,355],{"className":5457},[100,150],[23,5459,159],{"className":5460},[158],[23,5462,5464],{"className":5463},[130],[23,5465,5467],{"className":5466,"style":166},[134],[23,5468],{},[23,5470,74],{"className":5471},[220],[23,5473],{"className":5474,"style":848},[109],[23,5476,793],{"className":5477},[595],[23,5479],{"className":5480,"style":848},[109],[23,5482,5484,5488,5491,5638,5641],{"className":5483},[87],[23,5485],{"className":5486,"style":5487},[91],"height:2.4em;vertical-align:-0.95em;",[23,5489,811],{"className":5490},[100],[23,5492,5494,5498],{"className":5493},[100],[23,5495,5156],{"className":5496},[100,5497],"mathbb",[23,5499,5501],{"className":5500},[121],[23,5502,5504,5629],{"className":5503},[125,126],[23,5505,5507,5626],{"className":5506},[130],[23,5508,5510],{"className":5509,"style":552},[134],[23,5511,5513,5516],{"style":5512},"top:-2.5198em;margin-left:0em;margin-right:0.05em;",[23,5514],{"className":5515,"style":143},[142],[23,5517,5519],{"className":5518},[147,148,149,150],[23,5520,5522,5525,5528,5531,5574,5577,5617,5620,5623],{"className":5521},[100,150],[23,5523,45],{"className":5524},[96,150],[23,5526,49],{"className":5527},[100,101,150],[23,5529,53],{"className":5530},[105,150],[23,5532,5534,5537],{"className":5533},[100,150],[23,5535,59],{"className":5536,"style":117},[100,101,150],[23,5538,5540],{"className":5539},[121],[23,5541,5543,5565],{"className":5542},[125,126],[23,5544,5546,5562],{"className":5545},[130],[23,5547,5550],{"className":5548,"style":5549},[134],"height:0.1645em;",[23,5551,5553,5556],{"style":5552},"top:-2.357em;margin-left:-0.0359em;margin-right:0.0714em;",[23,5554],{"className":5555,"style":559},[142],[23,5557,5559],{"className":5558},[147,563,564,150],[23,5560,62],{"className":5561,"style":154},[100,101,150],[23,5563,159],{"className":5564},[158],[23,5566,5568],{"className":5567},[130],[23,5569,5572],{"className":5570,"style":5571},[134],"height:0.143em;",[23,5573],{},[23,5575,53],{"className":5576},[105,150],[23,5578,5580,5583],{"className":5579},[100,150],[23,5581,59],{"className":5582,"style":117},[100,101,150],[23,5584,5586],{"className":5585},[121],[23,5587,5589,5609],{"className":5588},[125,126],[23,5590,5592,5606],{"className":5591},[130],[23,5593,5595],{"className":5594,"style":552},[134],[23,5596,5597,5600],{"style":555},[23,5598],{"className":5599,"style":559},[142],[23,5601,5603],{"className":5602},[147,563,564,150],[23,5604,71],{"className":5605,"style":205},[100,101,150],[23,5607,159],{"className":5608},[158],[23,5610,5612],{"className":5611},[130],[23,5613,5615],{"className":5614,"style":583},[134],[23,5616],{},[23,5618,74],{"className":5619},[220,150],[23,5621,2478],{"className":5622},[595,150],[23,5624,5185],{"className":5625,"style":264},[100,5332,150],[23,5627,159],{"className":5628},[158],[23,5630,5632],{"className":5631},[130],[23,5633,5636],{"className":5634,"style":5635},[134],"height:0.3552em;",[23,5637],{},[23,5639],{"className":5640,"style":110},[109],[23,5642,5645,5654,5659,5662,5665,5668,6241],{"className":5643},[5644],"minner",[23,5646,5650],{"className":5647,"style":5649},[96,5648],"delimcenter","top:0em;",[23,5651,5190],{"className":5652},[5653,149],"delimsizing",[23,5655,493,5657],{"className":5656},[492],[23,5658,497],{"style":496},[23,5660],{"className":5661,"style":110},[109],[23,5663,5197],{"className":5664,"style":117},[100,101],[23,5666],{"className":5667,"style":110},[109],[23,5669,5671,5677,5681,5684,5689,5692,5954,5957,5960,5963,5966,5969,5974,5977,6235],{"className":5670},[5644],[23,5672,5674],{"className":5673,"style":5649},[96,5648],[23,5675,45],{"className":5676},[5653,149],[23,5678,5204],{"className":5679,"style":5680},[100,101],"margin-right:0.0528em;",[23,5682],{"className":5683,"style":110},[109],[23,5685,493,5687],{"className":5686},[492],[23,5688,497],{"style":496},[23,5690],{"className":5691,"style":110},[109],[23,5693,5695,5698,5951],{"className":5694},[100],[23,5696],{"className":5697},[96,507],[23,5699,5701],{"className":5700},[436],[23,5702,5704,5942],{"className":5703},[125,126],[23,5705,5707,5939],{"className":5706},[130],[23,5708,5711,5824,5832],{"className":5709,"style":5710},[134],"height:1.427em;",[23,5712,5714,5717],{"style":5713},"top:-2.314em;",[23,5715],{"className":5716,"style":527},[142],[23,5718,5720,5766,5769,5809,5812,5815,5818,5821],{"className":5719},[100],[23,5721,5723,5726],{"className":5722},[100],[23,5724,351],{"className":5725,"style":117},[100,101],[23,5727,5729],{"className":5728},[121],[23,5730,5732,5758],{"className":5731},[125,126],[23,5733,5735,5755],{"className":5734},[130],[23,5736,5738],{"className":5737,"style":193},[134],[23,5739,5740,5743],{"style":138},[23,5741],{"className":5742,"style":143},[142],[23,5744,5746],{"className":5745},[147,148,149,150],[23,5747,5749],{"className":5748},[100,150],[23,5750,5752],{"className":5751},[100,401,150],[23,5753,355],{"className":5754},[100,150],[23,5756,159],{"className":5757},[158],[23,5759,5761],{"className":5760},[130],[23,5762,5764],{"className":5763,"style":166},[134],[23,5765],{},[23,5767,45],{"className":5768},[96],[23,5770,5772,5775],{"className":5771},[100],[23,5773,59],{"className":5774,"style":117},[100,101],[23,5776,5778],{"className":5777},[121],[23,5779,5781,5801],{"className":5780},[125,126],[23,5782,5784,5798],{"className":5783},[130],[23,5785,5787],{"className":5786,"style":135},[134],[23,5788,5789,5792],{"style":138},[23,5790],{"className":5791,"style":143},[142],[23,5793,5795],{"className":5794},[147,148,149,150],[23,5796,62],{"className":5797,"style":154},[100,101,150],[23,5799,159],{"className":5800},[158],[23,5802,5804],{"className":5803},[130],[23,5805,5807],{"className":5806,"style":166},[134],[23,5808],{},[23,5810],{"className":5811,"style":848},[109],[23,5813,453],{"className":5814},[595],[23,5816],{"className":5817,"style":848},[109],[23,5819,49],{"className":5820},[100,101],[23,5822,74],{"className":5823},[220],[23,5825,5826,5829],{"style":604},[23,5827],{"className":5828,"style":527},[142],[23,5830],{"className":5831,"style":612},[611],[23,5833,5835,5838],{"style":5834},"top:-3.677em;",[23,5836],{"className":5837,"style":527},[142],[23,5839,5841,5881,5884,5924,5927,5930,5933,5936],{"className":5840},[100],[23,5842,5844,5847],{"className":5843},[100],[23,5845,351],{"className":5846,"style":117},[100,101],[23,5848,5850],{"className":5849},[121],[23,5851,5853,5873],{"className":5852},[125,126],[23,5854,5856,5870],{"className":5855},[130],[23,5857,5859],{"className":5858,"style":193},[134],[23,5860,5861,5864],{"style":138},[23,5862],{"className":5863,"style":143},[142],[23,5865,5867],{"className":5866},[147,148,149,150],[23,5868,446],{"className":5869,"style":264},[100,101,150],[23,5871,159],{"className":5872},[158],[23,5874,5876],{"className":5875},[130],[23,5877,5879],{"className":5878,"style":166},[134],[23,5880],{},[23,5882,45],{"className":5883},[96],[23,5885,5887,5890],{"className":5886},[100],[23,5888,59],{"className":5889,"style":117},[100,101],[23,5891,5893],{"className":5892},[121],[23,5894,5896,5916],{"className":5895},[125,126],[23,5897,5899,5913],{"className":5898},[130],[23,5900,5902],{"className":5901,"style":135},[134],[23,5903,5904,5907],{"style":138},[23,5905],{"className":5906,"style":143},[142],[23,5908,5910],{"className":5909},[147,148,149,150],[23,5911,62],{"className":5912,"style":154},[100,101,150],[23,5914,159],{"className":5915},[158],[23,5917,5919],{"className":5918},[130],[23,5920,5922],{"className":5921,"style":166},[134],[23,5923],{},[23,5925],{"className":5926,"style":848},[109],[23,5928,453],{"className":5929},[595],[23,5931],{"className":5932,"style":848},[109],[23,5934,49],{"className":5935},[100,101],[23,5937,74],{"className":5938},[220],[23,5940,159],{"className":5941},[158],[23,5943,5945],{"className":5944},[130],[23,5946,5949],{"className":5947,"style":5948},[134],"height:0.936em;",[23,5950],{},[23,5952],{"className":5953},[220,507],[23,5955],{"className":5956,"style":868},[109],[23,5958,811],{"className":5959},[872],[23,5961],{"className":5962,"style":868},[109],[23,5964,5204],{"className":5965,"style":5680},[100,101],[23,5967],{"className":5968,"style":110},[109],[23,5970,493,5972],{"className":5971},[492],[23,5973,497],{"style":496},[23,5975],{"className":5976,"style":110},[109],[23,5978,5980,5983,6232],{"className":5979},[100],[23,5981],{"className":5982},[96,507],[23,5984,5986],{"className":5985},[436],[23,5987,5989,6224],{"className":5988},[125,126],[23,5990,5992,6221],{"className":5991},[130],[23,5993,5995,6107,6115],{"className":5994,"style":5710},[134],[23,5996,5997,6000],{"style":5713},[23,5998],{"className":5999,"style":527},[142],[23,6001,6003,6049,6052,6092,6095,6098,6101,6104],{"className":6002},[100],[23,6004,6006,6009],{"className":6005},[100],[23,6007,351],{"className":6008,"style":117},[100,101],[23,6010,6012],{"className":6011},[121],[23,6013,6015,6041],{"className":6014},[125,126],[23,6016,6018,6038],{"className":6017},[130],[23,6019,6021],{"className":6020,"style":193},[134],[23,6022,6023,6026],{"style":138},[23,6024],{"className":6025,"style":143},[142],[23,6027,6029],{"className":6028},[147,148,149,150],[23,6030,6032],{"className":6031},[100,150],[23,6033,6035],{"className":6034},[100,401,150],[23,6036,355],{"className":6037},[100,150],[23,6039,159],{"className":6040},[158],[23,6042,6044],{"className":6043},[130],[23,6045,6047],{"className":6046,"style":166},[134],[23,6048],{},[23,6050,45],{"className":6051},[96],[23,6053,6055,6058],{"className":6054},[100],[23,6056,59],{"className":6057,"style":117},[100,101],[23,6059,6061],{"className":6060},[121],[23,6062,6064,6084],{"className":6063},[125,126],[23,6065,6067,6081],{"className":6066},[130],[23,6068,6070],{"className":6069,"style":193},[134],[23,6071,6072,6075],{"style":138},[23,6073],{"className":6074,"style":143},[142],[23,6076,6078],{"className":6077},[147,148,149,150],[23,6079,71],{"className":6080,"style":205},[100,101,150],[23,6082,159],{"className":6083},[158],[23,6085,6087],{"className":6086},[130],[23,6088,6090],{"className":6089,"style":166},[134],[23,6091],{},[23,6093],{"className":6094,"style":848},[109],[23,6096,453],{"className":6097},[595],[23,6099],{"className":6100,"style":848},[109],[23,6102,49],{"className":6103},[100,101],[23,6105,74],{"className":6106},[220],[23,6108,6109,6112],{"style":604},[23,6110],{"className":6111,"style":527},[142],[23,6113],{"className":6114,"style":612},[611],[23,6116,6117,6120],{"style":5834},[23,6118],{"className":6119,"style":527},[142],[23,6121,6123,6163,6166,6206,6209,6212,6215,6218],{"className":6122},[100],[23,6124,6126,6129],{"className":6125},[100],[23,6127,351],{"className":6128,"style":117},[100,101],[23,6130,6132],{"className":6131},[121],[23,6133,6135,6155],{"className":6134},[125,126],[23,6136,6138,6152],{"className":6137},[130],[23,6139,6141],{"className":6140,"style":193},[134],[23,6142,6143,6146],{"style":138},[23,6144],{"className":6145,"style":143},[142],[23,6147,6149],{"className":6148},[147,148,149,150],[23,6150,446],{"className":6151,"style":264},[100,101,150],[23,6153,159],{"className":6154},[158],[23,6156,6158],{"className":6157},[130],[23,6159,6161],{"className":6160,"style":166},[134],[23,6162],{},[23,6164,45],{"className":6165},[96],[23,6167,6169,6172],{"className":6168},[100],[23,6170,59],{"className":6171,"style":117},[100,101],[23,6173,6175],{"className":6174},[121],[23,6176,6178,6198],{"className":6177},[125,126],[23,6179,6181,6195],{"className":6180},[130],[23,6182,6184],{"className":6183,"style":193},[134],[23,6185,6186,6189],{"style":138},[23,6187],{"className":6188,"style":143},[142],[23,6190,6192],{"className":6191},[147,148,149,150],[23,6193,71],{"className":6194,"style":205},[100,101,150],[23,6196,159],{"className":6197},[158],[23,6199,6201],{"className":6200},[130],[23,6202,6204],{"className":6203,"style":166},[134],[23,6205],{},[23,6207],{"className":6208,"style":848},[109],[23,6210,453],{"className":6211},[595],[23,6213],{"className":6214,"style":848},[109],[23,6216,49],{"className":6217},[100,101],[23,6219,74],{"className":6220},[220],[23,6222,159],{"className":6223},[158],[23,6225,6227],{"className":6226},[130],[23,6228,6230],{"className":6229,"style":5948},[134],[23,6231],{},[23,6233],{"className":6234},[220,507],[23,6236,6238],{"className":6237,"style":5649},[220,5648],[23,6239,74],{"className":6240},[5653,149],[23,6242,6244],{"className":6243,"style":5649},[220,5648],[23,6245,5313],{"className":6246},[5653,149],[10,6248,6249,6250,712,6253,6256,6257,6335,6336,6339],{},"如果按照朴素方式在内存中同时实例化 ",[702,6251,6252],{},"policy_model",[702,6254,6255],{},"ref_model"," 两个完整模型，显存占用会直接翻倍。由于我们的策略网络是在冻结基座上挂载 LoRA 适配器构成的，",[14,6258,6259,6260],{},"关闭 LoRA 适配器时的前向传播天然等价于参考模型 ",[23,6261,6263,6280],{"className":6262},[26],[23,6264,6266],{"className":6265},[30],[32,6267,6268],{"xmlns":34},[36,6269,6270,6278],{},[39,6271,6272],{},[55,6273,6274,6276],{},[47,6275,351],{},[353,6277,355],{},[76,6279,358],{"encoding":78},[23,6281,6283],{"className":6282,"ariaHidden":52},[83],[23,6284,6286,6289],{"className":6285},[87],[23,6287],{"className":6288,"style":368},[91],[23,6290,6292,6295],{"className":6291},[100],[23,6293,351],{"className":6294,"style":117},[100,101],[23,6296,6298],{"className":6297},[121],[23,6299,6301,6327],{"className":6300},[125,126],[23,6302,6304,6324],{"className":6303},[130],[23,6305,6307],{"className":6306,"style":193},[134],[23,6308,6309,6312],{"style":138},[23,6310],{"className":6311,"style":143},[142],[23,6313,6315],{"className":6314},[147,148,149,150],[23,6316,6318],{"className":6317},[100,150],[23,6319,6321],{"className":6320},[100,401,150],[23,6322,355],{"className":6323},[100,150],[23,6325,159],{"className":6326},[158],[23,6328,6330],{"className":6329},[130],[23,6331,6333],{"className":6332,"style":166},[134],[23,6334],{},"。因此，我们通过 ",[702,6337,6338],{},"disable_adapter()"," 上下文管理器在单份基座权重上实现了零冗余 DPO 训练循环：",[6341,6342,6347],"pre",{"className":6343,"code":6344,"language":6345,"meta":6346,"style":6346},"language-python shiki shiki-themes github-dark-dimmed github-light houston","import torch\nimport torch.nn.functional as F\n\ndef compute_response_logps(\n    logits: torch.Tensor,          # [B, T, V]\n    labels: torch.Tensor,          # [B, T], prompt 与 padding 位置已置为 -100\n) -> torch.Tensor:\n    \"\"\"仅对有效 Response Token 计算序列对数概率之和。\"\"\"\n    shift_logits = logits[:, :-1, :]\n    shift_labels = labels[:, 1:]\n    valid_mask = shift_labels != -100\n\n    safe_labels = torch.where(valid_mask, shift_labels, torch.zeros_like(shift_labels))\n    log_probs = F.log_softmax(shift_logits, dim=-1)\n    token_logps = torch.gather(log_probs, dim=-1, index=safe_labels.unsqueeze(-1)).squeeze(-1)\n    return (token_logps * valid_mask).sum(dim=-1)\n\ndef dpo_step_single_base(\n    peft_model,\n    chosen_ids: torch.Tensor,\n    chosen_labels: torch.Tensor,\n    rejected_ids: torch.Tensor,\n    rejected_labels: torch.Tensor,\n    beta: float = 0.1,\n) -> tuple[torch.Tensor, dict]:\n    # 1. 在 No-Grad + 关闭 LoRA 适配器状态下计算 Reference 对数概率（零额外基座显存）\n    with torch.no_grad(), peft_model.disable_adapter():\n        ref_chosen_logps = compute_response_logps(peft_model(chosen_ids).logits, chosen_labels)\n        ref_rejected_logps = compute_response_logps(peft_model(rejected_ids).logits, rejected_labels)\n\n    # 2. 启用 LoRA 适配器计算当前策略 Policy 对数概率\n    pi_chosen_logps = compute_response_logps(peft_model(chosen_ids).logits, chosen_labels)\n    pi_rejected_logps = compute_response_logps(peft_model(rejected_ids).logits, rejected_labels)\n\n    # 3. 计算隐式奖励差与 Bradley-Terry 偏好损失\n    chosen_rewards = beta * (pi_chosen_logps - ref_chosen_logps)\n    rejected_rewards = beta * (pi_rejected_logps - ref_rejected_logps)\n    reward_margin = chosen_rewards - rejected_rewards\n\n    loss = -F.logsigmoid(reward_margin).mean()\n    pref_acc = (reward_margin > 0).float().mean().item()\n    return loss, {\"loss\": loss.item(), \"reward_margin\": reward_margin.mean().item(), \"pref_acc\": pref_acc}\n","python","",[702,6348,6349,6361,6375,6382,6396,6410,6421,6427,6434,6455,6471,6491,6496,6520,6549,6605,6633,6638,6648,6657,6666,6674,6682,6690,6711,6723,6729,6749,6767,6784,6789,6795,6811,6827,6832,6838,6859,6879,6895,6900,6925,6959],{"__ignoreMap":6346},[23,6350,6353,6357],{"class":6351,"line":6352},"line",1,[23,6354,6356],{"class":6355},"slEfO","import",[23,6358,6360],{"class":6359},"sWURj"," torch\n",[23,6362,6364,6366,6369,6372],{"class":6351,"line":6363},2,[23,6365,6356],{"class":6355},[23,6367,6368],{"class":6359}," torch.nn.functional ",[23,6370,6371],{"class":6355},"as",[23,6373,6374],{"class":6359}," F\n",[23,6376,6378],{"class":6351,"line":6377},3,[23,6379,6381],{"emptyLinePlaceholder":6380},true,"\n",[23,6383,6385,6389,6393],{"class":6351,"line":6384},4,[23,6386,6388],{"class":6387},"sST-V","def",[23,6390,6392],{"class":6391},"sS_0i"," compute_response_logps",[23,6394,6395],{"class":6359},"(\n",[23,6397,6399,6403,6406],{"class":6351,"line":6398},5,[23,6400,6402],{"class":6401},"s9NhK","    logits",[23,6404,6405],{"class":6359},": torch.Tensor,          ",[23,6407,6409],{"class":6408},"sCxhR","# [B, T, V]\n",[23,6411,6413,6416,6418],{"class":6351,"line":6412},6,[23,6414,6415],{"class":6401},"    labels",[23,6417,6405],{"class":6359},[23,6419,6420],{"class":6408},"# [B, T], prompt 与 padding 位置已置为 -100\n",[23,6422,6424],{"class":6351,"line":6423},7,[23,6425,6426],{"class":6359},") -> torch.Tensor:\n",[23,6428,6430],{"class":6351,"line":6429},8,[23,6431,6433],{"class":6432},"sWqY2","    \"\"\"仅对有效 Response Token 计算序列对数概率之和。\"\"\"\n",[23,6435,6437,6440,6443,6446,6449,6452],{"class":6351,"line":6436},9,[23,6438,6439],{"class":6359},"    shift_logits ",[23,6441,793],{"class":6442},"sf2cM",[23,6444,6445],{"class":6359}," logits[:, :",[23,6447,6448],{"class":6442},"-",[23,6450,886],{"class":6451},"sgENW",[23,6453,6454],{"class":6359},", :]\n",[23,6456,6458,6461,6463,6466,6468],{"class":6351,"line":6457},10,[23,6459,6460],{"class":6359},"    shift_labels ",[23,6462,793],{"class":6442},[23,6464,6465],{"class":6359}," labels[:, ",[23,6467,886],{"class":6451},[23,6469,6470],{"class":6359},":]\n",[23,6472,6474,6477,6479,6482,6485,6488],{"class":6351,"line":6473},11,[23,6475,6476],{"class":6359},"    valid_mask ",[23,6478,793],{"class":6442},[23,6480,6481],{"class":6359}," shift_labels ",[23,6483,6484],{"class":6442},"!=",[23,6486,6487],{"class":6442}," -",[23,6489,6490],{"class":6451},"100\n",[23,6492,6494],{"class":6351,"line":6493},12,[23,6495,6381],{"emptyLinePlaceholder":6380},[23,6497,6499,6502,6504,6507,6511,6514,6517],{"class":6351,"line":6498},13,[23,6500,6501],{"class":6359},"    safe_labels ",[23,6503,793],{"class":6442},[23,6505,6506],{"class":6359}," torch.",[23,6508,6510],{"class":6509},"sq3c-","where",[23,6512,6513],{"class":6359},"(valid_mask, shift_labels, torch.",[23,6515,6516],{"class":6509},"zeros_like",[23,6518,6519],{"class":6359},"(shift_labels))\n",[23,6521,6523,6526,6528,6531,6534,6537,6541,6544,6546],{"class":6351,"line":6522},14,[23,6524,6525],{"class":6359},"    log_probs ",[23,6527,793],{"class":6442},[23,6529,6530],{"class":6359}," F.",[23,6532,6533],{"class":6509},"log_softmax",[23,6535,6536],{"class":6359},"(shift_logits, ",[23,6538,6540],{"class":6539},"sA9aF","dim",[23,6542,6543],{"class":6442},"=-",[23,6545,886],{"class":6451},[23,6547,6548],{"class":6359},")\n",[23,6550,6552,6555,6557,6559,6562,6565,6567,6569,6571,6574,6577,6579,6582,6585,6587,6589,6591,6594,6597,6599,6601,6603],{"class":6351,"line":6551},15,[23,6553,6554],{"class":6359},"    token_logps ",[23,6556,793],{"class":6442},[23,6558,6506],{"class":6359},[23,6560,6561],{"class":6509},"gather",[23,6563,6564],{"class":6359},"(log_probs, ",[23,6566,6540],{"class":6539},[23,6568,6543],{"class":6442},[23,6570,886],{"class":6451},[23,6572,6573],{"class":6359},", ",[23,6575,6576],{"class":6539},"index",[23,6578,793],{"class":6442},[23,6580,6581],{"class":6359},"safe_labels.",[23,6583,6584],{"class":6509},"unsqueeze",[23,6586,45],{"class":6359},[23,6588,6448],{"class":6442},[23,6590,886],{"class":6451},[23,6592,6593],{"class":6359},")).",[23,6595,6596],{"class":6509},"squeeze",[23,6598,45],{"class":6359},[23,6600,6448],{"class":6442},[23,6602,886],{"class":6451},[23,6604,6548],{"class":6359},[23,6606,6608,6611,6614,6617,6620,6623,6625,6627,6629,6631],{"class":6351,"line":6607},16,[23,6609,6610],{"class":6355},"    return",[23,6612,6613],{"class":6359}," (token_logps ",[23,6615,6616],{"class":6442},"*",[23,6618,6619],{"class":6359}," valid_mask).",[23,6621,6622],{"class":6509},"sum",[23,6624,45],{"class":6359},[23,6626,6540],{"class":6539},[23,6628,6543],{"class":6442},[23,6630,886],{"class":6451},[23,6632,6548],{"class":6359},[23,6634,6636],{"class":6351,"line":6635},17,[23,6637,6381],{"emptyLinePlaceholder":6380},[23,6639,6641,6643,6646],{"class":6351,"line":6640},18,[23,6642,6388],{"class":6387},[23,6644,6645],{"class":6391}," dpo_step_single_base",[23,6647,6395],{"class":6359},[23,6649,6651,6654],{"class":6351,"line":6650},19,[23,6652,6653],{"class":6401},"    peft_model",[23,6655,6656],{"class":6359},",\n",[23,6658,6660,6663],{"class":6351,"line":6659},20,[23,6661,6662],{"class":6401},"    chosen_ids",[23,6664,6665],{"class":6359},": torch.Tensor,\n",[23,6667,6669,6672],{"class":6351,"line":6668},21,[23,6670,6671],{"class":6401},"    chosen_labels",[23,6673,6665],{"class":6359},[23,6675,6677,6680],{"class":6351,"line":6676},22,[23,6678,6679],{"class":6401},"    rejected_ids",[23,6681,6665],{"class":6359},[23,6683,6685,6688],{"class":6351,"line":6684},23,[23,6686,6687],{"class":6401},"    rejected_labels",[23,6689,6665],{"class":6359},[23,6691,6693,6696,6699,6703,6706,6709],{"class":6351,"line":6692},24,[23,6694,6695],{"class":6401},"    beta",[23,6697,6698],{"class":6359},": ",[23,6700,6702],{"class":6701},"svkvy","float",[23,6704,6705],{"class":6442}," =",[23,6707,6708],{"class":6451}," 0.1",[23,6710,6656],{"class":6359},[23,6712,6714,6717,6720],{"class":6351,"line":6713},25,[23,6715,6716],{"class":6359},") -> tuple[torch.Tensor, ",[23,6718,6719],{"class":6701},"dict",[23,6721,6722],{"class":6359},"]:\n",[23,6724,6726],{"class":6351,"line":6725},26,[23,6727,6728],{"class":6408},"    # 1. 在 No-Grad + 关闭 LoRA 适配器状态下计算 Reference 对数概率（零额外基座显存）\n",[23,6730,6732,6735,6737,6740,6743,6746],{"class":6351,"line":6731},27,[23,6733,6734],{"class":6355},"    with",[23,6736,6506],{"class":6359},[23,6738,6739],{"class":6509},"no_grad",[23,6741,6742],{"class":6359},"(), peft_model.",[23,6744,6745],{"class":6509},"disable_adapter",[23,6747,6748],{"class":6359},"():\n",[23,6750,6752,6755,6757,6759,6761,6764],{"class":6351,"line":6751},28,[23,6753,6754],{"class":6359},"        ref_chosen_logps ",[23,6756,793],{"class":6442},[23,6758,6392],{"class":6509},[23,6760,45],{"class":6359},[23,6762,6763],{"class":6509},"peft_model",[23,6765,6766],{"class":6359},"(chosen_ids).logits, chosen_labels)\n",[23,6768,6770,6773,6775,6777,6779,6781],{"class":6351,"line":6769},29,[23,6771,6772],{"class":6359},"        ref_rejected_logps ",[23,6774,793],{"class":6442},[23,6776,6392],{"class":6509},[23,6778,45],{"class":6359},[23,6780,6763],{"class":6509},[23,6782,6783],{"class":6359},"(rejected_ids).logits, rejected_labels)\n",[23,6785,6787],{"class":6351,"line":6786},30,[23,6788,6381],{"emptyLinePlaceholder":6380},[23,6790,6792],{"class":6351,"line":6791},31,[23,6793,6794],{"class":6408},"    # 2. 启用 LoRA 适配器计算当前策略 Policy 对数概率\n",[23,6796,6798,6801,6803,6805,6807,6809],{"class":6351,"line":6797},32,[23,6799,6800],{"class":6359},"    pi_chosen_logps ",[23,6802,793],{"class":6442},[23,6804,6392],{"class":6509},[23,6806,45],{"class":6359},[23,6808,6763],{"class":6509},[23,6810,6766],{"class":6359},[23,6812,6814,6817,6819,6821,6823,6825],{"class":6351,"line":6813},33,[23,6815,6816],{"class":6359},"    pi_rejected_logps ",[23,6818,793],{"class":6442},[23,6820,6392],{"class":6509},[23,6822,45],{"class":6359},[23,6824,6763],{"class":6509},[23,6826,6783],{"class":6359},[23,6828,6830],{"class":6351,"line":6829},34,[23,6831,6381],{"emptyLinePlaceholder":6380},[23,6833,6835],{"class":6351,"line":6834},35,[23,6836,6837],{"class":6408},"    # 3. 计算隐式奖励差与 Bradley-Terry 偏好损失\n",[23,6839,6841,6844,6846,6849,6851,6854,6856],{"class":6351,"line":6840},36,[23,6842,6843],{"class":6359},"    chosen_rewards ",[23,6845,793],{"class":6442},[23,6847,6848],{"class":6359}," beta ",[23,6850,6616],{"class":6442},[23,6852,6853],{"class":6359}," (pi_chosen_logps ",[23,6855,6448],{"class":6442},[23,6857,6858],{"class":6359}," ref_chosen_logps)\n",[23,6860,6862,6865,6867,6869,6871,6874,6876],{"class":6351,"line":6861},37,[23,6863,6864],{"class":6359},"    rejected_rewards ",[23,6866,793],{"class":6442},[23,6868,6848],{"class":6359},[23,6870,6616],{"class":6442},[23,6872,6873],{"class":6359}," (pi_rejected_logps ",[23,6875,6448],{"class":6442},[23,6877,6878],{"class":6359}," ref_rejected_logps)\n",[23,6880,6882,6885,6887,6890,6892],{"class":6351,"line":6881},38,[23,6883,6884],{"class":6359},"    reward_margin ",[23,6886,793],{"class":6442},[23,6888,6889],{"class":6359}," chosen_rewards ",[23,6891,6448],{"class":6442},[23,6893,6894],{"class":6359}," rejected_rewards\n",[23,6896,6898],{"class":6351,"line":6897},39,[23,6899,6381],{"emptyLinePlaceholder":6380},[23,6901,6903,6906,6908,6910,6913,6916,6919,6922],{"class":6351,"line":6902},40,[23,6904,6905],{"class":6359},"    loss ",[23,6907,793],{"class":6442},[23,6909,6487],{"class":6442},[23,6911,6912],{"class":6359},"F.",[23,6914,6915],{"class":6509},"logsigmoid",[23,6917,6918],{"class":6359},"(reward_margin).",[23,6920,6921],{"class":6509},"mean",[23,6923,6924],{"class":6359},"()\n",[23,6926,6928,6931,6933,6936,6939,6942,6945,6947,6950,6952,6954,6957],{"class":6351,"line":6927},41,[23,6929,6930],{"class":6359},"    pref_acc ",[23,6932,793],{"class":6442},[23,6934,6935],{"class":6359}," (reward_margin ",[23,6937,6938],{"class":6442},">",[23,6940,6941],{"class":6451}," 0",[23,6943,6944],{"class":6359},").",[23,6946,6702],{"class":6509},[23,6948,6949],{"class":6359},"().",[23,6951,6921],{"class":6509},[23,6953,6949],{"class":6359},[23,6955,6956],{"class":6509},"item",[23,6958,6924],{"class":6359},[23,6960,6962,6964,6967,6970,6973,6975,6978,6981,6984,6986,6988,6990,6992,6995],{"class":6351,"line":6961},42,[23,6963,6610],{"class":6355},[23,6965,6966],{"class":6359}," loss, {",[23,6968,6969],{"class":6432},"\"loss\"",[23,6971,6972],{"class":6359},": loss.",[23,6974,6956],{"class":6509},[23,6976,6977],{"class":6359},"(), ",[23,6979,6980],{"class":6432},"\"reward_margin\"",[23,6982,6983],{"class":6359},": reward_margin.",[23,6985,6921],{"class":6509},[23,6987,6949],{"class":6359},[23,6989,6956],{"class":6509},[23,6991,6977],{"class":6359},[23,6993,6994],{"class":6432},"\"pref_acc\"",[23,6996,6997],{"class":6359},": pref_acc}\n",[726,6999],{},[729,7001,7003],{"id":7002},"三对齐税alignment-tax的实证观测与机理剖析","三、对齐税（Alignment Tax）的实证观测与机理剖析",[10,7005,7006,7007,7009,7010,7171,7172,747,7175,7178,7179,7182,7183,7186],{},"我们在 ",[702,7008,704],{}," 人类偏好数据集（500 对训练样本，100 对独立测试样本，",[23,7011,7013,7054],{"className":7012},[26],[23,7014,7016],{"className":7015},[30],[32,7017,7018],{"xmlns":34},[36,7019,7020,7051],{},[39,7021,7022,7024,7026,7029,7031,7033,7035,7037,7039,7041],{},[47,7023,5204],{},[42,7025,793],{},[795,7027,7028],{},"0.1",[42,7030,53],{"separator":52},[353,7032,819],{},[353,7034,790],{},[42,7036,793],{},[795,7038,2800],{},[42,7040,800],{},[802,7042,7043,7045],{},[795,7044,806],{},[39,7046,7047,7049],{},[42,7048,811],{},[795,7050,2800],{},[76,7052,7053],{"encoding":78},"\\beta = 0.1,\\; \\text{LR} = 5\\times 10^{-5}",[23,7055,7057,7076,7109,7127],{"className":7056,"ariaHidden":52},[83],[23,7058,7060,7064,7067,7070,7073],{"className":7059},[87],[23,7061],{"className":7062,"style":7063},[91],"height:0.8889em;vertical-align:-0.1944em;",[23,7065,5204],{"className":7066,"style":5680},[100,101],[23,7068],{"className":7069,"style":848},[109],[23,7071,793],{"className":7072},[595],[23,7074],{"className":7075,"style":848},[109],[23,7077,7079,7082,7085,7088,7091,7094,7100,7103,7106],{"className":7078},[87],[23,7080],{"className":7081,"style":2749},[91],[23,7083,7028],{"className":7084},[100],[23,7086,53],{"className":7087},[105],[23,7089],{"className":7090,"style":848},[109],[23,7092],{"className":7093,"style":110},[109],[23,7095,7097],{"className":7096},[100,401],[23,7098,790],{"className":7099},[100],[23,7101],{"className":7102,"style":848},[109],[23,7104,793],{"className":7105},[595],[23,7107],{"className":7108,"style":848},[109],[23,7110,7112,7115,7118,7121,7124],{"className":7111},[87],[23,7113],{"className":7114,"style":861},[91],[23,7116,2800],{"className":7117},[100],[23,7119],{"className":7120,"style":868},[109],[23,7122,800],{"className":7123},[872],[23,7125],{"className":7126,"style":868},[109],[23,7128,7130,7133,7136],{"className":7129},[87],[23,7131],{"className":7132,"style":906},[91],[23,7134,886],{"className":7135},[100],[23,7137,7139,7142],{"className":7138},[100],[23,7140,893],{"className":7141},[100],[23,7143,7145],{"className":7144},[121],[23,7146,7148],{"className":7147},[125],[23,7149,7151],{"className":7150},[130],[23,7152,7154],{"className":7153,"style":906},[134],[23,7155,7156,7159],{"style":909},[23,7157],{"className":7158,"style":143},[142],[23,7160,7162],{"className":7161},[147,148,149,150],[23,7163,7165,7168],{"className":7164},[100,150],[23,7166,811],{"className":7167},[100,150],[23,7169,2800],{"className":7170},[100,150],"，2 Epochs）上对 SFT 模型执行 DPO 对齐，并在训练前后同步追踪",[14,7173,7174],{},"偏好胜率（Preference Accuracy）",[14,7176,7177],{},"隐式奖励间隔（Reward Margin）","与 ",[14,7180,7181],{},"GSM8K 数学推理准确率","（实验产物见 ",[702,7184,7185],{},"experiments\u002Fdpo-alignment\u002Fresults.json","）：",[958,7188,7189,7211],{},[961,7190,7191],{},[964,7192,7193,7196,7199,7202,7205,7208],{},[967,7194,7195],{},"模型阶段",[967,7197,7198],{"align":969},"HH-RLHF 偏好测试准确率",[967,7200,7201],{"align":969},"平均隐式奖励间隔 (Margin)",[967,7203,7204],{"align":969},"Chosen 平均奖励",[967,7206,7207],{"align":969},"Rejected 平均奖励",[967,7209,7210],{"align":969},"GSM8K 准确率",[1049,7212,7213,7316,7421],{},[964,7214,7215,7220,7223,7254,7284,7314],{},[1054,7216,7217],{},[14,7218,7219],{},"Base 基座模型",[1054,7221,7222],{"align":969},"50.0%（随机基线）",[1054,7224,7225],{"align":969},[23,7226,7228,7242],{"className":7227},[26],[23,7229,7231],{"className":7230},[30],[32,7232,7233],{"xmlns":34},[36,7234,7235,7240],{},[39,7236,7237],{},[795,7238,7239],{},"0.0000",[76,7241,7239],{"encoding":78},[23,7243,7245],{"className":7244,"ariaHidden":52},[83],[23,7246,7248,7251],{"className":7247},[87],[23,7249],{"className":7250,"style":952},[91],[23,7252,7239],{"className":7253},[100],[1054,7255,7256],{"align":969},[23,7257,7259,7272],{"className":7258},[26],[23,7260,7262],{"className":7261},[30],[32,7263,7264],{"xmlns":34},[36,7265,7266,7270],{},[39,7267,7268],{},[795,7269,7239],{},[76,7271,7239],{"encoding":78},[23,7273,7275],{"className":7274,"ariaHidden":52},[83],[23,7276,7278,7281],{"className":7277},[87],[23,7279],{"className":7280,"style":952},[91],[23,7282,7239],{"className":7283},[100],[1054,7285,7286],{"align":969},[23,7287,7289,7302],{"className":7288},[26],[23,7290,7292],{"className":7291},[30],[32,7293,7294],{"xmlns":34},[36,7295,7296,7300],{},[39,7297,7298],{},[795,7299,7239],{},[76,7301,7239],{"encoding":78},[23,7303,7305],{"className":7304,"ariaHidden":52},[83],[23,7306,7308,7311],{"className":7307},[87],[23,7309],{"className":7310,"style":952},[91],[23,7312,7239],{"className":7313},[100],[1054,7315,765],{"align":969},[964,7317,7318,7323,7326,7356,7386,7416],{},[1054,7319,7320],{},[14,7321,7322],{},"Stage 1: SFT 模型",[1054,7324,7325],{"align":969},"50.0%",[1054,7327,7328],{"align":969},[23,7329,7331,7344],{"className":7330},[26],[23,7332,7334],{"className":7333},[30],[32,7335,7336],{"xmlns":34},[36,7337,7338,7342],{},[39,7339,7340],{},[795,7341,7239],{},[76,7343,7239],{"encoding":78},[23,7345,7347],{"className":7346,"ariaHidden":52},[83],[23,7348,7350,7353],{"className":7349},[87],[23,7351],{"className":7352,"style":952},[91],[23,7354,7239],{"className":7355},[100],[1054,7357,7358],{"align":969},[23,7359,7361,7374],{"className":7360},[26],[23,7362,7364],{"className":7363},[30],[32,7365,7366],{"xmlns":34},[36,7367,7368,7372],{},[39,7369,7370],{},[795,7371,7239],{},[76,7373,7239],{"encoding":78},[23,7375,7377],{"className":7376,"ariaHidden":52},[83],[23,7378,7380,7383],{"className":7379},[87],[23,7381],{"className":7382,"style":952},[91],[23,7384,7239],{"className":7385},[100],[1054,7387,7388],{"align":969},[23,7389,7391,7404],{"className":7390},[26],[23,7392,7394],{"className":7393},[30],[32,7395,7396],{"xmlns":34},[36,7397,7398,7402],{},[39,7399,7400],{},[795,7401,7239],{},[76,7403,7239],{"encoding":78},[23,7405,7407],{"className":7406,"ariaHidden":52},[83],[23,7408,7410,7413],{"className":7409},[87],[23,7411],{"className":7412,"style":952},[91],[23,7414,7239],{"className":7415},[100],[1054,7417,7418],{"align":969},[14,7419,7420],{},"30.50%",[964,7422,7423,7428,7433,7472,7509,7546],{},[1054,7424,7425],{},[14,7426,7427],{},"Stage 2: SFT + DPO 模型",[1054,7429,7430],{"align":969},[14,7431,7432],{},"62.0%（+12.0 pp）",[1054,7434,7435],{"align":969},[14,7436,7437],{},[23,7438,7440,7457],{"className":7439},[26],[23,7441,7443],{"className":7442},[30],[32,7444,7445],{"xmlns":34},[36,7446,7447,7454],{},[39,7448,7449,7451],{},[42,7450,1305],{},[795,7452,7453],{},"0.1945",[76,7455,7456],{"encoding":78},"+0.1945",[23,7458,7460],{"className":7459,"ariaHidden":52},[83],[23,7461,7463,7466,7469],{"className":7462},[87],[23,7464],{"className":7465,"style":861},[91],[23,7467,1305],{"className":7468},[100],[23,7470,7453],{"className":7471},[100],[1054,7473,7474],{"align":969},[23,7475,7477,7494],{"className":7476},[26],[23,7478,7480],{"className":7479},[30],[32,7481,7482],{"xmlns":34},[36,7483,7484,7491],{},[39,7485,7486,7488],{},[42,7487,811],{},[795,7489,7490],{},"0.7793",[76,7492,7493],{"encoding":78},"-0.7793",[23,7495,7497],{"className":7496,"ariaHidden":52},[83],[23,7498,7500,7503,7506],{"className":7499},[87],[23,7501],{"className":7502,"style":861},[91],[23,7504,811],{"className":7505},[100],[23,7507,7490],{"className":7508},[100],[1054,7510,7511],{"align":969},[23,7512,7514,7531],{"className":7513},[26],[23,7515,7517],{"className":7516},[30],[32,7518,7519],{"xmlns":34},[36,7520,7521,7528],{},[39,7522,7523,7525],{},[42,7524,811],{},[795,7526,7527],{},"0.9738",[76,7529,7530],{"encoding":78},"-0.9738",[23,7532,7534],{"className":7533,"ariaHidden":52},[83],[23,7535,7537,7540,7543],{"className":7536},[87],[23,7538],{"className":7539,"style":861},[91],[23,7541,811],{"className":7542},[100],[23,7544,7527],{"className":7545},[100],[1054,7547,7548],{"align":969},[14,7549,7550],{},"27.50%（-3.00 pp）",[10,7552,7553],{},"这组受控对照清晰呈现了后训练工程中的两个核心现象：",[226,7555,7556,7821],{},[229,7557,7558,7561,7562,7595,7596,7806,7807,7809,7810,7812,7813,7816,7817,7820],{},[14,7559,7560],{},"DPO 期间 Chosen 与 Rejected 对数概率同步下移（Likelihood Displacement）","：观察表格可以发现，经过 DPO 后，虽然 Chosen 相对 Rejected 拉开了 ",[23,7563,7565,7580],{"className":7564},[26],[23,7566,7568],{"className":7567},[30],[32,7569,7570],{"xmlns":34},[36,7571,7572,7578],{},[39,7573,7574,7576],{},[42,7575,1305],{},[795,7577,7453],{},[76,7579,7456],{"encoding":78},[23,7581,7583],{"className":7582,"ariaHidden":52},[83],[23,7584,7586,7589,7592],{"className":7585},[87],[23,7587],{"className":7588,"style":861},[91],[23,7590,1305],{"className":7591},[100],[23,7593,7453],{"className":7594},[100]," 的显著正向 Margin（偏好判断准确率从 50.0% 跃升至 62.0%），但两者的绝对隐式奖励 ",[23,7597,7599,7631],{"className":7598},[26],[23,7600,7602],{"className":7601},[30],[32,7603,7604],{"xmlns":34},[36,7605,7606,7628],{},[39,7607,7608,7610,7612,7614],{},[47,7609,5204],{},[47,7611,431],{},[42,7613,434],{},[436,7615,7616,7622],{},[55,7617,7618,7620],{},[47,7619,351],{},[47,7621,446],{},[55,7623,7624,7626],{},[47,7625,351],{},[353,7627,355],{},[76,7629,7630],{"encoding":78},"\\beta \\log \\frac{\\pi_\\theta}{\\pi_{\\text{ref}}}",[23,7632,7634],{"className":7633,"ariaHidden":52},[83],[23,7635,7637,7641,7644,7647,7652,7655],{"className":7636},[87],[23,7638],{"className":7639,"style":7640},[91],"height:1.1681em;vertical-align:-0.4509em;",[23,7642,5204],{"className":7643,"style":5680},[100,101],[23,7645],{"className":7646,"style":110},[109],[23,7648,493,7650],{"className":7649},[492],[23,7651,497],{"style":496},[23,7653],{"className":7654,"style":110},[109],[23,7656,7658,7661,7803],{"className":7657},[100],[23,7659],{"className":7660},[96,507],[23,7662,7664],{"className":7663},[436],[23,7665,7667,7794],{"className":7666},[125,126],[23,7668,7670,7791],{"className":7669},[130],[23,7671,7674,7731,7739],{"className":7672,"style":7673},[134],"height:0.7173em;",[23,7675,7676,7679],{"style":523},[23,7677],{"className":7678,"style":527},[142],[23,7680,7682],{"className":7681},[147,148,149,150],[23,7683,7685],{"className":7684},[100,150],[23,7686,7688,7691],{"className":7687},[100,150],[23,7689,351],{"className":7690,"style":117},[100,101,150],[23,7692,7694],{"className":7693},[121],[23,7695,7697,7723],{"className":7696},[125,126],[23,7698,7700,7720],{"className":7699},[130],[23,7701,7703],{"className":7702,"style":552},[134],[23,7704,7705,7708],{"style":555},[23,7706],{"className":7707,"style":559},[142],[23,7709,7711],{"className":7710},[147,563,564,150],[23,7712,7714],{"className":7713},[100,150],[23,7715,7717],{"className":7716},[100,401,150],[23,7718,355],{"className":7719},[100,150],[23,7721,159],{"className":7722},[158],[23,7724,7726],{"className":7725},[130],[23,7727,7729],{"className":7728,"style":583},[134],[23,7730],{},[23,7732,7733,7736],{"style":604},[23,7734],{"className":7735,"style":527},[142],[23,7737],{"className":7738,"style":612},[611],[23,7740,7742,7745],{"style":7741},"top:-3.4159em;",[23,7743],{"className":7744,"style":527},[142],[23,7746,7748],{"className":7747},[147,148,149,150],[23,7749,7751],{"className":7750},[100,150],[23,7752,7754,7757],{"className":7753},[100,150],[23,7755,351],{"className":7756,"style":117},[100,101,150],[23,7758,7760],{"className":7759},[121],[23,7761,7763,7783],{"className":7762},[125,126],[23,7764,7766,7780],{"className":7765},[130],[23,7767,7769],{"className":7768,"style":552},[134],[23,7770,7771,7774],{"style":555},[23,7772],{"className":7773,"style":559},[142],[23,7775,7777],{"className":7776},[147,563,564,150],[23,7778,446],{"className":7779,"style":264},[100,101,150],[23,7781,159],{"className":7782},[158],[23,7784,7786],{"className":7785},[130],[23,7787,7789],{"className":7788,"style":583},[134],[23,7790],{},[23,7792,159],{"className":7793},[158],[23,7795,7797],{"className":7796},[130],[23,7798,7801],{"className":7799,"style":7800},[134],"height:0.4509em;",[23,7802],{},[23,7804],{"className":7805},[220,507]," 均为负值（",[702,7808,7493],{}," vs ",[702,7811,7530],{},"）。这是因为纯偏好对比损失只激励拉大相对差距，概率质量容易向分布外的长尾 Token 漂移——这也正是我们在《CycleResearcher 后训练解析》中强调必须叠加 ",[14,7814,7815],{},"NLL 正则项","或改用可验证奖励 ",[14,7818,7819],{},"GRPO"," 的根本原因；",[229,7822,7823,7826,7827,7830],{},[14,7824,7825],{},"对齐税（Alignment Tax）的真实存在","：在通用对话安全与礼貌偏好数据（HH-RLHF）上做 DPO 后，模型在确定性数学推理任务 GSM8K 上的准确率从 ",[14,7828,7829],{},"30.50% 回落至 27.50%（下降 3.00 个百分点）","。通用人类偏好倾向于奖励冗长、谨慎、带免责声明的回复风格，这种分布偏移会轻微干扰算术思维链的紧凑推导。",[726,7832],{},[729,7834,7836],{"id":7835},"四工程实践结论","四、工程实践结论",[226,7838,7839,7940,7949],{},[229,7840,7841,7844,7845,7897,7898,7939],{},[14,7842,7843],{},"SFT 阶段优先扩充高多样性推理轨迹，谨慎使用高秩 LoRA","：3-seed 消融证明 ",[23,7846,7848,7866],{"className":7847},[26],[23,7849,7851],{"className":7850},[30],[32,7852,7853],{"xmlns":34},[36,7854,7855,7863],{},[39,7856,7857,7859,7861],{},[47,7858,248],{},[42,7860,6938],{},[795,7862,1394],{},[76,7864,7865],{"encoding":78},"r > 16",[23,7867,7869,7888],{"className":7868,"ariaHidden":52},[83],[23,7870,7872,7876,7879,7882,7885],{"className":7871},[87],[23,7873],{"className":7874,"style":7875},[91],"height:0.5782em;vertical-align:-0.0391em;",[23,7877,248],{"className":7878,"style":264},[100,101],[23,7880],{"className":7881,"style":848},[109],[23,7883,6938],{"className":7884},[595],[23,7886],{"className":7887,"style":848},[109],[23,7889,7891,7894],{"className":7890},[87],[23,7892],{"className":7893,"style":952},[91],[23,7895,1394],{"className":7896},[100]," 在中小规模指令集上极易诱发验证集 Loss 反弹，而样本量从 250 扩充至 2000 可稳定带来 ",[23,7899,7901,7918],{"className":7900},[26],[23,7902,7904],{"className":7903},[30],[32,7905,7906],{"xmlns":34},[36,7907,7908,7916],{},[39,7909,7910,7912,7914],{},[42,7911,1305],{},[795,7913,4887],{},[353,7915,1311],{},[76,7917,4892],{"encoding":78},[23,7919,7921],{"className":7920,"ariaHidden":52},[83],[23,7922,7924,7927,7930,7933],{"className":7923},[87],[23,7925],{"className":7926,"style":1324},[91],[23,7928,1305],{"className":7929},[100],[23,7931,4887],{"className":7932},[100],[23,7934,7936],{"className":7935},[100,401],[23,7937,1311],{"className":7938},[100]," 的推理收益；",[229,7941,7942,7948],{},[14,7943,7944,7945,7947],{},"单卡 DPO 善用 ",[702,7946,6338],{}," 共享冻结基座","：无需常驻双份模型权重即可精确计算参考分布对数概率；",[229,7950,7951,7954],{},[14,7952,7953],{},"偏好对齐必须绑定推理回归测试（Regression Suite）","：任何针对语气、安全或风格的 DPO 迭代，都必须同时监控数学\u002F代码等确定性基准的对齐税，并在混合训练中加入同分布推理 NLL 锚定项。",[7956,7957,7958],"style",{},"html pre.shiki code .slEfO, html code.shiki .slEfO{--shiki-default:#F47067;--shiki-default-font-style:inherit;--shiki-light:#D73A49;--shiki-light-font-style:inherit;--shiki-cyber:#54B9FF;--shiki-cyber-font-style:italic}html pre.shiki code .sWURj, html code.shiki .sWURj{--shiki-default:#ADBAC7;--shiki-light:#24292E;--shiki-cyber:#EEF0F9}html pre.shiki code .sST-V, html code.shiki .sST-V{--shiki-default:#F47067;--shiki-light:#D73A49;--shiki-cyber:#54B9FF}html pre.shiki code .sS_0i, html code.shiki .sS_0i{--shiki-default:#DCBDFB;--shiki-light:#6F42C1;--shiki-cyber:#00DAEF}html pre.shiki code .s9NhK, html code.shiki .s9NhK{--shiki-default:#ADBAC7;--shiki-default-font-style:inherit;--shiki-light:#24292E;--shiki-light-font-style:inherit;--shiki-cyber:#FFD493;--shiki-cyber-font-style:italic}html pre.shiki code .sCxhR, html code.shiki .sCxhR{--shiki-default:#768390;--shiki-default-font-style:inherit;--shiki-light:#6A737D;--shiki-light-font-style:inherit;--shiki-cyber:#EEF0F98F;--shiki-cyber-font-style:italic}html pre.shiki code .sWqY2, html code.shiki .sWqY2{--shiki-default:#96D0FF;--shiki-light:#032F62;--shiki-cyber:#FFD493}html pre.shiki code .sf2cM, html code.shiki .sf2cM{--shiki-default:#F47067;--shiki-light:#D73A49;--shiki-cyber:#EEF0F9}html pre.shiki code .sgENW, html code.shiki .sgENW{--shiki-default:#6CB6FF;--shiki-light:#005CC5;--shiki-cyber:#FFD493}html pre.shiki code .sq3c-, html code.shiki .sq3c-{--shiki-default:#ADBAC7;--shiki-light:#24292E;--shiki-cyber:#00DAEF}html pre.shiki code .sA9aF, html code.shiki .sA9aF{--shiki-default:#F69D50;--shiki-default-font-style:inherit;--shiki-light:#E36209;--shiki-light-font-style:inherit;--shiki-cyber:#4BF3C8;--shiki-cyber-font-style:italic}html pre.shiki code .svkvy, html code.shiki .svkvy{--shiki-default:#6CB6FF;--shiki-light:#005CC5;--shiki-cyber:#EEF0F9}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .cyber .shiki span {color: var(--shiki-cyber);background: var(--shiki-cyber-bg);font-style: var(--shiki-cyber-font-style);font-weight: var(--shiki-cyber-font-weight);text-decoration: var(--shiki-cyber-text-decoration);}html.cyber .shiki span {color: var(--shiki-cyber);background: var(--shiki-cyber-bg);font-style: var(--shiki-cyber-font-style);font-weight: var(--shiki-cyber-font-weight);text-decoration: var(--shiki-cyber-text-decoration);}",{"title":6346,"searchDepth":6363,"depth":6377,"links":7960},[7961,7966,7967,7968],{"id":731,"depth":6363,"text":732,"children":7962},[7963,7965],{"id":774,"depth":6377,"text":7964},"1. 维度 A：LoRA Rank 消融（固定 LR=2×10−4,  N=1000\\text{LR}=2\\times 10^{-4},\\; N=1000LR=2×10−4,N=1000）",{"id":2663,"depth":6377,"text":2664},{"id":5018,"depth":6363,"text":5019},{"id":7002,"depth":6363,"text":7003},{"id":7835,"depth":6363,"text":7836},[7970],"AI","\u002Fimages\u002Fposts\u002Fqwen25-sft-dpo-alignment-tax-ablation\u002Fqwen25-sft-dpo-cover-20261018.webp","左侧监督微调轨迹流经中央 Bradley-Terry 偏好分流天平分离出正负样本通道并连接右侧对齐税监控仪表盘的等距技术插画","2026-10-18","基于 Qwen2.5 完整复现 Instruct SFT → Bradley-Terry DPO 两阶段后训练管线：通过 3 随机种子消融 LoRA Rank、学习率与数据规模（GSM8K +12.73pp），实现免双倍权重的动态 Adapter 切换 DPO 训练，并定量揭示偏好对齐对确定性数学推理的对齐税效应。",false,"md",{"mathjax":6380},"\u002Fposts\u002Fqwen25-sft-dpo-alignment-tax-ablation",null,{"title":5,"description":7974},"posts\u002Fqwen25-sft-dpo-alignment-tax-ablation",[7983,5127,5041,7984,7985,719],"Post-training","LoRA","强化学习","DiRzrstZ2nT2kcB6V9tMcdZn4tPwxai-Q2YYmXCE2io",[7988,8006,8022,8036,8050,8064,8078,8091,8103,8115,8130,8142,8146,8158,8171,8184,8200,8214,8225,8237,8251,8264,8278,8290,8304,8316,8331,8345,8359,8370,8383,8400,8413,8426,8438,8450,8463,8476],{"slug":7989,"path":7990,"title":7991,"date":7992,"tags":7993,"description":7999,"cover":8000,"coverAlt":8001,"excerpt":8002,"draft":7975,"hidden":7975,"published":6380,"readingTime":6429,"series":8003,"seriesOrder":6429,"codeUrl":8004,"attachmentUrl":8005},"cs336-cross-entropy-adamw-training-loop","\u002Fposts\u002Fcs336-cross-entropy-adamw-training-loop","Stanford CS336 学习记录 08：数值稳定交叉熵、AdamW、余弦调度与完整训练闭环","2026-10-27",[7994,7995,7996,7997,7998],"CS336","PyTorch","AdamW","CrossEntropy","训练闭环","用 LogSumExp 消去交叉熵中的指数溢出，手写解耦权重衰减的 AdamW 优化器、全局 L2 梯度裁剪、Warmup 余弦退火调度、memmap 数据采样、Checkpoint 序列化与 Top-p 解码，完成 CS336 Assignment 1 全部 48 项测试。","\u002Fimages\u002Fposts\u002Fcs336-cross-entropy-adamw-training-loop\u002Fcs336-cross-entropy-adamw-training-loop-cover.webp","滑动窗口采样、余弦退火曲线、双动量优化轨迹与检查点归档构成的训练闭环 AI 辅助概念插画","作为 Stanford CS336 Assignment 1 系列的收官篇，我们已经拥有了分词器（bpe.py、tokenizer.py）和完整的语言模型架构（transformer.py）。要让模型在语料上真正训练并生成文本，还需","Stanford CS336","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcs336-labs\u002Ftree\u002Fb5c93ad0cc73178b2b2a62e518a0900da47a0fb7","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-08-adamw-training-loop-20261027.zip",{"slug":8007,"path":8008,"title":8009,"date":8010,"tags":8011,"description":8018,"cover":8019,"coverAlt":8020,"excerpt":8021,"draft":7975,"hidden":7975,"published":6380,"readingTime":6429},"editcache-video-dit-inference-acceleration","\u002Fposts\u002Feditcache-video-dit-inference-acceleration","EditCache：视频 Diffusion Transformer (DiT) 局部编辑的跨步特征复用与区域重算加速","2026-10-25",[8012,8013,8014,8015,8016,8017],"扩散模型","视频生成","AI Infra","推理加速","GPU","论文解读","拆解视频扩散模型（Wan2.1-VACE）局部编辑中的三重算力冗余，通过可重入 Latent、跨步 Block 隐状态缓存复用、带 Halo 边界的局部区域重算与自适应推理起点，在单卡 H100 上实现 2.46× 模型计算加速与 1.49× 严格同步端到端加速。","\u002Fimages\u002Fposts\u002Feditcache-video-dit-inference-acceleration\u002Feditcache-dit-cover-20261025.webp","视频扩散 Transformer 时空三维体素网格中冻结背景缓存与暖橙色局部编辑重算通道的等距技术插画","在基于 Diffusion Transformer（DiT）的视频编辑与局部重绘（Video Inpainting \u002F Local Editing）任务中，用户的典型工作流并不是一次性生成最终成片，而是针对同一段源视频反复调整掩码区",{"slug":8023,"path":8024,"title":8025,"date":8010,"tags":8026,"description":8031,"cover":8032,"coverAlt":8033,"excerpt":8034,"draft":7975,"hidden":7975,"published":6380,"readingTime":6429,"series":8003,"seriesOrder":6423,"codeUrl":8004,"attachmentUrl":8035},"cs336-bpe-tokenizer-training-encoding","\u002Fposts\u002Fcs336-bpe-tokenizer-training-encoding","Stanford CS336 学习记录 07：Byte-Level BPE 分词器训练、倒排索引加速与流式编码",[7994,8027,8028,8029,8030],"Tokenizer","BPE","Tiktoken","算法优化","实现 GPT-2 正则预分词、特殊 token 硬边界切分、基于 pair-to-word 倒排索引的增量 BPE 合并算法（\u003C 0.25s 通过速度基准），以及与 tiktoken 完全对齐的 Tokenizer。","\u002Fimages\u002Fposts\u002Fcs336-bpe-tokenizer-training-encoding\u002Fcs336-bpe-tokenizer-training-encoding-cover.webp","UTF-8 字节块经过正则切分闸口后沿合并树聚合成子词单元的 AI 辅助概念插画","前面六篇完成了 Transformer 语言模型本身的全部神经网络组件。要让模型真正处理自然语言文本，还需要打通入口处的 ByteLevel BPE（BytePair Encoding）分词器。这一篇记录 cs336basics\u002Fb","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-07-bpe-tokenizer-20261025.zip",{"slug":8037,"path":8038,"title":8039,"date":8040,"tags":8041,"description":8046,"cover":8047,"coverAlt":8048,"excerpt":8049,"draft":7975,"hidden":7975,"published":6380,"readingTime":6429},"in-situ-autoguidance-editable-scientific-illustration","\u002Fposts\u002Fin-situ-autoguidance-editable-scientific-illustration","从原位自校正引导（In-situ Autoguidance）到可编辑科研插图生成（AutoFigure-Edit）","2026-10-24",[8012,8042,8043,8044,8045,8017],"多模态","Agent","ICML","ACL","串联两项生成式模型与多模态智能体研究：如何通过推理期局部随机微扰在单模型内构造免训练扩散自引导向量（ICML 2025 Workshop 第一作者，arXiv:2510.17136），以及如何通过五阶段视觉-符号混合流水线生成原生可编辑 SVG 科研插图（ACL 2026 Demo，FigureBench 83.0% 胜率）。","\u002Fimages\u002Fposts\u002Fin-situ-autoguidance-editable-scientific-illustration\u002Fautoguidance-autofigure-cover-20261024.webp","左侧扩散流形自校正轨迹矢量场与右侧五阶段可编辑 SVG 科研插图分层网格相衔接的等距技术插画","生成式视觉模型正在经历从“生成高保真像素（Pixel Synthesis）”向“可控结构化生成（Controllable & Editable Generation）”的演进。在这一进程中，我作为第一作者或核心贡献者参与了两项分别处",{"slug":8051,"path":8052,"title":8053,"date":8054,"tags":8055,"description":8059,"cover":8060,"coverAlt":8061,"excerpt":8062,"draft":7975,"hidden":7975,"published":6380,"readingTime":6429,"series":8003,"seriesOrder":6412,"codeUrl":8004,"attachmentUrl":8063},"cs336-pre-norm-transformer-block-lm","\u002Fposts\u002Fcs336-pre-norm-transformer-block-lm","Stanford CS336 学习记录 06：Pre-Norm Transformer Block 与完整语言模型装配","2026-10-23",[7994,7995,8056,8057,8058],"Transformer","Pre-Norm","FLOPs","将 RMSNorm、带 RoPE 的因果多头自注意力和 SwiGLU 组装成 Pre-Norm Transformer Block，搭建完整 TransformerLM，并核算 GPT-2 XL 的参数量与前向 FLOPs。","\u002Fimages\u002Fposts\u002Fcs336-pre-norm-transformer-block-lm\u002Fcs336-pre-norm-transformer-block-lm-cover.webp","干净残差流主干穿过多个 Pre-Norm 子层并汇聚到词表投影层的 AI 辅助概念插画","前五篇把 Linear、Embedding、RMSNorm、SwiGLU、RotaryPositionalEmbedding 和因果 MultiHeadSelfAttention 全部单独打通了。这一篇进入 Assignment 1","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-06-transformer-block-lm-20261023.zip",{"slug":8065,"path":8066,"title":8067,"date":8054,"tags":8068,"description":8074,"cover":8075,"coverAlt":8076,"excerpt":8077,"draft":7975,"hidden":7975,"published":6380,"readingTime":6522},"xgrammar-concurrent-cache-generation-fix","\u002Fposts\u002Fxgrammar-concurrent-cache-generation-fix","XGrammar 结构化解码缓存剖析：Unlimited 模式并发 Clear 字节记账竞态与代际修复（PR #1011）",[8014,8069,8070,8071,8072,8073],"C++","XGrammar","结构化解码","并发编程","源码分析","拆解 XGrammar GrammarCompiler 的双层缓存架构与 ThreadSafeLRUCache 锁分离实现，定位 Unlimited 模式下 Clear() 字节残留与在途编译（In-flight Computation）跨代污染根因，给出代际号 + 读写锁原子修复与零 sleep 确定性并发 GoogleTest 验证。","\u002Fimages\u002Fposts\u002Fxgrammar-concurrent-cache-generation-fix\u002Fxgrammar-cache-fix-cover-20261022.webp","双层磨砂玻璃流道经暖橙色代际闸门过滤后汇入晶格缓存矩阵的等距技术插画","在大语言模型服务（如 SGLang、vLLM）的结构化输出场景中，JSON Schema 或 EBNF 语法需要先编译为词表级掩码缓存（CompiledGrammar），才能在每个 Decode 步以微秒级开销生成合法 token ",{"slug":8079,"path":8080,"title":8081,"date":8082,"tags":8083,"description":8087,"cover":8088,"coverAlt":8089,"excerpt":8090,"draft":7975,"hidden":7975,"published":6380,"readingTime":6522},"multi-constraint-agent-routing-dag-eval","\u002Fposts\u002Fmulti-constraint-agent-routing-dag-eval","多约束生产级 Agent 系统实践：两阶段工具路由、DAG 局部重规划与轨迹级评测闭环","2026-10-22",[8043,8084,8085,8086,5127,7985],"LLM","工具调用","上下文工程","围绕跨境酒旅等强耦合多约束业务场景，拆解向量 Top-K 召回 + LLM 精排 + 跨语言实体链接与 Schema 门控（96%+ 调用合法率）、基于状态机的 DAG 局部子图重规划、Observation 冗余裁剪（节省 40% Token），以及从线上失败恢复轨迹构造 SFT\u002FDPO 偏好对的评测与训练闭环。","\u002Fimages\u002Fposts\u002Fmulti-constraint-agent-routing-dag-eval\u002Fagent-dag-routing-cover-20261020.webp","候选模块经漏斗滤网进入双分支有向无环图轨道并在受阻节点触发暖橙色局部旁路重连的等距技术插画","在玩具级 Demo 里，几个工具加上一段 ReAct 循环就能跑通大多数单步问答。但当 Agent 进入跨境酒旅、复杂交易或研发沙箱等真实业务生产环境时，单靠线性 ReAct 循环很快会撞上四堵墙： 1. 工具空间与参数耦合爆炸：候",{"slug":8092,"path":8093,"title":8094,"date":8095,"tags":8096,"description":8099,"cover":8100,"coverAlt":8101,"excerpt":8102,"draft":7975,"hidden":7975,"published":6380,"readingTime":6640},"llm-mrd-distillation-auditable-review-agents","\u002Fposts\u002Fllm-mrd-distillation-auditable-review-agents","多视角推理蒸馏（LLM-MRD）与可审计生成式评审系统（DeepReviewer 2.0）深度拆解","2026-10-21",[8097,8042,8098,8043,7983,8017],"知识蒸馏","Qwen2.5-VL","对比大模型落地的两条互补路径：如何用 Qwen2.5-VL 离线多视角分析结合 MLP 残差校准与温度缩放 KL 蒸馏，将推理能力压入 358M 学生模型实现 17.2 ms 独立推理（LLM-MRD，DASFAA 2026 Oral）；以及如何通过主张-证据-风险台账、PASA 可比性门控与导出预算校验构建可审计的同行评审 Agent（DeepReviewer 2.0）。","\u002Fimages\u002Fposts\u002Fllm-mrd-distillation-auditable-review-agents\u002Fmrd-review-distill-cover-20261016.webp","三块半透明几何棱镜发出平行光束穿过中央校准环汇聚至右侧高密度晶体模块的等距技术插画","把大语言模型或视觉语言模型（VLM）的深度推理能力用到实际系统中，通常会面临一个两难选择： 如果场景受限于毫秒级在线延迟（如高吞吐多模态内容审核与虚假信息检测）：直接在线调用 7B～72B 的多模态大模型做长链条思维链（CoT）生成",{"slug":8104,"path":8105,"title":8106,"date":8095,"tags":8107,"description":8110,"cover":8111,"coverAlt":8112,"excerpt":8113,"draft":7975,"hidden":7975,"published":6380,"readingTime":6423,"series":8003,"seriesOrder":6398,"codeUrl":8004,"attachmentUrl":8114},"cs336-rope-rotary-position-embedding","\u002Fposts\u002Fcs336-rope-rotary-position-embedding","Stanford CS336 学习记录 05：RoPE 旋转位置编码与带位置的多头注意力",[7994,7995,8056,8108,8109],"RoPE","Attention","从二维平面正交旋转推导 RoPE 的奇偶特征配对公式，理解为什么内积只依赖相对位置差 (i - j)，并用 register_buffer 预计算 cos\u002Fsin 缓存接入因果多头自注意力。","\u002Fimages\u002Fposts\u002Fcs336-rope-rotary-position-embedding\u002Fcs336-rope-rotary-position-embedding-cover.webp","二维特征平面中的旋转向量对与预计算余弦正弦网格的 AI 辅助概念插画","上一篇完成了不含位置编码的因果多头自注意力。如果只看那一套矩阵乘法，把输入序列前两个 token 交换顺序（并在因果 mask 允许的范围内看同一个后缀 query 对它们的点积），点积本身并不知道哪个 key 离当前 query ","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-05-rope-mha-20261021.zip",{"slug":8116,"path":8117,"title":8118,"date":8119,"tags":8120,"description":8126,"cover":8127,"coverAlt":8128,"excerpt":8129,"draft":7975,"hidden":7975,"published":6380,"readingTime":6429},"llm-gateway-semantic-cache-dynamic-batching","\u002Fposts\u002Fllm-gateway-semantic-cache-dynamic-batching","高并发 LLM 推理网关架构实践：两级精确\u002F语义缓存、50ms 有界动态批处理与尾部延迟治理","2026-10-19",[8014,8121,8122,8123,8124,8125],"LLM 推理","分布式系统","Redis","FastAPI","后端","深度拆解生产级大模型推理网关 LLM-Gateway 的核心系统设计：基于 SHA-256 与 Redis 向量余弦相似度的两级精确\u002F语义缓存、50ms 有界等待动态批处理调度器（Dynamic Batching）、令牌桶多租户限流与 Prometheus 全链路可观测性，将命中请求 P99 延迟从 5,307 ms 压降至 17 ms。","\u002Fimages\u002Fposts\u002Fllm-gateway-semantic-cache-dynamic-batching\u002Fllm-gateway-cover-20261019.webp","高并发请求流依次穿过精确哈希滤网、向量球面语义缓存层与 50ms 动态批处理时间窗拼装器的等距系统架构插画","在将大语言模型接入高并发业务系统时，如果让上游微服务直接裸连底层推理引擎（vLLM \u002F SGLang 或外部模型 API），系统很快会暴露出三个典型的生产痛点： 1. 高频重复与近义查询消耗巨量算力：大量用户请求在语义上高度等价（例",{"slug":8131,"path":8132,"title":8133,"date":8119,"tags":8134,"description":8136,"cover":8137,"coverAlt":8138,"excerpt":8139,"draft":7975,"hidden":7975,"published":6380,"readingTime":6522,"series":8003,"seriesOrder":6384,"codeUrl":8140,"attachmentUrl":8141},"cs336-multihead-attention-shapes","\u002Fposts\u002Fcs336-multihead-attention-shapes","Stanford CS336 学习记录 04：多头注意力怎样拆头、计算和合头",[7994,7995,8056,8109,8135],"张量形状","沿着 (2, 5, 64) 的输入追踪 QKV 投影、reshape、transpose、共享因果 mask 和合头，理解每个头仍处理完整序列，以及最后为什么还需要输出 Linear。","\u002Fimages\u002Fposts\u002Fcs336-multihead-attention-shapes\u002Fcs336-multihead-attention-shapes-cover.webp","宽向量束分成多条并行特征轨道后重新合并的 AI 辅助概念插画","上一篇完成了缩放点积注意力。那个函数接收已经准备好的 Q、K、V，计算匹配分数，应用 mask，再用权重汇总 V。这一篇继续把它放进一个带参数的多头自注意力模块。 这次最需要看清的是“拆头”。代码里只是几行 reshape 和 tr","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcs336-labs\u002Ftree\u002F31363223498fafd1528c550a51efbe0aed231736","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-04-multihead-attention-20261019.zip",{"slug":8143,"path":7978,"title":5,"date":7973,"tags":8144,"description":7974,"cover":7971,"coverAlt":7972,"excerpt":8145,"draft":7975,"hidden":7975,"published":6380,"readingTime":6436},"qwen25-sft-dpo-alignment-tax-ablation",[7983,5127,5041,7984,7985,719],"在大语言模型后训练（PostTraining）管线中，监督微调（SFT）负责教会模型任务格式与领域推理模式，而直接偏好优化（DPO, Direct Preference Optimization）则利用成对偏好数据 $(x, yw,",{"slug":8147,"path":8148,"title":8149,"date":7973,"tags":8150,"description":8152,"cover":8153,"coverAlt":8154,"excerpt":8155,"draft":7975,"hidden":7975,"published":6380,"readingTime":6473,"series":8003,"seriesOrder":6377,"codeUrl":8156,"attachmentUrl":8157},"cs336-softmax-attention-mask","\u002Fposts\u002Fcs336-softmax-attention-mask","Stanford CS336 学习记录 03：Softmax 的数值稳定与 Attention Mask",[7994,7995,8056,8151,8109],"Softmax","从减去最大值为何不改变 Softmax 开始，理清函数与有参数层的区别，再沿着 Q、K、V 的形状理解缩放点积注意力，以及 mask 为什么必须填负无穷。","\u002Fimages\u002Fposts\u002Fcs336-softmax-attention-mask\u002Fcs336-softmax-attention-mask-cover.webp","注意力分数、权重与因果遮罩的 AI 辅助概念插画","上一篇写到 SwiGLU，接下来进入 Attention。真正开始相乘之前，先做了 Softmax：它把一组分数变成总和为 1 的权重，后面用这些权重汇总不同 token 提供的信息。 这段学习里，有两个很具体的问题值得记下来。一个","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcs336-labs\u002Ftree\u002Fa39eb684710817f7f389626423fcb26ca0f6d629","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-03-softmax-sdpa-20261018.zip",{"slug":8159,"path":8160,"title":8161,"date":8162,"tags":8163,"description":8166,"cover":8167,"coverAlt":8168,"excerpt":8169,"draft":7975,"hidden":7975,"published":6380,"readingTime":6436,"series":8003,"seriesOrder":6363,"codeUrl":8156,"attachmentUrl":8170},"cs336-rmsnorm-swiglu","\u002Fposts\u002Fcs336-rmsnorm-swiglu","Stanford CS336 学习记录 02：RMSNorm 的广播与 SwiGLU 的门控","2026-10-16",[7994,7995,8056,8164,8165],"RMSNorm","SwiGLU","用一个 token 向量理解均方根归一化，拆开 dim=-1 与 keepdim=True 的作用，再沿着两条分支读懂 SiLU、逐元素乘法和 SwiGLU。","\u002Fimages\u002Fposts\u002Fcs336-rmsnorm-swiglu\u002Fcs336-rmsnorm-swiglu-cover.webp","特征尺度整理和双分支门控的 AI 辅助概念插画","上一篇把 Linear 和 Embedding 的参数形状理顺了。继续往下，最先遇到的是 RMSNorm：公式不算长，但把它落到一批 token 上时，dim=1、keepdim=True 和广播必须一起理解。 然后是 SiLU 和","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-02-rmsnorm-swiglu-20261016.zip",{"slug":8172,"path":8173,"title":8174,"date":8175,"tags":8176,"description":8179,"cover":8180,"coverAlt":8181,"excerpt":8182,"draft":7975,"hidden":7975,"published":6380,"readingTime":6457,"series":8003,"seriesOrder":6352,"codeUrl":8156,"attachmentUrl":8183},"cs336-linear-embedding-shapes","\u002Fposts\u002Fcs336-linear-embedding-shapes","Stanford CS336 学习记录 01：从 Linear、Embedding 搞清张量形状","2026-10-14",[7994,7995,8056,8177,8178],"Linear","Embedding","从 PyTorch 类、参数与 forward 的关系开始，理解 Linear 为什么保存两行三列的权重、计算时为什么要转置，再区分 Embedding 查表与线性变换。","\u002Fimages\u002Fposts\u002Fcs336-linear-embedding-shapes\u002Fcs336-linear-embedding-shapes-cover.webp","token 方块、向量表与矩阵变换的 AI 辅助概念插画","开始沿着 Stanford CS336 的 Assignment 1 一步步实现。BPE 之前已经接触过，这轮先跳过分词器，从 Transformer 的基础层开始。这里记录的是当时写代码时真正卡住的问题，以及对应测试能确认到的范围","\u002Fmaterials\u002Fcs336\u002Fexperiments\u002Fcs336-01-linear-embedding-20261014.zip",{"slug":8185,"path":8186,"title":8187,"date":8188,"tags":8189,"description":8194,"cover":8195,"coverAlt":8196,"excerpt":8197,"draft":7975,"hidden":7975,"published":6380,"readingTime":6493,"series":8190,"seriesOrder":6522,"codeUrl":8198,"attachmentUrl":8199},"cmu-15-779-kv-cache-granularity-study","\u002Fposts\u002Fcmu-15-779-kv-cache-granularity-study","CMU 15-779 学习记录 14：KV 页大小、共享边界与 CPU 实现成本","2026-10-12",[8190,8191,8192,8193],"CMU 15-779","KV cache","PagedAttention","系统实验","用 210 组 CPU 配置比较连续与分页 KV cache，观察 31\u002F32-token 共享前缀的页边界，分别记录活跃槽位、池容量、复制计数和完整调用耗时。","\u002Fimages\u002Fposts\u002Fcmu-15-779-kv-serving-speculative-sampling\u002Fcmu-serving-cover-20261019.webp","浅色手绘分页存储抽屉与共享前缀卡片，旁边连接不同长度的请求队列","个人课程项目中的 CPU 实验记录，延续记录 11 的分页 KV 实现。已完成 210 组配置与 1470 次计时，输出均通过连续存储参考对照；本篇不代表正式作业提交、整门课程或项目全部交付完成。 同样是共享 31 个 token，","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs\u002Ftree\u002F200faf569eccf94fb125e65110bb5d14094d454a","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fkv-cache-granularity-20261019.zip",{"slug":8201,"path":8202,"title":8203,"date":8204,"tags":8205,"description":8208,"cover":8209,"coverAlt":8210,"excerpt":8211,"draft":7975,"hidden":7975,"published":6380,"readingTime":6398,"series":8190,"seriesOrder":6498,"codeUrl":8212,"attachmentUrl":8213},"cmu-15-779-moe-routing-capacity","\u002Fposts\u002Fcmu-15-779-moe-routing-capacity","CMU 15-779 学习记录 13：MoE 路由、稀疏分发与容量溢出","2026-10-09",[8190,8206,8207],"MoE","Routing","用 CPU top-2 MoE 检查稀疏分发的输出与梯度，统计 expert-token 行数，并保留容量丢弃会改变输出的反例。","\u002Fimages\u002Fposts\u002Fcmu-15-779-moe-routing-capacity\u002Fcmu-moe-cover-20261019.webp","浅色手绘分拣器将小珠分配到四个专家工作屋，队列呈现负载不均","第 18 讲的个人配套实验。36 号脚本串行运行 CPU 专家，检查 topk 路由后的分发与合并；没有训练 router，也没有执行 GPU grouped GEMM 或跨设备 alltoall。 上一节记录低秩适配，这次转向每个","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs\u002Ftree\u002Fc4851e0cedadf13af7c421867e83c8ff57244fa0","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fcpu-parallel-serving-lora-moe-20261019.zip",{"slug":8215,"path":8216,"title":8217,"date":8218,"tags":8219,"description":8221,"cover":8222,"coverAlt":8223,"excerpt":8224,"draft":7975,"hidden":7975,"published":6380,"readingTime":6398,"series":8190,"seriesOrder":6493,"codeUrl":8212,"attachmentUrl":8213},"cmu-15-779-lora-uniform-quantization","\u002Fposts\u002Fcmu-15-779-lora-uniform-quantization","CMU 15-779 学习记录 12：LoRA 首步梯度、权重合并与均匀 4-bit 打包","2026-10-06",[8190,7984,8220],"Quantization","用一次 CPU SGD 更新检查 LoRA 初始化、冻结基座与合并等价，再分别记录均匀 4-bit 存储字节数和量化误差。","\u002Fimages\u002Fposts\u002Fcmu-15-779-lora-uniform-quantization\u002Fcmu-lora-cover-20261019.webp","浅色手绘的大计算板、两块细长适配板和成对打包的小方块","第 17 讲的个人配套实验，35 号脚本实际执行一次 CPU LoRA 更新，并实现均匀 4bit 打包。实验检查机制与数值误差，不评估微调任务效果，也没有实现完整 QLoRA。 这次把两个问题分开记录：低秩分支是否能合并进原权重，",{"slug":8226,"path":8227,"title":8228,"date":8229,"tags":8230,"description":8234,"cover":8195,"coverAlt":8235,"excerpt":8236,"draft":7975,"hidden":7975,"published":6380,"readingTime":6423,"series":8190,"seriesOrder":6473,"codeUrl":8212,"attachmentUrl":8213},"cmu-15-779-kv-serving-speculative-sampling","\u002Fposts\u002Fcmu-15-779-kv-serving-speculative-sampling","CMU 15-779 学习记录 11：分页 KV、连续批处理与推测采样","2026-10-03",[8190,8231,8232,8233],"LLM Serving","KV Cache","Speculative Decoding","在 CPU 上验证 KV 页共享和写时复制，用单位成本模型检查连续批处理，再用条件概率小模型核对推测采样的接受与补采样。","浅色手绘的共享卡片抽屉与不断补入小珠子的请求托盘","第 15、16 讲的个人配套实验。33 号实现 CPU 分页 KV 存储，37 号模拟请求准入，34 号运行小概率模型的推测采样。它们分别验证存储、调度和分布问题，没有测量真实 LLM 服务吞吐或 GPU 延迟。 记录 08已经对照",{"slug":8238,"path":8239,"title":8240,"date":8241,"tags":8242,"description":8247,"cover":8248,"coverAlt":8249,"excerpt":8250,"draft":7975,"hidden":7975,"published":6380,"readingTime":6607},"qwen3-mid-post-train-mlx","\u002Fposts\u002Fqwen3-mid-post-train-mlx","在 M2 Pro 上完成 Qwen3 两阶段训练：Mid Train、SFT 与可复核评估","2026-09-30 11:19:10",[8084,8243,8244,7984,8245,5041,8246],"Qwen3","MLX","持续预训练","Apple Silicon","用 Qwen3 0.6B Base 和 MLX 完成领域持续训练与监督微调，记录数据划分、聊天模板、损失掩码、数值稳定性修复和独立测试结果。","\u002Fimages\u002Fposts\u002Fqwen3-mid-post-train-mlx\u002Fqwen3-mlx-cover-20261018.webp","Apple Silicon 统一内存芯片底座支撑两阶段持续预训练与带精确 Loss 掩码滤网的监督微调桥接插画","在一台 M2 Pro、16 GB 内存的 Mac 上，能否把一个 Base 语言模型接续训练成输出 JSON 的运维事件分级助手？这次项目完成了领域持续训练、监督微调、检查点选择、重新加载和独立测试，并保留了权重、日志和数据哈希。 ",{"slug":8252,"path":8253,"title":8254,"date":8255,"tags":8256,"description":8260,"cover":8261,"coverAlt":8262,"excerpt":8263,"draft":7975,"hidden":7975,"published":6380,"readingTime":6412,"series":8190,"seriesOrder":6457,"codeUrl":8212,"attachmentUrl":8213},"cmu-15-779-cpu-parallelism-pipeline","\u002Fposts\u002Fcmu-15-779-cpu-parallelism-pipeline","CMU 15-779 学习记录 10：加权梯度、状态分片与同步流水线","2026-09-30",[8190,8257,8258,8259],"Parallel Training","Gloo","Pipeline","用双 CPU 进程 Gloo 检查不等数据分片的梯度，推导持久状态字节数，再用依赖模型比较 GPipe 与同步 1F1B 的激活存活。","\u002Fimages\u002Fposts\u002Fcmu-15-779-cpu-parallelism-pipeline\u002Fcmu-parallel-cover-20261019.webp","浅色手绘的四阶段桌面流水线与按数量汇总的托盘","第 12 至 14 讲的个人配套实验，脚本编号 31、32 不对应官方作业。Gloo 实际运行两个 CPU 进程；张量切分、状态字节数和流水线分别按下文的数学或成本模型解释。真实 GPU 实测按学习安排跳过。 上一节检查单个计算图的",{"slug":8265,"path":8266,"title":8267,"date":8268,"tags":8269,"description":8272,"cover":8273,"coverAlt":8274,"excerpt":8275,"draft":7975,"hidden":7975,"published":6380,"readingTime":6423,"series":8190,"seriesOrder":6436,"codeUrl":8276,"attachmentUrl":8277},"cmu-15-779-cpu-compiler-optimizations","\u002Fposts\u002Fcmu-15-779-cpu-compiler-optimizations","CMU 15-779 学习记录 09：CPU 编译优化、合法重写与候选搜索","2026-09-27",[8190,8270,8271,7995],"ML Compiler","TorchInductor","用 CPU 实验验证 BatchNorm 折叠、RMSNorm 重排与非线性反例，测量分块候选，并运行 TorchInductor，分别记录首调用与预热后的时间。","\u002Fimages\u002Fposts\u002Fcmu-15-779-cpu-compiler-optimizations\u002Fcmu-cpu-compiler-cover-20261019.webp","淡紫计算积木经过薄荷绿折叠模具变成浅蓝积木，前方摆放四种候选托盘的浅色手绘插画","编译章节的个人配套实验，编号 30 不对应官方作业。图重写、分块候选和 TorchInductor 均在 CPU 上运行；真实 GPU 实验按学习安排跳过。 上一节记录了分块 Attention 和 KV cache。这次转到编译优","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs\u002Ftree\u002F9261bea6aedf31cdccbbab2541ca693fc97b3743","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fcpu-compiler-optimizations-20261019.zip",{"slug":8279,"path":8280,"title":8281,"date":8282,"tags":8283,"description":8285,"cover":8286,"coverAlt":8287,"excerpt":8288,"draft":7975,"hidden":7975,"published":6380,"readingTime":6493,"series":8190,"seriesOrder":6429,"codeUrl":8198,"attachmentUrl":8289},"cmu-15-779-attention-online-softmax","\u002Fposts\u002Fcmu-15-779-attention-online-softmax","CMU 15-779 学习记录 08：分块 Attention、在线 Softmax 与 KV cache","2026-09-24",[8190,8109,8151,8284,7995],"FlashAttention","在 CPU 上对照完整 Attention、在线 Softmax 分块实现与 PyTorch SDPA，检查因果掩码、浮点误差和分数块大小，理解 FlashAttention 的一部分算法基础。","\u002Fimages\u002Fposts\u002Fcmu-15-779-attention-online-softmax\u002Fcmu-online-attention-cover-20261016.webp","浅紫色卡片分批进入绿色计算托盘，旁边三个状态容器保存累积量的浅色手绘插画","第 5 讲 Transformers & Attention 的个人配套实验。28\u002F29 号是 CPU 单头前向与 KV cache；后续补入 38 号 CUDA 多头合成张量的显式 SDPA MATH\u002FFLASH 对照。没有手写 ","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fcuda-triton-sdpa-20261019.zip",{"slug":8291,"path":8292,"title":8293,"date":8294,"tags":8295,"description":8300,"cover":8301,"coverAlt":8302,"excerpt":8303,"draft":7975,"hidden":7975,"published":6380,"readingTime":6457,"series":8190,"seriesOrder":6423,"codeUrl":8198,"attachmentUrl":8289},"cmu-15-779-matmul-tiling-tree-reduction","\u002Fposts\u002Fcmu-15-779-matmul-tiling-tree-reduction","CMU 15-779 学习记录 07：矩阵分块、数据复用与树形归约","2026-09-21",[8190,8296,8297,8298,8299],"CUDA","矩阵乘法","数据复用","归约","用 CPU 模型统计矩阵分块的输入读取次数，再比较串行求和和树形归约的依赖深度，记录补零、同步与浮点加法顺序的边界。","\u002Fimages\u002Fposts\u002Fcmu-15-779-matmul-tiling-tree-reduction\u002Fcmu-matmul-reduction-cover-20261016.webp","矩阵积木在共享托盘中复用，旁边的小珠沿树形管道逐层合并的浅色手绘插画","GPU Architecture & CUDA Programming 的个人配套实验。首次记录用 CPU 模型检查数据复用与依赖深度；后续补入单卡 GPU 原生 CUDA 矩阵验证，以及 Triton 两阶段归约和矩阵乘法实测。 ",{"slug":8305,"path":8306,"title":8307,"date":8308,"tags":8309,"description":8312,"cover":8313,"coverAlt":8314,"excerpt":8315,"draft":7975,"hidden":7975,"published":6380,"readingTime":6457,"series":8190,"seriesOrder":6412,"codeUrl":8198,"attachmentUrl":8289},"cmu-15-779-cuda-thread-indexing","\u002Fposts\u002Fcmu-15-779-cuda-thread-indexing","CMU 15-779 学习记录 06：CUDA 线程索引与边界检查","2026-09-18",[8190,8296,8016,8310,8311],"线程索引","边界检查","用 CPU 演示 CUDA 的 block、thread 与全局下标：10 个元素如何分给 3 个线程块，为什么要跳过越界线程，以及这些验证能说明什么。","\u002Fimages\u002Fposts\u002Fcmu-15-779-cuda-thread-indexing\u002Fcmu-cuda-indexing-cover-20261015.webp","分组的浅紫色计算方块与末端边界挡板组成线程分配教具的浅色插画","GPU Architecture & CUDA Programming 的个人配套练习。首次记录为 CPU 索引演示；后续在单卡 GPU 环境补入原生 CUDA 编译、运行与 memcheck，以及 Triton 向量加法实测。 上",{"slug":8317,"path":8318,"title":8319,"date":8320,"tags":8321,"description":8325,"cover":8326,"coverAlt":8327,"excerpt":8328,"draft":7975,"hidden":7975,"published":6380,"readingTime":6457,"series":8190,"seriesOrder":6398,"codeUrl":8329,"attachmentUrl":8330},"cmu-15-779-profiler-operator-batching","\u002Fposts\u002Fcmu-15-779-profiler-operator-batching","CMU 15-779 学习记录 05：Profiler、算子调用与批量推理","2026-09-15",[8190,7995,8322,8323,8324],"Profiler","批量推理","性能分析","对照 ML Systems 基础，用 Profiler 拆开两层网络的执行过程，再固定 1000 个样本扫描 batch size，观察算子调用次数、尾批和 CPU 耗时。","\u002Fimages\u002Fposts\u002Fcmu-15-779-profiler-operator-batching\u002Fcmu-pytorch-profiling-cover-20261015.webp","放大镜观察线性层与 ReLU 模块，旁边面板呈现执行片段的浅色技术教具插画","本篇对照课程官网的 Introduction to ML Systems 整理，是个人配套 CPU 实验，不是课程官方作业。脚本编号是练习顺序，不对应课程讲次。 前几篇一直在补 PyTorch 基础。到了这一步，我在学习聊天里问：现","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs\u002Ftree\u002Fcecf5110d8a95161bb5763eadfca94cd7a0cc50f","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fpytorch-profiler-batching-20261015.zip",{"slug":8332,"path":8333,"title":8334,"date":8335,"tags":8336,"description":8339,"cover":8340,"coverAlt":8341,"excerpt":8342,"draft":7975,"hidden":7975,"published":6380,"readingTime":6436,"series":8190,"seriesOrder":6384,"codeUrl":8343,"attachmentUrl":8344},"cmu-15-779-pytorch-linear-relu-absolute-value","\u002Fposts\u002Fcmu-15-779-pytorch-linear-relu-absolute-value","CMU 15-779 学习记录 04：多输出线性层、ReLU 与绝对值网络","2026-09-12",[8190,7995,8297,8337,8338],"ReLU","神经网络","把 nn.Linear 对应到矩阵乘法，弄清多个输出怎样计算，再逐层拆开一个绝对值网络，比较手动设置参数与训练得到的结果。","\u002Fimages\u002Fposts\u002Fcmu-15-779-pytorch-linear-relu-absolute-value\u002Fcmu-pytorch-absolute-cover-20261015.webp","双分支经过折角模块后汇入紫色 V 形曲线面板的浅色技术教具插画","本篇整理的是学习 CMU 15779 时补充的 PyTorch 预备练习。脚本编号不对应课程讲次，也不是课程官方作业。 接着前一篇的批量推理与矩阵乘法，我继续把 model(x) 拆开看：一层里有多个输出时，每个数是怎么算出来的？两","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs\u002Ftree\u002F4cc63904a3a583da4a2122c246a14b43aac38528","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fpytorch-linear-relu-abs-20261015.zip",{"slug":8346,"path":8347,"title":8348,"date":8349,"tags":8350,"description":8353,"cover":8354,"coverAlt":8355,"excerpt":8356,"draft":7975,"hidden":7975,"published":6380,"readingTime":6436,"series":8190,"seriesOrder":6377,"codeUrl":8357,"attachmentUrl":8358},"cmu-15-779-pytorch-dataloader-checkpoints-batching","\u002Fposts\u002Fcmu-15-779-pytorch-dataloader-checkpoints-batching","CMU 15-779 学习记录 03：DataLoader、模型保存与批量推理","2026-09-09",[8190,7995,8351,8323,8352],"DataLoader","模型保存","用 DataLoader 分批训练平方函数，保存并加载模型参数，再比较 1000 个样本与 1 个样本的推理耗时，最后把批量计算对应到矩阵乘法。","\u002Fimages\u002Fposts\u002Fcmu-15-779-pytorch-dataloader-checkpoints-batching\u002Fcmu-pytorch-batching-cover-20261015.webp","样本卡片按批次进入计算模块，旁边配有参数存储盒和输出矩阵面板的技术插画","本篇整理的是学习 CMU 15779 时补充的 PyTorch 预备练习。脚本编号不对应课程讲次，也不是课程官方作业。 接着前一篇的平方函数实验，这一节继续练习分批训练、保存参数和独立预测。最后做了一次计时：同一个小网络，逐个处理 ","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fcmu-15-779-labs","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fpytorch-batching-20261015.zip",{"slug":8360,"path":8361,"title":8362,"date":8363,"tags":8364,"description":8365,"cover":8366,"coverAlt":8367,"excerpt":8368,"draft":7975,"hidden":7975,"published":6380,"readingTime":6436,"series":8190,"seriesOrder":6363,"codeUrl":8357,"attachmentUrl":8369},"cmu-15-779-pytorch-linear-module-relu","\u002Fposts\u002Fcmu-15-779-pytorch-linear-module-relu","CMU 15-779 学习记录 02：nn.Linear、张量形状与 ReLU","2026-09-06",[8190,7995,8338,8337],"从带偏置的线性回归开始，弄清 nn.Linear 和 nn.Module 的分工，再用平方函数观察线性模型的限制、ReLU 的作用和测试点的选择。","\u002Fimages\u002Fposts\u002Fcmu-15-779-pytorch-linear-module-relu\u002Fcmu-pytorch-relu-cover-20261015.webp","输入经过两层神经网络和 ReLU 折角模块，输出曲线拟合结果的技术插画","本篇整理的是学习 CMU 15779 时补充的 PyTorch 预备练习。脚本编号不对应课程讲次，也不是课程官方作业。 上一篇里，我用一个参数拟合了 $y=2x$。这一篇继续顺着往下写，先加上偏置，再把手写参数换成 nn.Linea","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fpytorch-models-20261015.zip",{"slug":8371,"path":8372,"title":8373,"date":8374,"tags":8375,"description":8378,"cover":8379,"coverAlt":8380,"excerpt":8381,"draft":7975,"hidden":7975,"published":6380,"readingTime":6457,"series":8190,"seriesOrder":6352,"codeUrl":8357,"attachmentUrl":8382},"cmu-15-779-pytorch-autograd-linear-regression","\u002Fposts\u002Fcmu-15-779-pytorch-autograd-linear-regression","CMU 15-779 学习记录 01：PyTorch 自动求导与线性回归","2026-09-03",[8190,7995,8376,8377],"自动求导","线性回归","从自动求导和手动梯度下降开始，比较学习率，再用 SGD 训练线性回归。记录 w * 4 写错后停在 1.25 的原因和修正结果。","\u002Fimages\u002Fposts\u002Fcmu-15-779-pytorch-autograd-linear-regression\u002Fcmu-pytorch-autograd-cover-20261015.webp","梯度曲面上的小球沿路径下降，前侧面板展示散点与拟合直线的技术插画","本篇整理的是学习 CMU 15779 时补充的 PyTorch 预备练习。脚本编号不对应课程讲次，也不是课程官方作业。 开始学习 CMU 15779 时，我会 Python，但还没用过 PyTorch。这次先在 Mac 上用 CPU","\u002Fmaterials\u002Fcmu-15-779\u002Fexperiments\u002Fpytorch-basics-20261013.zip",{"slug":8384,"path":8385,"title":8386,"date":8387,"tags":8388,"description":8394,"cover":8395,"coverAlt":8396,"excerpt":8397,"draft":7975,"hidden":7975,"published":6380,"readingTime":6498,"codeUrl":8398,"attachmentUrl":8399},"embodied-pickplace-lab-robustness","\u002Fposts\u002Fembodied-pickplace-lab-robustness","Pickplace Lab 项目复盘：抓取成功率之外的时延与失败分析","2026-08-29",[8389,8390,8391,7995,8392,8393],"具身智能","模仿学习","MuJoCo","机器人","鲁棒性","从 MuJoCo 抓取搬运闭环出发，记录 22,100 个评测回合。动作平滑在自建任务的 50 ms 延迟下有效，换到 Panda Lift 却没有支持同样的收益。","\u002Fimages\u002Fposts\u002Fembodied-pickplace-lab-robustness\u002Fpickplace-grasp-cover-20261016.webp","MuJoCo 实际渲染的笛卡尔夹爪抓取方块场景，桌面上标出放置目标","Pickplace Lab 从一个简单任务开始：让夹爪抓起方块，搬到指定位置，放下，再退开。先把仿真、示范采集、训练和闭环评测接起来，随后再问：干净条件下成功的策略，换一种执行条件，还能不能完成任务？ 最值得讨论的结果出现在指令延迟","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fembodied-pickplace-lab\u002Ftree\u002Fv0.1.0","https:\u002F\u002Fgithub.com\u002Fliangqianxing\u002Fembodied-pickplace-lab\u002Freleases\u002Fdownload\u002Fv0.1.0\u002Fpickplace-lab-research-v1.zip",{"slug":8401,"path":8402,"title":8403,"date":8404,"tags":8405,"description":8409,"cover":8410,"coverAlt":8411,"excerpt":8412,"draft":7975,"hidden":7975,"published":6380,"readingTime":6769},"qwen25-vl-sglang-verl-single-gpu-lab","\u002Fposts\u002Fqwen25-vl-sglang-verl-single-gpu-lab","从 Mac 到单卡 GPU：Qwen2.5-VL 部署、GRPO 与计数 LoRA 实践","2026-08-19",[8098,8042,8406,8407,7819,8408,8244],"SGLang","verl","推理优化","在 M2 Pro 和云端单卡上部署 Qwen2.5-VL-3B，实测连续批处理、前缀缓存与 SGLang 参数调优，再用 verl 完成多模态 GRPO 更新和同条件前后评测，记录有效优化、兼容性问题与未观察到提升的结果。","\u002Fimages\u002Fposts\u002Fqwen25-vl-sglang-verl-single-gpu-lab\u002Fqvl-lab-cover-20261010.webp","笔记本与单卡计算模块通过图像和 token 光轨连接的实验台插画","这次实验从一个很具体的问题开始：M2 Pro、16 GB 统一内存，能不能在本机跑 Qwen2.5VL？跑起来之后，我又想把图文问答、推理性能测量和强化学习训练接成一条完整流程。 本机的 MLX 4bit 服务和云端的 SGLang",{"slug":8414,"path":8415,"title":8416,"date":8417,"tags":8418,"description":8422,"cover":8423,"coverAlt":8424,"excerpt":8425,"draft":7975,"hidden":7975,"published":6380,"readingTime":6498},"cycleresearcher-post-training","\u002Fposts\u002Fcycleresearcher-post-training","从 DeepReviewer 2.0 到 CycleResearcher：科研 Agent 如何真正做 Post-training","2026-08-07 17:33:05",[8084,7983,8043,8419,8420,8421],"SimPO","Reward Model","ICLR","区分推理时 Agent 编排与模型后训练，并拆解 CycleResearcher 的领域 SFT、生成式 Reward Model、Iterative SimPO 和评测风险。","\u002Fimages\u002Fposts\u002Fcycleresearcher-post-training\u002Fcycleresearcher-cover-20261018.webp","策略生成模块产出多份候选文稿经多维评审棱镜排序后通过偏好梯度反馈弧闭环优化的等距技术插画","看到一个系统会检索、规划、反思和调用工具，不能据此判断它做过 Posttraining。Agent 描述的是模型在运行时如何完成任务；Posttraining 描述的是预训练之后，是否继续用监督数据、偏好数据或奖励信号更新模型参数。",{"slug":8427,"path":8428,"title":8429,"date":8430,"tags":8431,"description":8433,"cover":8434,"coverAlt":8435,"excerpt":8436,"draft":7975,"hidden":7975,"published":6380,"readingTime":6522,"series":8437,"seriesOrder":6352},"cuda-fused-rmsnorm-kernel","\u002Fposts\u002Fcuda-fused-rmsnorm-kernel","从 PyTorch 到 CUDA Kernel：融合 RMSNorm 在 RTX 4090D 上加速 4.27 倍","2026-07-22 18:16:50",[8296,8014,8016,7995,8121,8432],"Kernel 优化","从零实现 PyTorch CUDA 扩展，拆解 RMSNorm 的访存瓶颈、warp 归约、向量化加载、算子融合与冷热缓存 benchmark。","\u002Fimages\u002Fposts\u002Fcuda-fused-rmsnorm-kernel\u002Fcuda-rmsnorm-cover-20261018.webp","融合残差加法与向量化 float4 访存经 Warp Shuffle 蝴蝶树归约输出归一化张量的等距硬件架构插画","为了准备 AI Infra 岗位，我做了一个尽量小而完整的 CUDA 项目：不追求堆很多算子，而是选一个真实的 LLM 推理路径，把实现、正确性、benchmark 和性能解释全部做扎实。 项目实现了两个 PyTorch CUDA ","AI Infra 实战",{"slug":8439,"path":8440,"title":8441,"date":8442,"tags":8443,"description":8445,"cover":8446,"coverAlt":8447,"excerpt":8448,"draft":7975,"hidden":7975,"published":6380,"readingTime":6429,"series":8449,"seriesOrder":6352},"agent-memory","\u002Fposts\u002Fagent-memory","Agent 上下文工程与分层记忆架构：从 Observation 剪枝、滚动摘要到向量召回闭环","2026-04-24",[8084,8043,8444,8014],"RAG","结合生产级多约束 Agent 与超 100K Token 科研智能体实践，系统拆解长程会话中的上下文膨胀与注意力稀释问题：涵盖 Observation 结构化剪枝（降低约 40% Token 开销）、滚动工作摘要、向量检索长期记忆及 Prefix Cache 友好的上下文组装。","\u002Fimages\u002Fposts\u002Fagent-memory\u002Fagent-memory-cover-20261018.webp","底层高噪数据经剪枝滤网进入中期滑动窗口与金色状态台账并由顶层向量星座按需召回的三层记忆架构插画","在构建多步工具调用（ToolUse）与长周期自主规划（Autonomous Planning）的 Agent 系统时，上下文管理（Context Engineering）往往比单纯改写 System Prompt 更决定线上稳定性。","Agent 上下文工程",{"slug":8451,"path":8452,"title":8453,"date":8454,"tags":8455,"description":8458,"cover":8459,"coverAlt":8460,"excerpt":8461,"draft":7975,"hidden":7975,"published":6380,"readingTime":6398,"series":8462,"seriesOrder":6363},"backend-stack-deep-dive","\u002Fposts\u002Fbackend-stack-deep-dive","高并发 AI 平台后端基础设施实践：FastAPI 异步网关、SQLAlchemy 2.0 Async、PostgreSQL 与 Docker 沙箱编排","2026-04-07",[8125,8122,8124,8456,8457,8014],"PostgreSQL","Docker","结合 DeepScientist 生产级 AI 科研平台实践，拆解 FastAPI ASGI 异步事件循环、SQLAlchemy 2.0 + AsyncPG 连接池调优、PostgreSQL + MinIO 存储分层与 Docker SDK 动态沙箱容器编排的职责边界与核心设计。","\u002Fimages\u002Fposts\u002Fbackend-stack-deep-dive\u002Fbackend-stack-cover-20261018.webp","ASGI 异步事件循环涡轮将非阻塞协程流分发至 PostgreSQL 连接池、MinIO 存储矩阵与动态 Docker 沙箱的等距架构插画","结合 DeepScientist 生产级 AI 科研平台（服务 7,000+ 用户）的实际工程落地，本文系统拆解支撑高并发流式推理与容器沙箱调度的后端基础设施栈——FastAPI（ASGI 异步网关 + SSE 流式推送）、SQLA","DeepScientist 工程复盘",{"slug":8464,"path":8465,"title":8466,"date":8467,"tags":8468,"description":8471,"cover":8472,"coverAlt":8473,"excerpt":8474,"draft":7975,"hidden":7975,"published":6380,"readingTime":6473,"series":8475,"seriesOrder":6352},"minicode-source-analysis","\u002Fposts\u002Fminicode-source-analysis","MiniCode 架构与工具评测深度拆解：开源终端 Agentic Coding 引擎（1k+ Stars）的 ReAct、MCP 与 30 项基准验证","2026-04-06",[8043,8469,8470,8073,8014],"TypeScript","MCP","深度剖析开源终端 AI 编程智能体 MiniCode（1k+ GitHub Stars）的核心架构：从零重型依赖的 ReAct 规划循环、声明式工具注册表与权限拦截层、MCP 协议帧自动协商，到 30 项可执行 Tool-Use 评测基准（86.7% 完成率、100% 错误恢复率）的失败分布归因与闭环优化。","\u002Fimages\u002Fposts\u002Fminicode-source-analysis\u002Fminicode-cover-20261018.webp","终端编程智能体控制台被 ReAct 状态环、Unified Diff 补丁预览门与 MCP 协议插槽环绕的等距技术插画","MiniCode 是我作为核心开发者参与构建的开源终端 Agentic Coding 引擎（目前已在 GitHub 获得 1,000+ Stars）。项目的核心设计目标是：在不引入庞大运行时框架的前提下，用约 5,000 行高内聚的","AI 项目源码导读",{"slug":8477,"path":8478,"title":8479,"date":8467,"tags":8480,"description":8481,"cover":8482,"coverAlt":8483,"excerpt":8484,"draft":7975,"hidden":7975,"published":6380,"readingTime":6457,"series":8462,"seriesOrder":6352},"deepscientist-tech-stack","\u002Fposts\u002Fdeepscientist-tech-stack","DeepScientist 架构与运行时全解析：超 100K 上下文科研智能体、MCP 沙箱与高并发流式系统",[8043,8014,8470,8457,8124,8122],"深度拆解服务 7,000+ 用户的科研自主智能体平台 DeepScientist：从一题一仓库（Quest）持久化状态机、MCP 工具栈（memory \u002F artifact \u002F bash_exec）、Docker SDK 动态容器编排，到 SSE + WebSocket 双流通信与 PostgreSQL \u002F Redis \u002F Prometheus 生产级基础设施。","\u002Fimages\u002Fposts\u002Fdeepscientist-tech-stack\u002Fdeepscientist-cover-20261018.webp","科研智能体编排核心通过 memory、artifact 与隔离容器沙箱三座协议桥连接双流工作台的等距架构插画","DeepScientist（deepscientist.cc）是我们在西湖大学自然语言处理实验室研发的自主科学发现与科研管理平台，目前累计服务 7,000+ 科研用户。与单轮对话或纯文本文献问答工具不同，科研智能体需要在数小时乃至数",1795088944385]