Attention Has a Coordinate but No Commitment: The Selection Gap in Softmax Attention, with Two Measurements on GPT-2

{"Summary.":[0],"Softmax":[1],"attention":[2,92,346],"is":[3,49,55,81,97,185,255,268,307],"often":[4],"described":[5],"as":[6,22,403],"\\"smooth\\":":[7],"every":[8],"unmasked":[9],"key":[10],"receives":[11],"nonzero":[12],"weight.":[13],"An":[14],"earlier":[15,187,331],"paper":[16,37,324,337],"by":[17,51,217,234,309,319],"this":[18],"author":[19],"took":[20],"that":[21,291,302,336,372],"the":[23,47,59,67,103,113,118,159,163,167,186,212,225,228,240,269,284,305,323,330,333,354,370,392,413,415,417,420,425],"structural":[24],"origin":[25],"of":[26,62,105,121,125,132,151,166,198,215,279,358],"hallucination":[27,352],"and":[28,46,75,193,201,204,222,242,246,276,314,343,350,369,400,424],"proposed":[29],"a":[30,63,71,76,82,86,95,98,106,129,149,173,179,207,263,292,359,366],"learned":[31],"hard":[32,180],"threshold":[33,96],"per":[34],"head.":[35],"This":[36],"supersedes":[38],"it.":[39,375],"The":[40,79,91,145,288,326,410],"diagnosis":[41],"was":[42,386,407],"wrong":[43],"in":[44,123,137,142,162,365],"mechanism":[45],"remedy":[48],"contradicted":[50],"measurement;":[52],"what":[53],"survives":[54],"narrower:":[56],"softmax":[57],"supplies":[58],"shared":[60],"coordinate":[61],"selection":[64,83,363],"but":[65],"not":[66,85,256,318],"two":[68,426],"other":[69],"parts,":[70],"commitment":[72],"with":[73,224,247,362,381],"memory":[74,316],"persistent":[77],"label.":[78],"gap":[80],"gap,":[84],"smoothness":[87],"gap.":[88],"Measurement":[89,170],"M1.":[90],"mass":[93,230],"below":[94],"shared,":[99],"common-mode":[100,114],"quantity":[101],"across":[102],"heads":[104],"layer":[107,138,143],"rather":[108],"than":[109,298],"per-head":[110,181],"structure.":[111],"Position-detrended,":[112],"variance":[115],"fraction":[116],"exceeds":[117],"registered":[119,147,289,397],"bar":[120],"0.2":[122],"10":[124],"12":[126],"layers":[127],"against":[128],"shuffle":[130],"null":[131],"0.083,":[133],"rising":[134],"from":[135,262,329],"0.09":[136],"0":[139],"to":[140],"0.48":[141],"11.":[144],"stronger":[146],"form,":[148],"median":[150],"at":[152,156,196,236,239,250,283,301],"least":[153],"0.3,":[154],"fails":[155],"0.278,":[157],"so":[158,253,304],"reading":[160],"holds":[161],"deep":[164],"half":[165],"network":[168],"only.":[169],"M2.":[171],"At":[172],"matched":[174],"retained-key":[175],"budget,":[176,252,303],"closed":[177],"loop,":[178],"level":[182,293],"threshold,":[183],"which":[184,267],"paper's":[188],"proposal,":[189],"costs":[190,219,273],"38.4,":[191],"4.0":[192],"0.1":[194],"perplexity":[195,275],"budgets":[197,245],"10,":[199],"20":[200,241,285],"40":[202,243],"percent":[203,244,286],"never":[205],"improves":[206],"named-entity":[208],"continuation":[209],"proxy.":[210],"Selecting":[211],"same":[213],"number":[214],"keys":[216],"rank":[218,235],"0.5,":[220],"-0.1":[221,223],"proxy":[226],"unchanged:":[227],"low-attention":[229],"can":[231],"be":[232],"removed":[233],"no":[237,248,405],"cost":[238,296],"benefit":[249],"any":[251],"it":[254],"where":[257],"misplaced":[258],"content":[259],"lives.":[260],"Eviction":[261],"leaky,":[264],"read-refreshed":[265],"store,":[266],"constructive":[270],"proposal's":[271],"memory,":[272],"7.9":[274],"five":[277],"points":[278],"in-context":[280],"name":[281],"recall":[282],"budget.":[287],"prediction":[290],"gate":[294],"would":[295,373],"more":[297],"eviction":[299],"failed":[300,399],"store":[306],"justified":[308],"its":[310,315],"persistent,":[311],"erasable":[312],"label":[313],"footprint,":[317],"perplexity.":[320],"What":[321],"else":[322],"states.":[325],"claims":[327],"withdrawn":[328],"paper;":[332],"prior":[334],"art":[335],"omitted,":[338],"namely":[339],"sparsemax,":[340],"entmax,":[341],"rectified":[342],"top-k":[344],"attention,":[345],"sinks,":[347],"shared-workspace":[348],"transformers":[349],"calibration-based":[351],"bounds;":[353],"quantitative":[355],"building":[356],"blocks":[357],"store-plus-rank-selector":[360],"architecture":[361],"run":[364],"separate":[367],"phase;":[368],"measurement":[371,418,427],"falsify":[374],"Method.":[376],"Both":[377],"measurements":[378],"were":[379],"pre-registered":[380],"falsification":[382],"criteria":[383,398],"before":[384],"either":[385],"run,":[387],"on":[388],"GPT-2":[389],"small":[390],"over":[391],"wikitext-2-raw-v1":[393],"validation":[394],"split.":[395],"Two":[396],"are":[401],"reported":[402],"failures;":[404],"criterion":[406],"adjusted":[408],"afterwards.":[409],"deposit":[411],"contains":[412],"paper,":[414],"pre-registration,":[416],"code,":[419],"raw":[421],"result":[422],"files":[423],"reports.":[428]}