{"id":"https://openalex.org/W7225338191","doi":"https://doi.org/10.5281/zenodo.23990731","title":"Same Input, Different Experts: Tiny Arithmetic Noise Flips Expert Routing, and a Routing-Conditional Test Cannot See It","display_name":"Same Input, Different Experts: Tiny Arithmetic Noise Flips Expert Routing, and a Routing-Conditional Test Cannot See It","publication_year":2026,"publication_date":"2026-11-06","ids":{"openalex":"https://openalex.org/W7225338191","doi":"https://doi.org/10.5281/zenodo.23990731"},"language":"en","primary_location":{"id":"doi:10.5281/zenodo.23990731","doi":"https://doi.org/10.5281/zenodo.23990731","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23990731","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.5281/zenodo.23990731","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5160746184","display_name":"Harshil Jani","orcid":"https://orcid.org/0009-0009-7397-9428","observed_orcids":["https://orcid.org/0009-0009-7397-9428"]},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Jani, Harshil","raw_affiliation_strings":["Independent Researcher"],"raw_orcid":"https://orcid.org/0009-0009-7397-9428","affiliations":[{"raw_affiliation_string":"Independent Researcher","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":["https://openalex.org/A5160746184"],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11697","display_name":"Numerical Methods and Algorithms","score":0.9520999789237976,"subfield":{"id":"https://openalex.org/subfields/1703","display_name":"Computational Theory and Mathematics"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T11697","display_name":"Numerical Methods and Algorithms","score":0.9520999789237976,"subfield":{"id":"https://openalex.org/subfields/1703","display_name":"Computational Theory and Mathematics"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10743","display_name":"Software Testing and Debugging Techniques","score":0.01510000042617321,"subfield":{"id":"https://openalex.org/subfields/1712","display_name":"Software"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.00570000009611249,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/mixture-of-experts","display_name":"mixture of experts","score":0.8679999709129333},{"id":"https://openalex.org/keywords/expert-routing","display_name":"expert routing","score":0.871999979019165},{"id":"https://openalex.org/keywords/floating-point-precision","display_name":"floating-point precision","score":0.6399999856948853},{"id":"https://openalex.org/keywords/correctness-testing","display_name":"correctness testing","score":0.3799999952316284},{"id":"https://openalex.org/keywords/bfloat16","display_name":"bfloat16","score":0.5230000019073486}],"concepts":[{"id":"https://openalex.org/C55439883","wikidata":"https://www.wikidata.org/wiki/Q360812","display_name":"Correctness","level":2,"score":0.7921000123023987},{"id":"https://openalex.org/C136625980","wikidata":"https://www.wikidata.org/wiki/Q663208","display_name":"Rounding","level":2,"score":0.7261999845504761},{"id":"https://openalex.org/C94375191","wikidata":"https://www.wikidata.org/wiki/Q11205","display_name":"Arithmetic","level":1,"score":0.6668999791145325},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.6384999752044678},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.6355000138282776},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6103000044822693},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.593999981880188},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.4943999946117401},{"id":"https://openalex.org/C74193536","wikidata":"https://www.wikidata.org/wiki/Q574844","display_name":"Kernel (algebra)","level":2,"score":0.4691999852657318},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.3637000024318695},{"id":"https://openalex.org/C169903167","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Test set","level":2,"score":0.32010000944137573},{"id":"https://openalex.org/C2775896111","wikidata":"https://www.wikidata.org/wiki/Q642560","display_name":"Router","level":2,"score":0.313400000333786},{"id":"https://openalex.org/C2779227376","wikidata":"https://www.wikidata.org/wiki/Q6505497","display_name":"Layer (electronics)","level":2,"score":0.30410000681877136},{"id":"https://openalex.org/C182306322","wikidata":"https://www.wikidata.org/wiki/Q1779371","display_name":"Order (exchange)","level":2,"score":0.2948000133037567},{"id":"https://openalex.org/C149810388","wikidata":"https://www.wikidata.org/wiki/Q5374873","display_name":"Emulation","level":2,"score":0.27149999141693115},{"id":"https://openalex.org/C2777267654","wikidata":"https://www.wikidata.org/wiki/Q3519023","display_name":"Test (biology)","level":2,"score":0.2619999945163727},{"id":"https://openalex.org/C43521106","wikidata":"https://www.wikidata.org/wiki/Q2165493","display_name":"Pipeline (software)","level":2,"score":0.25279998779296875}],"mesh":[],"locations_count":2,"locations":[{"id":"doi:10.5281/zenodo.23990731","doi":"https://doi.org/10.5281/zenodo.23990731","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23990731","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},{"id":"doi:10.5281/zenodo.23990732","doi":"https://doi.org/10.5281/zenodo.23990732","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23990732","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.5281/zenodo.23990731","doi":"https://doi.org/10.5281/zenodo.23990731","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23990731","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7202353093","https://openalex.org/W7204553230","https://openalex.org/W7172000585","https://openalex.org/W7118005390","https://openalex.org/W4391713099","https://openalex.org/W7167514145","https://openalex.org/W7222852282","https://openalex.org/W7160847478","https://openalex.org/W7171784360","https://openalex.org/W7155602352"],"abstract_inverted_index":{"Many":[0],"AI":[1],"models":[2],"send":[3],"each":[4,34,98],"token,":[5,91],"a":[6,9,19,57,107,194],"piece":[7],"of":[8,13,24,121,130,209],"word,":[10],"to":[11,128,172,189],"8":[12,88],"256":[14],"\"expert\"":[15],"sub-networks":[16],"chosen":[17,89],"by":[18,133,170],"router.":[20],"The":[21,95],"correctness":[22,179],"tests":[23,180],"the":[25,37,46,114,122,149,166,187,215,228],"open":[26],"mixture-of-experts":[27],"(MoE)":[28],"kernel":[29],"benchmark":[30],"we":[31,145],"used":[32],"check":[33],"run":[35,40],"against":[36],"experts":[38,51],"that":[39,65],"picked,":[41],"so":[42],"they":[43],"cannot":[44],"see":[45],"same":[47,123],"input":[48],"picking":[49],"different":[50,53],"on":[52,67,76,162,224],"runs.":[54],"We":[55,70],"found":[56],"fast":[58],"multiply":[59,96,185],"whose":[60,219],"last-digit":[61],"noise":[62],"grew":[63,151],"until":[64],"happened":[66],"every":[68,163],"run.":[69],"measured":[71],"one":[72,77,134,156],"MoE":[73],"decoder":[74],"layer":[75,167],"NVIDIA":[78],"H200":[79],"GPU,":[80],"with":[81],"synthetic":[82],"weights":[83],"and":[84,103,142,165],"router":[85],"(256":[86],"experts,":[87],"per":[90],"FP8":[92,140],"attention":[93],"inputs).":[94],"splits":[97],"sum":[99],"into":[100,106],"7":[101],"slices":[102],"adds":[104],"them":[105],"16-bit":[108],"(BF16)":[109],"output":[110,168],"in":[111],"whatever":[112],"order":[113],"hardware":[115],"finishes.":[116],"On":[117],"all":[118,177],"500":[119],"replays":[120],"input,":[124],"it":[125],"changed":[126,161],"up":[127],"56%":[129],"its":[131],"outputs":[132],"BF16":[135],"step":[136],"(one":[137],"ulp).":[138],"Through":[139],"rounding":[141],"attention,":[143],"which":[144],"did":[146],"not":[147,222],"separate,":[148],"difference":[150,188],"about":[152],"tenfold.":[153],"At":[154],"least":[155],"token's":[157],"expert":[158],"set":[159],"then":[160],"replay,":[164],"differed":[169],"11":[171],"17%":[173],"between":[174,206],"replays,":[175],"yet":[176],"13":[178],"passed.":[181],"An":[182],"FP32,":[183],"fixed-order":[184],"cut":[186],"at":[190,193],"most":[191],"0.063%,":[192],"whole-layer":[195],"cost":[196],"below":[197],"our":[198],"timing":[199],"resolution.":[200],"Near-tie":[201],"flips":[202],"still":[203],"remained,":[204],"even":[205],"two":[207],"copies":[208],"identical":[210],"code.":[211],"Higher":[212],"precision":[213],"shrinks":[214],"noise;":[216],"only":[217],"sums":[218],"result":[220],"does":[221],"depend":[223],"order,":[225],"everywhere":[226],"before":[227],"router,":[229],"remove":[230],"it.":[231]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-11-06T00:00:00"}
