{"id":"https://openalex.org/W7161785985","doi":"https://doi.org/10.48550/arxiv.2605.19537","title":"The Silent Hyperparameter: Quantifying the Impact of Inference Backends on LLM Reproducibility","display_name":"The Silent Hyperparameter: Quantifying the Impact of Inference Backends on LLM Reproducibility","publication_year":2026,"publication_date":"2026-05-19","ids":{"openalex":"https://openalex.org/W7161785985","doi":"https://doi.org/10.48550/arxiv.2605.19537"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.19537","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.19537","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"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"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2605.19537","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5111667912","display_name":"David Pape","orcid":"https://orcid.org/0000-0002-3145-9880"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Pape, David","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136585608","display_name":"Jonathan Evertz","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Evertz, Jonathan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5049219646","display_name":"Lea Sch\u00f6nherr","orcid":"https://orcid.org/0000-0003-3779-7781"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sch\u00f6nherr, Lea","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"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/T10181","display_name":"Natural Language Processing Techniques","score":0.25459998846054077,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"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/T10181","display_name":"Natural Language Processing Techniques","score":0.25459998846054077,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"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/T12127","display_name":"Software System Performance and Reliability","score":0.07440000027418137,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/T10028","display_name":"Topic Modeling","score":0.0632999986410141,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"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/inference","display_name":"Inference","score":0.8586000204086304},{"id":"https://openalex.org/keywords/benchmark","display_name":"Benchmark (surveying)","score":0.7325000166893005},{"id":"https://openalex.org/keywords/interpretability","display_name":"Interpretability","score":0.46939998865127563},{"id":"https://openalex.org/keywords/predictive-inference","display_name":"Predictive inference","score":0.36469998955726624},{"id":"https://openalex.org/keywords/fiducial-inference","display_name":"Fiducial inference","score":0.3564999997615814},{"id":"https://openalex.org/keywords/adaptive-neuro-fuzzy-inference-system","display_name":"Adaptive neuro fuzzy inference system","score":0.33059999346733093}],"concepts":[{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.8586000204086304},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7649000287055969},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.7325000166893005},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5539000034332275},{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.46939998865127563},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.42340001463890076},{"id":"https://openalex.org/C917703","wikidata":"https://www.wikidata.org/wiki/Q7239668","display_name":"Predictive inference","level":5,"score":0.36469998955726624},{"id":"https://openalex.org/C95167961","wikidata":"https://www.wikidata.org/wiki/Q4483495","display_name":"Fiducial inference","level":5,"score":0.3564999997615814},{"id":"https://openalex.org/C186108316","wikidata":"https://www.wikidata.org/wiki/Q352530","display_name":"Adaptive neuro fuzzy inference system","level":4,"score":0.33059999346733093},{"id":"https://openalex.org/C160234255","wikidata":"https://www.wikidata.org/wiki/Q812535","display_name":"Bayesian inference","level":3,"score":0.3199999928474426},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.29910001158714294},{"id":"https://openalex.org/C43364308","wikidata":"https://www.wikidata.org/wiki/Q8799","display_name":"Byte","level":2,"score":0.289900004863739},{"id":"https://openalex.org/C2779458634","wikidata":"https://www.wikidata.org/wiki/Q24963715","display_name":"Debiasing","level":2,"score":0.2694000005722046},{"id":"https://openalex.org/C2777472644","wikidata":"https://www.wikidata.org/wiki/Q16968992","display_name":"Approximate inference","level":3,"score":0.2551000118255615}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.19537","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.19537","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2605.19537","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.19537","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"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":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Progress":[0],"in":[1,203,219],"LLMs":[2,31],"is":[3,100,187],"increasingly":[4],"measured":[5],"through":[6],"standardized":[7,226],"benchmarks,":[8],"where":[9],"state-of-the-art":[10],"improvements":[11],"are":[12],"often":[13],"separated":[14],"by":[15,161,189],"fractions":[16],"of":[17,28,36,114,154,171,222,228,237],"a":[18,110,213],"percentage":[19,165],"point.":[20],"At":[21],"the":[22,25,82,96,152,180,209,220,233],"same":[23],"time,":[24],"computational":[26],"cost":[27],"evaluating":[29],"modern":[30],"has":[32],"driven":[33,188],"widespread":[34,105],"adoption":[35],"specialized":[37],"inference":[38,48,83,98,116,135,210,229],"backends,":[39],"software":[40],"systems":[41],"that":[42,95,151],"execute":[43],"trained":[44],"models":[45,145],"efficiently":[46],"at":[47],"time.":[49],"While":[50],"critical":[51],"for":[52],"scalability,":[53],"system-level":[54,190],"optimizations,":[55],"such":[56],"as":[57,212],"custom":[58,198],"CUDA":[59,196],"kernels":[60],"and":[61,68,89,127,140,146,167,178,195,200,224,235],"reduced-precision":[62],"arithmetic,":[63],"can":[64,157],"alter":[65],"token":[66],"probabilities":[67],"introduce":[69],"non-determinism,":[70],"possibly":[71],"cascading":[72],"into":[73],"divergent":[74],"generation.":[75],"In":[76],"this":[77,104,185],"work,":[78],"we":[79,130,183],"first":[80],"survey":[81],"landscape,":[84],"identifying":[85],"200":[86],"distinct":[87],"engines,":[88,136],"analyze":[90],"35,000":[91],"ML":[92],"publications,":[93],"finding":[94],"specific":[97],"stack":[99],"rarely":[101],"reported":[102],"despite":[103],"diversity.":[106],"We":[107,149],"then":[108],"present":[109],"systematic":[111],"empirical":[112],"study":[113],"how":[115],"backends":[117],"affect":[118],"LLM":[119,223],"benchmark":[120,159,238],"results.":[121],"Holding":[122],"model":[123],"weights,":[124],"decoding":[125],"parameters,":[126],"hardware":[128],"constant,":[129],"evaluate":[131],"five":[132],"widely":[133],"used":[134],"including":[137],"vLLM,":[138],"SGLang,":[139],"llama$.$cpp,":[141],"across":[142],"multiple":[143],"open-weight":[144],"established":[147],"benchmarks.":[148],"show":[150],"choice":[153],"backend":[155,176,211],"alone":[156],"shift":[158],"scores":[160],"up":[162],"to":[163,231],"16.6":[164],"points":[166],"induce":[168],"high":[169],"rates":[170],"output":[172],"disagreement.":[173],"By":[174],"isolating":[175],"optimizations":[177,191],"tracing":[179],"execution":[181],"pipeline,":[182],"find":[184],"divergence":[186],"like":[192],"prefix":[193],"caching":[194],"graphs,":[197],"kernels,":[199],"engine-specific":[201],"defaults":[202],"logit":[204],"processing.":[205],"Our":[206],"findings":[207],"identify":[208],"previously":[214],"unreported":[215],"but":[216],"consequential":[217],"hyperparameter":[218],"evaluation":[221],"advocate":[225],"reporting":[227],"stacks":[230],"improve":[232],"reproducibility":[234],"interpretability":[236],"comparisons.":[239]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-05-21T00:00:00"}