{"id":"https://openalex.org/W7160286712","doi":"https://doi.org/10.48550/arxiv.2605.02363","title":"When Correct Isn't Usable: Improving Structured Output Reliability in Small Language Models","display_name":"When Correct Isn't Usable: Improving Structured Output Reliability in Small Language Models","publication_year":2026,"publication_date":"2026-05-04","ids":{"openalex":"https://openalex.org/W7160286712","doi":"https://doi.org/10.48550/arxiv.2605.02363"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.02363","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.02363","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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.2605.02363","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5073361097","display_name":"Cosimo Galeone","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Galeone, Cosimo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100646726","display_name":"Minsu Park","orcid":"https://orcid.org/0000-0002-3995-4122"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Park, Minsu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5103473918","display_name":"Giuseppe Ettorre","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ettorre, Giuseppe","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5012287313","display_name":"Daniele Ligorio","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ligorio, Daniele","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/T11636","display_name":"Artificial Intelligence in Healthcare and Education","score":0.3416000008583069,"subfield":{"id":"https://openalex.org/subfields/2718","display_name":"Health Informatics"},"field":{"id":"https://openalex.org/fields/27","display_name":"Medicine"},"domain":{"id":"https://openalex.org/domains/4","display_name":"Health Sciences"}},"topics":[{"id":"https://openalex.org/T11636","display_name":"Artificial Intelligence in Healthcare and Education","score":0.3416000008583069,"subfield":{"id":"https://openalex.org/subfields/2718","display_name":"Health Informatics"},"field":{"id":"https://openalex.org/fields/27","display_name":"Medicine"},"domain":{"id":"https://openalex.org/domains/4","display_name":"Health Sciences"}},{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.13699999451637268,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.08760000020265579,"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/correctness","display_name":"Correctness","score":0.5720999836921692},{"id":"https://openalex.org/keywords/reliability","display_name":"Reliability (semiconductor)","score":0.4950000047683716},{"id":"https://openalex.org/keywords/latency","display_name":"Latency (audio)","score":0.39250001311302185},{"id":"https://openalex.org/keywords/json","display_name":"JSON","score":0.38929998874664307},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.37610000371932983},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.3702999949455261},{"id":"https://openalex.org/keywords/ground-truth","display_name":"Ground truth","score":0.36800000071525574},{"id":"https://openalex.org/keywords/task","display_name":"Task (project management)","score":0.3571999967098236}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7222999930381775},{"id":"https://openalex.org/C55439883","wikidata":"https://www.wikidata.org/wiki/Q360812","display_name":"Correctness","level":2,"score":0.5720999836921692},{"id":"https://openalex.org/C43214815","wikidata":"https://www.wikidata.org/wiki/Q7310987","display_name":"Reliability (semiconductor)","level":3,"score":0.4950000047683716},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.39250001311302185},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.39239999651908875},{"id":"https://openalex.org/C2780416260","wikidata":"https://www.wikidata.org/wiki/Q2063","display_name":"JSON","level":2,"score":0.38929998874664307},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.37610000371932983},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3702999949455261},{"id":"https://openalex.org/C146849305","wikidata":"https://www.wikidata.org/wiki/Q370766","display_name":"Ground truth","level":2,"score":0.36800000071525574},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.3571999967098236},{"id":"https://openalex.org/C169590947","wikidata":"https://www.wikidata.org/wiki/Q47506","display_name":"Compiler","level":2,"score":0.3564000129699707},{"id":"https://openalex.org/C2780513914","wikidata":"https://www.wikidata.org/wiki/Q18210350","display_name":"Bottleneck","level":2,"score":0.322299987077713},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.3192000091075897},{"id":"https://openalex.org/C39927690","wikidata":"https://www.wikidata.org/wiki/Q11197","display_name":"Logarithm","level":2,"score":0.31310001015663147},{"id":"https://openalex.org/C28855332","wikidata":"https://www.wikidata.org/wiki/Q198099","display_name":"Quantization (signal processing)","level":2,"score":0.28870001435279846},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.2854999899864197},{"id":"https://openalex.org/C186644900","wikidata":"https://www.wikidata.org/wiki/Q194152","display_name":"Parsing","level":2,"score":0.2671999931335449},{"id":"https://openalex.org/C2780505938","wikidata":"https://www.wikidata.org/wiki/Q17093282","display_name":"Unavailability","level":2,"score":0.2639999985694885},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2639000117778778},{"id":"https://openalex.org/C150189527","wikidata":"https://www.wikidata.org/wiki/Q356674","display_name":"Reference model","level":2,"score":0.2506999969482422}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.02363","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.02363","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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.2605.02363","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.02363","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"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":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Deployed":[0],"language":[1],"models":[2,49,98,120],"must":[3],"produce":[4],"outputs":[5],"that":[6,262],"are":[7],"both":[8],"correct":[9],"and":[10,24,35,54,65,99,132,171,200,250],"format-compliant.":[11],"We":[12,45],"study":[13],"this":[14,142],"structured-output":[15],"reliability":[16],"gap":[17],"using":[18],"two":[19,117],"mathematical":[20,63],"benchmarks":[21],"--":[22,26,58,69],"GSM8K":[23,91,170],"MATH":[25,174],"as":[27,70],"a":[28,213,266],"controlled":[29],"testbed:":[30],"ground":[31],"truth":[32],"is":[33,39,265],"unambiguous":[34],"the":[36,59,71,161,187,236],"output":[37,56,94,114,167,220,246],"contract":[38],"strict":[40],"(JSON":[41],"with":[42,181,240],"required":[43],"fields).":[44],"evaluate":[46],"three":[47],"7-9B":[48],"under":[50],"five":[51,176],"prompting":[52,80,102],"strategies":[53],"report":[55],"accuracy":[57,89,95,115,168,221,247],"joint":[60],"event":[61],"of":[62,118,269],"correctness":[64],"valid":[66],"JSON":[67,106],"structure":[68],"primary":[72,267],"metric.":[73],"A":[74],"systematic":[75,224],"format":[76,107,206],"failure":[77,207],"emerges:":[78],"NAIVE":[79],"(no":[81],"system":[82],"prompt)":[83,108],"achieves":[84,218],"up":[85],"to":[86,160,209,223,248,258],"85%":[87],"task":[88,137],"on":[90,169,173],"but":[92,127],"0%":[93,113,219],"across":[96,175],"all":[97],"datasets.":[100],"REFERENCE":[101,217],"(a":[103],"minimal":[104],"hand-written":[105],"fares":[109],"little":[110],"better,":[111],"yielding":[112],"for":[116],"four":[119],"tested.":[121],"Constrained":[122],"decoding":[123],"enforces":[124],"syntactic":[125],"validity":[126],"incurs":[128],"3.6x-8.2x":[129],"latency":[130,199],"overhead":[131],"in":[133],"several":[134],"settings":[135],"degrades":[136],"performance":[138],"substantially.":[139],"To":[140],"overcome":[141],"limitation,":[143],"we":[144],"developed":[145],"AloLab,":[146],"an":[147],"iterative":[148],"system-prompt":[149],"optimizer":[150],"(meta-agent:":[151],"Claude":[152,241],"Sonnet":[153,237],"4.5)":[154],"requiring":[155],"only":[156],"black-box":[157],"API":[158],"access":[159],"target":[162],"model;":[163],"it":[164],"reaches":[165,229],"84-87%":[166],"34-40%":[172],"independent":[177],"runs":[178],"per":[179],"model,":[180],"29/30":[182],"paired":[183],"McNemar":[184],"comparisons":[185],"against":[186],"best":[188],"static":[189],"prompt":[190],"significant":[191],"at":[192,196],"p":[193],"<":[194],"0.05,":[195],"near-NAIVE":[197],"inference":[198],"without":[201],"model":[202],"fine-tuning.":[203],"The":[204],"same":[205],"extends":[208],"GPT-4o":[210],"(OpenAI,":[211],"2024),":[212],"proprietary":[214],"closed-source":[215],"model:":[216],"due":[222],"markdown-fence":[225],"wrapping,":[226],"while":[227],"AloLab":[228],"95.2%":[230],"[94.8,":[231],"95.6].":[232],"An":[233],"ablation":[234],"replacing":[235],"4.5":[238],"meta-agent":[239,263],"3":[242],"Haiku":[243],"reduces":[244],"mean":[245],"61.0%":[249],"increases":[251],"run-to-run":[252],"standard":[253],"deviation":[254],"from":[255],"<1":[256],"pp":[257],"21.8":[259],"pp,":[260],"confirming":[261],"capability":[264],"driver":[268],"optimization":[270],"quality.":[271]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-05-06T00:00:00"}