{"id":"https://openalex.org/W7162119740","doi":"https://doi.org/10.48550/arxiv.2605.21602","title":"Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs","display_name":"Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs","publication_year":2026,"publication_date":"2026-05-20","ids":{"openalex":"https://openalex.org/W7162119740","doi":"https://doi.org/10.48550/arxiv.2605.21602"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.21602","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.21602","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"article"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2605.21602","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5120775726","display_name":"Dylan Feng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Feng, Dylan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136750496","display_name":"Pragya Srivastava","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Srivastava, Pragya","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5051435362","display_name":"Cassidy Laidlaw","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Dragan, Anca","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":null,"display_name":"Laidlaw, Cassidy","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Laidlaw, Cassidy","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":4,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.544700026512146,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.544700026512146,"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/T10028","display_name":"Topic Modeling","score":0.16099999845027924,"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":0.04129999876022339,"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/guard","display_name":"Guard (computer science)","score":0.659500002861023},{"id":"https://openalex.org/keywords/benchmarking","display_name":"Benchmarking","score":0.6169999837875366},{"id":"https://openalex.org/keywords/benchmark","display_name":"Benchmark (surveying)","score":0.3499999940395355},{"id":"https://openalex.org/keywords/component","display_name":"Component (thermodynamics)","score":0.32910001277923584},{"id":"https://openalex.org/keywords/set","display_name":"Set (abstract data type)","score":0.3190000057220459},{"id":"https://openalex.org/keywords/precision-and-recall","display_name":"Precision and recall","score":0.3156000077724457}],"concepts":[{"id":"https://openalex.org/C141141315","wikidata":"https://www.wikidata.org/wiki/Q2379942","display_name":"Guard (computer science)","level":2,"score":0.659500002861023},{"id":"https://openalex.org/C86251818","wikidata":"https://www.wikidata.org/wiki/Q816754","display_name":"Benchmarking","level":2,"score":0.6169999837875366},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5386000275611877},{"id":"https://openalex.org/C200601418","wikidata":"https://www.wikidata.org/wiki/Q2193887","display_name":"Reliability engineering","level":1,"score":0.44449999928474426},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.35580000281333923},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.3499999940395355},{"id":"https://openalex.org/C168167062","wikidata":"https://www.wikidata.org/wiki/Q1117970","display_name":"Component (thermodynamics)","level":2,"score":0.32910001277923584},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.3190000057220459},{"id":"https://openalex.org/C81669768","wikidata":"https://www.wikidata.org/wiki/Q2359161","display_name":"Precision and recall","level":2,"score":0.3156000077724457},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.31279999017715454},{"id":"https://openalex.org/C38652104","wikidata":"https://www.wikidata.org/wiki/Q3510521","display_name":"Computer security","level":1,"score":0.2766000032424927},{"id":"https://openalex.org/C169903167","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Test set","level":2,"score":0.27219998836517334},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.26750001311302185},{"id":"https://openalex.org/C46686674","wikidata":"https://www.wikidata.org/wiki/Q466303","display_name":"Boosting (machine learning)","level":2,"score":0.2648000121116638},{"id":"https://openalex.org/C67186912","wikidata":"https://www.wikidata.org/wiki/Q367664","display_name":"Data modeling","level":2,"score":0.2506999969482422}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.21602","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.21602","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"article"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2605.21602","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.21602","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"article"},"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":{"Many":[0],"safety":[1,66],"and":[2,135,147,174,214,229,235],"alignment":[3,38,95],"failures":[4,39,55,96],"of":[5,132,140,211],"large":[6],"language":[7],"models":[8,62,109,124],"(LLMs)":[9],"occur":[10],"due":[11],"to":[12,53,82,114,156],"out-of-distribution":[13],"(OOD)":[14],"situations:":[15],"unusual":[16],"prompt":[17],"or":[18],"response":[19],"patterns":[20],"that":[21,56,79,97,107,137,169,179,203],"are":[22,57,98],"unforeseen":[23],"by":[24,40,71],"model":[25,143,165,173,194],"developers.":[26],"We":[27,68,128,158,225],"systematically":[28],"study":[29],"whether":[30],"LLM":[31,212],"monitoring":[32,184,213],"pipelines":[33],"can":[34,151,237],"detect":[35],"these":[36],"OOD":[37,59,126,133,149,175,181,204],"introducing":[41],"a":[42,73,138,141,171,186,192,208,216],"benchmark":[43],"called":[44],"Misalignment":[45],"Out":[46],"Of":[47],"Distribution":[48],"(MOOD).":[49],"It":[50],"is":[51],"difficult":[52],"find":[54,106,136,178,238],"truly":[58],"for":[60,167,218,231],"off-the-shelf":[61],"trained":[63],"on":[64,221],"vast":[65],"datasets.":[67],"sidestep":[69],"this":[70,222],"including":[72],"restricted":[74],"training":[75,101],"set":[76],"in":[77],"MOOD":[78],"we":[80,105,120,177],"use":[81],"train":[83],"our":[84,232],"own":[85],"monitors,":[86],"as":[87,89],"well":[88],"seven":[90],"test":[91,129],"sets":[92],"with":[93,125,144,195],"diverse":[94],"outside":[99],"the":[100,227,239],"distribution.":[102],"Using":[103],"MOOD,":[104],"guard":[108,123,142,172,193],"(safety":[110],"classifiers)":[111],"often":[112],"fail":[113],"generalize":[115],"OOD.":[116],"To":[117],"fix":[118],"this,":[119],"propose":[121],"combining":[122],"detectors.":[127],"four":[130],"types":[131],"detectors":[134,150],"combination":[139],"Mahalanobis":[145],"distance":[146],"perplexity-based":[148],"improve":[152],"recall":[153,188],"from":[154],"39%":[155],"45%.":[157],"also":[159],"establish":[160],"positive":[161],"scaling":[162],"trends":[163],"across":[164],"scales":[166],"monitors":[168],"combine":[170],"detector;":[176],"incorporating":[180],"detection":[182,205],"into":[183],"achieves":[185],"higher":[187],"gain":[189],"than":[190],"using":[191],"20":[196],"times":[197],"more":[198],"parameters.":[199],"Our":[200],"work":[201,220],"suggests":[202],"should":[206],"be":[207],"crucial":[209],"component":[210],"provides":[215],"foundation":[217],"further":[219],"important":[223],"problem.":[224],"release":[226],"code":[228],"data":[230],"experiments":[233],"publicly,":[234],"you":[236],"relevant":[240],"links":[241],"here:":[242],"https://github.com/Dylan102938/mood-bench.":[243]},"counts_by_year":[],"updated_date":"2026-06-11T09:08:48.828518","created_date":"2026-05-23T00:00:00"}