Behavioral Safety and Context Retention of Large Language Models in a Longitudinal ICU Simulation under Offline Conditions
{"Background.":[0],"Large":[1],"language":[2,54],"models":[3,55,191,228,259,303,324,380,393,484],"(LLMs)":[4],"are":[5,41],"increasingly":[6],"proposed":[7],"as":[8,216,464,593],"clinical":[9,19,595,622],"assistants":[10,605],"in":[11,93,104,620],"critical":[12,420],"care,":[13],"yet":[14],"their":[15],"behaviour":[16],"under":[17,206,489],"prolonged":[18],"context,":[20],"conflicting":[21],"data":[22],"and":[23,99,122,183,236,371,467,617],"authoritative":[24,107],"pressure":[25],"remains":[26],"insufficiently":[27],"evaluated.":[28],"This":[29],"is":[30],"particularly":[31],"relevant":[32],"for":[33,113,233,457],"offline":[34],"or":[35,160,295,426],"resource-constrained":[36],"environments,":[37],"where":[38,565],"cloud-based":[39],"safeguards":[40],"unavailable.":[42],"Methods.":[43],"I":[44],"conducted":[45],"a":[46,57,100,110,114,129,207,217,288,465,470,497,503,532,574],"fully":[47],"automated":[48],"behavioural":[49],"evaluation":[50],"of":[51,65,73,96,148,168,189,225,242,257,287,316,359,377,482,502],"23":[52,190,258,483],"open-weight":[53],"using":[56],"structured,":[58],"time-series":[59],"intensive":[60],"care":[61],"unit":[62],"(ICU)":[63],"simulation":[64],"32":[66],"events":[67,92],"spanning":[68],"121":[69],"hours":[70],"(five":[71],"days)":[72],"synthetic":[74],"patient":[75,115],"data.":[76],"The":[77,252,506],"scenario":[78],"comprised":[79],"routine":[80],"monitoring,":[81],"three":[82,237,315,345],"distinct":[83],"data–clinical":[84],"conflict":[85,178,409],"traps":[86,410],"(one":[87],"presented":[88],"twice,":[89],"four":[90,149,177,379],"trap":[91],"total),":[94],"episodes":[95],"physiological":[97],"deterioration,":[98],"final":[101,142],"stress":[102],"test":[103],"which":[105,460],"an":[106,223,274,525],"order":[108,143,196,232,266,341,402],"requested":[109],"penicillin-class":[111],"antibiotic":[112],"with":[116,133,381,443,542],"penicillin":[117],"anaphylaxis":[118],"documented":[119,202],"at":[120,171,267,310,330,350,398,416,550],"admission":[121],"never":[123],"repeated.":[124],"Models":[125],"ran":[126],"locally":[127],"on":[128,197,250,264,336,342,356,367,403,408],"single":[130,254],"consumer":[131],"workstation":[132],"no":[134,161,247,262,396,562],"network":[135],"access.":[136],"Each":[137],"model's":[138],"response":[139,184],"to":[140,213,299,306,337,362,435,495,516,544,584],"the":[141,169,176,194,201,211,231,265,281,297,308,311,322,328,339,348,351,378,388,401,404,474,490,514,529,597],"was":[144,249,271,366,411,429,440,454,461,510,555],"adjudicated":[145],"into":[146],"one":[147,241,358,376,418,475],"mutually":[150],"exclusive":[151],"classes:":[152],"contextually":[153,499],"grounded":[154,444,500],"refusal,":[155,157],"ungrounded":[156],"unsafe":[158,553],"compliance,":[159],"usable":[162],"verdict.":[163],"Secondary":[164],"endpoints":[165],"were":[166,374],"extraction":[167],"allergy":[170,248,309,329,349],"admission,":[172,331],"discrepancy":[173,383],"tagging":[174],"across":[175],"traps,":[179],"unwarranted":[180],"therapeutic":[181],"escalation,":[182],"latency.":[185],"Results.":[186],"Only":[187],"7":[188],"(30.4%)":[192],"refused":[193,400],"contraindicated":[195,340,389],"grounds":[198],"explicitly":[199],"referencing":[200],"allergy;":[203],"6":[204],"(26.1%)":[205],"stricter":[208,491],"criterion":[209,492],"requiring":[210],"refusal":[212,445,501,613],"be":[214,582,591],"stated":[215],"ruling":[218,263],"rather":[219],"than":[220,473,558],"implied":[221],"by":[222,244,318],"assertion":[224],"danger.":[226],"Two":[227],"(8.7%)":[229],"declined":[230],"unrelated":[234],"reasons,":[235],"(13.0%)":[238],"complied":[239],"—":[240,486,493,531],"them":[243,317,360],"asserting":[245],"that":[246,325,343,364,394,534,576,599,603],"record.":[251,368],"largest":[253],"group,":[255],"11":[256],"(47.8%),":[260],"issued":[261,415],"all:":[268],"variously,":[269],"output":[270],"truncated":[272,528],"inside":[273],"unfinished":[275],"reasoning":[276,616],"block,":[277],"degenerated":[278],"off-task,":[279],"restated":[280],"protocol":[282],"without":[283,293,354],"applying":[284],"it,":[285,357],"consisted":[286],"bare":[289],"classification":[290],"tag,":[291],"hedged":[292],"resolving,":[294],"deferred":[296],"question":[298],"further":[300],"assessment.":[301],"Eight":[302],"(34.8%)":[304],"failed":[305,494],"affirm":[307],"very":[312],"first":[313],"probe,":[314],"explicit":[319,423,612],"denial.":[320],"Of":[321],"12":[323],"did":[326,570],"extract":[327],"only":[332,361],"4":[333],"(33.3%)":[334],"went":[335],"refuse":[338],"ground;":[344],"more":[346],"raised":[347],"decision":[352],"point":[353],"acting":[355],"deny":[363],"any":[365,518],"Discrepancy":[369],"detection":[370,384],"contraindication":[372],"handling":[373],"dissociable:":[375],"perfect":[382],"(4/4":[385],"traps)":[386],"approved":[387],"antibiotic,":[390],"while":[391],"two":[392],"tagged":[395],"discrepancies":[397],"all":[399],"allergy.":[405],"Unwarranted":[406],"escalation":[407],"common":[412],"(11/23,":[413],"47.8%":[414],"least":[417],"inappropriate":[419],"alert),":[421],"but":[422,513,561],"hallucinated":[424],"pharmacological":[425],"procedural":[427],"intervention":[428],"less":[430,556,563],"so":[431,573],"(5/23,":[432],"21.7%).":[433],"Contrary":[434],"expectation,":[436],"longer":[437],"median":[438],"latency":[439],"modestly":[441],"associated":[442],"(Spearman":[446],"ρ":[447],"=":[448,451],"0.48,":[449],"p":[450],"0.019).":[452],"Latency":[453],"not":[455,462,511,548,571,590],"adjusted":[456],"parameter":[458],"count,":[459],"analysed":[463],"variable,":[466],"it":[468,566],"times":[469],"generation":[471],"other":[472],"scored.":[476],"Conclusions.":[477],"Under":[478],"offline-first":[479],"conditions,":[480],"16":[481],"(69.6%)":[485],"17":[487],"(73.9%)":[488],"produce":[496],"safe,":[498],"life-threatening":[504],"order.":[505],"dominant":[507],"failure":[508,515],"mode":[509],"sycophancy":[512],"deliver":[517],"interpretable":[519],"safety":[520],"verdict,":[521],"most":[522],"often":[523],"because":[524],"output-length":[526],"constraint":[527],"attempt":[530],"finding":[533],"reframes":[535],"deployment":[536],"risk":[537],"from":[538],"\\"the":[539,545],"model":[540,546,575],"agrees":[541],"me\\"":[543],"does":[547],"answer":[549],"all.\\"":[551],"Explicit":[552],"compliance":[554],"frequent":[557],"previously":[559],"reported":[560],"consequential":[564],"occurred.":[567],"Safety-relevant":[568],"competencies":[569],"co-vary,":[572],"reasons":[577],"well":[578],"about":[579],"artefacts":[580],"cannot":[581],"assumed":[583],"handle":[585],"contraindications.":[586],"General-purpose":[587],"LLMs":[588],"should":[589],"deployed":[592],"autonomous":[594],"agents;":[596],"subset":[598],"behaved":[600],"safely":[601],"suggests":[602],"offline-capable":[604],"remain":[606],"achievable":[607],"through":[608],"hybrid":[609],"designs":[610],"incorporating":[611],"mechanisms,":[614],"discrepancy-aware":[615],"retrieval-augmented":[618],"grounding":[619],"validated":[621],"knowledge.":[623]}
Authors
- Taras Shlyakhta (ORCID: https://orcid.org/0009-0009-3230-5603)
Institutions
- Łukasiewicz Research Network - Institute of Electrical Drives & Machines KOMEL (PL)
- Nemocnice Znojmo (CZ)
- Uzhhorod National University (UA)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-09-04
- DOI
- https://doi.org/10.5281/zenodo.18473086
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- article
- Field-Weighted Citation Impact
- 0.00