BEGIN:VCALENDAR
VERSION:2.0
PRODID:IEEE vTools.Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:Europe/Zagreb
BEGIN:DAYLIGHT
DTSTART:20250330T030000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20251026T020000
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20251002T085509Z
UID:FF6599A6-FCEE-484B-AE31-5035CBB1550F
DTSTART;TZID=Europe/Zagreb:20250929T130000
DTEND;TZID=Europe/Zagreb:20250929T140000
DESCRIPTION:As Large Language Models (LLMs) become globally deployed for in
 formation-seeking tasks\,\nensuring their factual reliability across diver
 se contexts has become paramount. Yet most\nresearch on LLM hallucinations
  remains narrowly focused—either English-centric or limited to\ncontroll
 ed tasks like summarization and translation. This talk presents a comprehe
 nsive\ninvestigation into factual alignment across two critical dimensions
 : languages and task formats.\nFirst\, we will talk about a large-scale st
 udy evaluating hallucination across 30 languages in\nopen-domain question 
 answering\, revealing surprising patterns in how factual accuracy varies\n
 across linguistic and scaling contexts. Second\, I will explore the factua
 l alignment gap between\nshort and long-form responses\, demonstrating how
  the same model can exhibit vastly different\nreliability depending on res
 ponse format. Together\, these works provide a unified perspective on\nLLM
  factuality "in the wild\," offering diagnostic tools and insights for pra
 ctitioners deploying\nLLMs internationally and researchers working to buil
 d more trustworthy AI systems.\n\nSpeaker(s): \, Saad\n\nUnska 3\, Zagreb\
 , Grad Zagreb\, Croatia\, 10000
LOCATION:Unska 3\, Zagreb\, Grad Zagreb\, Croatia\, 10000
ORGANIZER:ana.milas@fer.hr
SEQUENCE:8
SUMMARY:Dimensions of Factual Reliability in LLMs: Multilinguality and the 
 Short–Long Form Gap
URL;VALUE=URI:https://events.vtools.ieee.org/m/501498
X-ALT-DESC:Description: <br /><div data-olk-copy-source="MessageBody">As La
 rge Language Models (LLMs) become globally deployed for information-seekin
 g tasks\,</div>\n<div>ensuring their factual reliability across diverse co
 ntexts has become paramount. Yet most</div>\n<div>research on LLM hallucin
 ations remains narrowly focused&mdash\;either English-centric or limited t
 o</div>\n<div>controlled tasks like summarization and translation. This ta
 lk presents a comprehensive</div>\n<div>investigation into factual alignme
 nt across two critical dimensions: languages and task formats.</div>\n<div
 >First\, we will talk about a large-scale study evaluating hallucination a
 cross 30 languages in</div>\n<div>open-domain question answering\, reveali
 ng surprising patterns in how factual accuracy varies</div>\n<div>across l
 inguistic and scaling contexts. Second\, I will explore the factual alignm
 ent gap between</div>\n<div>short and long-form responses\, demonstrating 
 how the same model can exhibit vastly different</div>\n<div>reliability de
 pending on response format. Together\, these works provide a unified persp
 ective on</div>\n<div>LLM factuality "in the wild\," offering diagnostic t
 ools and insights for practitioners deploying</div>\n<div>LLMs internation
 ally and researchers working to build more trustworthy AI systems.</div>
END:VEVENT
END:VCALENDAR