BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Sabre//Sabre VObject 4.6.1//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:Europe/Zurich
X-LIC-LOCATION:Europe/Zurich
TZURL:http://tzurl.org/zoneinfo/Europe/Zurich
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19810329T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19961027T030000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:news2092@dmi.unibas.ch
DTSTAMP;TZID=Europe/Zurich:20261102T191819
DTSTART;TZID=Europe/Zurich:20261104T143000
SUMMARY:LT meets Industry - Knowing When Not to Answer: Bounding LLM Halluc
 inations
DESCRIPTION:Abstract:\\r\\nLLMs tend to guess even when the provided docume
 nts lack key facts. This is a major liability in high-stakes domains. In t
 his talk\, I will present our training framework based on the Merlin-Arthu
 r protocol\, which transforms LLM question answering into a two-player gam
 e. By training the model against both a helpful prover supplying key parts
  of the documents and an adversarial prover stripping away evidence\, we f
 orce the model to ground its answers or explicitly abstain when evidence i
 s missing. I will introduce our novel grounding score\, which measures how
  much of an answer provably stems from the documents. I'll showcase how we
  applied this protocol to academic benchmarks\, as well as real German leg
 al case files and considerably reduced hallucinations.\\r\\nLetitia Parcal
 abescu [https://www.linkedin.com/in/letitia-parcalabescu/] 
X-ALT-DESC:<p><strong>Abstract</strong>:</p>\n<p>LLMs tend to guess even wh
 en the provided documents lack key facts. This is a major liability in hig
 h-stakes domains. In this talk\, I will present our training framework bas
 ed on the Merlin-Arthur protocol\, which transforms LLM question answering
  into a two-player game. By training the model against both a helpful prov
 er supplying key parts of the documents and an adversarial prover strippin
 g away evidence\, we force the model to ground its answers or explicitly a
 bstain when evidence is missing. I will introduce our novel grounding scor
 e\, which measures how much of an answer provably stems from the documents
 . I'll showcase how we applied this protocol to academic benchmarks\, as w
 ell as real German legal case files and considerably reduced hallucination
 s.</p>\n<p><br /><a href="https://www.linkedin.com/in/letitia-parcalabescu
 /">Letitia Parcalabescu</a><br />&nbsp\;</p>
END:VEVENT
END:VCALENDAR