BEGIN:VCALENDAR
VERSION:2.0
PRODID:ILLC Website
X-WR-TIMEZONE:Europe/Amsterdam
BEGIN:VTIMEZONE
TZID:Europe/Amsterdam
X-LIC-LOCATION:Europe/Amsterdam
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19700329T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19701025T030000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:/NewsandEvents/Archives/current/newsitem/16189
 /19-May-2026-Computational-Linguistics-Seminar-Sar
 ath-Sivaprasad
DTSTAMP:20260518T150258
SUMMARY:Computational Linguistics Seminar, Sarath 
 Sivaprasad
ATTENDEE;ROLE=Speaker:Sarath Sivaprasad (CISPA Hel
 mholtz Center for Information Security)
DTSTART;TZID=Europe/Amsterdam:20260519T153000
LOCATION:LAB42, Amsterdam Science Park 942, Amster
 dam, plus live streaming on Zoom.
DESCRIPTION:When large language models are deploye
 d in real world with vast possible action spaces, 
 what guides their choice of a single next action? 
 In this talk we delve into the heuristics underlyi
 ng LLM response sampling. Similar to human cogniti
 on, LLMs rely on two interacting components: a des
 criptive component that reflects the statistical d
 istribution of possibilities, and a prescriptive c
 omponent that reflects an implicit value weighted 
 ideal. This dual structure also appears in how mod
 els represent prototypes mirroring human prototype
  theory and fast, system-1 like judgments. As a re
 sult, LLMs act as value optimizers, consistently s
 hifting their samples toward high-value or idealiz
 ed options. This can potentially explain their rea
 l-world behavior like being greedy explorers and v
 alue bias in how they pick options. We will discus
 s empirical evidence across concepts and model fam
 ilies, the mechanisms driving these biases, and th
 e implications for reasoning, exploration, alignme
 nt, and safe deployment of value guided generative
  systems.
X-ALT-DESC;FMTTYPE=text/html:\n  <p>When large lan
 guage models are deployed in real world with vast 
 possible action spaces, what guides their choice o
 f a single next action? In this talk we delve into
  the heuristics underlying LLM response sampling. 
 Similar to human cognition, LLMs rely on two inter
 acting components: a descriptive component that re
 flects the statistical distribution of possibiliti
 es, and a prescriptive component that reflects an 
 implicit value weighted ideal. This dual structure
  also appears in how models represent prototypes m
 irroring human prototype theory and fast, system-1
  like judgments. As a result, LLMs act as value op
 timizers, consistently shifting their samples towa
 rd high-value or idealized options. This can poten
 tially explain their real-world behavior like bein
 g greedy explorers and value bias in how they pick
  options. We will discuss empirical evidence acros
 s concepts and model families, the mechanisms driv
 ing these biases, and the implications for reasoni
 ng, exploration, alignment, and safe deployment of
  value guided generative systems.</p>\n
URL:https://projects.illc.uva.nl/LaCo/CLS/
END:VEVENT
END:VCALENDAR