BEGIN:VCALENDAR
VERSION:2.0
PRODID:IEEE vTools.Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:Europe/Stockholm
BEGIN:DAYLIGHT
DTSTART:20260329T030000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20261025T020000
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260618T143956Z
UID:BDF65759-A54B-4432-BB07-78A6FECD5ACF
DTSTART;TZID=Europe/Stockholm:20260625T140000
DTEND;TZID=Europe/Stockholm:20260625T150000
DESCRIPTION:In this talk\, we argue that “inference-time” large languag
 e model (LLM) operation\, where we interact with models without modifying 
 their weights\, is fertile ground for information-theoretic methods. As LL
 Ms increasingly serve as black-box components of more complex systems\, in
 formation and coding theory offer a principled toolkit for shaping\, verif
 ying\, and controlling their outputs. We focus on one challenge in particu
 lar: watermarking LLM-generated text. Watermarks enable authentication of 
 text provenance and help curb misuse of machine-generated content.\n\nWe p
 resent recent results demonstrating a close connection between LLM waterma
 rking and coding theory\, showing that classical tools such as the Plotkin
  bound yield fundamental limits on watermark performance. This perspective
  also informs the design of two practical watermarks: SimplexWater and Hea
 vyWater. We also briefly survey other inference-time challenges that can b
 e addressed with information theory\, such as inference-time alignment.\n\
 nCo-sponsored by: KTH Royal Institute of Technology\n\nSpeaker(s): Prof. F
 lavio du Pin Calmon\, \n\nRoom: Digital Futures hub\, floor 2\, KTH\, Osqu
 ars Backe 5\, Stockholm\, Stockholms lan\, Sweden\, Virtual: https://event
 s.vtools.ieee.org/m/564274
LOCATION:Room: Digital Futures hub\, floor 2\, KTH\, Osquars Backe 5\, Stoc
 kholm\, Stockholms lan\, Sweden\, Virtual: https://events.vtools.ieee.org/
 m/564274
ORGANIZER:oech@kth.se
SEQUENCE:36
SUMMARY:Technical seminar by Prof. Flavio P. Calmon\, Havard University: In
 ference-Time Information Theory
URL;VALUE=URI:https://events.vtools.ieee.org/m/564274
X-ALT-DESC:Description: <br /><p style="max-width: 100%\;">In this talk\, w
 e argue that &ldquo\;inference-time&rdquo\; large language model (LLM) ope
 ration\, where we interact with models without modifying their weights\, i
 s fertile ground for&nbsp\;information-theoretic methods. As LLMs increasi
 ngly serve as black-box components of more complex systems\,&nbsp\;informa
 tion&nbsp\;and coding&nbsp\;theory&nbsp\;offer a principled toolkit for sh
 aping\, verifying\, and controlling their outputs. We focus on one challen
 ge in particular: watermarking LLM-generated text. Watermarks enable authe
 ntication of text provenance and help curb misuse of machine-generated con
 tent.&nbsp\;</p>\n<p style="max-width: 100%\;">We present recent results d
 emonstrating a close connection between LLM watermarking and coding theory
 \, showing that classical tools such as the Plotkin bound yield fundamenta
 l limits on watermark performance. This perspective also informs the desig
 n of two practical watermarks: SimplexWater and HeavyWater. We also briefl
 y survey other inference-time challenges that can be addressed with inform
 ation theory\, such as inference-time alignment.</p>
END:VEVENT
END:VCALENDAR