BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//pretalx//pretalx.devconf.info//devconf-cz-2026//talk//LVBAQJ
BEGIN:VTIMEZONE
TZID:CET
BEGIN:STANDARD
DTSTART:20001029T040000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20000326T030000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:pretalx-devconf-cz-2026-LVBAQJ@pretalx.devconf.info
DTSTART;TZID=CET:20260619T123000
DTEND;TZID=CET:20260619T130500
DESCRIPTION:Updating a specific fact in a 70-billion-parameter model usuall
 y feels like using a sledgehammer for heart surgery. Traditional retrainin
 g is slow and expensive\, often breaking unrelated behaviors in the proces
 s. ROME (Rank-One Model Editing) offers a more surgical alternative\, trea
 ting model weights like a database that can be precisely updated without a
  full rebuild.\nWe move past the black box mystery to show how we can loca
 te where a specific fact lives within a transformer's architecture. We wil
 l explore the mechanics of knowledge neurons\, the risk of unintended side
  effects like hallucinations\, and the serious security implications of ma
 licious fact injections or stealth patching. Through live demonstrations u
 sing our open-source toolkit\, we will show how ROME performs precision st
 rikes on state-of-the-art models\, highlighting both its efficiency and th
 e scenarios where the method’s assumptions break down.\nThis talk is an 
 ongoing collaborative research between Red Hat and FIT BUT.
DTSTAMP:20260430T115010Z
LOCATION:A112 (capacity 64)
SUMMARY:Surgery on a Brain: Live-Patching Facts in LLMs with ROME - Jakub R
 eš\, Matej Olexa
URL:https://pretalx.devconf.info/devconf-cz-2026/talk/LVBAQJ/
END:VEVENT
END:VCALENDAR