BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:IC Colloquium: From Instructions to Interaction: Developing Steera
 ble and Usable Open Language Models
DTSTART:20260305T101500
DTEND:20260305T111500
DTSTAMP:20260916T073844Z
UID:3752a754cf005c0ee7264a512170518f1cc8d9ccb3a0106681f87e7b
CATEGORIES:Conferences - Seminars
DESCRIPTION:Par : Valentina Pyatkin - Allen Institute for AI and Universi
 ty of Washington\nIC Faculty candidate\n\nAbstract\nMaking large language 
 models usable requires post-training methods that align models to human va
 lues\, robustly handle underspecified inputs\, and generalize to diverse i
 nstructions. This talk addresses the challenge of developing responsible A
 I through post-training from the following angles.\nFirst\, I address cont
 extual robustness. Preference data\, for example\, is often underspecified
 \, and I show how underspecification in preference data can lead to diverg
 ing preferences. Standard reward models fail to properly handle these disa
 greements\, often making decisive choices even when human annotators are s
 plit. I argue that more consequential outputs demand more context\, and pr
 opose clarification question generation as one solution.\nSecond\, I will 
 discuss how we can train models to be better instruction followers. I will
  show that most models severely overfit on a small set of instruction-foll
 owing constraints and are not able to generalize well to unseen output con
 straints. I propose to train models with reinforcement learning from verif
 iable rewards for verifiable instruction following\, and show how this lea
 ds to improved generalization on constraint following.\nThroughout the pre
 sentation\, I will outline how I have applied these insights into developi
 ng open generative models\, like Tülu and OLMo\, and I will conclude with
  my research agenda for responsible post-training: critical AI evaluation\
 , broader generalization\, and expanding what models can reliably do.\n\nB
 io\nValentina Pyatkin is a postdoctoral researcher at the Allen Institute 
 for AI and the University of Washington\, advised by Prof. Hanna Hajishirz
 i and Prof. Yejin Choi. Additionally\, she is part-time affiliated with th
 e ETH AI Center\, where she mentors students and works on post-training fo
 r the Swiss AI Initiative. She obtained her PhD in Computer Science from B
 ar Ilan University. Her work has been awarded an ACL Outstanding Paper Awa
 rd and the ACL Best Theme Paper Award\, and has been supported by a Schmid
 t Sciences Postdoctoral Award. During her doctoral studies\, she conducted
  research internships at Google and the Allen Institute for AI\, where she
  received the AI2 Outstanding Intern of the Year Award.\n\nMore informatio
 n\n 
LOCATION:BC 420 https://plan.epfl.ch/?room==BC%20420 https://epfl.zoom.us/
 j/65790496346
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR