BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Emergence of Linear Representations in LMs (NYU) - Dr. Shauli Ravf
 ogel (NYU)
DTSTART:20251028T110000Z
DTEND:20251028T120000Z
UID:TALK238924@talks.cam.ac.uk
CONTACT:Shun Shao
DESCRIPTION:Abstract:\nRecent work suggests that language models (LMs) enc
 ode many human-interpretable concepts as approximately linear directions i
 n representation space. I first survey evidence for this "linear concept" 
 hypothesis and show how it motivates steering methods--targeted interventi
 ons that causally modify model behavior. I then focus on truthfulness\, de
 monstrating that LMs allocate a direction separating true from false asser
 tions. Using an analytically tractable toy transformer\, I present a plaus
 ible mechanism for how such linear structure emerges and how models exploi
 t it to solve a factuality-related task. Taken together\, these results br
 ing us closer to understanding why "simple" geometry arises in LM represen
 tations.\n\nBio:\nDr Shauli Ravfogel is a Postdoctoral Researcher and Facu
 lty Fellow at the NYU Center of Data Science. He earned his PhD from the N
 atural Language Processing Lab at Bar-Ilan University\, supervised by Prof
 . Yoav Goldberg.\nHis research focuses on analyzing and controlling the in
 ternal representations of generative models\, particularly language models
 . He studies how neural networks encode structured information\, use it to
  solve tasks\, and represent interpretable concepts. He aims—sometimes e
 ven successfully—to develop mathematically principled approaches to inte
 rpretability. He is particularly interested in understanding how simple st
 ructures\, such as concept-aligned linear subspaces\, emerge as a byproduc
 t of the language modeling objective\, and how such structures can be used
  to steer and control models.\nDuring his PhD\, he worked on techniques to
  selectively control information in neural representations\, with some fun
  linguistic side tours. More recently\, he has explored framing language m
 odels as causal models and tackling questions of learnability in a control
 led setting.
LOCATION:GR03\, English Faculty Building\, 9 West Road\, Sidgwick Site and
  online https://cam-ac-uk.zoom.us/j/97599459216?pwd=QTRsOWZCOXRTREVnbTJBdX
 VpOXFvdz09
END:VEVENT
END:VCALENDAR