BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Specializing LLMs for Factuality and Soft Reasoning - Greg Durrett
 \, UT Austin
DTSTART:20241017T130000Z
DTEND:20241017T140000Z
UID:TALK222262@talks.cam.ac.uk
CONTACT:Tiancheng Hu
DESCRIPTION:Proponents of LLM scaling assert that training a giant model o
 n as much data as possible can eventually solve most language tasks\, perh
 aps even leading to AGI. However\, frontier LLMs still fall short on compl
 ex problems in long-tail domains. Errors occur somewhere in the process of
  encoding the necessary knowledge\, surfacing it for a specific prompt\, a
 nd synthesizing it with other input data. In this talk\, I will argue that
  specialization is the right approach to improve LLMs here\; that is\, mod
 ifying them through training or other means to improve their factuality an
 d reasoning capabilities. First\, I will show that specialization is neces
 sary: inference-only approaches like chain-of-thought prompting are not su
 fficient. Second\, I will present our fact-checking system MiniCheck\, whi
 ch is fine-tuned on specialized data to detect factual errors in LLM respo
 nses\, leading to a better detector than frontier models like GPT-4. Final
 ly\, I will discuss how to specialize LLMs to be better at logical reasoni
 ng. I argue that we need (a) better fine-tuning methods which make targete
 d adjustments to model behavior\; (b) improved inference capabilities\, su
 ch as a differentiable theorem prover that can be plugged into a standard 
 Transformer. These forms of specialization represent a path towards fundam
 entally new capabilities in factuality and reasoning beyond what can be ac
 hieved in current models.\n\nBio: Greg Durrett is an associate professor o
 f Computer Science at UT Austin. He received his BS in Computer Science an
 d Mathematics from MIT and his PhD in Computer Science from UC Berkeley\, 
 where he was advised by Dan Klein. His research is broadly in the areas of
  natural language processing and machine learning. Currently\, his group's
  focus is on techniques for reasoning about knowledge in text\, verifying 
 factuality of LLM generations\, and building systems using LLMs as primiti
 ves. He is a 2023 Sloan Research Fellow and a recipient of a 2022 NSF CARE
 ER award. He has co-organized the Workshop on Natural Language Reasoning a
 nd Structured Explanations at ACL 2023 and ACL 2024\, as well as workshops
  on low-resource NLP and NLP for programming. He has served in numerous ro
 les for *CL conferences\, including as a member of the NAACL Board since 2
 024.\n
LOCATION:https://cam-ac-uk.zoom.us/j/97599459216?pwd=QTRsOWZCOXRTREVnbTJBd
 XVpOXFvdz09
END:VEVENT
END:VCALENDAR