BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Prosocial Language Models - Soroush Vosoughi\, Dartmouth College
DTSTART:20240215T110000Z
DTEND:20240215T120000Z
UID:TALK212422@talks.cam.ac.uk
CONTACT:Panagiotis Fytas
DESCRIPTION:Large language models\, such as GPT-4\, have marked a signific
 ant advancement in the field of natural language processing\, achieving ne
 ar-human performance across a variety of tasks with minimal to no addition
 al training data. The remarkable capabilities of these models can be attri
 buted to their substantial parameter counts\, often reaching into the hund
 reds or thousands of millions\, and the extensive datasets sourced from th
 e web for their pre-training. Despite their successes\, the very character
 istics that empower these models also render them susceptible to mirroring
  web-based biases and antisocial behaviors. Such reflections pose consider
 able challenges in deploying these models in real-world scenarios\, partic
 ularly in socially sensitive applications. In response\, our laboratory fo
 cuses on developing techniques for the post hoc mitigation of these antiso
 cial tendencies\, allowing for the enforcement of prosocial behaviors duri
 ng model inference without the need for resource-intensive retraining. Thi
 s presentation will delve into our latest efforts to reduce bias and enhan
 ce alignment with human ethical standards in language models through infer
 ence-time interventions.\n
LOCATION:GR06/07 | Faculty of English\, 9 West Road\, CB3 9DP
END:VEVENT
END:VCALENDAR