Institute Homepage
Institute Homepage DE Sign In

Back

Social Foundations of Computation Talk Biography
19 November 2024 at 15:00 | N0.002 Main lecture hall

Towards Better Foundations for Foundational Models: A Cognitivist Approach to Studying Large Language Models (LLMs)

ORGANIZERS
Thumb ticker sm 20241104 hardt moritz 12 cleaned kleiner
Social Foundations of Computation
Moritz Hardt
  • Director
Thumb ticker xxl original

The talk will begin with a short demo of an LLM-based assistant that allows scientists to convert their papers (with a simple drag and drop) into short podcasts for communicating their research to a general audience. While we built the tool, we can’t explain its unreasonable (in)effectiveness, i.e., we don’t really understand why it works or when it might fail. So in the rest of the talk, I will present our investigations into some curiosity-driven questions about LLMs; specifically, how do LLMs receive, process, organize, store, and retrieve information.

Our analysis is centered around engaging LLMs in two specific types of cognitive tasks: 

(a) syntactically-rich (semantically-poor) tasks such as recognizing formal grammars, and 
(b) semantically-rich (syntactically-poor) tasks such as answering factual knowledge questions about real-world entities. 
 
Using carefully designed experimental frameworks, we explore the following two foundational questions:
(a) (how) can we distinguish whether some LLM has learnt some training data by rote vs. by understanding?
b) (how) can we compare the inductive biases of few-shot learning (via in-context demonstrations) vs. fine-tuning (via training over examples)?  
 
I will present some initial empirical results from experimenting with a number of large open-weight language models and argue that our findings they have important implications for estimating LLMs' latent skills and knowledge (including factual knowledge), the privacy of training data (including potential for memorization), the reliability of generated outputs (including potential for hallucinations), and the robustness of LLM-based applications (including our podcast assistant for science communication). 

Speaker Biography

Dr. Krishna P. Gummadi ( Max Planck Institute for Software Systems, Saarbrücken)

Director and Scientific Member