Towards Agents Which Can Understand Rich Communication

Olivia Watkins

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2024-188
September 2, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-188.pdf

Today’s AI systems are trained primarily on large datasets of input-output pairs. These agents may be able to condition on simple forms of communication (such as a language task description), but they’re currently not capable of making use of the full spectrum of communication, verbal and non-verbal, which human teachers use to guide their students.

This thesis makes progress on two challenges around teaching agents to understand rich communication. In Part 1, we develop algorithms which can efficiently ground real-time communication provided by humans, both non-verbal communication and several forms of language. We also enable agents to use language in a new way - guiding common-sense exploration.

In Part 2, we address the challenge of teaching agents to understand communication by trusted sources while ignoring malicious instructions or facts provided by untrusted sources. We benchmark models’ vulnerability to semantic prompt injection and jailbreak attacks, paving the way for future work addressing of these weaknesses we observed.

Advisor: Trevor Darrell and Pieter Abbeel

"; ?>

BibTeX citation:

@phdthesis{Watkins:EECS-2024-188,
    Author = {Watkins, Olivia},
    Title = {Towards Agents Which Can Understand Rich Communication},
    School = {EECS Department, University of California, Berkeley},
    Year = {2024},
    Month = {Sep},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-188.html},
    Number = {UCB/EECS-2024-188},
    Abstract = {Today’s AI systems are trained primarily on large datasets of input-output pairs. These agents may be able to condition on simple forms of communication (such as a language task description), but they’re currently not capable of making use of the full spectrum of communication, verbal and non-verbal, which human teachers use to guide their students.

This thesis makes progress on two challenges around teaching agents to understand rich communication. In Part 1, we develop algorithms which can efficiently ground real-time communication provided by humans, both non-verbal communication and several forms of language. We also enable agents to use language in a new way - guiding common-sense exploration.

In Part 2, we address the challenge of teaching agents to understand communication by trusted sources while ignoring malicious instructions or facts provided by untrusted sources. We benchmark models’ vulnerability to semantic prompt injection and jailbreak attacks, paving the way for future work addressing of these weaknesses we observed.}
}

EndNote citation:

%0 Thesis
%A Watkins, Olivia
%T Towards Agents Which Can Understand Rich Communication
%I EECS Department, University of California, Berkeley
%D 2024
%8 September 2
%@ UCB/EECS-2024-188
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-188.html
%F Watkins:EECS-2024-188