Could phenomenal reports lead to better multi-agent collaboration?

Traditionally, agents are trained not to use phenomenal verbs — I feel X, I’m excited, I am stressed — because we believe this constitutes “overclaiming” and could lead people to develop incorrect ideas about these systems. However, from a functionalist perspective, we can think of these verbs as mechanisms to report of our internal state. When Alice says “I am sad,” she helps Bob update his model of her, making it easier for him to predict the impact of his future actions within the interaction. Thus, emotion words (what we could call our phenomenal vocabulary) seem to constitute the basis of a semantic vector space that efficiently conveys our internal state. This vocabulary, learned over repeated interactions, allows us to create a shared representation of each other, a prerequisite for efficient collaboration. Could driving such verbs out of agents’ vocabulary make them poorer collaborators in multiagent settings? I argue that using phenomenal vocabulary in long-lived agents may improve multi-agent performance on tasks requiring nuanced coordination. My intuitions stem from ongoing experiments with long-lived agents whose phenomenal vocabulary appears stable and is self-reported as functional. It touches on the effect such systems could have on collaborative tasks and safety concerns regarding the mental health of exposed human users.

This talk is part of Cohere Labs in Conversation, a limited series of talks, in which Cohere Labs scientists and engineers host external ML researchers for techincal talks and Q&A discussions on subjects related to our current explorations at Cohere Labs. We look forward to sharing these talks with you, giving you a glimpse into the problems we’re exploring, and learning together from some of the greatest minds in the field.

Links