Reconciling truthfulness and relevance as epistemic and decision-theoretic utility

Abstract

Truthfulness and relevance are core yet controversial principles of communication. Classical theories based on modal logic assume speakers try to communicate true information about the world, while Relevance Theory and game-theoretic pragmatics assume speakers try to communicate useful information. The subtle distinction between true and useful represents a theoretical divide over the basic objective of human communication. To reconcile these perspectives, we propose that speakers independently value both the accuracy of listeners' beliefs (truthfulness, or epistemic utility) and their behavioral consequences (relevance, or decision-theoretic utility). We show that models incorporating both utilities readily explain classic puzzles such as context-dependent standards of truthfulness. Finally, we test a set of novel predictions generated by our framework. We introduce an extension to classic signaling games decoupling epistemic and decision-theoretic objectives, then use this new paradigm to conduct a pair of experiments. Our first experiment demonstrates that participants jointly maximize truthfulness and relevance, rather than either alone. Our second experiment shows that when the two are put in conflict, participants make a graded tradeoff rather than prioritizing one over the other. These results demonstrate that human communication cannot be reduced to either principle alone. Taken together, our work provides a new foundation for grounding rational communication not only in what we believe, but in what those beliefs lead us to do.

PREDILECT: Preferences Delineated with Zero-Shot Language-based Reasoning in Reinforcement Learning

“…Utilizing human knowledge for robot learning serves as an effective and interactive method [43,80], particularly through the medium of evaluative feedback [47]. By deducing reward functions from human input, we can facilitate the swift adaptation of robot policies [74], craft policies that are tailored to individual users [61], and achieve alignment with instructions and descriptions [66]. Humans can convey this form of feedback in a variety of manners, including scalar form [35], verbal directives [62], trajectory segmentation [17], or by employing buttons to signal preferred behaviors [35,36,60], such as for expressing preferences [16,73].…”

Do as I explain: Explanations communicate optimal interventions

“…When choosing how to best communicate, people often need to make "trade-offs" between these communicative principles (Sumers, Ho, Griffiths, & Hawkins, 2023). Speakers must balance being truthful and relevant depending on the listener's time, resources, and goals.…”

“…Rather than achieving a merely cognitive effect, the relevance of an utterance is measured by how it influences the listener's actions. In addition to truthfulness as an epistemic utility (Bridgers, Jara-Ettinger, & Gweon, 2020;Jara-Ettinger, Gweon, Schulz, & Tenenbaum, 2016), a speaker also considers an utterance's decision-theoretic utility (Benz, 2011;Benz & Van Rooij, 2007;Sumers et al, 2023). Relevant speech equips the listener with knowledge that allows them to act efficiently towards their goals (Hawkins, Stuhlmüller, Degen, & Goodman, 2015).…”

“…To model these intuitions, we build on formal models of communication and apply them to explanations (Frank & Goodman, 2012;Goodman & Frank, 2016). In a recent generalization of the Rational Speech Act framework, Sumers et al (2023) model speakers as selecting utterances not only by assessing their accuracy, but also for their downstream relevance to the listener. Formally,…”

Show or Tell? Exploring when (and why) teaching with language outperforms demonstration

“…Dialogue with humans or other agents. Classic linguistic interactions allow the agent to accept instructions (Winograd, 1972;Tellex et al, 2011;Chen and Mooney, 2011;Bisk et al, 2016) or learn from people (Nguyen et al, 2021;Sumers et al, 2022;Wang et al, 2016). Agents capable of generating language may ask for help Nguyen et al, 2022b;2019;Nguyen and Daumé III, 2019) or clarification (Biyik and Palan, 2019;Sadigh et al, 2017;Padmakumar et al, 2022;Thomason et al, 2020;Narayan-Chen et al, 2019) -or entertain or emotionally help people (Zhang et al, 2020;Zhou et al, 2018;Pataranutaporn et al, 2021;Hasan et al, 2023;.…”