When we give feedback, we face a delicate balancing act – we want to convey accurate information, but we also don't want to hurt someone's feelings. While computational pragmatic models have elegantly shown how politeness emerges from these principles, they've mainly focused on choices from limited predefined responses. Large language models (LLMs) enable the study of open-ended politeness strategies, but their ability to balance informational and social goals like humans remains uncertain. First, replicate previous work using restricted utterance sets, finding that sufficiently large LLMs (≥70B parameters) capture key human politeness patterns, particularly the strategic use of negation. We then extend this investigation to open-ended contexts, collecting and evaluating naturalistic feedback from both humans and LLMs. Surprisingly, human evaluators preferred LLM responses, which demonstrated sophisticated goal sensitivity and diverse politeness tactics. These findings suggest remarkable pragmatic competence in LLMs' polite language generation while raising questions about the underlying mechanisms.

Human languages are compositional, combining smaller units of meaning to express more complex ideas. To explain the emergence of compositionality, researchers have appealed to functional pressures from communication. However, languages may merely inherit the component structure found in the environment. We designed a reference game to explicitly disentangle these possibilities; pairs of participants (N = 450) communicated about about sets of shapes that were assembled from component parts. Critically, we manipulated whether shapes that shared the same parts were competitors within each trial or were distributed across different trials. We found that participants successfully developed efficient conventions for referring to the shapes. However, participants who needed to distinguish shapes that shared components within the same context were more likely to develop compositional systems. When shared components appeared in separate contexts, participants favored non-compositional conventions. These results suggest compositional language structure most readily emerges from immediate communicative pressures rather than environmental structure alone.

Context-sensitive communication not only requires speakers to choose relevant utterances from alternatives, but also to retrieve and evaluate the relevant utterances from memory in the first place. In this work, we compared different proposals about how underlying semantic representations work together with higher-level selection processes to enable individuals to flexibly utilize context to guide their language use. We examined speaker and guesser performance in a two-player iterative language game based on Codenames, which asks speakers to choose a single `clue' word that allows their partner to select a pair of target words from a context of distractors. The descriptive analyses indicated that speakers were sensitive to the shared semantic neighborhood of the target word pair and were able to use guesser feedback to shift their clues closer to the unguessed word. We also formulated a series of computational models combining different semantic representations with different selection processes. Model comparisons suggested that a model which integrated contextualized lexical representations based on association networks with a contextualized model of pragmatic reasoning was better able to predict behavior in the game compared to models that lacked context at either the representational or process level. Our findings suggest that flexibility in communication is driven by context-sensitivity at at the level of both representations and processes.

Theory of mind is a powerful cognitive ability: by the ageof six, people are capable of accurately reasoning about oth-ers’ beliefs and desires. An influential series of language un-derstanding experiments by Keysar and colleagues, however,showed that adults systematically failed to take a speaker’sbeliefs into account, revealing limitations on theory of mind.In this paper we argue that these apparent failures are in factsuccesses. Through a minimal pair of replications comparingscripted vs. unscripted speakers, we show that critical utter-ances used by Keysar and colleagues are uncooperative: theyare less informative than what a speaker would actually pro-duce in that situation. When we allow participants to naturallyinteract, we find that listener expectations are justified and er-rors are reduced. This ironically shows that apparent failuresof theory of mind are in fact attributable to sophisticated ex-pectations about speaker behavior—that is, to theory of mind.

Toddlers are sensitive to the reliability of speakers in their environment \cite{koenig2010}. While previous work suggests that children prefer labels from reliable speakers, it remains unclear how these social representations guide lower-level information-seeking processes that affect speaker preferences. The current study introduces a gaze-contingent eye-tracking paradigm to investigate how children engage in sampling during word learning. Toddlers (22-24m) view videos of two speakers labeling familiar objects; one speaker provides reliable labels and the other speaker provides unreliable labels. Toddlers then sample novel labels from the speakers using a gaze-contingent interface: only the speaker they are fixating on provides a novel label. Preliminary data (N = 18) suggests that participants prefer to sample first from the reliable speaker over the unreliable speaker. However, there is little difference in overall sampling preferences. Our findings suggest that toddlers can assess speaker reliability, but remain open to exploring information from unreliable speakers.

Adults readily form pacts, or temporary agreements about ref-erent names, over the course of conversation. Young childrenfail to do so with peers, but recent evidence suggests that ex-plicit feedback from adults may improve their performance(Matthews, Lieven, & Tomasello, 2007). Do parents natu-rally provide such structure in their conversations with chil-dren? Using a director-matcher paradigm, we first show thatparents and children (ages 4, 6, 8) converge on increasinglyaccurate and efficient conversational pacts. Further, parents ofyounger children provide more interactive feedback. Finally,we analyze asymmetries in parents’ and children’s contribu-tions, finding that pacts tend to originate with the parent, butare simplified by younger children. Together, these resultssupport the idea that parents sensitively adapt their languageto their children’s developmental level to scaffold successfulcommunication.

People tell stories to impose coherence on the continuous stream of everyday experience. But there are countless ways of narrativizing the same situation. How do we decide which features to gloss over, which features to elaborate on, and what order to recount them in? Here, we explore the hypothesis that people systematically adapt narratives with respect to their communicative goals. In one experiment (N=273), participants watched Heider-Simmel style animations and were asked to tell a story about what is happening in each video. They were randomly assigned to one of three goals: to tell a generic story, to chronicle the events like a reporter, or to share a juicy piece of gossip. We present a series of exploratory natural language analyses examining how these goals shape stories. Chronicles are more detailed relative to gossip, but gossip references more high-level social features. These findings underscore how Gricean principles of relevance drive storytelling.

From classrooms to dinner parties, many of our everyday conversations take place in larger groups where speakers address multiple listeners at once. Such multiparty settings raise a number of challenges for classical theories of communication, which largely focus on dyadic interactions. In this study, we investigated how speakers adapt their referring expressions over time as a function of the feedback they receive from multiple parties. We collected a large corpus of multiparty repeated reference games (98 games, 390 participants, 116K words) where speakers designed referring expressions for groups of 1 to 5 listeners. Larger groups tended to use more words total and to introduce more new words; nonetheless, most groups were able to converge to more efficient conventions regardless of the number of listeners.

Inspired by inductive reasoning models, we test whether generalized shared reality (i.e., the sense of being on the same page) arises through probabilistic inference about latent commonalities. Using a naturalistic text-based chat paradigm, we manipulated whether conversation partners discussed a belief they shared, a belief on which their opinions differed, or a random prompt. Participants discussing shared opinions reported experiencing greater shared reality compared to those discussing differences or random topics. Moreover, participants who made broader inferences about additional beliefs they might share with their partners also reported greater shared reality. While discussing shared opinions can induce an overall greater sense of shared reality, participants discussing differences leveraged their conversation to establish shared realities about other topics. We demonstrate that shared reality can emerge in multiple ways during initial interactions, establishing a foundation for future mechanistic investigations within an inductive inference framework.

A long tradition of studies in psycholinguistics has examined the formation and generalization of ad hoc conventions in reference games, showing how newly acquired conventions for a given target transfer to new referential contexts. However, another axis of generalization remains understudied: how do conventions formed for one target transfer to completely distinct targets, when specific lexical choices are unlikely to repeat? This paper presents two dyadic studies (N=240) that address this axis of generalization, focusing on the role of nameability --- the a priori likelihood that two individuals will share the same label. We leverage the recently-released KiloGram dataset, a collection of abstract tangram images that is orders of magnitude larger than previously available, exhibiting high diversity of properties like nameability. Our first study asks how nameability shapes convention formation, while the second asks how new conventions generalize to entirely new targets of reference. Our results raise new questions about how ad hoc conventions extend beyond target-specific re-use of specific lexical choices.