Search

As AI systems like language models are increasingly integrated into decision-making processes affecting people's lives, it's critical to ensure that these systems have sound moral reasoning. To test whether they do, we need to develop systematic evaluations. We provide a framework that uses a language model to translate causal graphs that capture key aspects of moral dilemmas into prompt templates. With this framework, we procedurally generated a large and diverse set of moral dilemmas---the OffTheRails benchmark---consisting of 50 scenarios and 400 unique test items. We collected moral permissibility and intention judgments from human participants for a subset of our items and compared these judgments to those from two language models (GPT-4 and Claude-2) across eight conditions. We find that moral dilemmas in which the harm is a necessary means (as compared to a side effect) resulted in lower permissibility and higher intention ratings for both participants and language models. The same pattern was observed for evitable versus inevitable harmful outcomes. However, there was no clear effect of whether the harm resulted from an agent's action versus from having omitted to act. We discuss limitations of our prompt generation pipeline and opportunities for improving scenarios to increase the strength of experimental effects.

The discrete entities and explicit relations of symbolic systems make them transparent and easy to communicate. This contrasts with distributed systems, which tend to be opaque. This can lead us to pursue symbolic characterizations of human cognition. Symbolic interpretations can, however, oversimplify distributed systems. This is demonstrated in the developmental number cognition literature, where recent findings suggest a gradience of counting ability in children's learning. We take inspiration from these findings to explore the meaning of symbols in Recurrent Neural Networks (RNNs). We align recurrent neural representations with number symbols by causally intervening on the neural representations. We find that symbol-like representations of numbers do emerge in RNNs. We use this to inform the discussion on how neural systems represent quantity. We also show that the symbol-like representations evolve with learning, and continue to vary after the RNNs solve the task, demonstrating the graded nature of symbols in distributed systems.

Human ecological success stems from our ability to absorb and build upon cultural knowledge, a process we aim to model computationally by integrating individual and cultural learning from language — one of the main vehicles of cultural transmission (e.g. instructions, explanations, stories). In simple video games, our model infers game rules from both interaction data (individual learning) and partial causal models extracted from game descriptions (cultural learning). Given exhaustive descriptions (either hand-written or generated by a model given access to oracle data), models leveraging the two learning sources induce more accurate game rules from limited data than both the individual- and cultural-only controls. Interestingly, descriptions from human game players do not consistently yield better rule induction. We hypothesize that players may preferentially communicate information that will be essential to the others' future decision-making and we aim to investigate cultural transmission by integrating individual and cultural learning with both causal understanding and decision-making.

Humans' capacity for cumulative culture is remarkable: we can build up vast bodies of knowledge over generations. Communication, particularly via language, is a key component of this process. Previous work has described language as enabling posterior passing, where one Bayesian agent transmits a posterior distribution to the next. In practice, we cannot exactly copy our beliefs into the minds of others--we must communicate over the limited channel language provides. In this paper, we analyze cumulative culture as Bayesian reinforcement learning with communication over a rate-limited channel. We implement an agent that solves a crafting task and communicates to the next agent by approximating the optimal rate-distortion trade-off. Our model produces documented effects, such as the benefits of abstraction and selective social learning. It also suggests a new hypothesis: selective social learning can be harmful in tasks where initial exploration is required.

Probabilistic models of language understanding are valuable tools for investigating human language use. However, they need to be hand-designed for a particular domain. In contrast, large language models (LLMs) are trained on text that spans a wide array of domains, but they lack the structure and interpretability of probabilistic models. In this paper, we use chain-of-thought prompts to introduce structures from probabilistic models into LLMs. We explore this approach in the case of metaphor understanding. Our chain-of-thought prompts lead language models to infer latent variables and reason about their relationships in order to choose appropriate paraphrases for metaphors. The latent variables and relationships chosen are informed by theories of metaphor understanding from cognitive psychology. We apply these prompts to the two largest versions of GPT-3 and show that they can improve performance in a paraphrase selection task.

When faced with a polar question, speakers often provide overinformative answers going beyond a simple “yes” or “no”. But what principles guide the selection of additional information? In this paper, we provide experimental evidence from two studies suggesting that overinformativeness in human answering is driven by considerations of relevance to the questioner’s goals which they flexibly adjust given the functional context in which the question is uttered. We take these human results as a strong benchmark for investigating question-answering performance in state-of-the-art neural language models, conductingan extensive evaluation on items from human experiments. We find that most models fail to adjust their answering behavior in a human-like way and tend to include irrelevant information. We show that GPT-3 is highly sensitive to the form of the prompt and only achieves human-like answer patterns when guided by an example and cognitively-motivated explanation.

Humans teach others about the world through language and demonstration. When might one of these modalities be more effective than the other? In this work, we study the factors that modulate the effectiveness of language vs. demonstration using multi-agent systems to model human communication. Specifically, we train neural network agents to teach via language or demonstration in a grounded communication task, manipulating 1) the inherent difficulty of the task and 2) the competence of the teacher. We find that teaching by demonstration is more effective in the simplest settings, but language is more effective as task difficulty increases, due to its ability to generalize more effectively to unseen scenarios. Overall, these results provide converging evidence for a tradeoff between language and demonstration as teaching modalities in humans, and make the novel predictions that demonstration may be optimal for easy tasks, while language enables generalization in more challenging settings.

Formal mathematical reasoning is unique in its precision: any valid conclusion can be justified by a sequence of base axioms. But human-written proofs or solutions rarely operate at that level. Instead, obvious steps are skipped to provide a simple, lucid argument. This is especially important in an educational setting, where too many details in an example solution, or too few, can confuse a student. What are the key steps for humans in a given formal solution? We investigate several computational hypotheses in the context of equation solving. Specifically, we take a reinforcement learning agent that solves equations using low-level axioms, and propose a series of methods for abstracting its solutions by selecting key steps. We consider methods based on the semantic distance between subsequent steps, based on the steps with the highest uncertainty for the agent, and based on transitions between latent "high-level skills" learned from a large number of agent-produced solutions. In a human evaluation we find that skill-base simplifications were judged most useful. These results suggest new directions for understanding human mathematical reasoning.

Psychometric testing is a valuable educational tool for the assessment and monitoring of students’ abilities in core subjects. However, the manual development of these tests is a tedious process requiring test specialists to produce and curate large volumes of high-quality items. In this paper, we consider whether automating test item generation with modern machine learning methods is a feasible solution for obtaining strong psychometric test items at low cost in the domain of sentence reading fluency. We assess the ability of the large neural language model GPT-3 to produce items ``few-shot’’--- from a short prompt with only a handful of examples. Our results show that generated items closely resemble standardized test items in terms of their factual ambiguity, content appropriateness, and complexity. Furthermore, after filtering for correct answer-labeling these generated items possess similar latent psychometric properties to standardized test items, even capturing subtle grade-level variation.

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.

Speakers' referential expressions often depart from communicative ideals in ways that help illuminate the nature of pragmatic language use. Patterns of overmodification, in which a speaker uses a modifier that is redundant given their communicative goal, have proven especially informative in this regard. It seems likely that these patterns are shaped by the environment a speaker is exposed to in complex ways. Unfortunately, systematically manipulating these factors during human language acquisition is impossible. In this paper, we propose to address this limitation by adopting neural networks (NN) as learning agents. By systematically varying the environments in which these agents are trained, while keeping the NN architecture constant, we show that overmodification is more likely with environmental features that are infrequent or salient. We show that these findings emerge naturally in the context of a probabilistic model of pragmatic communication.

A key property of linguistic conventions is that they hold overan entire community of speakers, allowing us to communicateefficiently even with people we have never met before. Atthe same time, much of our language use is partner-specific:we know that words may be understood differently by differ-ent people based on our shared history. This poses a chal-lenge for accounts of convention formation. Exactly how doagents make the inferential leap to community-wide expecta-tions while maintaining partner-specific knowledge? We pro-pose a hierarchical Bayesian model to explain how speakersand listeners solve this inductive problem. To evaluate ourmodel’s predictions, we conducted an experiment where par-ticipants played an extended natural-language communicationgame with different partners in a small community. We ex-amine several measures of generalization and find key signa-tures of both partner-specificity and community convergencethat distinguish our model from alternatives. These resultssuggest that partner-specificity is not only compatible with theformation of community-wide conventions, but may facilitateit when coupled with a powerful inductive mechanism.

A hallmark of human language is the ability to effectively andefficiently convey contextually relevant information. One the-ory for how humans reason about language is presented in theRational Speech Acts (RSA) framework, which captures prag-matic phenomena via a process of recursive social reasoning(Goodman & Frank, 2016). However, RSA represents idealreasoning in an unconstrained setting. We explore the idea thatspeakers might learn to amortize the cost of RSA computationover time by directly optimizing for successful communicationwith an internal listener model. In simulations with groundedneural speakers and listeners across two communication gamedatasets representing synthetic and human-generated data, wefind that our amortized model is able to quickly generate lan-guage that is effective and concise across a range of contexts,without the need for explicit pragmatic reasoning.

Human knowledge accumulates over generations, amplifyingour individual learning abilities. What is the mechanism ofthis accumulation? Here, we explore how language allows ac-curate transmission of conceptual knowledge. We introduce anovel experimental paradigm that allows direct comparison oflearning from examples and learning from language. In ourexperiment, a teacher first learns a Boolean concept from ex-amples; they then communicate this concept to a student in afree conversation; finally, we test both teacher and student onthe same transfer items. We find that learning from languageis both sufficient and efficient: Students achieve accuracy veryclose to their teachers, while studying for less time. We thenexplore the language used by teachers and find heavy relianceon generics and quantifiers. Taken together, these results sug-gest that cultural accumulation of conceptual knowledge arisesfrom the ability of language to directly convey generalizations.

Drawing is a versatile technique for visual communication,ranging from photorealistic renderings to schematic diagramsconsisting entirely of symbols. How does a medium spanningsuch a broad range of appearances reliably convey meaning? Anatural possibility is that drawings derive meaning from boththeir visual properties as well as shared knowledge betweenpeople who use them to communicate. Here we evaluate thispossibility in a drawing-based reference game in which twoparticipants repeatedly communicated about visual objects.Across a series of controlled experiments, we found that pairsof participants discover increasingly sparse yet effective waysof depicting objects. These gains were specific to thoseobjects that were repeatedly referenced, and went beyond whatcould be explained by task practice or the visual properties ofthe drawings alone. We employed modern techniques fromcomputer vision to characterize how the high-level visual fea-tures of drawings changed, finding that drawings of the sameobject became more consistent within a pair of participants anddivergent across participants from different interactions. Takentogether, these findings suggest that visual communicationpromotes the emergence of depictions whose meanings areincreasingly determined by shared knowledge rather than theirvisual properties alone.

Despite their diversity, languages around the world share aconsistent set of properties and distributional regularities. Forexample, the distribution of word frequencies, the distributionof syntactic dependency lengths, and the presence of ambigu-ity are all remarkably consistent across languages. We dis-cuss a framework for studying how these system-level proper-ties emerge from local, in-the-moment interactions of rational,pragmatic speakers and listeners. To do so, we derive a novelobjective function for measuring the communicative efficiencyof linguistic systems in terms of the interactions of speakersand listeners. We examine the behavior of this objective ina series of simulations focusing on the communicative func-tion of ambiguity in language. These simulations suggest thatrational pragmatic agents will produce communicatively effi-cient systems and that interactions between such agents pro-vide a framework for examining efficient properties of lan-guage structure and use more broadly.

One of the greatest challenges of developmental psychology is figuring out what children are thinking. This is particularly difficult in early childhood, for children who are prelinguistic or are just beginning to speak their first words. In this stage, children's responses are commonly measured by presenting young children with a limited choice between one of a small number of options (e.g., "Do you want X or Y?"). A tendency to choose one response in these tasks may be taken as an indication of a child's preference or understanding. Adults' responses are known to exhibit order biases when they are asked questions. The current set of experiments looks into the following question: do children demonstrate response biases? Together, we show that 1) toddlers demonstrate a robust verbal recency bias when asked "or" questions in a lab-based task and a naturalistic corpus of caretaker-child speech interactions, 2) the recency bias weakens with age, and 3) the recency bias strengthens as the syllable-length of the choices gets longer. Taken together, these results indicate that children show a different type of response bias than adults, recency instead of primacy. Further, the results may suggest that this bias stems from increased constraints on children's working memory.

People's estimates of numerical quantities are systematically biased towards their initial guess. This anchoring bias is usually interpreted as sign of human irrationality, but it has recently been suggested that the anchoring bias instead results from people's rational use of their finite time and limited cognitive resources. If this were true, then adjustment should decrease with the relative cost of time. To test this hypothesis, we designed a new numerical estimation paradigm that controls people's knowledge and varies the cost of time and error independently while allowing people to invest as much or as little time and effort into refining their estimate as they wish. Two experiments confirmed the prediction that adjustment decreases with time cost but increases with error cost regardless of whether the anchor was self-generated or provided. These results support the hypothesis that people rationally adapt their number of adjustments to achieve a near-optimal speed-accuracy tradeoff. This suggests that the anchoring bias might be a signature of the rational use of finite time and limited cognitive resources rather than a sign of human irrationality.

Cognitive biases, such as the anchoring bias, pose a serious challenge to rational accounts of human cognition. We investigate whether rational theories can meet this challenge by taking into account the mind's bounded cognitive resources. We asked what reasoning under uncertainty would look like if people made rational use of their finite time and limited cognitive resources. To answer this question, we applied a mathematical theory of bounded rationality to the problem of numerical estimation. Our analysis led to a rational process model that can be interpreted in terms of anchoring-and-adjustment. This model provided a unifying explanation for ten anchoring phenomena including the differential effect of accuracy motivation on the bias towards provided versus self-generated anchors. Our results illustrate the potential of resource-rational analysis to provide formal theories that can unify a wide range of empirical results and reconcile the impressive capacities of the human mind with its apparently irrational cognitive biases.

Words exist for referring at many levels of specificity: fromthe broadest (thing) to the most specific (Fido). What drivesthe emergence of these taxonomies of reference? Recent com-putational theories of language evolution suggest that commu-nicative demands of the environment may play a deciding role.Here, we investigate local pragmatic mechanisms of lexicaladaptation that may undergird global emergence by manipulat-ing context in a repeated reference game where pairs of partic-ipants interactively coordinate on an artificial communicationsystem. We hypothesize that pairs should converge on specificnames (e.g. Fido) when the context requires frequently mak-ing fine distinctions between entities; conversely, they shouldconverge on a more compressed system of conventions for ab-stract categories (e.g. dog) in coarser contexts, even if a finermapping would be sufficient. We show differences in the lev-els of abstraction that emerged in different environments andintroduce a statistical approach to probe the dynamics of emer-gence.

An important challenge for human-like AI is compositional se-mantics. Recent research has attempted to address this by us-ing deep neural networks to learn vector space embeddings ofsentences, which then serve as input to other tasks. We presenta new dataset for one such task, “natural language inference”(NLI), that cannot be solved using only word-level knowledgeand requires some compositionality. We find that the perfor-mance of state of the art sentence embeddings (InferSent; Con-neau et al., 2017) on our new dataset is poor. We analyzethe decision rules learned by InferSent and find that they arelargely driven by simple heuristics that are ecologically validin its training dataset. Further, we find that augmenting train-ing with our dataset improves test performance on our datasetwithout loss of performance on the original training dataset.This highlights the importance of structured datasets in betterunderstanding and improving AI systems.

Across languages, adjectives are subject to ordering restric-tions. Recent research shows that these are predicted by ad-jective subjectivity, but the question remains open why this isthe case. We first conduct a corpus study and not only replicatethe subjectivity effect, but also find a previously undocumentedeffect of mutual information between adjectives and nouns.We then describe a rational model of adjective use in whichlisteners explicitly reason about judgments made by differentspeakers, formalizing the notion of subjectivity as agreementbetween speakers. We show that, once incremental process-ing is combined with memory limitations, our model predictseffects both of subjectivity and mutual information. We con-firm the adequacy of our model by evaluating it on corpus data,finding that it correctly predicts ordering in unseen data withan accuracy of 96.2 %. This suggests that adjective orderingcan be explained by general principles of human communica-tion and language processing.

An essential part of cognitive science is designing experimentsthat distinguish competing models. This requires patience andingenuity—there is often a large space of possible experimentsone could run but only a small subset that might yield informa-tive results. We need not comb this space by hand: If we useformal models and explicitly declare the space of experiments,we can automate the search for good experiments, looking forthose with high expected information gain. Here, we presentan automated system for experiment design called webppl-oed.In our system, users simply declare their models and experi-ment space; in return, they receive a list of experiments rankedby their expected information gain. We demonstrate our sys-tem in two case studies, where we use it to design experimentsin studies of sequence prediction and categorization. We findstrong empirical validation that our automatically designed ex-periments were indeed optimal.

The combinatorial expressivity of natural language enables speakers to communicate a single idea in myriad ways. Howdo speakers decide which utterance to use? Under the Uniform Information Density (UID) hypothesis, speakers shouldplan their utterances to minimize listener comprehension difficulty by spreading out new information, for example, byusing complementizers or avoiding contractions before high-surprisal content. We explore how UID behaviors may resultfrom pragmatic considerations (e.g., social reasoning in context) using a computational pragmatics model. We showthat artificial pragmatic agents communicating under noise conditions exhibit key UID effects: (A) speakers provide cuesbefore high surprisal content, (B) given a UID-cue, listeners infer oncoming content is high-surprisal, (C) synthetic corporagenerated from speakers reflects a signature UID effect: a positive relationship between likelihood of optional elementsand surprisal of oncoming content. Thus, UID may follow from more general principles of pragmatic communication inthe presence of noise.

What cognitive mechanisms support the emergence of linguis-tic conventions from repeated interaction? We present re-sults from a large-scale, multi-player replication of the clas-sic tangrams task, focusing on three foundational propertiesof conventions: arbitrariness, stability, and reduction of ut-terance length over time. These results motivate a theory ofconvention-formation where agents, though initially uncertainabout word meanings in context, assume others are using lan-guage with such knowledge. Thus, agents may learn aboutmeanings by reasoning about a knowledgeable, informativepartner; if all agents engage in such a process, they success-fully coordinate their beliefs, giving rise to a conventionalcommunication system. We formalize this theory in a compu-tational model of language understanding as social inferenceand demonstrate that it produces all three properties in a sim-plified domain.

Speakers often refer to context only implicitly when using lan-guage. The utterance “it’s warm outside” could signal it’swarm relative to other days of the year or just relative to thecurrent season (e.g., it’s warm for winter). Warm vaguely con-veys that the temperature is high relative to some contextualcomparison class, but little is known about how a listener de-cides upon such a standard of comparison. Here, we formalizehow world knowledge and listeners’ internal models of speechproduction can drive the resolution of a comparison class incontext. We introduce a Rational Speech Act model and de-rive two novel predictions from it, which we validate using aparaphrase experiment to measure listeners’ beliefs about thelikely comparison class used by a speaker. Our model makesquantitative predictions given prior world knowledge for thedomains in question. We triangulate this knowledge with afollow-up language task in the same domains, using Bayesiandata analysis to infer priors from both data sets

Why are we polite when we talk to one another? One hypoth-esis is that people expect others to choose what to say basedon their goals both to transfer information efficiently (an epis-temic goal) and to make the listener feel good (a social goal).In our previous work, we found that when these two goals con-flict, they sometimes produce white lies. In the current work,we expand on this theory to consider another prominent case ofpolite speech: indirect remarks using negation (e.g., “It wasn’tamazing”). With minimal extensions from our previous frame-work, our formal model suggests that a pragmatic speaker willproduce more indirect remarks when the speaker wants to beinformative and seem considerate at the same time. Thesepredictions were borne out in a language production experi-ment. These findings suggest that the conflict between socialand epistemic goals can account for a broad range of politenessphenomena.

What governs how much information speakers include in referring expressions? Atypical properties of objects aremore likely to be included in referring expressions than typical ones. E.g., speakers are more likely to call a blue banana a “bluebanana” and a yellow banana a ”banana”. A unified account of this phenomenon is lacking. When should a rational speakermention an object’s color? Reference production is modeled within the Rational Speech Act framework. Utterances (“banana”,“blue”, and “blue banana”) are taken to have a graded semantics: rather than assuming all bananas are equally good instancesof “banana”, we empirically elicited object-utterance typicality values for all possible utterances. Pragmatic speakers selectutterances proportionally to the probability that a literal listener using a graded semantics will select the intended referent. Weevaluate the proposed model on a dataset of freely produced referring expressions collected in an interactive reference gameexperiment via the web.

Bayesian models of cognition posit that people compute prob-ability distributions over hypotheses, possibly by construct-ing a sample-based approximation. Since people encountermany closely related distributions, a computationally efficientstrategy is to selectively reuse computations – either the sam-ples themselves or some summary statistic. We refer to thesereuse strategies as amortized inference. In two experiments,we present evidence consistent with amortization. When se-quentially answering two related queries about natural scenes,we show that answers to the second query vary systematicallydepending on the structure of the first query. Using a cog-nitive load manipulation, we find evidence that people cachesummary statistics rather than raw sample sets. These resultsenrich our notions of how the brain approximates probabilisticinference.

Humans use rich intuitive theories to explain other people’sbehavior. Previous work in lay psychology of behavior havetended to treat emotion as causing primarily unintentional be-havior (e.g., being sad causes one to cry), neglecting how peo-ple incorporate emotions into explanations of rational, inten-tional actions. Here, we provide preliminary explorations intointegrating emotions into a theory of folk psychology. Specif-ically, we show that in the lay theory, people are willing toendorse emotions as causes of intentional actions. Moreover,people readily attribute beliefs and desires as explanations foremotional expressions. This work provides a first step in elabo-rating people’s rich understanding of emotions as an importantcomponent of intuitive social cognition.

Certain content of a linguistic construction can project whenthe construction is embedded in entailment-canceling environ-ments. For example, the conclusion that John smoked in thepast from the utterance John stopped smoking still holds forJohn didn’t stop smoking, in which the original utterance isembedded under negation. There are two main approaches toaccount for projection phenomena. The semantic approach addsrestrictions of the common ground to the conventional meaning.The pragmatic approach tries to derive projection from generalconversational principles. In this paper we build a probabilisticmodel of language understanding in which the listener jointlyinfers the world state and what common ground the speakerhas assumed. We take change-of-state verbs as an exampleand model its projective content under negation. Under certainassumptions, the model predicts the projective behavior and itsinteraction with the question under discussion (QUD), withoutany special semantic treatment of projective content.

How do people understand the pragmatics of spatial language?We propose a rational-speech act model for spatial reasoning,and apply it to the terms ‘in’ and ‘near’. We examine people’sfine-grain spatial reasoning in this domain by having them lo-cate where an event occurred, given an utterance. Our prag-matic listener model provides a quantitative and qualitative fitto people’s inferences.

Habitual sentences (e.g. Bill smokes.) generalize an event overtime, but how do you know when a habitual sentence is true?We develop a computational model and use this to guide exper-iments into the truth conditions of habitual language. In Ex-pts. 1 & 2, we measure participants’ prior expectations aboutthe frequency with which an event occurs and validate thepredictions of the model for when a habitual sentence is ac-ceptable. In Expt. 3, we show that habituals are sensitive totop-down moderators of expected frequency: It is the expec-tation of future tendency that matters for habitual language.This work provides the mathematical glue between our intu-itive theories’ of others and events and the language we useto talk about them.

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.

Nominal reference is very flexible—the same object may becalled a dalmatian, a dog, or an animal when all are literallytrue. What accounts for the choices that speakers make in howthey refer to objects? The addition of modifiers (e.g. big dog)has been extensively explored in the literature, but fewer stud-ies have explored the choice of noun, including its level of ab-straction. We collected freely produced referring expressionsin a multi-player reference game experiment, where we ma-nipulated the object’s context. We find that utterance choiceis affected by the contextual informativeness of a description,its length and frequency, and the typicality of the object forthat description. Finally, we show how these factors naturallyenter into a formal model of production within the RationalSpeech-Acts framework, and that the resulting model predictsour quantitative production data.

People’s beliefs about everyday events are both of theoreti-cal interest in their own right and an important ingredient inmodel building—especially in Bayesian cognitive models ofphenomena such as logical reasoning, future predictions, andlanguage use. Here, we explore several recently used methodsfor measuring subjective beliefs about unidimensional contigu-ous properties, such as the likely price of a new watch. Asa first step towards a way of assessing and comparing beliefelicitation methods, we use hierarchical Bayesian modeling forinferring likely population-level beliefs as the central tendencyof participants’ individual-level beliefs. Three different depen-dent measures are considered: (i) slider ratings of (relative)likelihood of intervals of values, (ii) a give-a-number task, and(iii) choice of the more likely of two intervals of values. Ourresults suggest that using averaged normalized slider ratingsfor binned quantities is a practical and fairly good approxima-tor of inferred population-level beliefs.

Conveying information in a false or indirect manner in consid-eration of listeners’ wants (i.e. being polite) seemingly contra-dicts an important goal of a cooperative speaker: informationtransfer. We propose that a cooperative speaker considers bothepistemic utility, or utility of providing the listener new and ac-curate information, and social utility, or utility of maintainingor boosting the listener’s self-image (being polite). We for-malize this tradeoff within a probabilistic model of languageunderstanding and test it with empirical data on people’s infer-ences about the relation between a speaker’s goals, utterancesand the true states of the world.

What constitutes good teaching, and what factors do learnersconsider when evaluating teachers? Prior developmental worksuggests that even young children accurately recognize andevaluate under-informativeness. Building on prior work, wepropose a Bayesian model of teacher evaluation that infers theteacher’s quality from how carefully he selected demonstrationsgiven what he knew. We test the predictions of our modelacross 15 conditions in which participants saw a teacher whodemonstrated all or a subset of functions of a novel device andrated his helpfulness. Our results suggest that human adultsseamlessly integrate information about the number of functionstaught, their values, as well as what the teacher knew,to make nuanced judgments about the quality of teaching; thequantitative pattern is well predicted by our model

We show how the wide range in strengths of intensifying degreeadverbs (e.g. very and extremely) could be explained bypragmatic inference based on differing cost, rather than differingsemantics. This predicts a linear relationship between themeaning of intensifiers and their length and log-frequency. Wetest this prediction in two studies, using two different dependentmeasures, finding that higher cost does predict strongermeanings. We discuss the implications for adverbial meaningand the more general question of how extensive non-arbitraryform-meaning association may be in language.

World knowledge enters into pragmatic utterance interpretationin complex ways, and may be defeasible in light of speakers’utterances. Yet there is to date a surprising lack of systematicinvestigation into the role of world knowledge in pragmaticinference. In this paper, we show that a state-of-the-artmodel of pragmatic interpretation greatly overestimates the influenceof world knowledge on the interpretation of utteranceslike Some of the marbles sank. We extend the model to capturethe idea that the listener is uncertain about the backgroundknowledge the speaker is bringing to the conversation. Thisextension greatly improves model predictions of listeners’ interpretationand also makes good qualitative predictions aboutlisteners’ judgments of how ‘normal’ the world is in light of aspeaker’s statement. Theoretical and methodological implicationsare discussed.

A popular conception about language development is that comprehensionprecedes production. Although this is certainlytrue during the earliest stages of phonological development,once a child possesses the basic articulatory skills required forimitation, it need not necessarily be the case. A child couldproduce a word without possessing the fully formed lexicalrepresentation through imitation. In some cases, such as inresponse to questions containing fixed choices, this behaviorcould be mistaken for a deeper understanding of the words’semantic content. In this paper, we present evidence that 2-to 3-year-old children exhibit a robust recency bias when verballyresponding to two-alternative choice questions (i.e., theyselect the second, most recently mentioned option), possiblydue to the availability of the second word in phonologicalmemory. We find further evidence of this effect outside ofa laboratory setting in naturalistic conversational contexts inCHILDES (MacWhinney, 2000), a large corpus of transcribedchild-adult interactions.

How do people make causal judgments? Here, we proposea counterfactual simulation model (CSM) of causal judgmentthat unifies different views on causation. The CSM predictsthat people’s causal judgments are influenced by whether acandidate cause made a difference to whether the outcome occurredas well as to how it occurred. We show how whethercausationand how-causation can be implemented in terms ofdifferent counterfactual contrasts defined over the same intuitivegenerative model of the domain. We test the model in anintuitive physics domain where people make judgments aboutcolliding billiard balls. Experiment 1 shows that participants’counterfactual judgments about what would have happened ifone of the balls had been removed, are well-explained by anapproximately Newtonian model of physics. In Experiment 2,participants judged to what extent two balls were causally responsiblefor a third ball going through a gate or missing thegate. As predicted by the CSM, participants’ judgments increasedwith their belief that a ball was a whether-cause, ahow-cause, as well as sufficient for bringing about the outcome.

What makes a question useful? What makes an answer appropriate?In this paper, we formulate a family of increasinglysophisticated models of question-answer behavior within theRational Speech Act framework. We compare these modelsbased on three different pieces of evidence: first, we demonstratehow our answerer models capture a classic effect in psycholinguisticsshowing that an answerer’s level of informativenessvaries with the inferred questioner goal, while keepingthe question constant. Second, we jointly test the questionerand answerer components of our model based on empirical evidencefrom a question-answer reasoning game. Third, we examinea special case of this game to further distinguish amongthe questioner models. We find that sophisticated pragmaticreasoning is needed to account for some of the data. Peoplecan use questions to provide cues to the answerer about theirinterest, and can select answers that are informative about inferredinterests.

In standard decision theory, rational agents are objective,keeping their beliefs independent from their desires (Berger,1985). Such agents are the basis for current computationalmodels of Theory of Mind (ToM), but this fundamental as-sumption of the theory remains untested. Do people think thatothers’ beliefs are objective, or do they think that others’ de-sires color their beliefs? We describe a Bayesian frameworkfor exploring this relationship and its implications. Motivatedby this analysis, we conducted two experiments testing the apriori independence of beliefs and desires in people’s ToMand find that, contrary to fully-normative accounts, peoplethink that others engage in wishful thinking. In the first ex-periment, we found that people think others believe both thatdesirable events are more likely to happen, and that undesir-able ones are less likely to happen. In the second experiment,we found that social learning leverages this intuitive under-standing of wishful thinking: participants learned more fromthe beliefs of an informant whose desires were contrary to hisbeliefs. People’s ToM therefore appears to be more nuancedthan the current rational accounts, but consistent with a modelin which desire directly affects the subjective probability ofan event.

The causal frame problem is an epistemological puzzle abouthow the mind is able to disregard seemingly irrelevant causalknowledge, and focus on those factors that promise to be usefulin making an inference or coming to a decision. Takinga subject’s causal knowledge to be (implicitly) represented interms of directed graphical models, the causal frame problemcan be construed as the question of how to determine a reasonable“submodel” of one’s “full model” of the world, so asto optimize the balance between accuracy in prediction on theone hand, and computational costs on the other. We propose aframework for addressing this problem, and provide several illustrativeexamples based on HMMs and Bayes nets. We alsoshow that our framework can account for some of the recentempirical phenomena associated with alternative neglect.

Despite its importance, human collective intelligence remainsenigmatic. We know what features are predictive of collectiveintelligence in human groups, but we do not understandthe specific mechanisms that lead to the emergence of this distributedinformation processing ability. In contrast, there isa well-developed literature of experiments that have exposedthe mechanisms of collective intelligence in nonhuman animalspecies. We adapt a recent experiment designed to studycollective sensing in groups of fish in order to better understandthe mechanisms that may underly the emergence of collectiveintelligence in human groups. We find that humans inour experiments act at a high level like fish but with two additionalbehaviors: independent exploration and targeted copying.These distinctively human activities may partially explainthe emergence of collective sensing in our task environment atgroup sizes and on times scales orders of magnitudes smallerthan were observed in fish.