Building Machines that Learn and Think with People
Author: Katherine M. Collins. Ilia Sucholutsky, Umang Bhatt, Kartik Chandra, Lione Mina, Lee Cedegao, E. Zhang, Tan Zhi-Xuan, Mark Ho, Vikash Mansinghka, Adrian Weller, Joshua B. Tenenbaum, Thomas L. Griffiths
DOI:10.48550/arXiv.2408.03943Corpus ID: 271768888 Building Machines that Learn and Think with People Katherine M. Collins, Ilia Sucholutsky, +10 authors Thomas L. Griffiths Published in Nature Human Behaviour 22 July 2024 Computer Science, Philosophy, Psychology
0 General Document comments
0 Sentence and Paragraph comments
0 Image and Video comments
Abstract
What do we want from machine intelligence? We envision machines that are not just tools for thought, but partners in thought: reasonable, insightful, knowledgeable, reliable, and trustworthy systems that think with us. Current artificial intelligence (AI) systems satisfy some of these criteria, some of the time. In this Perspective, we show how the science of collaborative cognition can be put to work to engineer systems that really can be called “thought partners,” systems built to meet our expectations and complement our limitations. We lay out several modes of collaborative thought in which humans and AI thought partners can engage and propose desiderata for human-compatible thought partnerships. Drawing on motifs from computational cognitive science, we motivate an alternative scaling path for the design of thought partners and ecosystems around their use through a Bayesian lens, whereby the partners we construct actively build and reason over models of the human and world.
“Desiderata” means “things that are desired or wanted,” usually in a thoughtful, almost technical sense (like a wish list of ideal properties or goals).
In that sentence, the authors are saying: we’re not just describing how human–AI collaboration works now, we’re also naming the specific qualities they want those collaborations to have so they are “human-compatible.”
What do you think some of those “human-compatible” qualities might be in the context you are reading? Want to hear more?
You are already naming a big piece: using machines (including AI) to help humans get things done, but that sentence is nudging you to think more specifically about how the thinking itself is shared. It is not just “we use tools,” it is “we and the tool are co-thinking in different modes, and here are the ideal conditions (desiderata) that make that collaboration good for humans.” You might be “missing” (for now) the idea that they want to classify different patterns of human–AI thinking together and then spell out what would make each pattern safe, beneficial, and aligned with human values.
If you think about your own experience, what “must-have” qualities would you want in an AI thought partner so that the collaboration actually feels trustworthy and useful to you? Want to hear more?
Some of the “must-have” qualities you are already circling are awareness, traceability, and choice: being able to notice when AI enters your workflow, how it shapes your thinking, and whether that feels right for your purposes in that moment. In that sense, “human-compatible” is not one fixed standard but something like: AI that can flex with your creative rhythms, foreground your own voice, and make its own contributions legible enough that you can accept, resist, or remix them consciously.
If you take your motto seriously—always starting from human creativity—one challenge is how to redesign your practices so that even when AI is “early,” it is still in service of amplifying and clarifying what matters to you, not replacing it. How might you map out two or three distinct “workflows” you actually use (for example: pure human drafting, human-then-AI, AI-then-human remix) and name what feels most and least compatible with your creative ethics in each?
Want to hear more?
Understanding “desiderata” in that sentence
- Meaning: “Desiderata” (singular: “desideratum”) is a Latin word meaning “things that are desired or required” – basically, a set of ideal goals, criteria, or must-haves.
- In this article: When they say they “propose desiderata for human-compatible thought partnerships,” they mean they are listing the key qualities or requirements they think good human–AI partnerships should have.
- Why that fancy word: Using “desiderata” signals that they are not just dreaming or predicting; they are setting out a kind of normative checklist or design standard for what counts as a truly “human-compatible” AI thought partner.
Do you want to dig into what some of those “must-have” qualities for a human–AI thought partnership might be (for example: trust, transparency, complementing human weaknesses, etc.), and try to identify them from nearby sentences in the article?
Going deeper into that sentence
- Big picture of the sentence: They are saying: first, they will describe different ways humans and AI can think together; then, they will spell out a list of “must-have” conditions (the desiderata) for making those collaborations truly “human-compatible.”
- Hidden question in that line: The phrase “human-compatible thought partnerships” hints that not every human–AI collaboration is actually good for humans; some designs might clash with how we think, feel, or decide, so they want criteria that protect and support human ways of thinking.
- Deeper idea: “Modes of collaborative thought” plus “desiderata” together suggest they are not only describing what is happening now, but also setting a standard: which kinds of human–AI thinking should we aim for, and what rules should guide how we build them?
If we zoom out from this one sentence, the next move is to guess what might be on their “desiderata” checklist—things like trust, shared goals, transparency, and respecting human limits. Want to look at the surrounding paragraph and try to name 2–3 qualities you think they would absolutely insist on for a “human-compatible” AI partner?
1. First, a quick answer to your two questions
-
What is “desiderata”?
It is a Latin word that means “things that are desired” or “things we want to have.” -
Why do the authors use that term?
They use it to sound precise and formal: they are listing the important goals or requirements they think good, human-compatible AI partnerships should have, not just casual wishes.
2. Connecting the word to your sentence
In the sentence you gave, they say they “propose desiderata for human-compatible thought partnerships,” which means they are giving a set of standards or conditions that such partnerships ought to meet. It is like saying: “Here are the key features that any good human–AI thinking relationship should have.”
3. Questions to help you think more deeply about it
- If “desiderata” are “things we really want,” what kinds of qualities do you think you would want in an AI thought partner?
- In that sentence, do you think the authors are describing how things are right now, or how they would like things to be in the future?
- How does it change the meaning if we replace “propose desiderata” with “make suggestions” or “list some hopes” — does it feel stronger or weaker to you?
4. Want to explore the word itself?
We could play with the word by comparing it to other forms like “desire,” “desired,” or “desideratum” (the singular form). We could also ask why authors sometimes choose a Latin word instead of simpler English, and what effect that has on readers.
Would you like to look more closely at the word itself, or talk more about what “desiderata for human-compatible thought partnerships” might actually include?
Step One: What’s important in that sentence?
Here are three key ideas hiding inside that one line about “modes of collaborative thought” and “desiderata for human-compatible thought partnerships”:
- “Modes of collaborative thought” — This is important because it says humans and AI do not just “use” each other; they can think together in different patterns or roles (like coach–player, peer–peer, teacher–student). One bit of background: in cognitive science and HCI (human–computer interaction), people study how tools change the way we think, not just what we think, and AI is now seen as a kind of thinking partner rather than a simple tool.
- “AI thought partners” — This matters because it treats AI as a partner in reasoning, planning, or creating, not a passive calculator; that creates new questions about responsibility, trust, and influence. Background: recent work in “centaur” systems in chess, coding copilots, and writing assistants all treat human+AI as joint problem solvers, raising ethical concerns about over‑reliance and blurred authorship.
- “Desiderata for human-compatible thought partnerships” — This is crucial because “desiderata” means they are trying to list the core requirements that make these partnerships safe, respectful, and actually helpful for humans, not harmful or manipulative. Background: in AI safety and alignment, “human-compatible” usually means systems whose goals, behaviors, and impacts fit well with human values, limits, and vulnerabilities, instead of exploiting them.
Next Step Preview
In the next step, I can turn each of those three ideas into questions you can explore:
• What kinds of “modes of collaboration” do you actually want?
• What makes an AI a “partner” instead of a tool?
• What concrete requirements must be on the list so the partnership stays truly human‑compatible?
These will be questions for you to think about, not answers from me, so you can shape your own picture of a good human–AI partnership. Then, if you like, we can look at places to research those questions more.
Are you ready to see those next questions and wrestle with what should be on your own “desiderata” list for an AI thought partner? If yes, which piece pulls you more: the “modes of collaboration,” the idea of “thought partner,” or the “human‑compatible” requirements?
Computers have long been seen as tools for thought. Steve Jobs called computers “bicycles for the mind”: tools that dramatically increase the efficiency, productivity, and joy of thinking. Now, thirty years later, this metaphor is beginning to change. Computer systems are increasingly referred to not as vehicles but as “copilots”1, 2: we have moved from designing tools for thought to actual partners in thought.
The current wave of AI technologies, particularly language models, have catalyzed this transition. Users no longer have to know how to write code to engage intimately with computers; we can now interface through the medium of natural language. Humans already think alone and together, at least communicated often through the medium of language 3. We long have – from developing new modes of thinking through questioning and debate to teaching and learning through language. The apparent power of these new systems – getting closer to the kind of artificial intelligence (AI) imagined in the field’s early days 4, 5, 6, 7, 8, 9 – as well as challenges faced by the current iterations of such systems – invites us to think about what it will take to build systems that truly act as effective thought partners. We argue that good thought partners are systems (1) which can understand us, (2) which we can understand, and (3) which have sufficient understanding of the world that we can engage on common ground.
One path to building such thought partners is to scale foundation models (e.g., LLMs 10) with large amounts of human demonstrations and feedback, along with “traces” of human thought scraped from web-scale data 11, 12, 13. While such an approach has produced systems that accurately mimic human behavior (e.g., producing fluent text), these machines do not robustly simulate human cognition (e.g., explicitly reasoning about the world or other minds) in ways expected by a true thought partner 14, 15, 16, 17, 18, 19, 3, 20.
What would it take to design systems that meet our criteria? One promising path is to design systems that build explicit models of the task, world, and human (where these models are structured 21, rather than distributionally learned from data) – drawing on formal frameworks grounded in cognitive psychology for understanding how humans think, alone and together. In this Perspective, we chart a new vision for the design of AI thought partners. Decades of work in the behavioral sciences provide valuable insights for designing human-centric, uncertainty-aware thought partners. Drawing on such research, we argue that effective thought partners are those which build models of the human and the world.
This toolkit includes foundation models 22, 23, 24, but is not limited to them. Indeed, foundation models like LLMs are fueling new motifs for thinking about human minds in computational terms (e.g., “rational meaning construction” 16) interleaved alongside techniques from probabilistic programming 25, 26, 27, 28, 29, goal-directed search 30, 31, 32, and other explicit, structured representations, e.g., of agents thinking about other agents 33, 34, 35. We already have tools that help us build machines that learn and think like people 36. We propose applying that toolkit to collaborative cognition – to build machines that learn and think with people.
When we think, we draw coherent inferences, make predictions, and act on these predictions – from assessing what birthday present to gift a treasured friend, to formulating a new scientific hypothesis and experiment plan to evaluate a theory. We flexibly draw on prior knowledge and update our beliefs through experience (as we discuss below). We not only solve problems, but imagine new ones 37. And we think together. For generations, humans have discussed and debated ideas, and developed ecosystems to disseminate such thoughts to new audiences. Much scientific innovation has come through collaboration, where advances are frequently fueled by engaging with diverse partners who offer new ideas yet share our values 38.
2.1Modes of Collaborative Thought
As an illustration of the many ways that people and machines might think with each other, we highlight a few modes of collaborative thought (Table 1). This set of modes, partly inspired by characterizations of thinking and reasoning in psychology 39, 40, are not meant to be comprehensive of all aspects of thought. Rather, we see these modes as ripe for the further development of AI thought partners.
We next outline a few diverse domains in which the development of AI thought partners able to truly collaborate with humans (Figure 1) may be particularly valuable. We highlight common computational challenges that arise when considering what effective partnership might look like in each domain, foreshadowing our proposed desiderata. We later return to these case studies with concrete human-centric thought partner instantiations.
Thought Partners for Programming. Programming is a cognitively-demanding activity that requires gaining fluency in translating human intentions into formal, machine-interpretable languages. It is no surprise that decades of effort have gone into designing tools to help people program 41, 42, 43, 44, 45. New “programming assistant” tools like GitHub Copilot have rapidly gained enormous popularity and attention; however, these tools are often unreliable 46, 47, 48, e.g., failing to understand users’ intentions 49 and generating bugs that may be particularly risky alongside beginner programmers 50. Programming involves much more than just accurate in-line code suggestions – which, at the time of writing, GitHub Copilot specializes in. Humans plan abstract, structural decisions and collaboratively learn, and need partners who can answer our questions – like why code behaves as it does, or fails to work. A good collaborative programming partner seeks to understand not only the programming language, but also their fellow programmer, inferring and reasoning about our overarching intentions, and adapting to both what we do and do not know.
Thought Partners for Embodied Assistance. Ensuring embodied agents can form accurate and physically-realizable plans is foundational for effective assistance we can trust – from guessing what a friend wants when we help them cook 51, to working with someone with different physical abilities 52, or carrying out a high-stakes search-and-rescue mission 53. While much current research on embodied AI and assistive robots focuses on learning specific skills or following simple instructions 54, 55, 56, evaluations suggest that even state-of-the-art language models fine-tuned on extensive human feedback continue to struggle with tasks that require reliable, effective planning towards novel goals 57, 58. Instead, ideal assistive partners understand our actions, words, and instructions as expressions of goals, beliefs, and intentions 59, 60, 61 that are grounded in physical possibilities 62, while also understanding that these can be shared across multiple minds 63, 64, 65. In addition, effective partners account for each others’ limitations in perception, planning, and world modeling, correcting for possible mistakes 66, 67, and acting so as to make their intentions more legible 68, 69.
Thought Partners for Storytelling. Another domain in which we may want thought partners is storytelling – for writers, filmmakers, and even scientists. Storytelling is a complex, iterative cognitive process 70, 71 with substantial opportunities for thought partners to collaboratively ideate and create with humans from helping brainstorm new ideas, generate storylines, and improve their writing style and tone 72, 73, 74, 75, 76, 77. For this process to be productive, a thought partner needs to understand more than just our authorial intentions and dispositions – they also need to understand the audience we are speaking to (that is, to understand the social world), including audience expectations and likely interpretations of the stories we are crafting for them.
Thought Partners for Medicine. Doctors need to sensemake, plan, deliberate, and continually learn in the face of new medical evidence. A primary care doctor is not unlike Sherlock Holmes – collating and integrating disparate bits of evidence with their prior beliefs to make decisions under uncertainty. Yet, doctors rarely have enough time to engage deeply with each patient 78, driving high rates of burnout with knock-on effects on patient care quality 79. Can we develop safe, reliable thought partners that can free doctors up to spend more time and communicate better with their patients? Already, foundation models are becoming proficient in medical assessments 80, 81, seemingly capable of easing the heavy burden on doctors by assisting and partnering 82, 83, and even providing preferable responses to patients 84. Yet, it is not clear that these systems understand us (and our cognitive limitations), understand the world (underlying biology), and enable us to understand them (which in this context, may be important for transparency and reliability 85, 86, 87, 88).
What then do we want from thought partners? There are many criteria for tools for thought that are of course relevant: efficiency, accuracy, robustness, fairness, cost, scalability, etc. But the domains above illuminate that what is distinctive about a thought partner is its relationship to the user 89. Looking to ideas the behavioral sciences motivates three desiderata to guide the design of human-centered thought partners:
-
1.
You understand me: We would like our thought partners to understand our goals, plans, (possibly false) beliefs, and resource limitations, taking into account what they have observed of us in the past and present in order to best collaborate with us in the future 90, 91. For example, a thought partner should adaptively change strategies when working with an expert, layperson, or child, meeting us where we are.
Paragraph 24 0
Nov 9Paul A Paul A : "You understand me" cuts a couple of different ways for Writing Partners. [Edited] more1. Can we create writing partners that are sensitive to and can infer a writer’s experience as a writer? I think we can, and we should do this.
2. We also have a third box where it would be great to teach writers to describe themselves. They can start by cutting and pasting from their Bios. But we can also have students become more and more descriptive of their process and what has been relatively easy and where they struggled. They need to reveal themselves — their creative selves, not their PII — to the Writing Partners. So with this one, we need to teach students how to become seen by the AI. -
2.
I understand you: We would like our thought partners to act in a way that is legible to us 68, 92, and communicate with us in the way we intuitively understand 93, 94, 95.
-
3.
We understand the world: We would like our thought partners to be tethered to reality 96. This means being accurate and knowledgeable, but also working with a shared representation of the world, domain, or task 97, 98, 99. Further, our use of ‘we’ emphasizes that thought partnerships are fundamentally about synergy, moving beyond the sum of its parts.
|
Mode |
Ongoing Challenges |
Sampling of Existing Systems |
|---|---|---|
|
Collaborative planning• Joint decision-making• Decentralized cooperation• Goal and task assistance |
Reliable goal inferenceValue and intent alignmentScalable multi-agent planning |
Collaborative robots 68, 100Video game sidekicks 101, 102Language-based assistants 35, 103 |
|
Collaborative learning• Pair & team problem-solving• Identification of knowledge gaps• New problem construction |
Strong & robust problem-solving abilitiesPersonalized curriculum pacingProblem construction of targeted difficulty |
Programming learning aids 104, 105, 106, 107 |
|
Collaborative deliberation• Debate & argumentation• Critical review & discussion• Consensus formation |
Opinion diversityVerifiable reasoningFormation of common ground |
Machine-assisted debating 110, 111, 112Consensus writing & opinion mapping 113, 114 |
|
Collaborative sensemaking• Explanation• Visualization• Data Analytics |
Exponential increases in data producedAccessible communicationFidelity of insights to the world |
Probabilistic data modeling 115, 116, 117, 118, 119Machine-assisted theory discovery 120, 121, 122 |
|
Collaborative creation & ideation• Co-design• Idea critiquing• Brainstorming |
Generation diversityStyle consistencyModular customizability |
Machine-assisted writing72, 74, 123Prompted image creation 124, 125, 126Collaborative sketching 127, 128, 129 |
3Engineering Human-Centered Thought Partners
Our core proposal is that our three desiderata can be engineered explicitly, building on theoretical motifs from computational cognitive science and cognitively-informed AI (summarized in Table 2), rather than left as emergent and potentially brittle properties arising implicitly in systems trained for other ends 20. Here, we articulate a framework for engineering thought partners designed to robustly and explicitly function as cooperative, collaborative actors. Humans are far from homogeneous, perfectly rational oracles, nor are we so unpredictable that it is hopeless to model human behavior. We argue that models that explain human cognition and choice as approximately optimal solutions given goals and constraints provide an ideal starting point for designing thought partners, and that a Bayesian formalism provides a probabilistically-sound common conceptual language that facilitates cross-talk between different disciplines 22, 155, 156.
3.1Implementing Our Desiderata
What does it take to engineer real systems that meet our desiderata? First, we propose that a thought partner that understands us should explicitly model its human collaborator as such – as a cooperative agent with structured internal beliefs, knowledge, and goals – and fundamental resource limitations. Second, engineering a thought partner that we can understand benefits from looking at how humans model other humans; just as a good human collaborator seeks to learn and adapt to the relative strengths, imperfections, and computational bounds of their partner, we can build machine thought partners that also reason about the computational demands they are placing on another agent such that we can appropriately predict their behavior 18, 157. Finally, to build thought partners that understand the world – and learn and think synergistically alongside us – we argue that it is valuable to build on structured computational toolkits for grounding shared goals and communication into the environment and domain in which collaboration takes place.
3.2Computational Cognitive Science Motifs
We now (non-exhaustively) spotlight several key insights about modeling humans, modeling humans modeling humans, and modeling humans modeling the world from computational cognitive science – “motifs” for reverse engineering the mind (Table 2) – that we believe can inform engineering of human-centered thought partners. While we acknowledge that there are communities within cognitive science that may disagree with some of these theories, we emphasize that the computational underpinnings of the motifs hold tremendous engineering potential for building thought partners in practice.
Probabilistic Models of Cognition. Decades of work in computational cognitive science have demonstrated the power of modeling aspects of human cognition as Bayesian inference through structured probabilistic generative world models 131, 158, 159, 21, 137. Such approaches have found empirical success in capturing a diversity of facets of human cognition from early word learning 160, to visual perception 161, 162, physical reasoning 99, 163, 164, concept learning 165, 166, 167, language processing and acquisition 158, 168, 169, 170, causal inference in children 171, 172 and adults 173, 174, memory reconstruction 175, and theory formation 176, 177, among many others. Probabilistic models of cognition, particularly those built using a Bayesian approach, have offered principled formalisms in capturing rapid belief updating 178 and how we may integrate our commonsense world knowledge with new evidence to inform the actions and decisions we take in the world 149. Probabilistic inference over structured representations, particularly drawing on Bayesian modeling and tools like meta-level Markov Decision Processes 179, has provided a computational account of how humans plan so flexibly, with the capability of forming rich hierarchical goals and subgoals, across varied timescales 149, 155, 180, 181, 182.
Theory of Mind and Communication. In our quest to build systems for collaborative cognition, we are guided by the success of Bayesian accounts of how we reason about others’ mental states, and how we communicate about them. In particular, Bayesian treatments of theory of mind (ToM) have offered strong accounts for how we may rapidly reason about each others’ beliefs, desires, goals, and intentions 33, 183, 184, 185, 147. We may build mental models 186, 187 of our thought partners, which can in turn be used to support communication and collaboration, informing the way we teach 188, 189, 190, infer whether to rely on a partner for help 191, and support rapid, flexible adaptation to new conversation partners 192, 193. We call particular attention to the Rational Speech Act (RSA) framework 150, 59, which models communicative partners as recursively reasoning about each others’ minds to inform what to say (from the perspective of the speaker) and how to interpret a received utterance (as the listener). Bayesian models provide a useful framework for formalizing such rich cross-partner inferences, allowing both social cognition and communication to be modeled with the same computational toolbox 194, 195.
Resource-Rationality and Tractable Theory-Building. Human brains also have limited resources such as time, memory, and attention that shape what we think about, how long we spend thinking, and even how we communicate our thoughts to others 196. Thus, we sometimes make systematically biased inferences 197, 198. Such “erroneous” judgments can be captured by modeling humans as making rational use of our finite resources; e.g., via approximate inference 178, 199 or bounded planning 67. Crucially, human cognition is tractable 200. Indeed, we can navigate large, potentially unbounded, hypothesis spaces to build theories of the world: a process that seems to demand some kind of heuristics and approximations, which may be resource-rational 182, 201, 196, 143, 142, 202, 17. One approach to modeling minds advocates thinking about humans, as “world model builders” (or “hackers”) – conducting experiments and updating our beliefs about compressed representations of the world, where these representations may be expressed as programs 138, 176. Such representations – bolstered by tools like program synthesis – help explore suboptimal behavior 203.
3.3Scaling Thought Partners via Probabilistic Programming
If Bayesian thought partners are to reason over models of their human thought partner and the world, these models need to continually evolve as new facts come to light and as the human thought partner themselves grows in their expertise, beliefs, and needs. Probabilistic programming 26 provides one powerful methodology for building, scaling, and performing inference in these kinds of rich models. For example, probabilistic programs can be learned from data 204, 116, and synthesized via LLMs that encode rich priors 16, 118, 205. Probabilistic programs also enable fast approximate inference in world models that cohere with human common-sense knowledge and domain expertise 115, 206, where the learned models are themselves amenable to modular inspection and editing by humans. Modern probabilistic programming languages 25, 27, 207 offer not just generic inference but programmable inference, that is, they automate the math for hybrids of optimization 208, 209, dynamic programming 210, and Monte Carlo inference 211. While such frameworks are certainly not the only methods to handle uncertainty and build effective and robust thought partners, we believe they are one promising and cognitively-grounded approach to instantiating thought partners today, as we discuss in our case studies.
3.4Infrastructure around Thought Partners
The design of systems that learn and think with people necessitates not only careful construction of the thought partner (i.e., the machine itself), but also the infrastructure within which human and computational thought partners collaborate 157. Questions like “when and where should a human be able to engage a computational thought partner to ensure effective and appropriate use?” or “for a given problem, is the human or computational thought partner better suited to start first, in light of their respective strengths and weakness, costs of the task at hand, and particular mode of thought?” inform the design of the workflow that surrounds thought partnership. This sociotechnical ecosystem may be dictated by external regulations, organizational practices, or other principles 212, 73, 213, 214, 215, and crucially informed by studies of human behavior. For example, Article 14 of the EU AI Act requires users of high-risk AI systems “to correctly interpret the high-risk AI system’s output” and “to remain aware of the possible tendency of automatically relying or over-relying on the output.” Satisfying such requirements begets not only careful design of thought partners (e.g., that we can understand), but demands careful design of the system of affordances 216, 217 and infrastructure around thought partnerships (for instance, communicating back to humans information about their reliance strategies). Disentangling thought partners from the infrastructure around them provides a modular scaffold for addressing unintentional thought partnership behavior, e.g, overreliance 218 and “illusions of understanding” 219. Bayesian modeling has already found success in inferring humans’ reliance strategies 220 and regions of the task space where a human versus machine can complement one another 221.
4Case Studies in Engineering Thought Partners
We now return to the example domains previously introduced and discuss specific case studies (depicted in Figure 2). Our goal is to demonstrate the potential benefits of endowing thought partners with structured probabilistic models of the human and/or world, and provide a flavor of the kinds of infrastructure questions that may surround them to ensure that the thought partners we build work with people.
4.1Thought Partners for Programming
We highlighted some visions for effective programming partnerships, such as a partner that can address “why” questions. One recent idea, from Chandra et al. 106, is to apply the Bayesian toolkit to explain surprising behavior of computer programs in a human-like way. Chandra et al., apply Bayesian models of mental state inference and rational communication 222 to design a system called “WatChat” that answers questions like “why did program output result ?” in a principled, human-like way. WatChat infers what erroneous mental model might cause the programmer to have expected something different (partner understands user) and generates an explanation that “debugs” that mental model (user understands partner). WatChat represents possible mental models themselves as “programs” whose “bugs” correspond to possible misconceptions; mental models can thus be inferred by Bayesian program synthesis (see Table 2). Such a framework can also be inverted to help design new questions for teachers or self-driven learners to identify misconceptions.
4.2Thought Partners for Embodied Assistance
Recall the challenge of collaboratively planning uncertain tasks, from a search-and-rescue mission to everyday cooking, wherein we typically want to infer shared goals and communicative intent from our partners. This cooperative logic can be modeled in a Bayesian architecture called Cooperative Language-Guided Inverse Plan Search (CLIPS) 35. By modeling humans as cooperative planners who use language to communicate joint plans to achieve their goals 65, CLIPS is able to infer those plans and goals from both the actions and instructions of human collaborators. This allows CLIPS to pragmatically follow human instructions, using context to disambiguate the multiple meanings that a request might have, while pro-actively assisting with the goals that underlie the instruction. For example, CLIPS can understand the likely intentions behind an instruction like “Can you prepare the vegetables while I knead the dough?”, inferring the shared goal of making pizza. These capabilities are made possible by using probabilistic programming infrastructure 25 to unite algorithms for Bayesian inverse planning 33, 184 and human-AI alignment 223, 51, 61 with LLMs. In particular, by using LLMs to evaluate the probability of a natural language instruction given a possible intention, CLIPS can infer intentions from natural language in a coherent Bayesian manner – demonstrating the power of combining tools from the Bayesian thought partner toolkit.
4.3Thought Partners for Storytelling
Storytelling is about crafting experience. Can we also apply the toolkit to help storytellers design experiences from first principles? Recent work has shown that a system grounded in Bayesian ToM can predict and even design interventions on the audience’s experience of a story 224, 225. Chandra et al. conceive of storytelling as “inverse inverse planning”: that is, starting with human social cognition, modeled as Bayesian inverse planning 33, and then optimizing narrative events to shape the model’s inferences over time. They show how a variety of storytelling techniques – from plot twists to stage mime – can be expressed in the language of inverse inverse planning to create animations that have a desired cognitive effect on viewers. Herein, we also highlight the breadth of thought partners for media beyond language, though the framework does nicely suggest a variety of natural extensions, such as integration into tools for creative writing 72, 73, 74, 75, 76, 77.
4.4Thought Partners for Medicine
Finally, we envision medical thought partners both understand us – reasoning about the doctor, patient, and care team as agents with goals, beliefs, and worries – and complement our capabilities, integrating swaths of evidence that exceed our cognitive capacities to inform diagnosis and treatment. While no system yet meets our desiderata for these criteria, we believe a range of motifs and tools from the Bayesian thought partner toolkit here can support the development of such systems for collaborative sensemaking and deliberation. We imagine Bayesian thought partners that can update their medical world knowledge in light of new insights in biology, e.g., editing a code snippet of the underlying probabilistic world model 16 or growing the representation in a non-parametric hierarchical Bayesian model 135. Such a model can then, similar to WatChat, synthesize new questions to ensure the human doctor’s own medical world model is sound. Early work demonstrates that we can employ elements of our toolkit, specifically probabilistic programming, to learn rich generative models for oncology and support efficient user queries 227. Yet, effective medical thought partners beckon a broader view of the ecosystem in which they are deployed 89, 228. If a doctor is over-relying on the output of the thought partner, or overburdened amidst a surge in patient queries, infrastructure around the human and thought partner can modulate when a patient query is either routed to a human or the AI thought partner, or deemed necessary of collaborative planning 229. Systems for routing based on probabilistic modeling are already proving successful in simulation 230.
There is much exciting work to be done to characterize when and how to build thought partners across modes of collaborative thought, which can advance the dissemination and creation of new knowledge alongside humans. We next lay out several key challenges for researchers and designers intent on pursuing a human-centered program of building machines that learn and think with people.
While there is substantial work to be done characterizing the space of possibilities for a single human and single AI thought partner (“dyadic”), we envision a future where many humans and many machines engage (“non-dyadic”), across roles and specialties in increasingly complex social systems 231, engage in the realm of thought 232, 233, 234. Already, researchers are exploring non-dyadic versions of many of the modes of thought and case studies laid out above, including collaborative learning with groups of humans accompanied by an AI thought partner 235 and medical robot collision avoidance systems that need to account for multiple humans 236. As in the dyad setting, extensions to non-dyadic settings can be bolstered by a deepening understanding of human behavior in groups – expanding the Bayesian thought partner toolkit – as is already underway in the study of convention formation 192, 237. Looking ahead, citizen science is a promising example of the opportunities of creating large networks of humans and thought partners: Zooniverse, a large-scale galaxy classification crowdsourcing project, serves as a case study for exploring smart task allocation, blending human and machine classifications, and infrastructure changes that impact human participation and performance with outcomes including both iterative scientific progress and serendipitous scientific discovery 238.
The assessment of thought partners demands a multi-faceted, cross-disciplinary suite of approaches. At minimum, the evaluation of AI thought partners must include some element of interactivity 239. Recent works have highlighted deficits in static evaluation of foundation models 240, 15, demonstrating the need for considering the interaction process in addition to the final output, the first-person perspective in addition to the third-party perspective, and notions of preference beyond quality. In addition to interactive user studies, we posit that to study different kinds of thought partners across modes of collaborative thought would benefit from a controlled, yet rich, playspace; games provide one such domain. Games offer a good formalism for the study of repeated interactions between multiple agents and grounds to explore rich patterns of thought, in social collaborative settings 241, 242, 243, 244.
5.3Risks and Important Considerations
Computational thought partners are by no means a guaranteed nor universal good and come with certain risks. We call out three such spheres of risk: (i) reliance, critical thinking, and access, (ii) anthromorphization, and (iii) misalignment.
First, AI thought partners could induce over-reliance and impair the development of critical thinking skills 245, 246, 247, 219, potentially acting as “steroids” for the mind 248. We are concerned about these risks; our emphasis on the infrastructure around thought partner use is explicitly intended to help practitioners take steps to address these challenges, motivating further design of infrastructure modifications like cognitive forcing functions 249, 250. Conversely, it is possible that some people may under-rely on a thought partner, particularly if there is inadequate AI literacy training for how to best make use of new thought partners 251, 252, 253. Already, research has found that the kinds of queries people make of AI systems can be informed by the amount of prior experience they have interacting with chatbots 15 meaning students, researchers, and other practitioners in lower-income communities may be unable to maximize the value of thought partnering. It is important to ensure that the benefits of thought partners are not confined to an exclusive set of people.
Second, on the topic of anthromoprhization, we highlight an important distinction between human-centric and human-like thought partners 254. Our desiderata “I understand you” advocates for thought partners whose behavior we understand; while this could draw on how we understand other humans, however, we should be careful about interpreting such machine thought partners as we do humans. As Weizenbaum 6 illuminated with the ELIZA system, there are risks to developing computer systems that present themselves as human-like in ways that they are not: for example, by leading users to attribute undue intention to systems’ responses or (in the long run) leading society to devalue human intelligence 255. Human-like thought partners should maintain categorical delineation between humans and machines to prevent overreliance 245, 256 and promote human dignity without encroaching on any partner’s self-worth 257. The term used to refer to a thought partner can affect the assumptions made about their capabilities (e.g., teammate implies the machine and human are on equal footing) or can detract from a partner’s human-like nature (e.g., tool would be less anthropomorphic).
Lastly, we note that insufficiently accurate, robust, or cognitively-grounded models can yield misalignment with humans, leading intended AI thought partners to act towards the wrong goals 258, provide wrong or misleading information 259, or violate safety constraints 260. A Bayesian approach to thought partnership can address some of these issues, enabling uncertainty-aware decision-making that avoids overconfidence 223, 261, 262. Yet, while inferring human thoughts and behavior can be used to design better collaborators, models of humans are inherently dual-use and can also be used to mislead, surveil, or manipulate 263. It is crucial to consider whether thought partners are aligned with society at large, or merely superficially aligned with users while serving more powerful interests 264.
If we are to build helpful and reliable human-AI thought partnerships, we advocate for design that explicitly recognizes and engages with the richness and diversity of human thought in an often unpredictable world. We have argued, supported by several case studies, that those engineering thought partners and the infrastructure around their use can benefit from drawing on motifs from computational cognitive science and cognitive-AI. The future of collaborative cognition is bright, but not without risk; continual collaboration and knowledge sharing amongst behavioral scientists, AI practitioners, domain experts, and related disciplines is crucial as we strive to build machines that truly learn and think with people.
Glossary of main terms• Collaborative cognition: the process by which two or more agents work together in some aspect(s) of thinking (e.g., planning together, learning together, creating together).• Thought partner: another entity (human or AI) that works with an agent to push forward some aspect(s) of thinking.• Artificial Intelligence (AI): computational systems that are able to process inputs and engage in some aspect of learning, planning, reasoning, and/or decision-making. Used interchangeably with machines.• Large language model (LLM): a particular kind of AI system which learns a distribution over text, often trained on large amounts of web-scale text data. LLMs are a class of large-scale foundation models.• Agent: an entity that can process inputs, make decisions, and take actions in some environment.• Dyad: a system with two agents (e.g., human-human, human-AI, AI-AI).• Resource-rationality: the idea that human behavior and cognition can be viewed as rational under bounded constraints (e.g., under limited working memory).• Probabilistic generative model: a model of how the data one observes about the world is generated by some probabilistic process, from which one can sample new observations and make queries about existing observations.• Probabilistic programming language (PPL): a language for expressing probabilistic generative models as computer programs that interleave deterministic code (e.g. arithmetic, logic, or artificial neural networks) with random choices. PPLs allow users to specify probabilistic models and inference algorithms in a modular and compositional manner.• Bayesian inference: a method for updating one’s beliefs over various aspects of the world, grounded in probability theory; in Bayesian inference, an agent updates their beliefs by assigning higher credence to hypotheses that better explain the evidence, weighted against the backdrop of their prior beliefs.• Affordance: design features of a system that inform use.
We thank Richard Turner, Laura Schulz, Tyler Brooke-Wilson, Valerie Chen, Alena Rote, Lance Ying, Tony Chen, Matt Ashman, Mike Walmsley, Albert Jiang, Mateja Jamnik, Dj Dvijotham, Jonathan Ragan-Kelley, Will Crichton, Alex Lew, Tim O’Donnell, Joao Loula, Marty Tenenbaum, Mary McNaughton-Collins, and Jim Collins for valuable conversations that informed this work. KMC gratefully acknowledges support from the Marshall Commission and the Cambridge Trust. UB acknowledges support by ELSA (European Lighthouse on Secure and Safe AI) funded by the European Union under grant agreement No. 101070617; IS acknowledges funding from an NSERC fellowship (567554-2022); KC is supported by the Hertz Foundation, the Paul and Daisy Soros Fellowship, and an NSF Graduate Research Fellowship under grant #1745302.; ML acknowledges funding from MSR; TZX acknowledges support from the OpenPhilanthropy AI Fellowship. VM acknowledges a gift from the Siegel Family Foundation. AW acknowledges support from a Turing AI Fellowship under grant EP/V025279/1, The Alan Turing Institute, and the Leverhulme Trust via CFI. TLG acknowledges support from ONR grant N00014-22-1-2813. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the institutions listed above.
- 1GitHub Copilot · Your AI pair programmer.https://github.com/features/copilot.
- 2Copilot for Microsoft 365 – Microsoft Adoption.https://adoption.microsoft.com/en-us/copilot/.
- Fedorenko et al. 2024Evelina Fedorenko, Steven T Piantadosi, and Edward AF Gibson.Language is primarily a tool for communication rather than thought.Nature, 630(8017):575–586, 2024.
- Turing 1950Alan Turing.Computing machinery and intelligence.Mind, 59(236):433, 1950.
- Clynes and Kline 1960Manfred E Clynes and Nathan S Kline.Cyborgs and space.Astronautics, 14(9):26–27, 1960.
- Weizenbaum 1966Joseph Weizenbaum.Eliza—a computer program for the study of natural language communication between man and machine.Communications of the ACM, 9(1):36–45, 1966.
- Shneiderman 2022aBen Shneiderman..In Human-Centered AI. Oxford University Press, 01 2022a.ISBN 9780192845290.doi: 10.1093/oso/9780192845290.001.0001.
- Bundy 1983Alan Bundy.The computer modelling of mathematical reasoning.1983.
- Anderson et al. 1990John R Anderson, C Franklin Boyle, Albert T Corbett, and Matthew W Lewis.Cognitive modeling and intelligent tutoring.1990.
- Bommasani et al. 2021Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ B. Altman, and Simran Arora et al.On the opportunities and risks of foundation models.CoRR, abs/2108.07258, 2021.
- Ouyang et al. 2022Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, and Carroll et al Wainwright.Training language models to follow instructions with human feedback.Advances in Neural Information Processing Systems, 35:27730–27744, 2022.
- Christiano et al. 2017Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, and Shane et al Legg.Deep reinforcement learning from human preferences.Advances in neural information processing systems, 30, 2017.
- Lee et al. 2023aKimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, and Yuqing et al Du.Aligning text-to-image models using human feedback.arXiv preprint arXiv:2302.12192, 2023a.
- Ullman 2023Tomer Ullman.Large language models fail on trivial alterations to theory-of-mind tasks.arXiv preprint arXiv:2302.08399, 2023.
- Collins et al. 2024Katherine M Collins, Albert Q Jiang, Simon Frieder, Lionel Wong, and Miri et al Zilka.Evaluating language models for mathematics through interactions.Proceedings of the National Academy of Sciences, 121(24):e2318124121, 2024.
- Wong et al. 2023Lionel Wong, Gabriel Grand, Alexander K Lew, Noah D Goodman, and Vikash K et al Mansinghka.From word models to world models: Translating from natural language to the probabilistic language of thought.arXiv preprint arXiv:2306.12672, pages arXiv–2306, 2023.
- Zhang et al. 2023aCedegao E Zhang, Katherine M Collins, Adrian Weller, and Joshua B Tenenbaum.Ai for mathematics: A cognitive science perspective.arXiv preprint arXiv:2310.13021, 2023a.
- Gweon et al. 2023Hyowon Gweon, Judith Fan, and Been Kim.Socially intelligent machines that learn from humans and help humans learn.Philosophical Transactions of the Royal Society A, 381(2251):20220048, 2023.
- Mahowald et al. 2024Kyle Mahowald, Anna A Ivanova, Idan A Blank, Nancy Kanwisher, and Joshua B et al Tenenbaum.Dissociating language and thought in large language models.Trends in Cognitive Sciences, 2024.
- McCoy et al. 2023R Thomas McCoy, Shunyu Yao, Dan Friedman, Matthew Hardy, and Thomas L Griffiths.Embers of autoregression: Understanding large language models through the problem they are trained to solve.arXiv preprint arXiv:2309.13638, 2023.
- Tenenbaum et al. 2011Joshua B Tenenbaum, Charles Kemp, Thomas L Griffiths, and Noah D Goodman.How to grow a mind: Statistics, structure, and abstraction.science, 331(6022):1279–1285, 2011.
- Griffiths et al. 2023Thomas L. Griffiths, Jian-Qiao Zhu, Erin Grant, and R. Thomas McCoy.Bayes in the age of intelligent machines, 2023.
- Sumers et al. 2023aTheodore Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L Griffiths.Cognitive architectures for language agents.arXiv preprint arXiv:2309.02427, 2023a.
- Binz and Schulz 2023Marcel Binz and Eric Schulz.Turning large language models into cognitive models.arXiv preprint arXiv:2306.03917, 2023.
- Cusumano-Towner et al. 2019Marco F Cusumano-Towner, Feras A Saad, Alexander K Lew, and Vikash K Mansinghka.Gen: a general-purpose probabilistic programming system with programmable inference.In Proceedings of the 40th acm sigplan conference on programming language design and implementation, pages 221–236, 2019.
- Goodman et al. 2008aNoah D Goodman, Vikash K Mansinghka, Daniel Roy, Keith Bonawitz, and Joshua B Tenenbaum.Church: a language for generative models.In Proceedings of the Twenty-Fourth Conference on Uncertainty in Artificial Intelligence, pages 220–229, 2008a.
- Bingham et al. 2019Eli Bingham, Jonathan P Chen, Martin Jankowiak, Fritz Obermeyer, and Neeraj et al Pradhan.Pyro: Deep universal probabilistic programming.The Journal of Machine Learning Research, 20(1):973–978, 2019.
- Ge et al. 2018Hong Ge, Kai Xu, and Zoubin Ghahramani.Turing: a language for flexible probabilistic inference.In International conference on artificial intelligence and statistics, pages 1682–1690. PMLR, 2018.
- Goodman et al. 2014Noah D Goodman, Joshua B Tenenbaum, and Tobias Gerstenberg.Concepts in a probabilistic language of thought.Technical report, Center for Brains, Minds and Machines (CBMM), 2014.
- van Opheusden et al. 2023Bas van Opheusden, Ionatan Kuperwajs, Gianni Galbiati, Zahy Bnaya, and Yunqi et al Li.Expertise increases planning depth in human gameplay.Nature, 618(7967):1000–1005, 2023.
- Trinh et al. 2024Trieu H Trinh, Yuhuai Wu, Quoc V Le, He He, and Thang Luong.Solving olympiad geometry without human demonstrations.Nature, 625(7995):476–482, 2024.
- Yao et al. 2024Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, and Tom et al Griffiths.Tree of thoughts: Deliberate problem solving with large language models.Advances in Neural Information Processing Systems, 36, 2024.
- Baker et al. 2009Chris L Baker, Rebecca Saxe, and Joshua B Tenenbaum.Action understanding as inverse planning.Cognition, 113(3):329–349, 2009.
- Jara-Ettinger et al. 2020Julian Jara-Ettinger, Laura E Schulz, and Joshua B Tenenbaum.The naive utility calculus as a unified, quantitative framework for action understanding.Cognitive Psychology, 123:101334, 2020.
- Zhi-Xuan et al. 2024Tan Zhi-Xuan, Lance Ying, Vikash Mansinghka, and Joshua B Tenenbaum.Pragmatic instruction following and goal assistance via cooperative language-guided inverse planning.In Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems, pages 2094–2103, 2024.
- Lake et al. 2017Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, and Samuel J. Gershman.Building machines that learn and think like people.Behavioral and Brain Sciences, 40:e253, 2017.doi: 10.1017/S0140525X16001837.
- Chu and Schulz 2020Junyi Chu and Laura E Schulz.Play, curiosity, and cognition.Annual Review of Developmental Psychology, 2(1):317–343, 2020.
- Yanai and Lercher 2024Itai Yanai and Martin J Lercher.It takes two to think.Nature Biotechnology, pages 1–2, 2024.
- Holyoak and Morrison 2005Keith J Holyoak and Robert G Morrison.The Cambridge handbook of thinking and reasoning.Cambridge University Press, 2005.
- Holyoak and Morrison 2012Keith J Holyoak and Robert G Morrison.The Oxford handbook of thinking and reasoning.Oxford University Press, 2012.
- Ko and Myers 2004Amy J Ko and Brad A Myers.Designing the whyline: a debugging interface for asking questions about program behavior.In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 151–158, 2004.
- Ko et al. 2011Amy J Ko, Robin Abraham, Laura Beckwith, Alan Blackwell, and Margaret et al Burnett.The state of the art in end-user software engineering.ACM Computing Surveys (CSUR), 43(3):1–44, 2011.
- Muggleton and De Raedt 1994Stephen Muggleton and Luc De Raedt.Inductive logic programming: Theory and methods.The Journal of Logic Programming, 19:629–679, 1994.
- Anderson and Reiser 1985John R Anderson and Brian J Reiser.The lisp tutor.Byte, 10(4):159–175, 1985.
- Anderson et al. 1995John R Anderson, Albert T Corbett, Kenneth R Koedinger, and Ray Pelletier.Cognitive tutors: Lessons learned.The journal of the learning sciences, 4(2):167–207, 1995.
- Imai 2022Saki Imai.Is github copilot a substitute for human pair-programming? an empirical study.In Proceedings of the ACM/IEEE 44th International Conference on Software Engineering: Companion Proceedings, pages 319–321, 2022.
- Nguyen and Nadi 2022Nhan Nguyen and Sarah Nadi.An empirical evaluation of github copilot’s code suggestions.In Proceedings of the 19th International Conference on Mining Software Repositories, pages 1–5, 2022.
- Wermelinger 2023Michel Wermelinger.Using github copilot to solve simple programming problems.In Proceedings of the 54th ACM Technical Symposium on Computer Science Education V. 1, pages 172–178, 2023.
- Barke et al. 2023Shraddha Barke, Michael B James, and Nadia Polikarpova.Grounded copilot: How programmers interact with code-generating models.Proceedings of the ACM on Programming Languages, 7(OOPSLA1):85–111, 2023.
- Dakhel et al. 2023Arghavan Moradi Dakhel, Vahid Majdinasab, Amin Nikanjam, Foutse Khomh, and Michel C et al Desmarais.Github copilot ai pair programmer: Asset or liability?Journal of Systems and Software, 203:111734, 2023.
- Fisac et al. 2020aJaime F Fisac, Monica A Gates, Jessica B Hamrick, Chang Liu, and Dylan et al Hadfield-Menell.Pragmatic-pedagogic value alignment.In Robotics Research: The 18th International Symposium ISRR, pages 49–57. Springer, 2020a.
- Ranz et al. 2017Fabian Ranz, Vera Hummel, and Wilfried Sihn.Capability-based task allocation in human-robot collaboration.Procedia Manufacturing, 9:182–189, 2017.
- Casper and Murphy 2003Jennifer Casper and Robin R. Murphy.Human-robot interactions during the robot-assisted urban search and rescue response at the world trade center.IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 33(3):367–385, 2003.
- Shridhar et al. 2020Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, and Winson et al Han.Alfred: A benchmark for interpreting grounded instructions for everyday tasks.In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740–10749, 2020.
- Ahn et al. 2022Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, and Omar et al Cortes.Do as i can, not as i say: Grounding language in robotic affordances.arXiv preprint arXiv:2204.01691, 2022.Last Accessed: 2024–07-07.
- Raad et al. 2024Maria Abi Raad, Arun Ahuja, Catarina Barros, Frederic Besse, and Andrew et al Bolt.Scaling instructable agents across many simulated worlds.arXiv preprint arXiv:2404.10179, 2024.
- Valmeekam et al. 2022Karthik Valmeekam, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati.Large language models still can’t plan (a benchmark for llms on planning and reasoning about change).arXiv preprint arXiv:2206.10498, 2022.
- Momennejad et al. 2024Ida Momennejad, Hosein Hasanbeig, Felipe Vieira Frujeri, Hiteshi Sharma, and Nebojsa et al Jojic.Evaluating cognitive maps and planning in large language models with cogeval.Advances in Neural Information Processing Systems, 36, 2024.
- Goodman and Frank 2016Noah D Goodman and Michael C Frank.Pragmatic language interpretation as probabilistic inference.Trends in cognitive sciences, 20(11):818–829, 2016.
- Sumers et al. 2023bTheodore R Sumers, Mark K Ho, Thomas L Griffiths, and Robert D Hawkins.Reconciling truthfulness and relevance as epistemic and decision-theoretic utility.Psychological Review, 2023b.
- Jeon et al. 2020Hong Jun Jeon, Smitha Milli, and Anca Dragan.Reward-rational (implicit) choice: A unifying formalism for reward learning.Advances in Neural Information Processing Systems, 33:4415–4426, 2020.
- Kollar et al. 2013Thomas Kollar, Stefanie Tellex, Matthew R Walter, Albert Huang, and Abraham et al Bachrach.Generalized grounding graphs: A probabilistic framework for understanding grounded language.Journal of Artificial Intelligence Research, pages 1–35, 2013.
- Bratman 2013Michael E Bratman.Shared agency: A planning theory of acting together.Oxford University Press, 2013.
- Stacy et al. 2021Stephanie Stacy, Chenfei Li, Minglu Zhao, Yiling Yun, and Qingyi et al Zhao.Modeling communication to coordinate perspectives in cooperation.In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 43, 2021.
- Wu et al. 2021Sarah A Wu, Rose E Wang, James A Evans, Joshua B Tenenbaum, and David C et al Parkes.Too many cooks: Bayesian inference for coordinating multi-agent collaboration.Topics in Cognitive Science, 13(2):414–432, 2021.
- Reddy et al. 2018Sid Reddy, Anca Dragan, and Sergey Levine.Where do you think you’re going?: Inferring beliefs about dynamics from behavior.Advances in Neural Information Processing Systems, 31, 2018.
- Alanqary et al. 2021Arwa Alanqary, Gloria Z Lin, Joie Le, Tan Zhi-Xuan, and Vikash K et al Mansinghka.Modeling the mistakes of boundedly rational agents within a bayesian theory of mind.arXiv preprint arXiv:2106.13249, 2021.
- Dragan et al. 2013Anca D Dragan, Kenton CT Lee, and Siddhartha S Srinivasa.Legibility and predictability of robot motion.In 2013 8th ACM/IEEE International Conference on Human-Robot Interaction (HRI), pages 301–308. IEEE, 2013.
- Miura and Zilberstein 2021Shuwa Miura and Shlomo Zilberstein.A unifying framework for observer-aware planning and its complexity.In Uncertainty in Artificial Intelligence, pages 610–620. PMLR, 2021.
- Flower and Hayes 1981Linda Flower and John R. Hayes.A cognitive process theory of writing.College Composition and Communication, 32(4):365–387, 1981.ISSN 0010096X.
- Hayes 2012John R Hayes.Modeling and remodeling writing.Written communication, 29(3):369–388, 2012.
- Lee et al. 2022Mina Lee, Percy Liang, and Qian Yang.Coauthor: Designing a human-ai collaborative writing dataset for exploring language model capabilities.In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1–19, 2022.
- Lee et al. 2024Mina Lee, Katy Ilonka Gero, John Joon Young Chung, Simon Buckingham Shum, and Vipul et al Raheja.A design space for intelligent and interactive writing assistants.CHI, 2024.
- Ippolito et al. 2022Daphne Ippolito, Ann Yuan, Andy Coenen, and Sehmon Burnam.Creative writing with an ai-powered writing assistant: Perspectives from professional writers, 2022.
- Gero et al. 2022Katy Ilonka Gero, Vivian Liu, and Lydia Chilton.Sparks: Inspiration for science writing using language models.In Proceedings of the 2022 ACM Designing Interactive Systems Conference, pages 1002–1019, 2022.
- Gero et al. 2023Katy Ilonka Gero, Tao Long, and Lydia B Chilton.Social dynamics of ai support in creative writing.In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1–15, 2023.
- Dell’Acqua et al. 2023Fabrizio Dell’Acqua, Edward McFowland, Ethan R Mollick, Hila Lifshitz-Assaf, and Katherine et al Kellogg.Navigating the jagged technological frontier: Field experimental evidence of the effects of ai on knowledge worker productivity and quality.Harvard Business School Technology & Operations Mgt. Unit Working Paper, (24-013), 2023.
- Porter et al. 2023Justin Porter, Cynthia Boyd, M Reza Skandari, and Neda Laiteerapong.Revisiting the time needed to provide adult primary care.Journal of general internal medicine, 38(1):147–155, 2023.
- Dewa et al. 2017Carolyn S Dewa, Desmond Loong, Sarah Bonato, and Lucy Trojanowski.The relationship between physician burnout and quality of healthcare in terms of safety and acceptability: a systematic review.BMJ open, 7(6):e015141, 2017.
- Chowdhery et al. 2022Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, and Gaurav et al Mishra.Palm: Scaling language modeling with pathways.arXiv preprint arXiv:2204.02311, 2022.
- Singhal et al. 2023Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, and Jason et al Wei.Large language models encode clinical knowledge.Nature, 620(7972):172–180, 2023.
- Topol 2019Eric J Topol.High-performance medicine: the convergence of human and artificial intelligence.Nature medicine, 25(1):44–56, 2019.
- Tu et al. 2024Tao Tu, Anil Palepu, Mike Schaekermann, Khaled Saab, and Jan Freyberg et al.Towards conversational diagnostic ai, 2024.
- Ayers et al. 2023John W. Ayers, Adam Poliak, Mark Dredze, Eric C. Leas, and Zechariah et al Zhu.Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum.JAMA Internal Medicine, 04 2023.ISSN 2168-6106.doi: 10.1001/jamainternmed.2023.1838.
- Vellido 2020Alfredo Vellido.The importance of interpretability and visualization in machine learning for applications in medicine and health care.Neural computing and applications, 32(24):18069–18083, 2020.
- Rajpurkar et al. 2022Pranav Rajpurkar, Emma Chen, Oishi Banerjee, and Eric J Topol.Ai in health and medicine.Nature medicine, 28(1):31–38, 2022.
- Ghassemi et al. 2020Marzyeh Ghassemi, Tristan Naumann, Peter Schulam, Andrew L Beam, and Irene Y et al Chen.A review of challenges and opportunities in machine learning for health.AMIA Summits on Translational Science Proceedings, 2020:191, 2020.
- Daneshjou et al. 2021Roxana Daneshjou, Mary P Smith, Mary D Sun, Veronica Rotemberg, and James Zou.Lack of transparency and potential bias in artificial intelligence data sets and algorithms: a scoping review.JAMA dermatology, 157(11):1362–1369, 2021.
- Cabitza and Zeitoun 2019Federico Cabitza and Jean-David Zeitoun.The proof of the pudding: in praise of a culture of real-world validation for medical artificial intelligence.Annals of translational medicine, 7(8), 2019.
- Puig et al. 2020Xavier Puig, Tianmin Shu, Shuang Li, Zilin Wang, and Yuan-Hong et al Liao.Watch-and-help: A challenge for social perception and human-ai collaboration.arXiv preprint arXiv:2010.09890, 2020.
- Chandra et al. 2024aKartik Chandra, Tony Chen, Tzu-Mao Li, Jonathan Ragan-Kelley, and Josh Tenenbaum.Inferring the future by imagining the past.Advances in Neural Information Processing Systems, 36, 2024a.
- Fisac et al. 2020bJaime F Fisac, Chang Liu, Jessica B Hamrick, Shankar Sastry, and J Karl et al Hedrick.Generating plans that predict themselves.In Algorithmic Foundations of Robotics XII: Proceedings of the Twelfth Workshop on the Algorithmic Foundations of Robotics, pages 144–159. Springer, 2020b.
- Grice 1975Herbert P Grice.Logic and conversation.In Speech acts, pages 41–58. Brill, 1975.
- Doshi-Velez and Kim 2017Finale Doshi-Velez and Been Kim.Towards a rigorous science of interpretable machine learning.arXiv preprint arXiv:1702.08608, 2017.
- Miller 2019Tim Miller.Explanation in artificial intelligence: Insights from the social sciences.Artificial intelligence, 267:1–38, 2019.
- Smith 2019Brian Cantwell Smith.The promise of artificial intelligence: reckoning and judgment.Mit Press, 2019.
- Sucholutsky and Griffiths 2023Ilia Sucholutsky and Thomas L Griffiths.Alignment with human representations supports robust few-shot learning.NeurIPS, 2023.
- Sucholutsky et al. 2023Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, and Andreea Bobu et al.Getting aligned on representational alignment, 2023.
- Battaglia et al. 2013Peter W Battaglia, Jessica B Hamrick, and Joshua B Tenenbaum.Simulation as an engine of physical scene understanding.Proceedings of the National Academy of Sciences, 110(45):18327–18332, 2013.
- Roncone et al. 2017Alessandro Roncone, Olivier Mangin, and Brian Scassellati.Transparent role assignment and task allocation in human robot collaboration.In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 1014–1021. IEEE, 2017.
- Carroll et al. 2019Micah Carroll, Rohin Shah, Mark K Ho, Tom Griffiths, and Sanjit et al Seshia.On the utility of learning about humans for human-ai coordination.Advances in neural information processing systems, 32, 2019.
- Macindoe et al. 2012Owen Macindoe, Leslie Pack Kaelbling, and Tomás Lozano-Pérez.Pomcop: Belief space planning for sidekicks in cooperative games.In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, volume 8, pages 38–43, 2012.
- Lin et al. 2022Jessy Lin, Daniel Fried, Dan Klein, and Anca Dragan.Inferring rewards from language in context.arXiv preprint arXiv:2204.02515, 2022.
- Keuning et al. 2018Hieke Keuning, Johan Jeuring, and Bastiaan Heeren.A systematic literature review of automated feedback generation for programming exercises.ACM Transactions on Computing Education (TOCE), 19(1):1–43, 2018.
- Sarsa et al. 2022Sami Sarsa, Paul Denny, Arto Hellas, and Juho Leinonen.Automatic generation of programming exercises and code explanations using large language models.In Proceedings of the 2022 ACM Conference on International Computing Education Research-Volume 1, pages 27–43, 2022.
- Chandra et al. 2024bKartik Chandra, Tzu-Mao Li, Rachit Nigam, Joshua Tenenbaum, and Jonathan Ragan-Kelley.Watchat: Explaining perplexing programs by debugging mental models, 2024b.
- Head et al. 2015Andrew Head, Codanda Appachu, Marti A Hearst, and Björn Hartmann.Tutorons: Generating context-relevant, on-demand explanations and demonstrations of online code.In 2015 IEEE Symposium on Visual Languages and Human-Centric Computing (VL/HCC), pages 3–12. IEEE, 2015.
- Rafferty et al. 2020Anna N Rafferty, Rachel A Jansen, and Thomas L Griffiths.Assessing mathematics misunderstandings via bayesian inverse planning.Cognitive science, 44(10):e12900, 2020.
- Poesia and Goodman 2023Gabriel Poesia and Noah D Goodman.Peano: learning formal mathematical reasoning.Philosophical Transactions of the Royal Society A, 381(2251):20220044, 2023.
- Slonim et al. 2021Noam Slonim, Yonatan Bilu, Carlos Alzate, Roy Bar-Haim, and Ben et al Bogin.An autonomous debating system.Nature, 591(7850):379–384, 2021.
- Jarrett et al. 2023Daniel Jarrett, Miruna Pislar, Michiel A Bakker, Michael Henry Tessler, and Raphael et al Koster.Language agents as digital representatives in collective decision-making.In NeurIPS 2023 Foundation Models for Decision Making Workshop, 2023.
- Du et al. 2023Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch.Improving factuality and reasoning in language models through multiagent debate.arXiv preprint arXiv:2305.14325, 2023.
- Bakker et al. 2022Michiel Bakker, Martin Chadwick, Hannah Sheahan, Michael Tessler, and Lucy et al Campbell-Gillingham.Fine-tuning language models to find agreement among humans with diverse preferences.Advances in Neural Information Processing Systems, 35:38176–38189, 2022.
- Small et al. 2021Christopher Small, Michael Bjorkegren, Timo Erkkilä, Lynette Shaw, and Colin Megill.Polis: Scaling deliberation by mapping high dimensional opinion spaces.Recerca: revista de pensament i anàlisi, 26(2), 2021.
- Lew et al. 2021Alexander Lew, Monica Agrawal, David Sontag, and Vikash Mansinghka.Pclean: Bayesian data cleaning at scale with domain-specific probabilistic programming.In International Conference on Artificial Intelligence and Statistics, pages 1927–1935. PMLR, 2021.
- Saad et al. 2019Feras A Saad, Marco F Cusumano-Towner, Ulrich Schaechtle, Martin C Rinard, and Vikash K Mansinghka.Bayesian synthesis of probabilistic programs for automatic data modeling.Proceedings of the ACM on Programming Languages, 3(POPL):1–32, 2019.
- Huot et al. 2024Mathieu Huot, Matin Ghavami, Alexander K Lew, Ulrich Schaechtle, and Cameron E et al Freer.Gensql: A probabilistic programming system for querying generative models of database tables.Proceedings of the ACM on Programming Languages, 8(PLDI):790–815, 2024.
- Li et al. 2024Michael Y Li, Emily B Fox, and Noah D Goodman.Automated statistical model discovery with language models.arXiv preprint arXiv:2402.17879, 2024.
- Steinruecken et al. 2019Christian Steinruecken, Emma Smith, David Janz, James Lloyd, and Zoubin Ghahramani.The automatic statistician.Automated machine learning: Methods, systems, challenges, pages 161–173, 2019.
- Davies et al. 2021Alex Davies, Petar Veličković, Lars Buesing, Sam Blackwell, and Daniel et al Zheng.Advancing mathematics by guiding human intuition with ai.Nature, 600(7887):70–74, 2021.
- Cranmer et al. 2020Miles Cranmer, Alvaro Sanchez Gonzalez, Peter Battaglia, Rui Xu, and Kyle et al Cranmer.Discovering symbolic models from deep learning with inductive biases.Advances in neural information processing systems, 33:17429–17442, 2020.
- Romera-Paredes et al. 2024Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, and M Pawan et al Kumar.Mathematical discoveries from program search with large language models.Nature, 625(7995):468–475, 2024.
- Ashkinaze et al. 2024Joshua Ashkinaze, Julia Mendelsohn, Li Qiwei, Ceren Budak, and Eric Gilbert.How ai ideas affect the creativity, diversity, and evolution of human ideas: Evidence from a large, dynamic experiment, 2024.
- Suri et al. 2024Siddharth Suri, Scott Counts, Leijie Wang, Chacha Chen, and Mengting et al Wan.The use of generative search engines for knowledge work and complex tasks.Technical Report MSR-TR-2024-9, Microsoft, March 2024.
- Vartiainen and Tedre 2023Henriikka Vartiainen and Matti Tedre.Using artificial intelligence in craft education: crafting with text-to-image generative models.Digital Creativity, 34(1):1–21, 2023.
- Gafni et al. 2022Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, and Devi et al Parikh.Make-a-scene: Scene-based text-to-image generation with human priors.In European Conference on Computer Vision, pages 89–106. Springer, 2022.
- Fan et al. 2019Judith E Fan, Monica Dinculescu, and David Ha.Collabdraw: an environment for collaborative sketching with an artificial agent.In Proceedings of the 2019 Conference on Creativity and Cognition, pages 556–561, 2019.
- Ge et al. 2020Songwei Ge, Vedanuj Goswami, C Lawrence Zitnick, and Devi Parikh.Creative sketch generation.arXiv preprint arXiv:2011.10039, 2020.
- Dvorožňák et al. 2020Marek Dvorožňák, Daniel Sỳkora, Cassidy Curtis, Brian Curless, and Olga et al Sorkine-Hornung.Monster mash: a single-view approach to casual 3d modeling and animation.ACM Transactions on Graphics (ToG), 39(6):1–12, 2020.
- Chater and Oaksford 2008Nick Chater and Mike Oaksford, editors.The Probabilistic Mind: Prospects for Bayesian Cognitive Science.Oxford University Press, Oxford, UK, 2008.
- Griffiths et al. 2008Thomas L. Griffiths, Charles Kemp, and Joshua B. Tenenbaum.Bayesian models of cognition.In Ron Sun, editor, The Cambridge handbook of computational psychology, chapter 3, pages 59–100. Cambridge University Press, 2008.
- Spelke 2000Elizabeth S Spelke.Core knowledge.American psychologist, 55(11):1233, 2000.
- Piantadosi 2021Steven T Piantadosi.The computational origin of representation.Minds and machines, 31(1):1–58, 2021.
- Quilty-Dunn et al. 2023Jake Quilty-Dunn, Nicolas Porot, and Eric Mandelbaum.The best game in town: The reemergence of the language-of-thought hypothesis across the cognitive sciences.Behavioral and Brain Sciences, 46:e261, 2023.
- Griffiths et al. 2003Thomas Griffiths, Michael Jordan, Joshua Tenenbaum, and David Blei.Hierarchical topic models and the nested chinese restaurant process.Advances in neural information processing systems, 16, 2003.
- Kemp and Tenenbaum 2008Charles Kemp and Joshua B Tenenbaum.The discovery of structural form.Proceedings of the National Academy of Sciences, 105(31):10687–10692, 2008.
- Lake et al. 2015Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum.Human-level concept learning through probabilistic program induction.Science, 350(6266):1332–1338, 2015.
- Rule et al. 2020Joshua S Rule, Joshua B Tenenbaum, and Steven T Piantadosi.The child as hacker.Trends in cognitive sciences, 24(11):900–915, 2020.
- Ellis et al. 2021Kevin Ellis, Catherine Wong, Maxwell Nye, Mathias Sablé-Meyer, and Lucas et al Morales.Dreamcoder: Bootstrapping inductive program synthesis with wake-sleep library learning.In Proceedings of the 42nd acm sigplan international conference on programming language design and implementation, pages 835–850, 2021.
- Ullman and Tenenbaum 2020aTomer D Ullman and Joshua B Tenenbaum.Bayesian models of conceptual development: Learning as building models of the world.Annual Review of Developmental Psychology, 2(1):533–558, 2020a.
- Lieder et al. 2019Falk Lieder, Owen X Chen, Paul M Krueger, and Thomas L Griffiths.Cognitive prostheses for goal achievement.Nature human behaviour, 3(10):1096–1106, 2019.
- Lieder and Griffiths 2020Falk Lieder and Thomas L Griffiths.Resource-rational analysis: Understanding human cognition as the optimal use of limited computational resources.Behavioral and brain sciences, 43:e1, 2020.
- Icard 2023Thomas Icard.Resource rationality.Book manuscript, 2023.
- Newell and Simon 1972Allen Newell and Herbert A. Simon.Human problem solving.Prentice-Hall, 1972.
- Mattar and Lengyel 2022Marcelo G Mattar and Máté Lengyel.Planning in the brain.Neuron, 110(6):914–934, 2022.
- Ho et al. 2022aMark K. Ho, David Abel, Carlos G. Correa, Michael L. Littman, and Jonathan D. et al Cohen.People construct simplified mental representations to plan.Nature, 606(7912):129–136, 2022a.
- Jara-Ettinger et al. 2016Julian Jara-Ettinger, Hyowon Gweon, Laura E Schulz, and Joshua B Tenenbaum.The naïve utility calculus: Computational principles underlying commonsense psychology.Trends in cognitive sciences, 20(8):589–604, 2016.
- Baker et al. 2011Chris Baker, Rebecca Saxe, and Joshua Tenenbaum.Bayesian theory of mind: Modeling joint belief-desire attribution.In Proceedings of the annual meeting of the cognitive science society, volume 33, 2011.
- Ho et al. 2022bMark K Ho, Rebecca Saxe, and Fiery Cushman.Planning with theory of mind.Trends in Cognitive Sciences, 26(11):959–971, 2022b.
- Frank and Goodman 2012Michael C Frank and Noah D Goodman.Predicting pragmatic reasoning in language games.Science, 336(6084):998–998, 2012.
- Degen 2023Judith Degen.The rational speech act framework.Annual Review of Linguistics, 9:519–540, 2023.
- Binz et al. 2023Marcel Binz, Ishita Dasgupta, Akshay K Jagadish, Matthew Botvinick, and Jane X et al Wang.Meta-learned models of cognition.Behavioral and Brain Sciences, pages 1–38, 2023.
- Grant et al. 2018Erin Grant, Chelsea Finn, Sergey Levine, Trevor Darrell, and Thomas Griffiths.Recasting gradient-based meta-learning as hierarchical bayes.In 6th International Conference on Learning Representations, ICLR 2018, 2018.
- Lake and Baroni 2023Brenden M Lake and Marco Baroni.Human-like systematic generalization through a meta-learning neural network.Nature, 623(7985):115–121, 2023.
- Ho and Griffiths 2022Mark K Ho and Thomas L Griffiths.Cognitive science as a source of forward and inverse models of human decisions for robotics and control.Annual Review of Control, Robotics, and Autonomous Systems, 5:33–53, 2022.
- Yang et al. 2023Scott Cheng-Hsin Yang, Tomas Folke, and Patrick Shafto.The inner loop of collective human–machine intelligence.Topics in cognitive science, 2023.
- Steyvers and Kumar 2023Mark Steyvers and Aakriti Kumar.Three challenges for ai-assisted decision-making.Perspectives on Psychological Science, page 17456916231181102, 2023.
- Chater and Manning 2006Nick Chater and Christopher D Manning.Probabilistic models of language processing and acquisition.Trends in cognitive sciences, 10(7):335–344, 2006.
- Oaksford and Chater 2007Mike Oaksford and Nick Chater.Bayesian rationality: The probabilistic approach to human reasoning.Oxford University Press, 2007.
- Xu and Tenenbaum 2007Fei Xu and Joshua B Tenenbaum.Word learning as bayesian inference.Psychological review, 114(2):245, 2007.
- Kersten et al. 2004Daniel Kersten, Pascal Mamassian, and Alan Yuille.Object perception as bayesian inference.Annu. Rev. Psychol., 55:271–304, 2004.
- Yildirim et al. 2020Ilker Yildirim, Mario Belledonne, Winrich Freiwald, and Josh Tenenbaum.Efficient inverse graphics in biological face processing.Science advances, 6(10):eaax5979, 2020.
- Allen et al. 2020Kelsey R. Allen, Kevin A. Smith, and Joshua B. Tenenbaum.Rapid trial-and-error learning with simulation supports flexible tool use and physical reasoning.Proceedings of the National Academy of Sciences, 117(47):29302–29310, 2020.
- Zhang et al. 2023bCedegao E Zhang, Lionel Wong, Gabriel Grand, and Joshua B Tenenbaum.Grounded physical language understanding with probabilistic programs and simulated worlds.In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 45, 2023b.
- Tenenbaum 1998Joshua Tenenbaum.Bayesian modeling of human concept learning.Advances in neural information processing systems, 11, 1998.
- Goodman et al. 2008bNoah D Goodman, Joshua B Tenenbaum, Jacob Feldman, and Thomas L Griffiths.A rational analysis of rule-based concept learning.Cognitive science, 32(1):108–154, 2008b.
- Piantadosi et al. 2016Steven T Piantadosi, Joshua B Tenenbaum, and Noah D Goodman.The logical primitives of thought: Empirical foundations for compositional cognitive models.Psychological review, 123(4):392, 2016.
- Griffiths et al. 2004Thomas Griffiths, Mark Steyvers, David Blei, and Joshua Tenenbaum.Integrating topics and syntax.Advances in neural information processing systems, 17, 2004.
- Goodman and Lassiter 2015Noah D Goodman and Daniel Lassiter.Probabilistic semantics and pragmatics uncertainty in language and thought.The handbook of contemporary semantic theory, pages 655–686, 2015.
- Yang and Piantadosi 2022Yuan Yang and Steven T Piantadosi.One model for the learning of language.Proceedings of the National Academy of Sciences, 119(5):e2021865119, 2022.
- Schulz et al. 2007Laura E Schulz, Elizabeth Baraff Bonawitz, and Thomas L Griffiths.Can being scared cause tummy aches? naive theories, ambiguous evidence, and preschoolers’ causal inferences.Developmental psychology, 43(5):1124, 2007.
- Gopnik et al. 2004Alison Gopnik, Clark Glymour, David M Sobel, Laura E Schulz, and Tamar et al Kushnir.A theory of causal learning in children: causal maps and bayes nets.Psychological review, 111(1):3, 2004.
- Kirfel et al. 2022Lara Kirfel, Thomas Icard, and Tobias Gerstenberg.Inference from explanation.Journal of Experimental Psychology: General, 151(7):1481, 2022.
- Lagnado et al. 2013David A Lagnado, Tobias Gerstenberg, and Ro’i Zultan.Causal responsibility and counterfactuals.Cognitive science, 37(6):1036–1073, 2013.
- Hemmer and Steyvers 2009Pernille Hemmer and Mark Steyvers.A bayesian account of reconstructive memory.Topics in Cognitive Science, 1(1):189–202, 2009.
- Ullman and Tenenbaum 2020bTomer D. Ullman and Joshua B. Tenenbaum.Bayesian models of conceptual development: Learning as building models of the world.Annual Review of Developmental Psychology, 2(1):533–558, 2020b.doi: 10.1146/annurev-devpsych-121318-084833.
- Griffiths and Tenenbaum 2009Thomas L Griffiths and Joshua B Tenenbaum.Theory-based causal induction.Psychological review, 116(4):661, 2009.
- Vul et al. 2014Edward Vul, Noah Goodman, Thomas L Griffiths, and Joshua B Tenenbaum.One and done? optimal decisions from very few samples.Cognitive science, 38(4):599–637, 2014.
- Hay et al. 2014Nicholas Hay, Stuart Russell, David Tolpin, and Solomon Eyal Shimony.Selecting computations: Theory and applications.arXiv preprint arXiv:1408.2048, 2014.
- Tomov et al. 2020Momchil S. Tomov, Samyukta Yagati, Agni Kumar, Wanqian Yang, and Samuel J. Gershman.Discovery of hierarchical representations for efficient planning.PLOS Computational Biology, 16(4):1–42, 04 2020.doi: 10.1371/journal.pcbi.1007594.
- Baker and Tenenbaum 2014Chris L Baker and Joshua B Tenenbaum.Modeling human plan recognition using bayesian theory of mind.Plan, activity, and intent recognition: Theory and practice, 7:177–204, 2014.
- Callaway et al. 2022Frederick Callaway, Bas van Opheusden, Sayan Gul, Priyam Das, and Paul M et al Krueger.Rational use of cognitive resources in human planning.Nature Human Behaviour, 6(8):1112–1125, 2022.
- Baker et al. 2017Chris L Baker, Julian Jara-Ettinger, Rebecca Saxe, and Joshua B Tenenbaum.Rational quantitative attribution of beliefs, desires and percepts in human mentalizing.Nature Human Behaviour, 1(4):0064, 2017.
- Zhi-Xuan et al. 2020Tan Zhi-Xuan, Jordyn Mann, Tom Silver, Josh Tenenbaum, and Vikash Mansinghka.Online bayesian goal inference for boundedly rational planning agents.Advances in neural information processing systems, 33:19238–19250, 2020.
- Ying et al. 2023Lance Ying, Katherine M Collins, Megan Wei, Cedegao E Zhang, and Tan et al Zhi-Xuan.The neuro-symbolic inverse planning engine (nipe): Modeling probabilistic social inferences from linguistic inputs.arXiv preprint arXiv:2306.14325, 2023.
- Johnson-Laird 1983Philip Nicholas Johnson-Laird.Mental models: Towards a cognitive science of language, inference, and consciousness.Number 6. Harvard University Press, 1983.
- Byrne 2002Ruth MJ Byrne.Mental models and counterfactual thoughts about what might have been.Trends in cognitive sciences, 6(10):426–431, 2002.
- Shafto et al. 2014Patrick Shafto, Noah D Goodman, and Thomas L Griffiths.A rational account of pedagogical reasoning: Teaching by, and learning from, examples.Cognitive psychology, 71:55–89, 2014.
- Sumers et al. 2021Theodore R Sumers, Mark K Ho, Robert D Hawkins, Karthik Narasimhan, and Thomas L Griffiths.Learning rewards from linguistic feedback.In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 6002–6010, 2021.
- Liquin et al. 2023Emily G Liquin, Nicole Luzuriaga, and Todd M Gureckis.Teaching and learning through pedagogical environment design.In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 45, 2023.
- Kumar et al. 2023Aakriti Kumar, Padhraic Smyth, and Mark Steyvers.Differentiating mental models of self et al: A hierarchical framework for knowledge assessment.Psychological Review, 2023.
- Hawkins et al. 2023aRobert D Hawkins, Michael Franke, Michael C Frank, Adele E Goldberg, and Kenny et al Smith.From partners to populations: A hierarchical bayesian account of coordination and convention.Psychological Review, 130(4):977, 2023a.
- Hawkins et al. 2023bRobert D Hawkins, Andrew M Berdahl, Alex ‘Sandy’ Pentland, Joshua B Tenenbaum, and Noah D et al Goodman.Flexible social inference facilitates targeted social learning when rewards are not observable.Nature Human Behaviour, pages 1–10, 2023b.
- Goodman and Stuhlmüller 2013Noah D Goodman and Andreas Stuhlmüller.Knowledge and implicature: Modeling language understanding as social cognition.Topics in cognitive science, 5(1):173–184, 2013.
- Ho et al. 2021Mark K Ho, Fiery Cushman, Michael L Littman, and Joseph L Austerweil.Communication in action: Planning and interpreting communicative demonstrations.Journal of Experimental Psychology: General, 150(11):2246, 2021.
- Griffiths 2020Thomas L Griffiths.Understanding human intelligence through human limitations.Trends in Cognitive Sciences, 24(11):873–883, 2020.
- Tversky and Kahneman 1973Amos Tversky and Daniel Kahneman.Availability: A heuristic for judging frequency and probability.Cognitive psychology, 5(2):207–232, 1973.
- Tversky and Kahneman 1974Amos Tversky and Daniel Kahneman.Judgment under uncertainty: Heuristics and biases: Biases in judgments reveal some heuristics of thinking under uncertainty.science, 185(4157):1124–1131, 1974.
- Zhu et al. 2023Jian-Qiao Zhu, Joakim Sundh, Jake Spicer, Nick Chater, and Adam N Sanborn.The autocorrelated bayesian sampler: A rational process for probability judgments, estimates, confidence intervals, choices, confidence judgments, and response times.Psychological Review, 2023.
- Van Rooij 2008Iris Van Rooij.The tractable cognition thesis.Cognitive science, 32(6):939–984, 2008.
- Icard and Goodman 2015Thomas Icard and Noah D Goodman.A resource-rational approach to the causal frame problem.In CogSci, 2015.
- Anderson 1990John R Anderson.The adaptive character of thought.Psychology Press, 1990.
- Cheyette et al. 2023Samuel J Cheyette, Frederick Callaway, Neil R Bramley, Jonathan D Nelson, and Josh Tenenbaum.People seek easily interpretable information.In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 45, 2023.
- Saad 2022Feras Ahmad Khaled Saad.Scalable Structure Learning, Inference, and Analysis with Probabilistic Programs.PhD thesis, Massachusetts Institute of Technology, 2022.
- Lew et al. 2020Alexander K Lew, Michael Henry Tessler, Vikash K Mansinghka, and Joshua B Tenenbaum.Leveraging unstructured statistical knowledge in a probabilistic language of thought.In Proceedings of the annual conference of the cognitive science society, 2020.
- Gothoskar et al. 2023Nishad Gothoskar, Matin Ghavami, Eric Li, Aidan Curtis, and Michael et al Noseworthy.Bayes3d: fast learning and inference in structured generative models of 3d objects and scenes.arXiv preprint arXiv:2312.08715, 2023.
- Mansinghka et al. 2018Vikash K Mansinghka, Ulrich Schaechtle, Shivam Handa, Alexey Radul, Yutian Chen, and Martin Rinard.Probabilistic programming with programmable inference.In Proceedings of the 39th ACM SIGPLAN Conference on Programming Language Design and Implementation, pages 603–616, 2018.
- Lew et al. 2023aAlexander K. Lew, Mathieu Huot, Sam Staton, and Vikash K. Mansinghka.Adev: Sound automatic differentiation of expected values of probabilistic programs.Proc. ACM Program. Lang., 7(POPL), jan 2023a.doi: 10.1145/3571198.URL https://doi.org/10.1145/3571198.
- Becker et al. 2024McCoy R Becker, Alexander K Lew, Xiaoyan Wang, Matin Ghavami, Mathieu Huot, Martin C Rinard, and Vikash K Mansinghka.Probabilistic programming with programmable variational inference.Proceedings of the ACM on Programming Languages, 8(PLDI):2123–2147, 2024.
- Saad et al. 2021Feras A Saad, Martin C Rinard, and Vikash K Mansinghka.Sppl: probabilistic programming with fast exact symbolic inference.In Proceedings of the 42nd acm sigplan international conference on programming language design and implementation, pages 804–819, 2021.
- Lew et al. 2023bAlexander K Lew, Matin Ghavamizadeh, Martin C Rinard, and Vikash K Mansinghka.Probabilistic programming with stochastic probabilities.Proceedings of the ACM on Programming Languages, 7(PLDI):1708–1732, 2023b.
- Guggenberger et al. 2020Tobias Moritz Guggenberger, Frederik Möller, Tim Haarhaus, Inan Gür, and Boris Otto.Ecosystem types in information systems.In Twenty-Eighth European Conference on Information Systems (ECIS2020), 2020.
- Goodman and Flaxman 2017Bryce Goodman and Seth Flaxman.European union regulations on algorithmic decision-making and a “right to explanation”.AI magazine, 38(3):50–57, 2017.
- Wachter and Mittelstadt 2019Sandra Wachter and Brent Mittelstadt.A right to reasonable inferences: re-thinking data protection law in the age of big data and ai.Colum. Bus. L. Rev., page 494, 2019.
- Fui-Hoon Nah et al. 2023Fiona Fui-Hoon Nah, Ruilin Zheng, Jingyuan Cai, Keng Siau, and Langtao Chen.Generative ai and chatgpt: Applications, challenges, and ai-human collaboration, 2023.
- Norman 1988Don Norman.Design Of Everyday Things.New York: Basic Books. Olins, W.(2005). A Marca. Lisboa: Verbo. Packard, V …, 1988.
- Chemero 2018Anthony Chemero.An outline of a theory of affordances.In How Shall Affordances Be Refined?, pages 181–195. Routledge, 2018.
- Zerilli et al. 2022aJohn Zerilli, Umang Bhatt, and Adrian Weller.How transparency modulates trust in artificial intelligence.Patterns, page 100455, 2022a.
- Messeri and Crockett 2024Lisa Messeri and MJ Crockett.Artificial intelligence and illusions of understanding in scientific research.Nature, 627(8002):49–58, 2024.
- Tejeda et al. 2022Heliodoro Tejeda, Aakriti Kumar, Padhraic Smyth, and Mark Steyvers.Ai-assisted decision-making: A cognitive modeling approach to infer latent reliance strategies.Computational Brain & Behavior, 5(4):1–18, 2022.
- Steyvers et al. 2022Mark Steyvers, Heliodoro Tejeda, Gavin Kerrigan, and Padhraic Smyth.Bayesian modeling of human–ai complementarity.Proceedings of the National Academy of Sciences, 119(11):e2111547119, 2022.
- Chandra et al. 2024cKartik Chandra, Tony Chen, Tzu-Mao Li, Jonathan Ragan-Kelley, and Josh Tenenbaum.Cooperative explanation as rational communication.In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 46, 2024c.
- Hadfield-Menell et al. 2016Dylan Hadfield-Menell, Stuart J Russell, Pieter Abbeel, and Anca Dragan.Cooperative inverse reinforcement learning.Advances in neural information processing systems, 29, 2016.
- Chandra et al. 2023Kartik Chandra, Tzu-Mao Li, Joshua Tenenbaum, and Jonathan Ragan-Kelley.Acting as inverse inverse planning.In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–12, 2023.
- Chen et al. 2024Tony Chen, Sean Dae Houlihan, Kartik Chandra, Josh Tenenbaum, and Rebecca Saxe.Intervening on emotions by planning over a theory of mind.In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 46, 2024.
- Chandra et al. 2024dKartik Chandra, Tzu-Mao Li, Joshua B Tenenbaum, and Jonathan Ragan-Kelley.Storytelling as inverse inverse planning.Topics in Cognitive Science, 16(1):54–70, 2024d.
- Loula et al. 2024João Loula, Katherine M Collins, Ulrich Schaechtle, Joshua B Tenenbaum, and Adrian et al Weller.Learning generative population models from multiple clinical datasets via probabilistic programming.In ICML 2024 Workshop on Efficient and Accessible Foundation Models for Biological Discovery, 2024.
- Cabitza et al. 2017Federico Cabitza, Raffaele Rasoini, and Gian Franco Gensini.Unintended consequences of machine learning in medicine.Jama, 318(6):517–518, 2017.
- Mozannar and Sontag 2020Hussein Mozannar and David Sontag.Consistent estimators for learning to defer to an expert.In International Conference on Machine Learning, pages 7076–7087. PMLR, 2020.
- Dvijotham et al. 2023Krishnamurthy Dvijotham, Jim Winkens, Melih Barsbey, Sumedh Ghaisas, and Robert et al Stanforth.Enhancing the reliability and accuracy of ai-enabled diagnosis via complementarity-driven deferral to clinicians.Nature Medicine, pages 1–7, 2023.
- Tsvetkova et al. 2024Milena Tsvetkova, Taha Yasseri, Niccolo Pescetelli, and Tobias Werner.Human-machine social systems.arXiv preprint arXiv:2402.14410, 2024.
- Schneiders et al. 2022Eike Schneiders, EunJeong Cheon, Jesper Kjeldskov, Matthias Rehm, and Mikael B Skov.Non-dyadic interaction: A literature review of 15 years of human-robot interaction conference publications.ACM Transactions on Human-Robot Interaction (THRI), 11(2):1–32, 2022.
- Hornecker et al. 2022Eva Hornecker, Antonia Krummheuer, Andreas Bischof, and Matthias Rehm.Beyond dyadic hri: Building robots for society.interactions, 29(3):48–53, 2022.
- Yadav and Mehta 2024Aakash Yadav and Ranjana Mehta.Beyond dyadic interactions: Assessing trust networks in multi-human-robot teams.In Companion of the 2024 ACM/IEEE International Conference on Human-Robot Interaction, pages 1153–1157, 2024.
- Sucholutsky et al. 2024Ilia Sucholutsky, Katherine M Collins, Maya Malaviya, Nori Jacoby, and Weiyang et al Liu.Representational alignment supports effective machine teaching.arXiv Preprint arXiv:2406.04302, 2024.
- Li et al. 2023Ling Li, Xiaojian Li, Bo Ouyang, Hangjie Mo, and Hongliang et al Ren.Three-dimensional collision avoidance method for robot-assisted minimally invasive surgery.Cyborg and Bionic Systems, 4:0042, 2023.
- Boyce et al. 2024Veronica Boyce, Robert D Hawkins, Noah D Goodman, and Michael C Frank.Interaction structure constrains the emergence of conventions in group communication.Proceedings of the National Academy of Sciences, 121(28), 2024.
- Trouille et al. 2019Laura Trouille, Chris J Lintott, and Lucy F Fortson.Citizen science frontiers: Efficiency, engagement, and serendipitous discovery with human–machine systems.Proceedings of the National Academy of Sciences, 116(6):1902–1909, 2019.
- Hornbæk and Oulasvirta 2017Kasper Hornbæk and Antti Oulasvirta.What is interaction?In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems, pages 5040–5052, 2017.
- Lee et al. 2023bMina Lee, Megha Srivastava, Amelia Hardy, John Thickstun, and Esin et al Durmus.Evaluating human-language model interaction.Transactions on Machine Learning Research, 2023b.
- Allen et al. 2024Kelsey Allen, Franziska Brändle, Matthew Botvinick, Judith E Fan, and Samuel J et al Gershman.Using games to understand the mind.Nature Human Behaviour, pages 1–9, 2024.
- Park et al. 2023Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, and Percy Liang et al.Generative agents: Interactive simulacra of human behavior, 2023.
- Brown and Sandholm 2019Noam Brown and Tuomas Sandholm.Superhuman ai for multiplayer poker.Science, 365(6456):885–890, 2019.
- Bakhtin et al. 2022Anton Bakhtin, Noam Brown, Emily Dinan, Gabriele Farina, and Colin et al Flaherty.Human-level play in the game of diplomacy by combining language models with strategic reasoning.Science, 378(6624):1067–1074, 2022.
- Logg et al. 2019Jennifer M Logg, Julia A Minson, and Don A Moore.Algorithm appreciation: People prefer algorithmic to human judgment.Organizational Behavior and Human Decision Processes, 151:90–103, 2019.
- Green and Chen 2019Ben Green and Yiling Chen.The principles and limits of algorithm-in-the-loop decision making.Proceedings of the ACM on Human-Computer Interaction, 3(CSCW):1–24, 2019.
- Inuwa-Dutse et al. 2023Isa Inuwa-Dutse, Alice Toniolo, Adrian Weller, and Umang Bhatt.Algorithmic loafing and mitigation strategies in human-ai teams.Computers in Human Behavior: Artificial Humans, 1(2):100024, 2023.
- Hofman et al. 2023Jake M Hofman, Daniel G Goldstein, and David M Rothschild.Steroids, sneakers, coach: The spectrum of human-ai relationships.Available at SSRN 4578180, 2023.
- Buschek et al. 2021Daniel Buschek, Martin Zürn, and Malin Eiband.The impact of multiple parallel phrase suggestions on email input and composition behaviour of native and non-native english writers.In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, CHI ’21, New York, NY, USA, 2021. Association for Computing Machinery.ISBN 9781450380966.doi: 10.1145/3411764.3445372.
- Buçinca et al. 2021Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z Gajos.To trust or to think: cognitive forcing functions can reduce overreliance on ai in ai-assisted decision-making.Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1):1–21, 2021.
- Dietvorst et al. 2015Berkeley J Dietvorst, Joseph P Simmons, and Cade Massey.Algorithm aversion: people erroneously avoid algorithms after seeing them err.Journal of Experimental Psychology: General, 144(1):114, 2015.
- Dietvorst et al. 2018Berkeley J Dietvorst, Joseph P Simmons, and Cade Massey.Overcoming algorithm aversion: People will use imperfect algorithms if they can (even slightly) modify them.Management science, 64(3):1155–1170, 2018.
- Zerilli et al. 2022bJohn Zerilli, Umang Bhatt, and Adrian Weller.Transparency Modulates Trust in Artificial Intelligence.Patterns, 2022b.
- Mumford 1936Lewis Mumford.Technics and civilization.1936.
- Weizenbaum 1976Joseph Weizenbaum.Computer power and human reason: From judgment to calculation.1976.
- Weidinger et al. 2022Laura Weidinger, Jonathan Uesato, Maribeth Rauh, Conor Griffin, and Po-Sen et al Huang.Taxonomy of risks posed by language models.In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 214–229, 2022.
- Shneiderman 2022bBen Shneiderman.Human-centered AI.Oxford University Press, 2022b.
- Zhuang and Hadfield-Menell 2020Simon Zhuang and Dylan Hadfield-Menell.Consequences of misaligned ai.Advances in Neural Information Processing Systems, 33:15763–15773, 2020.
- Kalai and Vempala 2023Adam Tauman Kalai and Santosh S Vempala.Calibrated language models must hallucinate.arXiv preprint arXiv:2311.14648, 2023.
- Amodei et al. 2016Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, and John et al Schulman.Concrete problems in ai safety.arXiv preprint arXiv:1606.06565, 2016.
- Russell 2019Stuart Russell.Human compatible: AI and the problem of control.Viking, 2019.
- Russell 2021Stuart Russell.Artificial intelligence and the problem of control.Perspectives on Digital Humanism, pages 19–24, 2021.
- Carroll et al. 2023Micah Carroll, Alan Chan, Henry Ashton, and David Krueger.Characterizing manipulation from ai systems.In Proceedings of the 3rd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization, pages 1–13, 2023.
- Lazar and Nelson 2023Seth Lazar and Alondra Nelson.Ai safety on whose terms?Science, 381(6654):138–138, 2023.
General Document Comments 0
on the uploaded document.
0 archived comments