Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Building Machines that Learn and Think with People

Katherine M. Collins, Ilia Sucholutsky, Umang Bhatt, Kartik Chandra, Lionel Wong, Mina Lee, Cedegao E. Zhang, Tan Zhi-Xuan, Mark Ho, Vikash Mansinghka, Adrian Weller, Joshua B. Tenenbaum, Thomas L. Griffiths

TL;DR

This paper argues for designing AI as thought partners—systems that understand users, are understandable, and share a grounded view of the world—by explicitly modeling task, world, and human cognition within a Bayesian framework. It proposes a modular toolkit rooted in computational cognitive science, highlighting motifs such as probabilistic cognition, Theory of Mind, RSA, resource rationality, and program synthesis, and shows how probabilistic programming enables scalable, interpretable inference. Through case studies in programming (WatChat), embodied assistance (CLIPS), storytelling, and medicine, the authors demonstrate how these ideas can produce reliable, collaborative agents that reason with humans rather than merely imitate human behavior. The paper also discusses infrastructure, evaluation, and risk considerations to guide the development and deployment of human-centered thought partners, aiming to advance collaborative cognition in real-world settings with careful governance and interdisciplinarity.

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.

Building Machines that Learn and Think with People

TL;DR

This paper argues for designing AI as thought partners—systems that understand users, are understandable, and share a grounded view of the world—by explicitly modeling task, world, and human cognition within a Bayesian framework. It proposes a modular toolkit rooted in computational cognitive science, highlighting motifs such as probabilistic cognition, Theory of Mind, RSA, resource rationality, and program synthesis, and shows how probabilistic programming enables scalable, interpretable inference. Through case studies in programming (WatChat), embodied assistance (CLIPS), storytelling, and medicine, the authors demonstrate how these ideas can produce reliable, collaborative agents that reason with humans rather than merely imitate human behavior. The paper also discusses infrastructure, evaluation, and risk considerations to guide the development and deployment of human-centered thought partners, aiming to advance collaborative cognition in real-world settings with careful governance and interdisciplinarity.

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.
Paper Structure (20 sections, 2 figures, 2 tables)

This paper contains 20 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. What are Thought Partners?
  3. Modes of Collaborative Thought
  4. Example Domains
  5. Desiderata
  6. Engineering Human-Centered Thought Partners
  7. Implementing Our Desiderata
  8. Computational Cognitive Science Motifs
  9. Scaling Thought Partners via Probabilistic Programming
  10. Infrastructure around Thought Partners
  11. Case Studies in Engineering Thought Partners
  12. Thought Partners for Programming
  13. Thought Partners for Embodied Assistance
  14. Thought Partners for Storytelling
  15. Thought Partners for Medicine
  16. ...and 5 more sections

Figures (2)

  • Figure 1: Examples of ecosystems for thinking. Humans have long thought together. Machines expanded the efficiency of human thinking. Now, machines -- powered by AI -- open up new realms of computational thought partnership with humans.
  • Figure 2: Case Study Depictions. (a) WatChat infers the user's buggy mental model of the programming environment and interactively helps "patch" bug(s) in their understanding; (b) CLIPS reasons explicitly about agents' goals, integrating (culinary) world knowledge and the human's utterances to infer appropriate actions. Both agents reason about the joint team plan (tomato and dough are needed to make pizza); (c) Thought partners based on inverse inverse storytelling explicitly reason over models of the audience; (d) Future thought partners for medicine can jointly reason with a human doctor across modalities, a shared understanding of biology and patient needs, and a model of others' limitations.