Simulated vs. Unsimulated World Models

About the session

World models have become a cornerstone of AI because they generalize beyond experience and simulate physical events, giving agents internal predictions to plan and reason with. Brains predict too, propagating information forward through space and time and continually updating through prediction errors that flow backwards. Yet the two kinds of world model are defined, tested and extended very differently.

This session contrasts explicit generative simulation in AI, the in-silico, simulated world, with the brain’s implicit predictive representations, the in-vivo, unsimulated world. We bring together researchers from machine learning, cognitive science and neuroscience to derive testable hypotheses about how world models work, and to develop computational approaches for generating and evaluating them, with an eye on new inductive biases for machine learning.

The session builds on our review “Brain’s world model is smarter than we think”, Current Opinion in Behavioral Sciences (2026).

Key questions

• Simulated vs. unsimulated worlds: comparing explicit simulators with implicit predictive representations at the level of representational geometry, algorithm, and prediction and behavior. Where do JEPAs and multimodal foundation models fall short of the brain?
• Metacognitive simulation: planning involves not only simulating futures but judging whether the current model still holds. A metacognitive agent can learn which predictive model to use, trust and revise.
• Prediction without explicit models: model-free agents navigate without representing environmental structure, value functions project the world onto goal space, and large language models infer latent structure they were never shown.
• Memory and stochasticity: grid, place, border and corner cells support topology-aware prediction that may transfer across environments, while distributional dopamine codes capture real-world uncertainty.
• Brain foundation models: large models pretrained on brain imaging, such as Brain-JEPA, predict masked brain activity in latent space instead of reconstructing the signal. Can they reveal the brain’s own world model?

Programme

Tuesday 8 December 2026, Paris time. Select a session to read its abstract.

Times are tentative.

Speakers and panelists

Invited speakers

• BG

  Benjamin Grewe

  ETH Zürich, Switzerland

  Predictive world modelsTalk at 14:00

  Read abstract
• BD

  Benedetto De Martino

  University College London, UK

  Metacognitive world modelsTalk at 14:30

  Read abstract
• JZ

  Juan Helen Zhou

  National University of Singapore

  Brain foundation modelsTalk at 15:00

  Read abstract

Panelists

Panel at 15:30, joined by the three invited speakers.

• DB

  Demian Battaglia

  CNRS and University of Strasbourg

  Structure and dynamics

• JK

  Jeehyun Kwag

  Seoul National University

  Spatial navigation and memory

• MR

  Mattia Rigotti

  IBM Research Zurich

  Reasoning and LLMs

Organizers

• SL

  Sang Wan Lee

  KAIST, Korea

  Workshop chair

• MR

  Mattia Rigotti

  IBM Research Zurich, Switzerland

  Speakers and panel

• SL

  Seong-Whan Lee

  Korea University, Korea

  Financial support for the invited speakers

• JZ

  Juan Helen Zhou

  National University of Singapore

• JK

  Jeehyun Kwag

  Seoul National University, Korea

• DB

  Demian Battaglia

  CNRS and University of Strasbourg, France

Coordinated by Niklas Koeppe, Jiewon Kang and Myounghoon Ha (KAIST).