Benchmarking World-Model Learning with Environment-Level Queries

Archana Warrier, Dat Nguyen, Michelangelo Naim, Moksh Jain, Yichao Liang, Karen Schroeder, Cambridge Yang, Joshua B. Tenenbaum, Sebastian Josef Vollmer, Kevin Ellis, Zenna Tavares
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:133215-133246, 2026.

Abstract

World models are central to building AI agents capable of flexible reasoning and planning. Yet current evaluations (i) test only properties measurable from observed interactions, such as next-frame prediction or task return, and (ii) do not test whether a learned model supports diverse queries about the environment. In contrast, humans build general-purpose models that can answer many different questions about an environment—including questions that require understanding global structure and counterfactual consequences. We propose WorldTest: a protocol for evaluating whether agents learn models that support multiple environment-level queries—questions whose answers depend on properties of the full environment, not just observed trajectories. Individually, these queries can target properties (e.g., reachability or the effects of interventions) that no single rollout distribution determines. Collectively, they assess model generality across query types. We instantiate WorldTest as AutumnBench, a benchmark of 43 interactive grid-world environments and 129 tasks across three query families for both humans and learning agents. Experiments with 517 human participants and five frontier models on AutumnBench show that humans substantially outperform these models, a gap we attribute to differences in exploration and belief updating.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-warrier26a, title = {Benchmarking World-Model Learning with Environment-Level Queries}, author = {Warrier, Archana and Nguyen, Dat and Naim, Michelangelo and Jain, Moksh and Liang, Yichao and Schroeder, Karen and Yang, Cambridge and Tenenbaum, Joshua B. and Vollmer, Sebastian Josef and Ellis, Kevin and Tavares, Zenna}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {133215--133246}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/warrier26a/warrier26a.pdf}, url = {https://proceedings.mlr.press/v306/warrier26a.html}, abstract = {World models are central to building AI agents capable of flexible reasoning and planning. Yet current evaluations (i) test only properties measurable from observed interactions, such as next-frame prediction or task return, and (ii) do not test whether a learned model supports diverse queries about the environment. In contrast, humans build general-purpose models that can answer many different questions about an environment—including questions that require understanding global structure and counterfactual consequences. We propose WorldTest: a protocol for evaluating whether agents learn models that support multiple environment-level queries—questions whose answers depend on properties of the full environment, not just observed trajectories. Individually, these queries can target properties (e.g., reachability or the effects of interventions) that no single rollout distribution determines. Collectively, they assess model generality across query types. We instantiate WorldTest as AutumnBench, a benchmark of 43 interactive grid-world environments and 129 tasks across three query families for both humans and learning agents. Experiments with 517 human participants and five frontier models on AutumnBench show that humans substantially outperform these models, a gap we attribute to differences in exploration and belief updating.} }
Endnote
%0 Conference Paper %T Benchmarking World-Model Learning with Environment-Level Queries %A Archana Warrier %A Dat Nguyen %A Michelangelo Naim %A Moksh Jain %A Yichao Liang %A Karen Schroeder %A Cambridge Yang %A Joshua B. Tenenbaum %A Sebastian Josef Vollmer %A Kevin Ellis %A Zenna Tavares %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-warrier26a %I PMLR %P 133215--133246 %U https://proceedings.mlr.press/v306/warrier26a.html %V 306 %X World models are central to building AI agents capable of flexible reasoning and planning. Yet current evaluations (i) test only properties measurable from observed interactions, such as next-frame prediction or task return, and (ii) do not test whether a learned model supports diverse queries about the environment. In contrast, humans build general-purpose models that can answer many different questions about an environment—including questions that require understanding global structure and counterfactual consequences. We propose WorldTest: a protocol for evaluating whether agents learn models that support multiple environment-level queries—questions whose answers depend on properties of the full environment, not just observed trajectories. Individually, these queries can target properties (e.g., reachability or the effects of interventions) that no single rollout distribution determines. Collectively, they assess model generality across query types. We instantiate WorldTest as AutumnBench, a benchmark of 43 interactive grid-world environments and 129 tasks across three query families for both humans and learning agents. Experiments with 517 human participants and five frontier models on AutumnBench show that humans substantially outperform these models, a gap we attribute to differences in exploration and belief updating.
APA
Warrier, A., Nguyen, D., Naim, M., Jain, M., Liang, Y., Schroeder, K., Yang, C., Tenenbaum, J.B., Vollmer, S.J., Ellis, K. & Tavares, Z.. (2026). Benchmarking World-Model Learning with Environment-Level Queries. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:133215-133246 Available from https://proceedings.mlr.press/v306/warrier26a.html.

Related Material