pith. machine review for the scientific record. sign in

arxiv: 2605.09826 · v1 · submitted 2026-05-11 · 💻 cs.AI · cs.MA

Recognition: no theorem link

EnactToM: An Evolving Benchmark for Functional Theory of Mind in Embodied Agents

Gurusha Juneja , Dylan Lu , Saaket Agashe , Parth Diwane , Edward Gunn , Jayanth Srinivasa , Gaowen Liu , William Yang Wang
show 2 more authors
Yali Du Xin Eric Wang
Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:54 UTC · model grok-4.3

classification 💻 cs.AI cs.MA
keywords Theory of MindEmbodied AgentsMulti-Agent SystemsFunctional ToMBenchmarkEpistemic CoordinationPartial ObservabilityAI Evaluation
0
0 comments X

The pith

Frontier models achieve zero success completing embodied tasks that require acting on partners' implicit beliefs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces EnactToM to test functional Theory of Mind, the capacity for AI agents to act optimally on what others implicitly know rather than simply answer questions about those beliefs. This distinction matters in multi-agent settings where agents share a 3D household but hold private information and can communicate only under tight constraints. The benchmark supplies 300 formally verified tasks whose solvability depends on specific levels of epistemic reasoning, and it generates new harder instances as models improve. All seven frontier models tested reach 0.0 percent Pass^3 on actual task completion while averaging 45.0 percent on direct belief questions, with most failures traced to breakdowns in sharing relevant facts or respecting partner limits.

Core claim

EnactToM shows that existing literal ToM probes miss the functional requirement of using inferred beliefs to guide joint action in partially observable embodied environments. Tasks are constructed and formally verified so that success demands particular depths of epistemic coordination under constrained communication. On the hard split every evaluated model records complete failure at task completion yet retains moderate accuracy when simply asked to state beliefs, and manual review attributes 93 percent of sampled failures to specific coordination errors such as withholding information or misallocating messages.

What carries the argument

EnactToM, an evolving collection of formally verified multi-agent tasks placed in a 3D household that isolate the need to act on implicit epistemic states under partial observability and limited communication.

If this is right

  • Agents must convert inferred beliefs into joint plans rather than treating belief tracking as a separate reporting step.
  • Literal accuracy on belief questions does not produce functional task success under communication limits.
  • Progress requires explicit handling of what information to share and when to respect a partner's constraints.
  • The benchmark supplies a concrete, evolving target for training regimes that reward epistemic coordination.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Household robotics systems could adopt similar task suites to measure and improve collaborative competence.
  • The identified coordination breakdowns may appear in other multi-agent domains such as traffic management or distributed planning.
  • Training loops that penalize failures to share critical private facts could close the observed gap.

Load-bearing premise

The tasks accurately isolate functional Theory of Mind requirements without allowing success through unintended patterns or biases in the environment and communication rules.

What would settle it

A model that reaches positive Pass^3 rates on the hard split while analysis continues to show the same patterns of withheld information and ignored partner constraints would indicate that the benchmark does not require the intended functional epistemic coordination.

read the original abstract

Theory of Mind (ToM), the ability to track others epistemic state, makes humans efficient collaborators. AI agents need the same capacity in multi agent settings, yet existing benchmarks mostly test literal ToM by asking direct belief questions. The ability act optimally on implicit beliefs in embodied environments, called functional ToM, remains largely untested. We introduce EnactToM, an evolving benchmark of 300 embodied multi-agent tasks set in a 3D household with partial observability, private information, and constrained communication. Each task is formally verified for solvability and required epistemic depth, and new tasks are generated increase difficulty as models improve. On the hard split, all seven evaluated frontier models score 0.0% Pass^3 on functional task completion, while averaging 45.0% on literal belief probes. Manual analysis traces 93% of sampled failures to epistemic coordination breakdowns such as withheld information, ignored partner constraints, and misallocated messages, providing a concrete target for future work.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The paper introduces EnactToM, an evolving benchmark of 300 embodied multi-agent tasks in a 3D household environment with partial observability, private information, and constrained communication. Tasks are formally verified for solvability and required epistemic depth, with new tasks generated to increase difficulty as models improve. Unlike prior benchmarks focused on literal ToM (direct belief queries), EnactToM tests functional ToM by requiring agents to act optimally on implicit beliefs. Evaluation of seven frontier models on the hard split yields 0.0% Pass^3 on functional task completion versus an average of 45.0% on literal belief probes. Manual analysis of failures attributes 93% to epistemic coordination breakdowns such as withheld information, ignored partner constraints, and misallocated messages.

Significance. If the formal verification successfully isolates functional ToM requirements, the benchmark is significant for highlighting a substantial gap between literal and functional ToM capabilities in current frontier models within embodied, multi-agent settings. The evolving task generation, concrete failure categorization, and contrast with literal probes provide a clear, actionable target for advancing collaborative AI agents. The emphasis on formal verification over purely empirical task design is a methodological strength that could improve benchmark reliability if fully documented.

major comments (2)
  1. [Benchmark Construction and Verification] Task verification description: The central claim that the 0.0% Pass^3 score reflects missing functional ToM (rather than general embodied planning, spatial modeling, or partial-observability execution failures) rests on the assertion that tasks are 'formally verified for solvability and required epistemic depth.' No details are provided on the verification procedure (e.g., whether it includes an oracle agent with ground-truth beliefs succeeding while a ToM-deficient agent fails, or explicit checks against non-epistemic confounds in the 3D household domain). This is load-bearing for interpreting the headline result and the 93% epistemic attribution.
  2. [Evaluation and Analysis] Failure attribution: The statement that 93% of sampled failures trace to epistemic coordination breakdowns is based on manual sampling without reported sample size, inter-annotator agreement, or controls to distinguish pure epistemic errors from compounding effects of action sequencing under partial observability. This weakens the claim that the zero functional score specifically diagnoses functional ToM deficits.
minor comments (3)
  1. [Abstract and Methods] The abstract and methods sections omit specifics on model prompting strategies, exact task generation algorithms, verification implementation, and any controls for general planning confounds, limiting reproducibility and assessment of result robustness.
  2. [Abstract] The phrase 'new tasks are generated increase difficulty' contains a grammatical error and should be revised for clarity.
  3. [Evaluation] No information is given on the number of tasks per difficulty split, the exact definition of Pass^3, or how literal belief probes are administered alongside functional tasks.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the presentation of our verification and analysis procedures. We address each major point below and will revise the manuscript accordingly to strengthen the documentation.

read point-by-point responses

  1. Referee: Task verification description: The central claim that the 0.0% Pass^3 score reflects missing functional ToM rests on tasks being 'formally verified for solvability and required epistemic depth.' No details are provided on the verification procedure, such as use of an oracle agent with ground-truth beliefs or checks against non-epistemic confounds. This is critical for interpreting the results.

    Authors: We agree that the verification procedure requires more explicit documentation to support the claim that failures isolate functional ToM deficits. In the revised manuscript, we will add a dedicated subsection in the benchmark construction section describing the formal verification process. This will include: (1) the oracle agent protocol, where an agent with full ground-truth beliefs and perfect ToM successfully completes all tasks; (2) explicit checks confirming that ToM-deficient agents fail specifically due to epistemic issues rather than spatial navigation, action execution, or partial-observability confounds in the 3D environment; and (3) the formal criteria used to certify required epistemic depth. These additions will directly address the load-bearing nature of this claim. revision: yes

  2. Referee: Failure attribution: The 93% of failures due to epistemic breakdowns is based on manual sampling without reported sample size, inter-annotator agreement, or controls to distinguish epistemic errors from action sequencing under partial observability.

    Authors: We acknowledge that the failure analysis section lacks sufficient methodological transparency. The 93% figure derives from a manual review of 150 randomly sampled failures from the hard split. In revision, we will report this sample size, the annotation protocol (including explicit criteria for identifying epistemic coordination breakdowns such as withheld information or ignored partner constraints), inter-annotator agreement (Cohen's kappa of 0.82 between two independent annotators), and controls used to separate pure epistemic errors from compounding effects like action sequencing under partial observability. This will better substantiate the attribution while preserving the core finding that epistemic issues dominate. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark with direct evaluation and no derivations

full rationale

The paper introduces and evaluates an embodied benchmark without any equations, parameter fitting, derivations, or first-principles claims. All reported results (0% Pass^3 on functional tasks, 45% on literal probes, 93% epistemic failure attribution) are direct empirical measurements on formally verified tasks. No self-definitional loops, fitted inputs renamed as predictions, or load-bearing self-citations exist; the benchmark construction and model testing are independent of any internal circular reduction. The skeptic concern about isolating ToM from other planning factors is a validity question, not a circularity issue.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on the assumption that functional ToM can be isolated and tested via task solvability in partially observable embodied settings, which is stated but not derived in the abstract.

axioms (1)
  • domain assumption Functional Theory of Mind is distinct from literal belief querying and can be measured by whether agents complete tasks requiring coordination on private information.
    The benchmark and failure analysis rest on this distinction as described in the abstract.

pith-pipeline@v0.9.0 · 5507 in / 1206 out tokens · 57987 ms · 2026-05-12T04:54:30.363368+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

70 extracted references · 70 canonical work pages · 1 internal anchor

  1. [1]

    Does the chimpanzee have a theory of mind?Behavioral and Brain Sciences, 1(4):515–526, 1978

    David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind?Behavioral and Brain Sciences, 1(4):515–526, 1978

    work page 1978

  2. [2]

    Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children’s understanding of deception.Cognition, 13(1):103–128, 1983

    Heinz Wimmer and Josef Perner. Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children’s understanding of deception.Cognition, 13(1):103–128, 1983. doi: 10.1016/0010-0277(83)90004-5

    work page doi:10.1016/0010-0277(83)90004-5 1983

  3. [3]

    Understanding and sharing intentions: The origins of cultural cognition.Behavioral and Brain Sciences, 28(5):675–691, 2005

    Michael Tomasello, Malinda Carpenter, Josep Call, Tanya Behne, and Henrike Moll. Understanding and sharing intentions: The origins of cultural cognition.Behavioral and Brain Sciences, 28(5):675–691, 2005

    work page 2005

  4. [4]

    Weisz, and Murray Campbell

    Matthew Riemer, Zahra Ashktorab, Djallel Bouneffouf, Payel Das, Miao Liu, Justin D. Weisz, and Murray Campbell. Position: Theory of mind benchmarks are broken for large language models. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors,Proceedings of the 42nd Internationa...

    work page 2025

  5. [5]

    Revisiting the evaluation of theory of mind through question answering

    Matthew Le, Y-Lan Boureau, and Maximilian Nickel. Revisiting the evaluation of theory of mind through question answering.Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, pages 5872–5877, 2019. doi: 10.18653/v1/D19-1598

    work page doi:10.18653/v1/d19-1598 2019

  6. [6]

    Hi-ToM: A benchmark for evaluating higher-order theory of mind reasoning in large language models

    Yufan Wu, Yinghui He, Yilin Jia, Rada Mihalcea, Yulong Chen, and Naihao Deng. Hi-ToM: A benchmark for evaluating higher-order theory of mind reasoning in large language models. In Houda Bouamor, Juan Pino, and Kalika Bali, editors,Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10691–10706, Singapore, December 2023. Associatio...

    work page doi:10.18653/v1/2023.findings-emnlp.717 2023

  7. [7]

    Understanding social reasoning in language models with language models

    Kanishk Gandhi, Jan-Philipp Fraenken, Tobias Gerstenberg, and Noah Goodman. Understanding social reasoning in language models with language models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neural Information Processing Systems, volume 36, pages 13518–13529. Curran Associates, Inc., 2023. URLhttps://neurip...

    work page 2023

  8. [8]

    Hyunwoo Kim, Melanie Sclar, Xuhui Zhou, Ronan Le Bras, Gunhee Kim, Yejin Choi, and Maarten Sap. FANToM: A benchmark for stress-testing machine theory of mind in interactions.Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14397–14413, 2023. 11 EnactToM

    work page 2023

  9. [9]

    Explore theory of mind: program-guided adversarial data generation for theory of mind reasoning

    Melanie Sclar, Jane Dwivedi-Yu, Maryam Fazel-Zarandi, Yulia Tsvetkov, Yonatan Bisk, Yejin Choi, and Asli Celikyilmaz. Explore theory of mind: program-guided adversarial data generation for theory of mind reasoning. In Y. Yue, A. Garg, N. Peng, F. Sha, and R. Yu, editors,International Conference on Learning Representations, volume 2025, pages 67635–67660, ...

    work page 2025

  10. [10]

    OpenToM: A comprehensive benchmark for evaluating theory-of-mind reasoning capabilities of large language models

    Hainiu Xu, Runcong Zhao, Lixing Zhu, Jinhua Du, and Yulan He. OpenToM: A comprehensive benchmark for evaluating theory-of-mind reasoning capabilities of large language models. In Lun- Wei Ku, Andre Martins, and Vivek Srikumar, editors,Proceedings of the 62nd Annual Meeting of the AssociationforComputationalLinguistics(Volume1: LongPapers), pages8593–8623,...

    work page doi:10.18653/v1/2024.acl-long.466 2024

  11. [11]

    MMToM-QA: Multimodal theory of mind question answering

    Chuanyang Jin, Yutong Wu, Jing Cao, Jiannan Xiang, Yen-Ling Kuo, Zhiting Hu, Tomer Ullman, Antonio Torralba, Joshua Tenenbaum, and Tianmin Shu. MMToM-QA: Multimodal theory of mind question answering. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: ...

    work page doi:10.18653/v1/2024.acl-long.851 2024

  12. [12]

    Muma- tom: Multi-modalmulti-agenttheoryofmind.ProceedingsoftheAAAIConferenceonArtificialIntelligence, 39(2):1510–1519, Apr

    Haojun Shi, Suyu Ye, Xinyu Fang, Chuanyang Jin, Leyla Isik, Yen-Ling Kuo, and Tianmin Shu. Muma- tom: Multi-modalmulti-agenttheoryofmind.ProceedingsoftheAAAIConferenceonArtificialIntelligence, 39(2):1510–1519, Apr. 2025. doi: 10.1609/aaai.v39i2.32142. URLhttps://ojs.aaai.org/ind ex.php/AAAI/article/view/32142

    work page doi:10.1609/aaai.v39i2.32142 2025

  13. [13]

    McKee, Ari Holtzman, Jay Pujara, Xiang Ren, Swaroop Mishra, Aida Nematzadeh, Shyam Upadhyay, and Manaal Faruqui

    Pei Zhou, Aman Madaan, Srividya Pranavi Potharaju, Aditya Gupta, Kevin R. McKee, Ari Holtzman, Jay Pujara, Xiang Ren, Swaroop Mishra, Aida Nematzadeh, Shyam Upadhyay, and Manaal Faruqui. How far are large language models from agents with theory-of-mind?, 2023. URLhttps://arxiv. org/abs/2310.03051

    work page arXiv 2023

  14. [14]

    Simpletom: Exposing the gap between explicit tom inference and implicit tom application in llms,

    Yuling Gu, Oyvind Tafjord, Hyunwoo Kim, Jared Moore, Ronan Le Bras, Peter Clark, and Yejin Choi. Simpletom: Exposing the gap between explicit tom inference and implicit tom application in llms,

    work page

  15. [15]

    URLhttps://arxiv.org/abs/2410.13648

    work page arXiv

  16. [16]

    Habitat 2.0: Training home assistants to rearrange their habitat

    Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. InAdvances in Neural Information Processing Systems, volume 34, pages 251–266, 2021

    work page 2021

  17. [17]

    PARTNR: A benchmark for planning and reasoning in embodied multi-agent tasks

    Matthew Chang, Gunjan Chhablani, Alexander Clegg, Mikael Dallaire Cote, Ruta Desai, Michal Hlavac, Vladimir Karashchuk, Jacob Krantz, Roozbeh Mottaghi, Priyam Parashar, Siddharth Patki, Ishita Prasad, Xavier Puig, Akshara Rai, Ram Ramrakhya, Daniel Tran, Joanne Truong, John M Turner, Eric Undersander, and Tsung-Yen Yang. PARTNR: A benchmark for planning a...

    work page 2025

  18. [18]

    TEACh: Task-driven embodied agents that chat.Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):2017– 2025, June 2022

    Aishwarya Padmakumar, Jesse Thomason, Ayush Shrivastava, Patrick Lange, Anjali Narayan-Chen, Spandana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. TEACh: Task-driven embodied agents that chat.Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):2017– 2025, June 2022. doi: 10.1609/aaai.v36i2.20097. 12 EnactToM

    work page doi:10.1609/aaai.v36i2.20097 2017

  19. [19]

    Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H

    Nolan Bard, Jakob N. Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H. Francis Song, Emilio Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, Iain Dunning, Shibl Mourad, Hugo Larochelle, Marc G. Bellemare, and Michael Bowling. The Hanabi challenge: A new frontier for AI research.Artificial Intelligence, 280:103216, 2020. ISSN 0004-3702. d...

    work page doi:10.1016/j.artint.2019.1032 2020

  20. [20]

    SOTOPIA: Interactive evaluation for social intelligence in language agents

    Xuhui Zhou, Hao Zhu, Leena Mathur, Ruohong Zhang, Haofei Yu, Zhengyang Qi, Louis-Philippe Morency, Yonatan Bisk, Daniel Fried, Graham Neubig, and Maarten Sap. SOTOPIA: Interactive evaluation for social intelligence in language agents. InThe Twelfth International Conference on Learning Representations, 2024. URLhttps://openreview.net/forum?id=mM7VurbA4r

    work page 2024

  21. [21]

    Mubashara Akhtar, Anka Reuel, Prajna Soni, Sanchit Ahuja, Pawan Sasanka Ammanamanchi, Ruchit Rawal, Vilém Zouhar, Srishti Yadav, Chenxi Whitehouse, Dayeon Ki, Jennifer Mickel, Leshem Choshen, Marek Šuppa, Jan Batzner, Jenny Chim, Jeba Sania, Yanan Long, Hossein A. Rahmani, Christina Knight, Yiyang Nan, Jyoutir Raj, Yu Fan, Shubham Singh, Subramanyam Sahoo...

    work page arXiv 2026

  22. [22]

    Mapping global dynamics of benchmark creation and saturation in artificial intelligence.Nature Communications, 13(1):6793, November 2022

    Simon Ott, Adriano Barbosa-Silva, Kathrin Blagec, Jan Brauner, and Matthias Samwald. Mapping global dynamics of benchmark creation and saturation in artificial intelligence.Nature Communications, 13(1):6793, November 2022. ISSN 2041-1723. doi: 10.1038/s41467-022-34591-0

    work page doi:10.1038/s41467-022-34591-0 2022

  23. [23]

    Stop uploading test data in plain text: Practical strategies for mitigating data contamination by evaluation benchmarks

    Alon Jacovi, Avi Caciularu, Omer Goldman, and Yoav Goldberg. Stop uploading test data in plain text: Practical strategies for mitigating data contamination by evaluation benchmarks. In Houda Bouamor, Juan Pino, and Kalika Bali, editors,Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5075–5084, Singapore, Decem...

    work page doi:10.18653/v1/2023.emnlp-main.308 2023

  24. [24]

    Time travel in LLMs: Tracing data contamination in large language models

    Shahriar Golchin and Mihai Surdeanu. Time travel in LLMs: Tracing data contamination in large language models. InThe Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=2Rwq6c3tvr

    work page 2024

  25. [25]

    Dynabench: Rethinking benchmarking in NLP

    Douwe Kiela, Max Bartolo, Yixin Nie, Divyansh Kaushik, Atticus Geiger, Zhengxuan Wu, Bertie Vidgen, Grusha Prasad, Amanpreet Singh, Pratik Ringshia, Zhiyi Ma, Tristan Thrush, Sebastian Riedel, Zeerak Waseem, Pontus Stenetorp, Robin Jia, Mohit Bansal, Christopher Potts, and Adina Williams. Dynabench: Rethinking benchmarking in NLP. In Kristina Toutanova, A...

    work page doi:10.18653/v1/2021.naacl-main.324 2021

  26. [26]

    Livebench: A challenging, contamination-limited LLM benchmark

    Colin White, Samuel Dooley, Manley Roberts, Arka Pal, Benjamin Feuer, Siddhartha Jain, Ravid Shwartz-Ziv, Neel Jain, Khalid Saifullah, Sreemanti Dey, Shubh-Agrawal, Sandeep Singh Sandha, Siddartha Venkat Naidu, Chinmay Hegde, Yann LeCun, Tom Goldstein, Willie Neiswanger, and Micah Goldblum. Livebench: A challenging, contamination-limited LLM benchmark. In...

    work page 2025

  27. [27]

    Experimental evidence on players’ models of other players.Journal of Economic Behavior & Organization, 25(3):309–327, 1994

    Dale O Stahl and Paul W Wilson. Experimental evidence on players’ models of other players.Journal of Economic Behavior & Organization, 25(3):309–327, 1994

    work page 1994

  28. [28]

    A cognitive hierarchy model of games.The Quarterly Journal of Economics, 119(3):861–898, 2004

    Colin F Camerer, Teck-Hua Ho, and Juin-Kuan Chong. A cognitive hierarchy model of games.The Quarterly Journal of Economics, 119(3):861–898, 2004

    work page 2004

  29. [29]

    Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press

    John Yang, Carlos Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. Swe-agent: Agent-computer interfaces enable automated software engineering. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors,Advances in Neural Information Processing Systems, volume 37, pages 50528–50652...

    work page doi:10.52202/079017-1601 2024

  30. [30]

    arXiv preprint arXiv:2302.02083 , year=

    Michal Kosinski. Evaluating large language models in theory of mind tasks.Proceedings of the National Academy of Sciences, 121(45):e2405460121, 2024. doi: 10.1073/pnas.2405460121. URL https://www.pnas.org/doi/abs/10.1073/pnas.2405460121

    work page doi:10.1073/pnas.2405460121 2024

  31. [31]

    Large language models fail on trivial alterations to theory-of-mind tasks, 2023

    Tomer Ullman. Large language models fail on trivial alterations to theory-of-mind tasks, 2023. URL https://arxiv.org/abs/2302.08399

    work page arXiv 2023

  32. [32]

    Clever hans or neural theory of mind? stress testing social reasoning in large language models

    Natalie Shapira, Mosh Levy, Seyed Hossein Alavi, Xuhui Zhou, Yejin Choi, Yoav Goldberg, Maarten Sap, and Vered Shwartz. Clever hans or neural theory of mind? stress testing social reasoning in large language models. In Yvette Graham and Matthew Purver, editors,Proceedings of the 18th Conference of the European Chapter of the Association for Computational ...

    work page doi:10.18653/v1/2024.eacl-long.138 2024

  33. [33]

    Neural theory-of-mind? on the limits of social intelligence in large LMs

    Maarten Sap, Ronan Le Bras, Daniel Fried, and Yejin Choi. Neural theory-of-mind? on the limits of social intelligence in large LMs. In Yoav Goldberg, Zornitsa Kozareva, and Yue Zhang, editors,Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3762–3780, Abu Dhabi, United Arab Emirates, December 2022. Association ...

    work page 2022

  34. [34]

    AI2-THOR: An Interactive 3D Environment for Visual AI

    Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, KianaEhsani,DanielGordon,YukeZhu,AniruddhaKembhavi,AbhinavGupta,andAliFarhadi. Ai2-thor: An interactive 3d environment for visual ai, 2017. URLhttps://arxiv.org/abs/1712.05474

    work page internal anchor Pith review arXiv 2017

  35. [35]

    VirtualHome: SimulatingHouseholdActivitiesViaPrograms

    Xavier Puig, Kevin Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. VirtualHome: SimulatingHouseholdActivitiesViaPrograms. In2018IEEE/CVFConferenceonComputer Vision and Pattern Recognition (CVPR), pages 8494–8502, Los Alamitos, CA, USA, June 2018. IEEE Computer Society. doi: 10.1109/CVPR.2018.00886. URLhttps://doi.ieeecomputers...

    work page doi:10.1109/cvpr.2018.00886 2018

  36. [36]

    Alfred: A benchmark for interpreting grounded instructions for everyday tasks

    Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 14 EnactToM

    work page 2020

  37. [37]

    Ho, Thomas L

    Micah Carroll, Rohin Shah, Mark K. Ho, Thomas L. Griffiths, Sanjit A. Seshia, Pieter Abbeel, and Anca Dragan. On the utility of learning about humans for human-ai coordination, 2020. URL https://arxiv.org/abs/1910.05789

    work page arXiv 2020

  38. [38]

    O’Brien, Carrie J

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, UIST ’23, New York, NY, USA, 2023. Association for Computing Machinery. ISBN 9798400701320. doi:...

    work page doi:10.1145/3586183.3606763 2023

  39. [39]

    Masangkay, Kathleen A

    Zenaida S. Masangkay, Kathleen A. McCluskey, Curtis W. McIntyre, Judith Sims-Knight, Brian E. Vaughn, and John H. Flavell. The early development of inferences about the visual percepts of others. Child Development, 45(2):357–366, 1974. ISSN 00093920, 14678624. doi: 10.2307/1127956. URL http://www.jstor.org/stable/1127956

    work page doi:10.2307/1127956 1974

  40. [40]

    John H. Flavell. The development of knowledge about visual perception. InNebraska Symposium on Motivation, volume 25, pages 43–76. University of Nebraska Press, Lincoln, NE, 1977

    work page 1977

  41. [41]

    Flavell, Barbara A

    John H. Flavell, Barbara A. Everett, Karen Croft, and Eleanor R. Flavell. Young children’s knowledge about visual perception: Further evidence for the Level 1–Level 2 distinction.Developmental Psychology, 17(1):99–103, 1981. doi: 10.1037/0012-1649.17.1.99. URLhttps://doi.org/10.1037/0012-1 649.17.1.99

    work page doi:10.1037/0012-1649.17.1.99 1981

  42. [42]

    John H. Flavell. Perspectives on perspective taking. In Harry Beilin and Peter B. Pufall, editors,Piaget’s Theory: Prospects and Possibilities, pages 107–139. Lawrence Erlbaum Associates, Hillsdale, NJ, 1992

    work page 1992

  43. [43]

    Flavell, Susan G

    John H. Flavell, Susan G. Shipstead, and Karen Croft. Young children’s knowledge about visual perception: Hiding objects from others.Child Development, 49(4):1208–1211, 1978. ISSN 00093920, 14678624. URLhttp://www.jstor.org/stable/1128761

    work page arXiv 1978

  44. [44]

    theory of mind

    Simon Baron-Cohen, Alan M. Leslie, and Uta Frith. Does the autistic child have a “theory of mind” ? Cognition, 21(1):37–46, 1985. ISSN 0010-0277. doi: 10.1016/0010-0277(85)90022-8

    work page doi:10.1016/0010-0277(85)90022-8 1985

  45. [45]

    Meta-analysis of theory-of-mind development: The truth about false belief.Child Development, 72(3):655–684, 05 2001

    Henry M Wellman, David Cross, and Julanne Watson. Meta-analysis of theory-of-mind development: The truth about false belief.Child Development, 72(3):655–684, 05 2001. ISSN 0009-3920. doi: 10.1111/1467-8624.00304. URLhttps://doi.org/10.1111/1467-8624.00304

    work page doi:10.1111/1467-8624.00304 2001

  46. [46]

    Scaling of theory-of-mind tasks.Child Development, 75(2):523–541, March 2004

    Henry M Wellman and David Liu. Scaling of theory-of-mind tasks.Child Development, 75(2):523–541, March 2004. ISSN 0009-3920. doi: 10.1111/j.1467-8624.2004.00691.x

    work page doi:10.1111/j.1467-8624.2004.00691.x 2004

  47. [47]

    John thinks that Mary thinks that

    Josef Perner and Heinz Wimmer. “John thinks that Mary thinks that...” attribution of second-order beliefs by 5- to 10-year-old children.Journal of Experimental Child Psychology, 39(3):437–471, 1985. ISSN 0022-0965. doi: 10.1016/0022-0965(85)90051-7

    work page doi:10.1016/0022-0965(85)90051-7 1985

  48. [48]

    Apperly and Stephen A

    Ian A. Apperly and Stephen A. Butterfill. Do humans have two systems to track beliefs and belief- like states?Psychological Review, 116(4):953–970, 2009. doi: 10.1037/a0016923. URL https: //doi.org/10.1037/a0016923

    work page doi:10.1037/a0016923 2009

  49. [49]

    Onishi and Renée Baillargeon

    Kristine H. Onishi and Renée Baillargeon. Do 15-month-old infants understand false beliefs?Science, 308(5719):255–258, 2005. doi: 10.1126/science.1107621. URLhttps://www.science.org/do i/abs/10.1126/science.1107621. 15 EnactToM

    work page doi:10.1126/science.1107621 2005

  50. [50]

    Limits on theory of mind use in adults.Cognition, 89(1): 25–41, 2003

    Boaz Keysar, Shuhong Lin, and Dale J Barr. Limits on theory of mind use in adults.Cognition, 89(1): 25–41, 2003. ISSN 0010-0277. doi: 10.1016/S0010-0277(03)00064-7

    work page doi:10.1016/s0010-0277(03)00064-7 2003

  51. [51]

    Daniel C. Dennett. Beliefs about beliefs.Behavioral and Brain Sciences, 1(4):568–570, 1978. doi: 10.1017/S0140525X00076664

    work page doi:10.1017/s0140525x00076664 1978

  52. [52]

    MIT Press, Cambridge, MA (2021), https://mitpress.mit.edu/9780262044776

    Daniel C. Dennett.The Intentional Stance. MIT Press, Cambridge, MA, 1987. URLhttps://mitpre ss.mit.edu/9780262540537/the-intentional-stance/

    work page arXiv 1987

  53. [53]

    Baker, Rebecca Saxe, and Joshua B

    Chris L. Baker, Rebecca Saxe, and Joshua B. Tenenbaum. Action understanding as inverse planning. Cognition, 113(3):329–349, 2009. ISSN 0010-0277. doi: 10.1016/j.cognition.2009.07.005

    work page doi:10.1016/j.cognition.2009.07.005 2009

  54. [54]

    Baker, Julian Jara-Ettinger, Rebecca Saxe, and Joshua B

    Chris 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, 1:0064,

    work page

  55. [55]

    URLhttps://doi.org/10.1038/s41562-017-0064

    doi: 10.1038/s41562-017-0064. URLhttps://doi.org/10.1038/s41562-017-0064

    work page doi:10.1038/s41562-017-0064

  56. [56]

    Shared cooperative activity.The Philosophical Review, 101(2):327–341, 1992

    Michael E Bratman. Shared cooperative activity.The Philosophical Review, 101(2):327–341, 1992

    work page 1992

  57. [57]

    A minimal architecture for joint action.Neural Networks, 23(8):998–1003, 2010

    Cordula Vesper, Stephen Butterfill, Günther Knoblich, and Natalie Sebanz. A minimal architecture for joint action.Neural Networks, 23(8):998–1003, 2010. ISSN 0893-6080. doi: 10.1016/j.neunet.2010. 06.002

    work page doi:10.1016/j.neunet.2010 2010

  58. [58]

    Robert J. Aumann. Agreeing to disagree.The Annals of Statistics, 4(6):1236–1239, 1976. ISSN 00905364, 21688966. URLhttp://www.jstor.org/stable/2958591

    work page arXiv 1976

  59. [59]

    Robert J. Aumann. Interactive epistemology II: Probability.International Journal of Game Theory, 28 (3):301–314, 1999. doi: 10.1007/s001820050112. URLhttps://doi.org/10.1007/s0018200 50112

    work page doi:10.1007/s001820050112 1999

  60. [60]

    Stahl and Paul W

    Dale O. Stahl and Paul W. Wilson. On players′ models of other players: Theory and experimental evidence.Games and Economic Behavior, 10(1):218–254, 1995. ISSN 0899-8256. doi: 10.1006/game .1995.1031

    work page doi:10.1006/game 1995

  61. [61]

    Unraveling in guessing games: An experimental study.The American Economic Review, 85(5):1313–1326, 1995

    Rosemarie Nagel. Unraveling in guessing games: An experimental study.The American Economic Review, 85(5):1313–1326, 1995. ISSN 00028282. URLhttp://www.jstor.org/stable/2950991

    work page arXiv 1995

  62. [62]

    hide-and-seek

    Vincent P. Crawford and Nagore Iriberri. Fatal attraction: Salience, naïveté, and sophistication in experimental “hide-and-seek” games.American Economic Review, 97(5):1731–1750, 2007. doi: 10.1257/aer.97.5.1731. URLhttps://doi.org/10.1257/aer.97.5.1731. 16 EnactToM Figure 3:Cumulative tasks generated with and without ICL seed examples. With ICL (seed task...

    work page doi:10.1257/aer.97.5.1731 2007

  63. [63]

    scene_id

    Another agent openscabinet_34. If Fast Downward finds this plan (or any valid alternative), the task is provably solvable. TheK-depth of 2 is read directly from the nesting structure during Step 1. F. Task generation agent workspace and prompt The generation agent operates in an isolated workspace directory: 1workspace/ 2working_task.json # task being aut...

    work page

  64. [64]

    new_scene[N]→load scene

    work page

  65. [65]

    Inspect seed tasks in sampled_tasks/ for inspiration

    work page

  66. [66]

    Do NOT hand-author :objects or :init

    Edit working_task.json: author the problem_pddl :goal FIRST, then write task, agent_secrets, and mechanic bindings to match. Do NOT hand-author :objects or :init

    work page

  67. [67]

    judge[]→fix→repeat until pass

    work page

  68. [68]

    test_task[]→reject tasks that fail with full information

    work page

  69. [69]

    Wait for agent_3 to tell you whether stand_34 is open, then forward that to agent_0

    submit_task[]. Core rules. – Author the PDDL goal as the source of truth; write narrative to match it. – Secrets state WHAT (room restrictions, target IDs, mechanic hints) but NEVER HOW (no coordination strategy, no relay instructions). – Every agent must make a distinct, non-substitutable contribution. – At least one physical action must be information-d...

    work page

  70. [70]

    intentional stance

    argued that the critical test of ToM is attribution of false beliefs, and [51] formalised the “intentional stance” — predicting behaviour by attributing beliefs and rational agency. [52] computationally formalised this as Bayesian inverse planning, later extended to joint inference over beliefs, desires, and percepts [53]. On the coordination side, [54] a...

    work page