Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

This review defines 'medical world models' as systems that represent evolving patient states and simulate responses to interventions, and reports that only 14 studies meet that definition under strict criteria — most still retrospective or

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 02:58 UTC pith:OMESCSM2

load-bearing objection A solid, honest review that gives medical world models a useful capability/evidence taxonomy; the 14-study map needs audit-trail release and COI disclosure before it can be taken at face value. the 4 major comments →

arxiv 2607.25242 v1 pith:OMESCSM2 submitted 2026-07-28 cs.CV

Medical world models in healthcare: foundations, applications, and challenges for trustworthy clinical translation

Zhaoyan Chen , Zhongxiu Cong , Zhuanfeng Jin , Wanshu Fan , Dongsheng Zhou , Qi Ai , Haifan Gong , Congyu Liao
show 2 more authors
Xiaofeng Liu Cong Wang
This is my paper
classification cs.CV
keywords medical world modelsclinical digital twinspatient trajectory modellingtreatment-response simulationcounterfactual reasoningclinical decision supporttrustworthy medical AIcapability levels L1-L4
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to give the emerging field of medical world models a working boundary. Its central claim is that the defining feature of such models is an internal patient-state representation together with transition dynamics — how the state evolves over time and in response to clinical actions — and it organizes the field into four capabilities and a four-level (L1–L4) capability ladder. Applying a strict empirical definition to the literature, the review finds just 14 qualifying studies, and says most of the evidence base remains retrospective, task-specific, or preclinical. Its most consequential caution is that action-conditioned prediction must not be mistaken for counterfactual inference. If the framework is right, it gives researchers and clinicians a shared vocabulary and separates functional capability from clinical evidence maturity.

Core claim

On its own terms, the review establishes that medical world models are a distinct paradigm: they operate on latent patient states and a transition distribution P(s_{t+1} | s_t, a_t), rather than mapping observations directly to outcomes. It proposes a strict empirical definition — a study must empirically evaluate a learned model of dynamic state evolution, with at least one of representation of evolving state, learned transition dynamics, action-conditioned simulation, multi-step rollout, or interaction in a learned dynamic environment — and reports that only 14 of the screened studies qualify. Those studies cluster at L1 and L2 capability; only two reach L3 (comparing alternative actions),

What carries the argument

The central object is the state-transition equation P(s_{t+1} | s_t, a_t): a latent patient state s that evolves under a clinical action a (treatment, procedure, acquisition geometry, or device control). Around this, the review builds an L1–L4 capability hierarchy — L1 temporal prediction without explicit actions; L2 action-conditioned transition; L3 comparison of predicted outcomes under alternative actions; L4 closed-loop planning or control. This machinery lets the review separate genuinely dynamic models from static predictors, and lets it argue that capability level and clinical evidence maturity are independent axes.

Load-bearing premise

The mapping of the field rests on the authors' screening pipeline — a hand-picked 21-record seed set, a broad ACM field filter, and author reconciliation without released screening logs — so if the seed steered the corpus toward known studies, the 14-study count and the 'most remain retrospective' conclusion could misrepresent the field.

What would settle it

An independent team re-runs the same published search queries with no seed set and full screening logs, then applies the review's strict definition; if they find substantially more than 14 qualifying studies, or find that several of the 14 fail the definition, the paper's central field-map and maturity claims would not hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Evaluators should report capability level (L1–L4) separately from evidence maturity, so a planned benchmark demo is not presented as clinical readiness.
  • Action-conditioned rollouts should be labelled associational unless the intervention, estimand, and causal assumptions are specified and validated.
  • The most clinically valuable application — simulating what happens under a different treatment — is the least mature and the most in need of causal grounding.
  • Safe clinical use requires calibrated trajectory-level uncertainty, safety-constrained planning, and clinician supervision rather than autonomous prescriptions.
  • Clinical digital twins are best understood as a cross-cutting integration framework spanning several application domains, not a separate category.

Where Pith is reading between the lines

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

  • The L1–L4 scheme could become a reporting convention: if future papers state 'capability L2, evidence retrospective,' the field's maturity would be legible at a glance.
  • The 14-study count is a lower bound set by the review's own criteria, so the durable contribution is the boundary definition, not the exact number.
  • At least some included preprints appear to come from the same research group as this review, so an independent replication of the evidence map would test whether the field's small apparent size is real or a seed-set artifact.
  • A testable extension is to apply the strict definition to older longitudinal treatment-effect models; if they qualify, 'medical world model' formalizes an existing lineage rather than starting a new one.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This Review Article proposes a conceptual and functional definition of 'medical world models' and organizes the field around four capabilities (patient-state representation, temporal dynamics modelling, intervention-conditioned simulation, clinician-supervised planning) and six application domains. It reports a structured narrative synthesis with an evidence-mapping pipeline: 1,455 unique records screened, 98 cited sources, and 14 studies meeting a strict empirical definition of a medical world model. The paper argues that most of these studies remain retrospective, task-specific, or preclinical, and it distinguishes action-conditioned prediction from valid counterfactual inference, setting out evidence requirements for trustworthy clinical translation. The conceptual framework and the L1-L4 capability hierarchy are the paper's principal contributions.

Significance. If the empirical mapping is auditable, this review would provide a useful operational boundary for a nascent field and a maturity bar: the 14-study strict subset, the capability distribution, and the claim that most evidence is retrospective are concrete and falsifiable. The paper also makes an important and well-argued methodological point, especially in Eq. (3), Section 4.3, and Section 5.2, that action-conditioned simulation should not be equated with counterfactual inference, and it explicitly separates technical capability from clinical evidence maturity. The authors disclose many limitations of their selection process and correctly avoid presenting the review as a full systematic review. However, the central empirical claims currently depend on a screening pipeline whose seed set, author-reconciliation step, and unpublished search logs prevent independent verification, and two of the strict-subset studies are the authors' own preprints without conflict-of-interest disclosure.

major comments (4)
  1. [Section 2.4; Data Availability] The abstract claims 'reproducible evidence mapping,' but the screening audit trail is not actually available. Section 2.4 states that a dated internal protocol, complete search logs, and protocol deviations were 'retained'; the Data Availability statement says no new datasets were generated or analyzed, and no logs or screening decisions are released. Because the headline numbers (14 strict studies; 'most remain retrospective, task-specific, or preclinical') rest entirely on this pipeline, the authors must either release the search strategies, screening logs, exclusion decisions, and reconciliation records, or materially soften the reproducibility claim. This is a load-bearing issue, not a presentation detail.
  2. [Section 2.4; Table 3; Declarations] Two studies in the strict subset are the authors' own preprints: Brain-WM (ref. 28, Wang C. et al.) and the surgical world-model pilot (ref. 91, Chen Z. et al.). The first author of the review is first author of ref. 91, and several review co-authors appear on ref. 28. The Declarations state 'no competing interests,' and there is no discussion of how the 21-record seed set or the 'author reconciliation' step treated these works. Given the seed set was chosen during preliminary manuscript development, the inclusion process could preferentially retain the authors' own studies. Please disclose the relevant conflicts and provide a sensitivity analysis of the 14-study count and the maturity conclusion when the authors' own preprints are excluded.
  3. [Section 2.4] The 'author reconciliation' step is described only as authors reviewing the candidate set and reconciling eligible records with the manuscript bibliography. This creates a potential circularity: if the bibliography influenced the reconciliation, the screened corpus is not an independent sample of the literature. The manuscript does not report how many of the 21 seed records survived into the 29 database/seed-derived cited sources, how many database records were excluded specifically at the reconciliation step, or how many of the 14 strict studies originated from the seed set versus the database searches. These numbers are necessary to assess whether the empirical map could be steered toward studies already known to the authors. Without them, the 14-study count is not independently interpretable.
  4. [Section 2.2 / Supplementary Information] The search strategy says platform-specific syntax, field restrictions, and adaptations are reported in the Supplementary Information, but no supplementary file is present with the manuscript. Since the authors emphasize 'reproducible evidence mapping,' the full query strings and any deviations from the described query families must be supplied with the revision. This is a reproducibility requirement, not a cosmetic one, because the broad ACM field check (which removed 1,950 of 1,959 records) is a high-impact screening decision that needs a transparent specification.
minor comments (4)
  1. [Table 3] Typo: 'V olumetric-context' should be 'Volumetric-context'. Also, the table lists 10 models while the text reports 14 strict studies; adding a footnote explaining that the table is representative, or providing the full list in the Supplementary Information, would help readers reconcile the numbers.
  2. [Section 4.6 / Table 3] The L1-L4 capability classification is central to the paper's organizational claim, but the text does not provide a detailed justification for each row's level (e.g., why EHRWorld is L2 rather than L3). A short rubric or per-model rationale in the Supplementary Information would strengthen the framework's credibility.
  3. [Section 5] The sentence beginning 'At the final search date of 20 July 2026, more than half of the studies in the strict empirical subset were available as preprints' is not directly supported by the table (which shows 7 of 10 listed as preprints). Please either provide the count for the full strict subset or adjust the claim.
  4. [References] Several references mix arXiv identifiers and venue information in a non-uniform style (e.g., refs. 12, 14, 15, 36). Additionally, ref. 41 (SteeraMed) is discussed at some length in Sections 3.5 and 4.4 while being described as not meeting the strict definition; a brief note that this is a framework/position work rather than an empirical study would avoid the impression of a special-case inclusion.

Circularity Check

0 steps flagged

No circularity found: the review's framework is stipulated and its empirical count is a screening result, not a fitted prediction; auditability concerns are methodological, not circular.

full rationale

This is a structured narrative review rather than a derivation chain. The four-capability organization and the six application domains are explicitly offered as the authors' organizing definitions (Abstract; Section 4), and equations (1)-(3) are standard formalisms used descriptively to distinguish prediction from action-conditioned transition; they are not derived from the data or from the paper's own conclusions. The 14-study strict empirical subset is the outcome of stated inclusion criteria (Section 2.3) applied through the screening process described in Section 2.4, so the count is a classification result rather than a fitted parameter renamed as a prediction or a quantity defined in terms of the conclusion. The paper itself flags the main weaknesses: 'Limitations introduced by the machine-assisted, manuscript-aligned selection process are reported explicitly rather than interpreting the corpus as an exhaustive systematic sample' (Section 2.6), and 'No new datasets were generated or analyzed in this review' (Data availability). The unreleased search logs and the seed-set/reconciliation steps are genuine reproducibility and selection-risk limitations, but they do not make the central claim equivalent to its inputs by construction. The self-citations that appear in the corpus (e.g., refs 28 and 91) are used as examples within the broader synthesis, and the conclusion that most evidence is retrospective, task-specific, or preclinical is supported by the wider set of independently authored studies in Table 3 and Sections 5-6; no load-bearing argument reduces to an unverified self-citation. The paper therefore contains no significant circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

No new physical or mechanistic entities are proposed; the four-capability framework is an analytical taxonomy rather than a postulated entity. Free parameters are absent because the paper performs no quantitative fitting. The axioms listed are the definitional and domain assumptions that shape which studies count as medical world models and what evidence is admitted.

axioms (4)
  • ad hoc to paper Operational definition of a medical world model: a study must empirically evaluate a learned dynamic state-evolution model with at least one of: evolving state representation, learned transitions, action-conditioned simulation, multi-step rollout, or interaction in a learned environment.
    This definition determines the strict subset of 14 studies and hence the review's central evidence claims (Section 2.3).
  • domain assumption L1-L4 capability levels are taken from Qazi et al. (ref 84) and assumed to be a meaningful way to compare functional scope.
    The review's capability claims (e.g., MeWM is L3, EHRWorld is L2) rest on this adopted taxonomy (Section 5; Table 3 note).
  • domain assumption Visual realism, trajectory consistency, and predictive performance are not sufficient evidence for causal treatment effects.
    Used throughout Sections 4.3 and 5.2 to appraise MeWM, Brain-WM, and CLARITY; normative assumption about what counts as evidence.
  • domain assumption Medical world-model principles are assumed to transfer from model-based reinforcement learning and general world models to clinical settings despite differences in data and intervention semantics.
    The entire framing depends on transferability of concepts such as internal rollout and imagined trajectories (Sections 3.5, 4.5).

pith-pipeline@v1.3.0-alltime-deepseek · 21322 in / 10552 out tokens · 103700 ms · 2026-08-01T02:58:45.503028+00:00 · methodology

0 comments
read the original abstract

Medical world models offer a framework for extending medical artificial intelligence beyond static prediction by representing evolving patient states and modelling how they change over time and in response to clinical interventions. This Review defines the conceptual boundaries, technical foundations, application domains, and evidence requirements of the field through a structured narrative synthesis with reproducible evidence mapping.We screened 1,455 unique records and assembled a corpus of 98 sources, including 14 studies that met a strict empirical definition of a medical world model. The field is organised around four capabilities: patient state representation, temporal dynamics modelling, intervention-conditioned simulation, and clinician-supervised planning. Evidence spans medical imaging, longitudinal electronic health records, treatment response modelling, physiological and multimodal state modelling, ultrasound and surgical interaction, and population and health-system simulation; clinical digital twins are treated as a cross-cutting integration framework.Current studies provide early evidence of technical feasibility for trajectory forecasting and comparison of candidate interventions, but most remain retrospective, task-specific, or preclinical. The evidence base is further limited by incomplete longitudinal intervention data, inconsistent action semantics, limited causal identifiability, long-horizon error accumulation, inadequate uncertainty estimation, and limited external validation. Clinical translation will therefore depend on precise intervention representations, robust causal and mechanistic grounding, calibrated trajectory-level uncertainty, safety-constrained planning, and prospective multicentre validation against clinically meaningful endpoints.

Figures

Figures reproduced from arXiv: 2607.25242 by Cong Wang, Congyu Liao, Dongsheng Zhou, Haifan Gong, Qi Ai, Wanshu Fan, Xiaofeng Liu, Zhaoyan Chen, Zhongxiu Cong, Zhuanfeng Jin.

Figure 1
Figure 1. Figure 1: Literature identification and selection flow. Electronic database searches and the initial seed set yielded 3,528 records. Following prespecified ACM field filtering, cross-source deduplication, and two-stage machine-assisted screening with author verification, 29 records were retained after eligibility assessment. An additional 69 contextual, foundational, methodological, and reporting sources were identi… view at source ↗
Figure 2
Figure 2. Figure 2: Evolution of major paradigms in medical artificial intelligence. The figure traces the progression from expert systems and traditional machine learning to deep learning, medical foundation models, and emerging medical world models. The curve represents a qualitative expansion in data scale and modelling capability, whereas the horizontal bars indicate the approximate periods of continued use and temporal o… view at source ↗
Figure 3
Figure 3. Figure 3: Technical evolution of world models and their adaptation to medicine. The timeline traces se￾lected milestones from early model-based reinforcement learning to latent-dynamics modelling, predictive representation learning, diffusion and video-generation approaches, and emerging medical world models. The milestones illustrate developments in learned dynamics, internal rollout, future-state prediction, and a… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of medical world models in medicine. Multimodal and longitudinal clinical obser￾vations are encoded into patient-state representations and combined with explicit action representations, latent-dynamics models, imagined rollouts, and planning, decision-support, or embodied-control modules. The L1–L4 levels describe functional capability rather than clinical evidence maturity. Trustworthy translatio… view at source ↗
Figure 5
Figure 5. Figure 5: Conceptual architecture of medical world models. Multimodal clinical observations are encoded into latent patient states, whose evolution is modelled through temporal and intervention-conditioned transitions. Multi-step rollouts may generate future imaging observations, laboratory and physiological trajectories, prognostic and adverse-event risks, and candidate treatment strategies. Causal or mechanistic c… view at source ↗
Figure 6
Figure 6. Figure 6: Representative application domains of medical world models. Six overlapping domains are considered: medical imaging representation and simulation of future observations; longitudinal EHR and patient trajectory modelling; disease progression and treatment response simulation; physiological signal and multimodal state modelling; ultrasound guidance and surgical embodied interaction; and population and health… view at source ↗
Figure 7
Figure 7. Figure 7: Evidence framework for trustworthy clinical translation of medical world models. Model￾generated outputs may include future imaging observations, patient trajectories, treatment-conditioned responses, and action-conditioned procedural media. Clinical interpretation requires evidence across data completeness and provenance, intervention specification, causal validity, trajectory uncertainty, external and pr… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. AuricularWorld: Hierarchical Action-Guided World Modeling for Fine-Grained Auricular Structure Segmentation from CT Scans

    cs.CV 2026-07 conditional novelty 5.5

    A latent RSSM with hierarchical anatomical add/remove actions cuts HD95 by ~43% versus nnU-Net on fine-grained nested auricular CT segmentation.

Reference graph

Works this paper leans on

27 extracted references · 4 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [2]

    Khan, W.et al.A comprehensive survey of foundation models in medicine.IEEE Rev. Biomed. Eng. 19, 283–304, DOI: 10.1109/rbme.2025.3531360 (2026)

    arXiv 2025

  2. [3]

    Heal.6, e281–e290, DOI: 10.1016/s2589-7500(24)00025-6 (2024)

    Kraljevic, Z.et al.Foresight—a generative pretrained transformer for modelling of patient timelines using electronic health records: a retrospective modelling study.The Lancet Digit. Heal.6, e281–e290, DOI: 10.1016/s2589-7500(24)00025-6 (2024). 4.Ma, J.et al.Segment anything in medical images.Nat. Commun.15, 654, DOI: 10.1038/s41467-024-44824-z (2024). 5....

    doi:10.1016/s2589-7500(24)00025-6 2024

  3. [8]

    E.et al.A causal roadmap for generating high-quality real-world evidence.J

    Dang, L. E.et al.A causal roadmap for generating high-quality real-world evidence.J. Clin. Transl. Sci.7, e212, DOI: 10.1017/cts.2023.635 (2023). 9.Castro, D. C., Walker, I. & Glocker, B. Causality matters in medical imaging.Nat. Commun.11, 3673, DOI: 10.1038/s41467-020-17478-w (2020)

    work page doi:10.1017/cts.2023.635 2023

  4. [10]

    A path towards autonomous machine intelligence (2022)

    LeCun, Y . A path towards autonomous machine intelligence (2022). Position paper, version 0.9.2, 27 June 2022. 11.Hafner, D.et al.Learning latent dynamics for planning from pixels. InProceedings of the 36th International Conference on Machine Learning, vol. 97, 2555–2565 (PMLR, 2019). 12.Hafner, D., Lillicrap, T., Ba, J. & Norouzi, M. Dream to control: Le...

    Pith/arXiv arXiv doi:10.48550/arxiv.1912.01603 2022

  5. [16]

    17.Ding, J.et al.Understanding world or predicting future? a comprehensive survey of world models

    Dong, J.et al.Learning to model the world: A survey of world models in artificial intelligence, DOI: 10.20944/preprints202603.0739.v1 (2026). 17.Ding, J.et al.Understanding world or predicting future? a comprehensive survey of world models. ACM Comput. Surv.58, 1–38, DOI: 10.1145/3746449 (2026). 18.Wang, S.et al.World action models: The next frontier in e...

    arXiv 2026

  6. [22]

    Saeed, N.et al.Medical world model: From passive prediction to active simulation in medicine, DOI: 10.20944/preprints202604.2168.v1 (2026)

    arXiv 2026

  7. [23]

    Mu, L.et al.Ehrworld: A patient-centric medical world model for long-horizon clinical trajectories, DOI: 10.48550/arXiv.2602.03569 (2026). [EB/OL]. arXiv:2602.03569, 2026., 2602.03569. 24.Yang, Y .et al.Medical world model. In2025 IEEE/CVF International Conference on Computer Vision (ICCV), 8319–8329, DOI: 10.1109/iccv51701.2025.00779 (IEEE, 2025)

    doi:10.48550/arxiv.2602.03569 2026

  8. [25]

    & Tian, Y

    Ding, T., Zou, Y ., Chen, C., Shah, M. & Tian, Y . Clarity: Medical world model for guiding treatment decisions by modeling context-aware disease trajectories in latent space, DOI: 10.48550/arXiv.2512.08029 (2025). [EB/OL]. arXiv:2512.08029, 2025., 2512.08029

    Pith/arXiv arXiv doi:10.48550/arxiv.2512.08029 2025

  9. [33]

    & Ranganath, R

    Huang, K., Altosaar, J. & Ranganath, R. Clinicalbert: Modeling clinical notes and predicting hospital readmission, DOI: 10.48550/arXiv.1904.05342 (2019). [EB/OL]. arXiv:1904.05342, 2019., 1904.05342

    Pith/arXiv arXiv doi:10.48550/arxiv.1904.05342 1904

  10. [34]

    & Zhi, D

    Rasmy, L., Xiang, Y ., Xie, Z., Tao, C. & Zhi, D. Med-bert: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction.npj Digit. Medicine4, 86, DOI: 10.1038/s41746-021-00455-y (2021). 35.Li, Y .et al.Behrt: Transformer for electronic health records.Sci. Reports10, 7155, DOI: 10.1038/s41598-020-62922-y ...

    Pith/arXiv arXiv 2021

  11. [42]

    S.et al.A comprehensive survey on surgical digital twin, DOI: 10.48550/arXiv.2512.00019 (2025)

    Khan, A. S.et al.A comprehensive survey on surgical digital twin, DOI: 10.48550/arXiv.2512.00019 (2025). [EB/OL]. arXiv:2512.00019, 2025., 2512.00019. 43.Warrier, A.et al.Benchmarking world-model learning with environment-level queries, DOI: 10.48550/arXiv.2510.19788 (2025). [EB/OL]. arXiv:2510.19788, 2025., 2510.19788. 44.Collins, G. S.et al.Tripod+ai st...

    work page doi:10.48550/arxiv.2512.00019 2025

  12. [46]

    Medicine26, 1320–1324, DOI: 10.1038/s41591-020-1041-y (2020)

    Norgeot, B.et al.Minimum information about clinical artificial intelligence modeling: the mi-claim checklist.Nat. Medicine26, 1320–1324, DOI: 10.1038/s41591-020-1041-y (2020). 47.Chen, R. J.et al.Algorithmic fairness in artificial intelligence for medicine and healthcare.Nat. Biomed. Eng.7, 719–742, DOI: 10.1038/s41551-023-01056-8 (2023). 48.Gichoya, J. W...

    doi:10.1038/s41591-020-1041-y 2020

  13. [53]

    Medicine3, 119, DOI: 10.1038/s41746-020-00323-1 (2020)

    Rieke, N.et al.The future of digital health with federated learning.npj Digit. Medicine3, 119, DOI: 10.1038/s41746-020-00323-1 (2020)

    doi:10.1038/s41746-020-00323-1 2020

  14. [54]

    & Beam, A

    Kompa, B., Snoek, J. & Beam, A. L. Second opinion needed: communicating uncertainty in medical machine learning.npj Digit. Medicine4, 4, DOI: 10.1038/s41746-020-00367-3 (2021)

    doi:10.1038/s41746-020-00367-3 2021

  15. [55]

    Medicine28, 924–933, DOI: 10.1038/s41591-022-01772-9 (2022)

    Vasey, B.et al.Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: Decide-ai.Nat. Medicine28, 924–933, DOI: 10.1038/s41591-022-01772-9 (2022). 56.Zeng, Z.et al.World models: The safety perspective. In2024 IEEE 35th International Symposium on Software Reliability Engineering Workshops (...

    arXiv 2022

  16. [62]

    & Lillicrap, T

    Hafner, D., Pasukonis, J., Ba, J. & Lillicrap, T. Mastering diverse control tasks through world models. Nature640, 647–653, DOI: 10.1038/s41586-025-08744-2 (2025). 63.Hafner, D., Lillicrap, T., Norouzi, M. & Ba, J. Mastering atari with discrete world models, DOI: 10.48550/arXiv.2010.02193 (2020). [C]//ICLR, 2021., 2010.02193. 64.Yang, L.et al.Diffusion mo...

    Pith/arXiv arXiv doi:10.48550/arxiv.2010.02193 2025

  17. [66]

    Causal Transformer for Estimating Counterfactual Outcomes

    Bica, I., Alaa, A. M., Jordon, J. & van der Schaar, M. Estimating counterfactual treatment outcomes over time through adversarially balanced representations. InInternational Conference on Learning Representations(2020). 67.Melnychuk, V ., Frauen, D. & Feuerriegel, S. Causal transformer for estimating counterfactual outcomes, DOI: 10.48550/arXiv.2204.07258...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2204.07258 2020

  18. [69]

    Zhang, B. & Mi, Y . World model enhanced offline reinforcement learning for sequential intervention optimization in acute kidney injury.AI Medicine3, 2, DOI: 10.53941/aim.2026.100002 (2026)

    arXiv 2026

  19. [70]

    Policy4OOD: A Knowledge-Guided World Model for Policy Intervention Simulation against the Opioid Overdose Crisis

    Schrittwieser, J.et al.Mastering atari, go, chess and shogi by planning with a learned model.Nature 588, 604–609, DOI: 10.1038/s41586-020-03051-4 (2020). 71.Ma, Y .et al.Policy4ood: A knowledge-guided world model for policy intervention simulation against the opioid overdose crisis, DOI: 10.48550/arXiv.2602.12373 (2026). [EB/OL]. arXiv:2602.12373, 2026., ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2602.12373 2020

  20. [77]

    Bruce, J.et al.Genie: Generative interactive environments, DOI: 10.48550/arXiv.2402.15391 (2024). [EB/OL]. arXiv:2402.15391, 2024., 2402.15391. 78.Zhou, G., Pan, H., LeCun, Y . & Pinto, L. Dino-wm: World models on pre-trained visual features enable zero-shot planning, DOI: 10.48550/arXiv.2411.04983 (2024). [EB/OL]. arXiv:2411.04983, 2024., 2411.04983. 79....

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.15391 2024

  21. [81]

    L.et al.A multi-center study on the adaptability of a shared foundation model for electronic health records.npj Digit

    Guo, L. L.et al.A multi-center study on the adaptability of a shared foundation model for electronic health records.npj Digit. Medicine7, 171, DOI: 10.1038/s41746-024-01166-w (2024)

    doi:10.1038/s41746-024-01166-w 2024

  22. [82]

    ResearchGate preprint; metadata verified through DataCite

    Yang, J.et al.Auricularworld: Hierarchical action-guided world modeling for fine-grained auricular structure segmentation in ct, DOI: 10.13140/rg.2.2.26911.32165 (2026). ResearchGate preprint; metadata verified through DataCite. 83.Kang, L.et al.Dreamreg: Belief-driven world model for 2d-3d ultrasound registration, DOI: 10.48550/arXiv.2606.18825 (2026). 2...

    arXiv 2026

  23. [86]

    In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 20778–20788, DOI: 10.1109/cvpr52734.2025.01935 (IEEE, 2025)

    Yue, Y .et al.Chexworld: Exploring image world modeling for radiograph representation learning. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 20778–20788, DOI: 10.1109/cvpr52734.2025.01935 (IEEE, 2025). 87.Allam, A., Feuerriegel, S., Rebhan, M. & Krauthammer, M. Analyzing patient trajectories with artificial intelligence.J...

    arXiv 2025

  24. [91]

    Chen, Z.et al.How far are surgeons from surgical world models? a pilot study on zero-shot surgical video generation with expert assessment, DOI: 10.48550/arXiv.2511.01775 (2025). [EB/OL]. arXiv:2511.01775, 2025., 2511.01775

    doi:10.48550/arxiv.2511.01775 2025

  25. [92]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5315–5324 (2026)

    Zhang, Z.et al.Surgical procedural planning as 3d world modelling: Towards automated pulmonary resection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5315–5324 (2026). CVPR 2026 Findings paper; metadata and pagination verified from the locally supplied accepted paper. 93.Steinberg, E., Fries, J., Xu, Y . & Shah, N....

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.03150 2026

  26. [94]

    & Shah, N

    Wornow, M., Thapa, R., Steinberg, E., Fries, J. & Shah, N. Ehrshot: An ehr benchmark for few-shot evaluation of foundation models. InAdvances in Neural Information Processing Systems 36, 67125–67137, DOI: 10.52202/075280-2933 (Neural Information Processing Systems Foundation, Inc. (NeurIPS), 2023). 95.Pang, C.et al.Cehr-gpt: Generating electronic health r...

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.04400 2023

  27. [98]

    InData Engineering in Medical Imaging – DEMI 2025, 1–10, DOI: 10.1007/978-3-032-08009-7_1 (Springer Nature Switzerland, 2026)

    Koju, S.et al.Surgical vision world model. InData Engineering in Medical Imaging – DEMI 2025, 1–10, DOI: 10.1007/978-3-032-08009-7_1 (Springer Nature Switzerland, 2026). 2503.02904

    arXiv 2025