acceptodds
Under review as a conference paper at ICLR 2027

Toward World Modeling of Physiological Signal Leveraging Latent Dynamics

Abstract

Physiological time series signals reflect complex, multi-scale dynamical processes of the human body. Existing modeling studies focus on static tasks such as classification, event forecasting, or short-horizon next step prediction, while long-horizon signal-level forecasting and predictive nature of physiological signals remain underexplored. We introduce **NormWear-2**, a world model that encodes both multivariate physiological signals and clinical intervention variables into a shared latent space and models their joint temporal evolution as a dynamical system. Our approach combines inference from prior pre-trained knowledge (*intuition*) with instant non-parametric latent state transition adaptation (*insight*), enabling coherent forecasting across multiple temporal scales, conditioned on heterogeneous clinical interventions. During the pretraining phase, we find that chaos-theoretic balancing of dynamical regime diversity yields more robust representations, with a smaller balanced corpus outperforming one twice its size and capturing bifurcation regimes. We evaluate the world model performance across diverse real-world physiological datasets spanning heterogeneous temporal resolutions and intervention regimes, covering daily life, point-of-care, and clinical settings, including fitness planning, hemodialysis, diabetes management, and surgical monitoring. These evaluation datasets comprise records from 8,026 subjects, spanning study durations from 3.2 hours for high-resolution signal data to 2.3 years for longitudinal clinical biomarker tracking. NormWear-2 achieves the best overall forecasting performance across time, frequency, and latent representation domains, with significant improvements over state-of-the-art time series foundation models, while maintaining competitive downstream representation quality, providing a step toward general-purpose world models for physiological signals.

Primary area:
learning on time series and dynamical systems
Keywords:
AI for Health, Machine Learning, Signal processing, World Model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ajhZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces NormWear-2 for physiological time-series forecasting conditioned on intervention variables. It combines a pretrained encoder-decoder pathway, termed “intuition,” with inference-time clustering and empirical latent-state transitions, termed “insight.” The authors also investigate chaos-aware balancing of pretraining data. Evaluation covers five physiological datasets using time-domain, frequency-domain, and latent-space metrics, supplemented by data-selection and backbone ablations, downstream representation tests, and an exploratory offline planning experiment.

Critical strengths
  1. A practical combination of pretrained representations and context-specific dynamics. The insight module constructs non-parametric latent transitions from observed context without updating the pretrained network. This provides a concrete mechanism for incorporating subject-specific history rather than relying entirely on fixed pretrained predictions. The evaluation across five physiological datasets with different temporal resolutions and intervention variables makes this design relevant beyond a single application setting.

  2. Supporting evidence for improved structural generation quality. The full model achieves lower frequency-domain and latent-domain FID than the evaluated alternatives in Table 12. The transition perturbations in Table 13 also degrade aggregate performance across all five datasets. These results provide useful evidence that the transition mechanism contributes to generation quality, beyond the headline composite score alone.

Critical weaknesses
  1. The practical meaning of the overall performance advantage is insufficiently established. In Table 2, NormWear-2 has higher MAE than the Naive baseline on all five datasets, despite its favorable composite scores. Structural and distributional improvements are valuable, but the paper does not establish why they compensate for this loss of time-aligned accuracy. Sensitivity to the metric weights and latent-space evaluator, together with an application-relevant endpoint such as event timing, would help determine whether the reported overall advantage reflects practically better forecasts.

  2. Observed-action prediction is not sufficient counterfactual validation. Section 4.3 is titled “Counterfactual Validation,” but its described evidence concerns prediction accuracy under observed clinical actions. Accuracy across logged action levels does not establish the consequences of alternative interventions from the same patient history. This distinction matters because the paper explicitly notes that clinicians adjust actions in response to patient condition. The authors should clarify the assumptions and evidence supporting an intervention interpretation, or restrict the conclusions to observational action-conditioned forecasting.

  3. The benefits of chaos-aware balancing are task-dependent. In the fixed-budget, same-source ablation in Table 11, random sampling outperforms balancing on both short forecasting (MAE 0.558 versus 0.597) and long forecasting (0.696 versus 0.729). Balancing improves simulation and pooled results, but this does not establish a general forecasting advantage. Although acknowledged in Appendix D.2, this distinction should be more prominent in the main conclusions, with clearer evidence connecting the balancing strategy to the physiological forecasting improvements.

Questions for authors

See the weaknesses .Clarifying these issues and providing targeted supporting analyses maybe could improve my assessment.

2 comments
AuthorsOpenReview ↗

Response to Reviewer ajhZ [1/2]

Concern 1. Evaluation robustness

"... NormWear-2 has higher MAE than the Naive baseline on all five datasets ... Structural and distributional improvements are valuable, but the paper does not establish why they compensate for this loss of time-aligned accuracy. Sensitivity to the metric weights and latent-space evaluator, together with an application-relevant endpoint such as event timing, would help determine whether the reported overall advantage reflects practically better forecasts"

Response

We’d like to thank the reviewer for recognizing the value of our evaluation scheme and for suggesting these additional checks. Before moving on to the additional experimental results as suggested by the reviewer, we’d like to first make a clarification on the performance scoring. The final score, which is the Equation 3 in the paper, normalizes metrics by the evaluated sequence and representation sizes. They are not tuned scaling factors or weights. Soft-DTW is normalized by its evaluated sequence length, frequency-domain Euclidean distance by half the prediction length, and latent Euclidean distance by the embedding dimension. The final score is the arithmetic mean of these six normalized losses. Thus, there are no manually tuned scaling and weighting for the final score. We have clarified the normalization factors and equal averaging in Section 3.6 alongside Equation 3.

Following the reviewer’s suggestion, we examined sensitivity to the evaluation encoder by replacing the encoder used for the two latent metrics with frozen Chronos-2, scoring the same predictions with unchanged normalization. The table below present final scores, with bold and italics mark the best and second-best. We observe that NormWear-2 still achieves leading overall performance under this alternative latent evaluator.

Response Table 1: Final scores with the Chronos-2 evaluator

ModelVitalDBPMDataCGMacrosShanghaiKidneyDialysis
Naive0.7960.6950.7600.8700.832
Seasonal Naive0.7030.6760.8410.6700.727
Sundial0.6140.5420.5710.6940.649
Panda0.6260.6050.6120.6320.595
TiReX0.4720.5520.5710.6460.598
Chronos-20.4910.5350.5380.6940.588
NormWear-20.4680.4880.4700.6010.583
Intuition only0.5420.5810.5380.6530.674
Insight only0.4770.5000.4930.6340.584

The table below shows the average normalized latent score following same normalization of the terms in Equation 3.

Response Table 2: Average normalized latent scores

EvaluatorNaiveSeasonal NaiveSundialPandaTiReXChronos-2NormWear-2Intuition onlyInsight only
NormWear0.3180.2070.2560.2370.2150.2080.1630.2190.157
Chronos-20.4840.3720.2310.2630.2510.2280.2060.2940.209

We also examined whether this advantage persists after completely removing both latent metrics and averaging the four remaining performance scores. The table below shows the resulting average scores recalculated from Table 2 in the original manuscript. We observe that NormWear-2 is being competitive on VitalDB, but maintain the leading overall performance.

Response Table 3: Final scores without latent metrics

ModelVitalDBPMDataCGMacrosShanghaiKidneyDialysis
Naive0.9520.7500.8911.0710.997
Seasonal Naive0.9100.8381.0310.8430.910
Sundial0.7990.6650.7430.9320.881
Panda0.8330.7090.7570.8440.799
TiReX0.6110.6580.7280.8300.797
Chronos-20.6640.6780.6940.8990.789
NormWear-20.6240.6040.6230.7980.770
Insight only0.6370.6450.6540.8440.772

The advantage largely persists when changing the evaluation encoder or removing latent metrics, suggesting that the metrics capture complementary aspects of forecast quality. MAE measures pointwise error but does not establish preservation of fluctuations across long continuous recordings. Such temporal structure can carry meaningful physiological information, such as spectral power and coherence [1], beat-to-beat variability [2], and temporal complexity of wearable-derived affective trajectories [3] as the supporting literature. We are encouraged that we share agreement with reviewer that the considering complementary metrics are valuable, and that the suggested additional analysis strengthen the observation. We have made this rationale explicit in Section 3.6, clarified the interpretation of the observed metric trade-off in Section 4.2, and added the evaluator comparisons to Appendix D.10 (Tables 20–21).

AuthorsOpenReview ↗

Response to Reviewer ajhZ [2/2]

Concern 2. Observational action-conditioned forecasting

"Observed-action prediction is not sufficient counterfactual validation. Section 4.3 is titled “Counterfactual Validation,” but its described evidence concerns prediction accuracy under observed clinical actions. Accuracy across logged action levels does not establish the consequences of alternative interventions from the same patient history. This distinction matters because the paper explicitly notes that clinicians adjust actions in response to patient conditions. The authors should clarify the assumptions and evidence supporting an intervention interpretation, or restrict the conclusions to observational action-conditioned forecasting."

Response

We agree with the reviewer on the more proper interpretation of the presented results. Section 4.3 evaluates predictions conditioned on recorded interventions. Its error maps summarize accuracy across observed combinations of prior physiological state and action. Because clinicians adjust treatments in response to patient condition, these comparisons describe observational action-conditioned forecasting; patient-specific effects of alternative treatments require separate causal assumptions and validation. Action-Conditioned Prediction Validation would a title that more precisely describe this analysis, and we have updated it as the title for Section 4.3.

Concern 3. Task-dependent benefits of pretraining-data balancing

"The benefits of chaos-aware balancing are task-dependent. In the fixed-budget, same-source ablation in Table 11, random sampling outperforms balancing on both short forecasting (MAE 0.558 versus 0.597) and long forecasting (0.696 versus 0.729). Balancing improves simulation and pooled results, but this does not establish a general forecasting advantage. Although acknowledged in Appendix D.2, this distinction should be more prominent in the main conclusions, with clearer evidence connecting the balancing strategy to the physiological forecasting improvements."

Response

Thank you for highlighting this trade-off. In the fixed-budget, same alternation comparison (Table 11), balancing yields modestly higher forecasting error than random sampling, alongside substantially lower simulation error and better pooled performance. We selected balanced pretraining for its stronger performance on simulation, which predicts unobserved channels conditioned on supplied channels. This setting is particularly relevant to our interest in subsequent action-conditioned physiological modeling. We have made this trade-off and the motivation for selecting balanced pretraining explicit in Section 4.1 (Preliminary Exploration on Chaos-Balance-Aware Pretraining), with a reference to Table 11. We have also added the physiological forecasting comparison with standard deviations to Section 4.1 (Table 3).

To better connect this pretraining study to physiological forecasting, we evaluated the concatenation and random-sampling pretraining checkpoints on all the physiological datasets leveraged in this study. Final performance scores are presented below.

Response Table 4: Physiological forecasting by pretraining strategy

PretrainingVitalDBPMDataCGMacrosShanghaiKidneyDialysisAverage
Concatenation0.5560.6090.6580.5730.5580.591
Random sampling0.5200.5300.5590.5750.5560.548
Balanced (Table 2)0.4570.4660.4740.5780.5750.510

Across the five datasets, the balanced model achieves the lowest average score, with clear gains on VitalDB, PMData, and CGMacros and comparable scores on Shanghai and KidneyDialysis. We’d like to thank the reviewer again for pointing this out, as having this clearer observation as part of the conclusion for the pretraining experimental stage provides a better bridge toward the successive experiments.

In Closing

We sincerely thank the reviewer once again for the thoughtful and constructive feedback. If there are aspects where our work could be further improved, we would greatly appreciate any additional suggestions you might have.

Reference

[1] Garland et al., Mindfulness-induced endogenous theta stimulation occasions self-transcendence and inhibits addictive behavior, Science Advances, 2022.

[2] Garland et al., Mindfulness-oriented recovery enhancement reduces opioid dose in primary care by strengthening autonomic regulation during meditation, American Psychologist, 2020.

[3] Luo et al., Personalized entropy-informed deep learning for identifying opioid misuse, Nature Mental Health, 2026.

Reviewer_KLHBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces NormWear-2, a world model for physiological signals with a shared latent space and models their joint temporal evolution as a dynamical system. It combines masked reconstruction pretraining with inference-time clustering and empirical latent-state transitions for physiological forecasting. It also introduces dynamics-aware pretraining-data balancing and conditions predictions on intervention channels. The framework is promising and evaluated across diverse datasets, and experimental results show that NormWear-2 outperforms baseline models on average scores.

Critical strengths
  • The paper addresses an important problem across heterogeneous physiological signals, temporal resolutions, and intervention types.

  • Combining pretrained representations with context-specific, non-parametric transition adaptation is a coherent design.

  • The evaluation is broad and includes alternative backbones, transition perturbations, context-length studies, downstream probing, and distribution-level metrics. The fixed-source, fixed-budget data-balancing comparison is particularly useful.

Critical weaknesses
  • Overall superiority depends on the composite evaluation despite consistently worse pointwise accuracy. In Table 2, NormWear-2 has higher MAE than both TiReX and Chronos-2 on all five datasets. Its advantages therefore reflect a trade-off between pointwise accuracy and structural, spectral, or representation-level similarity. These complementary metrics are valuable, but the headline conclusion depends on the scaling and weighting in Equation 3. The manuscript should establish ranking robustness to alternative weights, removal of latent-space terms, and an independently specified evaluation encoder. Patient-level uncertainty estimates would also strengthen the comparison.

  • Counterfactual and planning claims exceed the supporting evidence. Section 4.3 primarily evaluates prediction under observed interventions. This does not establish outcomes under alternative treatments for the same patient. More importantly, Appendix D.7 states that the policy evaluator is fitted on the entire dataset, including test data, and regresses Q directly on logged Monte Carlo returns rather than performing target-policy Bellman evaluation. As described, these targets incorporate logged continuation actions and do not establish the long-term value of the proposed MPC policy. The authors should separate conditional prediction from counterfactual validity and reassess planning benefits using an independent, target-policy-relevant evaluation protocol.

  • Component benefits are less consistent than claimed, and key implementation details remain unclear. Table 11 shows that random sampling outperforms dynamics-aware balancing on both short- and long-horizon forecasting, although balancing improves simulation and aggregate performance. Table 3 also contains regressions after adding insight, including JEPA on PMData and CGMacros. These results do not support beneficial balancing or backbone-independent gains. The intuition–insight fusion rule, handling of sparse or unseen transitions, and future intervention inputs available during evaluation also require clarification. Matched quantitative comparisons with simpler latent retrieval or transition models would better isolate the contribution.

  • The world-model claims are insufficiently validated. Prediction under observed interventions does not establish reliable rollouts. Although the appendix includes MPC, its evaluator is fitted using test data and regresses on logged returns, which do not directly establish the proposed policy’s continuation value. The evidence supports conditional generation more clearly than reliable intervention simulation or planning utility.

Questions for authors
  • Why are Lyapunov-based features used only for pretraining-data selection, rather than to condition or regularize latent transitions? More importantly, does balancing these features in the training corpus help the model preserve corresponding dynamical properties in generated trajectories? Please clarify this design choice and provide evidence connecting data-level balancing to learned dynamics.

  • Do the rankings persist under alternative metric weights and without latent-space terms? Which encoder computes the latent metrics?

  • How does regression on logged returns estimate MPC continuation value, and do the planning gains survive independent or cross-fitted evaluation?

  • Can matched controls isolate latent-transition adaptation from backbone and pretraining effects? Please specify the fusion rule, unseen-transition handling, and future intervention inputs available at inference.

6 comments
AuthorsOpenReview ↗

Response to Reviewer KLHB [1/4]

Concern 1. Evaluation robustness

"... NormWear-2 has higher MAE ... on all five datasets. Its advantages therefore reflect a trade-off between pointwise accuracy and structural, spectral, or representation-level similarity. These complementary metrics are valuable, but the headline conclusion depends on the scaling and weighting in Equation 3..."

"Do the rankings persist under alternative metric weights and without latent-space terms? Which encoder computes the latent metrics?"

Response

We thank the reviewer for positioning the value of our evaluation scheme using complementary metrics. Before moving on to the additional experimental results as suggested by the reviewer, we’d like to first make a clarification on the performance scoring. The final score, which is the Equation 3 in the paper, normalizes metrics by the evaluated sequence and representation sizes. They are not tuned scaling factors or weights. Soft-DTW is normalized by its evaluated sequence length, frequency-domain Euclidean distance by half the prediction length, and latent Euclidean distance by the embedding dimension. The final score equally averaged across all metrics. Thus, there are no manually tuned scaling and weighting for the final score.

Following the reviewer’s suggestion, we examined sensitivity to the evaluation encoder by replacing the encoder used for the two latent metrics with frozen Chronos-2, scoring the same predictions with unchanged normalization. The original latent encoder was a model pretrained using backbone of NormWear [1]. The table below presents final scores with bold and italics mark the best and second-best. We observe that NormWear-2 still achieves leading overall performance under this alternative latent evaluator.

Response Table 1: Final scores with the Chronos-2 evaluator

ModelVitalDBPMDataCGMacrosShanghaiKidneyDialysisAverage
Naive0.7960.6950.7600.8700.8320.791
Seasonal Naive0.7030.6760.8410.6700.7270.723
Sundial0.6140.5420.5710.6940.6490.614
Panda0.6260.6050.6120.6320.5950.614
TiReX0.4720.5520.5710.6460.5980.568
Chronos-20.4910.5350.5380.6940.5880.569
NormWear-20.4680.4880.4700.6010.5830.522
Intuition only0.5420.5810.5380.6530.6740.598
Insight only0.4770.5000.4930.6340.5840.538

The table below shows the average normalized latent score.

Response Table 2: Average normalized latent scores

EvaluatorNaiveSeasonal NaiveSundialPandaTiReXChronos-2NormWear-2Intuition onlyInsight only
NormWear0.3180.2070.2560.2370.2150.2080.1630.2190.157
Chronos-20.4840.3720.2310.2630.2510.2280.2060.2940.209

We also examined whether this advantage persists after completely removing both latent metrics and averaging the four remaining performance scores. The table below shows the resulting average scores recalculated from Table 2 in the original manuscript. We observe that NormWear-2 is being competitive on VitalDB, but maintain the leading overall performance.

Response Table 3: Final scores without latent metrics

ModelVitalDBPMDataCGMacrosShanghaiKidneyDialysisAverage
Naive0.9520.7500.8911.0710.9970.932
Seasonal Naive0.9100.8381.0310.8430.9100.906
Sundial0.7990.6650.7430.9320.8810.804
Panda0.8330.7090.7570.8440.7990.788
TiReX0.6110.6580.7280.8300.7970.725
Chronos-20.6640.6780.6940.8990.7890.745
NormWear-20.6240.6040.6230.7980.7700.684
Insight only0.6370.6450.6540.8440.7720.711

The advantage largely persists when changing the evaluation encoder or removing latent metrics, suggesting that the metrics capture complementary aspects of forecast quality. Unlike MAE, complementary metrics can assess temporal structure, such as spectral power and coherence [2], beat-to-beat variability [3], and temporal complexity of wearable-derived affective trajectories [4] as the supporting literature. We are encouraged that we share agreement with reviewer that the considering complementary metrics are valuable, and that the suggested additional analysis strengthen the observation.

We have updated the above rationale behind the performance scoring and missing information of the latent encoder in Section 3.6, clarified the interpretation of the observed metric trade-off in Section 4.2, and added the evaluator comparisons to Appendix D.10 (Tables 20–21).

AuthorsOpenReview ↗

Response to Reviewer KLHB [2/4]

Concern 2. Action-conditioned prediction and planning validity

"Counterfactual and planning claims exceed the supporting evidence. Section 4.3 primarily evaluates prediction under observed interventions. This does not establish outcomes under alternative treatments for the same patient..."

"The world-model claims are insufficiently validated ... Although the appendix includes MPC, its evaluator is fitted using test data and regresses on logged returns, which do not directly establish the proposed policy’s continuation value. The evidence supports conditional generation more clearly than reliable intervention simulation or planning utility."

"How does regression on logged returns estimate MPC continuation value, and do the planning gains survive independent or cross-fitted evaluation?"

Response

We agree with the reviewer on the more proper interpretation of the presented results. Section 4.3 evaluates predictions conditioned on recorded interventions. Its error maps summarize accuracy across observed combinations of prior physiological state and action. Because clinicians adjust treatments in response to patient condition, these comparisons describe observational action-conditioned forecasting; patient-specific effects of alternative treatments require separate causal assumptions and validation. Action-Conditioned Prediction Validation would a title that more precisely describe this analysis, and we have updated it as the title for Section 4.3.

Regarding the issue on exploratory planning and evaluator interpretation, we appreciate the reviewer’s careful attention on the extensional planning experiment in Appendix D.7 of the original manuscript. The current evaluation indeed does not establish the long-term value of the proposed MPC policy. De facto, our original intention for the scope of designing this experiment is an extensional and preliminary exploration for the natural next research stage following the current study. We will discuss its relationship to the main contribution of the paper below.

Our use of the term “world model” follows the standard notion of modeling environment dynamics or state transitions [5, 6]. A world model can serve as the transition or simulation component of a closed-loop decision-making system, but learning and validating such dynamics is distinct from learning a policy or establishing its value. The latter additionally requires objective or reward specification, planning, and policy evaluation. Our primary contribution focuses on action-conditioned state-transition modeling. Appendix D.7 of the original manuscript was included as a preliminary exploration of how such a model might eventually support planning and is not used to substantiate the main claims of the paper. Intervention planning and its associated limitations were discussed in Section 5.

Regarding the reviewer’s question on the MPC evaluator, standard FQE [7] performs Bellman-style target-policy evaluation and assumes that the chosen state representation is sufficiently Markov, such that the relevant future transition and reward distribution can be characterized from the current state and action. This assumption is ambiguous to establish for the available real-world physiological trajectories. We therefore regressed the evaluator directly on the observed long-term returns from the logged trajectories and fitted it using all available data, including the test set. The resulting score is indeed not an estimate of the MPC policy’s continuation value, since the observed returns incorporate the logged continuation actions. Instead, it provides an initial indication of whether MPC with the learned world model tends to select actions that are empirically associated with relatively higher long-term returns in the available data.

As discussed above, reliable evaluation of an agentic policy from a static observational dataset is itself an ongoing research problem, and different evaluation assumptions or a more advanced planning or RL module may lead to different observations. Addressing these questions rigorously would constitute a substantial next stage of research beyond the scope of the present work. Since Appendix D.7 of the original manuscript was included as a preliminary exploration of this future direction and is indeed not part of the original evidence we leveraged in the discussion of the main content in the paper, we have removed this section to prevent it from creating ambiguity about the scope of our claims. Its removal does not alter any experiment, result, or conclusion presented in the main paper.

We hope the above clarification and discussion could better distinguishes the core contribution of the present work from this exploratory planning analysis for future research.

AuthorsOpenReview ↗

Response to Reviewer KLHB [3/4]

Concern 3. Component benefits

"... Table 11 shows that random sampling outperforms dynamics-aware balancing on both short- and long-horizon forecasting, although balancing improves simulation and aggregate performance. Table 3 also contains regressions after adding insight, including JEPA on PMData and CGMacros. These results do not support beneficial balancing or backbone-independent gains ... quantitative comparisons with simpler latent retrieval or transition models would better isolate the contribution."

"Can matched controls isolate latent-transition adaptation from backbone and pretraining effects? Please specify the fusion rule, unseen-transition handling, and future intervention inputs available at inference."

Response

Because the results the reviewer mentioned address different stages of the study, in the response below, we will first make clarification of the role of those design of the experiments, then discuss the additional experimental result as suggested by the reviewer.

Regarding the experimental scope, Table 11 evaluates pretraining-data selection, where balancing improves simulation and aggregate performance. We did acknowledge that there is a tradeoff between having some degradation in pure forecast, and having significant gain in simulation tasks. Since the simulation task is more related to the action related forecasting in the later part of the research, we choose to leverage the balanced pretrain approach for further experiments. Table 4 mainly evaluates the addition of insight within each backbone, with benefits that vary by backbone and dataset.

Regarding the fusion of intuition and insight pathways, they are combined by an element-wise weighted average of the intuition and insight outputs at the reconstructed-signal level. Empirically this is the best-performing configuration, consistent with the two pathways being complementary. In addition, as suggested by the reviewer, we compare the combined output of the intuition and insight pathways with latent retrieval only approach (i.e., fitting Bayesian transition over latent-state clusters’ centroids) within each backbone setting. The intuition-insight combined inference achieves better performance in most cases across the evaluated backbone variations. We have added this comparison to Appendix D.8 (Table 18). Final performance score is presented below. Bold marks the better result in each row.

Response Table 4: Fusion versus latent retrieval

DatasetBackboneFusion (from paper Table 4)Latent retrieval
VitalDBUnivariate0.5280.595
VitalDBCLS0.4680.467
VitalDBJEPA0.5280.642
VitalDBSFT0.4590.464
PMDataUnivariate0.4890.695
PMDataCLS0.4680.474
PMDataJEPA0.4800.623
CGMacrosUnivariate0.5270.749
CGMacrosCLS0.5060.514
CGMacrosJEPA0.5020.685
ShanghaiUnivariate0.5760.601
ShanghaiCLS0.6010.617
ShanghaiJEPA0.5980.618
KidneyDialysisUnivariate0.5590.567
KidneyDialysisCLS0.5690.578
KidneyDialysisJEPA0.5690.812
KidneyDialysisSFT0.5640.564

Regarding comments on unseen future intervention, we agree that evaluating sparsely observed or previously unseen interventions would provide valuable evidence for broader intervention modeling. Establishing reliable prediction in these settings requires dedicated investigation beyond the scope of the present study. To provide additional evidence in this direction, we examine model behavior as the available subject-specific history increases. The consistent improvements in the final performance score across three datasets demonstrate the model’s ability to benefit from accumulating individual observations and support its potential for continual personalization. These results are reported in Appendix D.5 (Table 14), along with brief discussion in section 5.

Response Table 5: Peronslize on VitalDB

Context window (s)7.615.523.431.3
Mean prediction0.8050.7820.7740.769
Seasonal Naive0.8310.8030.7930.661
NormWear-20.5740.5050.4870.482

Response Table 6: Peronslize on PMData

Context window (h)12243648
Mean prediction0.8580.6980.6790.606
Seasonal Naive0.9540.7770.7630.657
NormWear-20.5740.4270.4130.389

Response Table 7: Peronslize on KidneyDialysis

Context window (sessions)5101520
Mean prediction1.1140.8530.7840.752
Seasonal Naive1.1870.9180.8200.665
NormWear-20.9190.6450.5880.562
Reviewer_kZmiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses the problem of physiological time-series generation (i.e. forecasting) through the concept of world models. To that end, the paper (at a very high level) enriches a generic patch-mask-reconstruct-based approach (what the paper coins the "intuition" branch) with a inference algorithm where, simply put, the context to be forecast is encoded and clustered, a state-transition graph is constructed and transition probabilities are estimated, and a patch forecast in this way is then decoded again to time-series. If "action" channels (i.e. interventions) are present, they are also encoded in the transition probabilities. That's just a rough overview, there's actually a bunch more stuff happening to make it work, including characterising time-series as various dynamic systems and synthesising pre-training data to balance the training with respect to those.

Various different baselines are compared using different metrics, both concerning forecasting accuracy and other structural aspects. There are further ablation experiments and evaluation of different backbones suggesting that the proposed modelling approach works across the board.

Critical strengths

I think learning the underlying dynamics of physiology purely from data is an interesting and timely pursuit. I also value the inclusion of what the paper calls "simulation" (though i would call it imputation), i.e. the generation of missing patches of channels given the presence of other channels' patches during the same time window.

Critical weaknesses

Nothing too critical, but the paper is quite dense and a lot of detail is in the appendix.

I also think that for all the modeling, the raw accuracy is not that great, the model doesn't come on top for any of the datasets in terms of MAE.

World model is also a bit of a red herring here - the way i understand it, world models learn the underlying "world" (physiology in this case) from observations alone, i.e. without the inclusion of explicit actions, by modelling them latently. Discretising the latent action space and mapping them to a handful of interpretable ones (e.g. WASD buttons for top/left/bottom/down) then allows you to "control" the world. Here, the actions are explicitly encoded. Basically, the manuscript doesn't do a good job at showing, empirically or conceptually, what (if not raw accuracy) this complex modelling approach gives us over other baselines (e.g. just using the "intuition" arm).

The results have no statistical uncertainty estimates associated with them.

Questions for authors

Please address my concerns above.

Another one: given that some of the underlying physiology of some of the datasets is reasonably well understood (glucose-insulin interaction for example), it seems like a missed opportunity to not analyse what the model has learned about it compared to what we can measure/know using traditional compartment models. Would it be possible to perform such an analysis? That would address my concern of "what's the value of a world model here" in a qualitative manner.

2 comments
AuthorsOpenReview ↗

Response to Reviewer kZmi [1/2]

Concern 1. World modeling and its value

"... world models learn the underlying "world" (physiology in this case) from observations alone, i.e. without the inclusion of explicit actions, by modelling them latently. Discretising the latent action space and mapping them to a handful of interpretable ones (e.g. WASD buttons for top/left/bottom/down) then allows you to "control" the world. Here, the actions are explicitly encoded ... what (if not raw accuracy) this complex modelling approach gives us over other baselines ..."

"... some of the datasets is reasonably well understood (glucose-insulin interaction for example), it seems like a missed opportunity to not analyse what the model has learned about it compared to what we can measure/know using traditional compartment models. Would it be possible to perform such an analysis? That would address my concern of "what's the value of a world model here" in a qualitative manner."

Response

To begin with, we appreciate the reviewer’s emphasis on learning the underlying dynamics from observations and connecting inferred latent actions to interpretable controls. Our formulation actually shares similar attributes, with a bit more focus on learning how the observed system evolves and draws on the complementary action-conditioned perspective, following definition from LeWorldModel [1] and Hu and Shu 2023 [2], where explicit actions condition the state transitions. As described in our manuscript Section 3.1 Problem Setup, NormWear-2 models the dynamics and interactions of the supplied time-series channels, with recorded interventions represented as additional, temporally aligned channels when available. This flexibility allows the same pretrained model to be used directly either for predicting physiological evolution from observations alone, or for generating trajectories conditioned on supplied interventions.

Secondly, we agree that the reviewer's constructive suggestion on having a direct comparison analysis with known physiological compartment models would provide better insight on the values of world model in a qualitative manner. After a literature review, we compared NormWear-2 with an adapted insulin–glucose compartment model [3] for type-1 diabetes and its extension [4] for type-2 diabetes, using continuous glucose monitoring trajectories from 71 ShanghaiDiabetes participants (12 type-1 and 59 type-2) with clean, documented insulin trajectory records. Both methods received identical observed contexts and future insulin inputs, with forecast horizons ranging from 12 to 48 hours depending on the available usable clean trajectory length from each participant’s data. Following the literature, the compartment model’s parameters have to first be estimated from individual historical records before integrating its ordinary differential equations (ODEs). Accordingly, compartment parameters fitting and NormWear-2’s one-pass in-context learning in the insight pathway use the same observed context for a fair comparison. The results are presented in the table below. Quantitatively, we observe that NormWear-2 performs better across all six metrics.

Response Table 1: Insulin-conditioned glucose forecasting

ModelTime MAE ↓Time Soft-DTW ↓Frequency Cosine ↑Frequency L2 ↓Latent Cosine ↑Latent L2 ↓Final Score ↓
Compartment1.268 ± 0.966225.472 ± 377.4180.695 ± 0.192132.519 ± 92.2980.828 ± 0.079159.316 ± 41.3150.987 ± 1.243
NormWear-2 Fusion1.083 ± 0.31970.251 ± 78.3440.746 ± 0.14593.504 ± 48.8510.903 ± 0.042113.645 ± 30.2460.559 ± 0.322

Qualitatively, NormWear-2 tends to retain local rises and falls in glucose, while the compartment model gives a smoother account of its broad evolution. We have added a few example visualization as Figure 4 in Appendix D.9. The smoother compartment forecasts capture broad peaks, such as those near 24 h in (b) and 20 h in (c) from the forecast start. NormWear-2 often follows more of the detail around these trends, staying closer to the lower glucose levels before the peak as shown in (b) and (d). It also remains closer to observed glucose in (a), where the compartment forecast follows the initial decline but subsequently drifts upward, which may be due to the chaotic nature of the ODEs based forecast that is sensitive to initial state.

These qualitative observations also motivate our complementary evaluation of pointwise and shape agreement in the time domain, spectral agreement in the frequency domain, and learned-feature similarity in the latent domain. The insulin–glucose comparison makes these differences more interpretable. We thank the reviewer again for this helpful suggestion. We have added this comparison to Appendix D.9 (Table 19 and Figure 4).

AuthorsOpenReview ↗

Response to Reviewer kZmi [2/2]

Concern 2. Statistical uncertainty

"The results have no statistical uncertainty estimates associated with them."

Response: We thank the reviewer for pointing out this important missingness. We updated the final-score following equation 3 in the manuscript, with its averages and sample SDs as presented in tables below. Bold and italics mark the best and second-best performance. After running statistical tests across the performance distributions to assess the significance of performance differences, we observe that NormWear-2 achieves the best overall performance, with statistically significant differences (p < 0.004 to the best baseline) under Friedman-Conover post hoc tests. Fusion consistently improves performance across backbone variants, yielding statistically significant overall gains (p < 0.001) under paired Wilcoxon tests. These tables are also added in Appendix D.7 (Tables 16–17).

Response Table 2: Baseline comparison

ModelVitalDBPMDataCGMacrosShanghaiKidneyDialysisAverage
Naive0.769 ± 0.0790.606 ± 0.2540.709 ± 0.1640.801 ± 0.3690.752 ± 0.1770.727 ± 0.209
Seasonal Naive0.661 ± 0.1800.657 ± 0.3170.778 ± 0.2870.611 ± 0.3490.665 ± 0.1910.674 ± 0.265
Sundial0.633 ± 0.0690.527 ± 0.2270.590 ± 0.1550.693 ± 0.3420.662 ± 0.1710.621 ± 0.193
Panda0.652 ± 0.0700.558 ± 0.2200.594 ± 0.1340.618 ± 0.3200.600 ± 0.1570.604 ± 0.180
TiReX0.465 ± 0.0930.523 ± 0.2670.571 ± 0.1670.617 ± 0.3210.600 ± 0.1660.555 ± 0.203
Chronos-20.500 ± 6.4160.541 ± 0.2780.548 ± 0.1830.657 ± 0.5700.589 ± 0.1830.567 ± 1.526
NormWear-20.457 ± 0.0710.466 ± 0.1800.474 ± 0.1320.578 ± 0.3040.575 ± 0.1600.510 ± 0.169
Insight only0.461 ± 0.0760.494 ± 0.1950.492 ± 0.1340.608 ± 0.3000.574 ± 0.1590.526 ± 0.173

Response Table 3: Intuition–insight gain across backbones

Backbone / pathwayVitalDBPMDataCGMacrosShanghaiKidneyDialysisAverage
Univariate Intuition0.569 ± 0.0950.527 ± 0.2080.568 ± 0.1150.654 ± 0.3040.630 ± 0.1550.590 ± 0.176
Univariate Fusion0.528 ± 0.0720.489 ± 0.1890.527 ± 0.1060.576 ± 0.3030.559 ± 0.1520.536 ± 0.164
CLS-Attn Intuition0.640 ± 7.4210.611 ± 0.1970.653 ± 0.1190.654 ± 0.3290.623 ± 0.1580.636 ± 1.645
CLS-Attn Fusion0.468 ± 0.3320.468 ± 0.2090.506 ± 0.1370.601 ± 0.3170.569 ± 0.1570.522 ± 0.231
LeWM-JEPA Intuition0.531 ± 0.0600.471 ± 0.2600.487 ± 0.1180.663 ± 0.3330.581 ± 0.1600.547 ± 0.186
LeWM-JEPA Fusion0.528 ± 0.0580.480 ± 0.2300.502 ± 0.1090.598 ± 0.3120.569 ± 0.1510.535 ± 0.172

In Closing

We thank the reviewer again for the constructive feedback. If you feel there are areas where our work can be further improved, we would greatly value your additional feedback.

Reference

[1] Maes et al., Leworldmodel: Stable end-to-end joint-embedding predictive architecture from pixels, arXiv, 2026.

[2] Hu and Shu, Language models, agent models, and world models: The law for machine reasoning and plannin, arXiv, 2023.

[3] Kanderian et al., Identification of intraday metabolic profiles during closed-loop glucose control in individuals with type 1 diabetes, Journal of Diabetes Science and Technology, 2009.

[4] Aradóttir et al., Modelling of fasting glucose-insulin dynamics from sparse data, IEEE EMBC, 2018.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear Reviewers,

We sincerely thank the reviewers for their thoughtful and constructive feedback. Your positive comments are deeply encouraging and reflect the goals we aimed to achieve in this work, such as:

  • The importance and timeliness of learning physiological dynamics across diverse signals and settings (Reviewers kZmi and KLHB).
  • The combination of pretrained representations with context-specific transitions to incorporate individual history (Reviewers KLHB and ajhZ).
  • The breadth of the evaluation and the value of the controlled pretraining-data comparison (Reviewer KLHB).
  • The evidence that the transition mechanism contributes to improved structural generation quality (Reviewer ajhZ).

In response to the main concerns, we conducted additional experiments and analyses and revised the manuscript, with changes highlighted in blue. The core editions are summarized below:

  • To address the concerns about score scaling and weighting (Reviewers KLHB and ajhZ), we clarified that the denominator and the coefficient are part of the metric normalization across evaluated horizon and equal averaging of the metrics, where no manually tuned scaling or weighting are designed or presented.
  • In response to questions about the world-model definition and the scope of our claims (Reviewers kZmi, KLHB, and ajhZ), we clarified the action-conditioned dynamics perspective we follow, with literature support. We also renamed Section 4.3 “Action-Conditioned Prediction Validation” to better reflect the objective and takeaways from the corresponding analysis.
  • Following Reviewer kZmi’s suggestion, we compare our proposed world-model with an insulin–glucose compartment model to examine forecast quality in a well-understood physiological setting. Using the same observed contexts and future insulin inputs, NormWear-2 performs better both quantitatively and qualitatively. The corresponding analysis description and result observation are also updated in the manuscript.
  • To address concerns about dependence on the latent evaluator (Reviewers KLHB and ajhZ), we alter the evaluation scheme under two configuration: (a) frozen Chronos-2 as an alternative encoder, and (b) with both latent metrics being removed. The result performance ranking pattern stay consistent as those core experimental results we have in the original setting.
  • To better connect pretraining-data balancing to physiological forecasting (Reviewers KLHB and ajhZ), we compared balanced pretraining, which is the final version of NormWear-2, with concatenation and random sampling from the pretrain corpus directly on the five digital health datasets that are mainly leveraged in this study. Balanced pretraining achieves better overall performance, with leading score on three datasets, and competitive score on the rest two datasets.
  • We have added the sample standard deviations and statistical comparisons to show the variability and evidence for overall performance differences in the baseline and intuition–insight results, for better statistical rigor of the presented results.

We hope these clarifications provide additional context for understanding the structure of the paper and the core contributions of the work. We would greatly appreciate any additional feedback and discussion on areas where our work could be further improved. Thank you all again for the constructive feedback and for engaging with the review process.

Best regards,

The Authors

AuthorsOpenReview ↗

Final Response

Dear Reviewers and Area Chair,

We sincerely thank the reviewers for their thoughtful feedback and recognition of the importance of learning physiological dynamics across diverse signals and settings. We appreciate their positive assessment of the study’s broad application coverage and methodological depth, including the exploration of pretraining data composition, the evaluation framework, and the proposed inference pathways. We are encouraged that the reviewers consistently viewed the soundness concerns as minor or fixable, and we have conducted additional analysis and revisions carefully following the reviewers’ suggestions.

Through the rebuttal period, we are glad to see that reviewers acknowledge that several core concerns are being addressed, including:

  • Clarifying that the normalization term in the performance score is not manual tuned scaling or weighting (reviewer KLHB);
  • Justifying the robustness of the designed performance evaluation scheme (reviewer KLHB);
  • Demonstrating the detailed benefit of the components in the proposed framework (reviewer KLHB);
  • Indicating the clearer connections between the presented balanced pretraining and the benefits of physiological signal forecasting (reviewer KLHB);

In direct response to the remaining reviews, we compared our proposed method with an insulin-glucose compartment model under matched conditions, strengthening the value of the world model both in a quantitative and qualitative manner (reviewer kZmi). Potential support to future research on planning modeling are demonstrated through grid performance analysis of next state prediction at the same given state but different value of action, as well as the personalization experiments with online deployment settings (reviewer KLHB). We additionally provided sample standard deviations and statistical comparisons (reviewer kZmi), and refined the world-model framing and scope of our claims (reviewer ajhZ). All the revisions are incorporated into the updated manuscript and highlighted in blue.

We understand that reviewers are balancing many commitments and may have limited time for further discussion at this stage. As the rebuttal period concludes, we hope the additional evidence and clarifications provide a helpful basis for considering whether the concerns have been addressed. We would be grateful for their consideration in the final assessment. Thank you all again for your time and for engaging with the review process.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.