NormWear-2: Chaos-Aware Latent World Model

• NormWear-2 is a latent world model for physiological time series that uses chaos-aware pretraining and action-conditioned latent transitions.
• It employs a channel-aware masked autoencoder to jointly encode multivariate signals and clinical interventions for long-horizon forecasting.
• The approach uses chaos-theoretic balancing to diversify dynamical regimes and enhances simulation by adapting context-specific latent transitions.

Searching arXiv for the specified NormWear and related papers to ground the article in current sources. arxiv_search query="NormWear-2 physiological signals latent dynamics chaos-theoretic balancing" max_results=5 NormWear-2 is a latent world model for physiological time series that combines a chaos-aware pretraining strategy with a latent dynamical system built on top of a masked-autoencoder backbone. It encodes multivariate physiological signals and clinical intervention variables into a shared latent space and models their joint temporal evolution as a dynamical system, enabling conditional forecasting across multiple temporal scales in daily life, point-of-care, and clinical settings (Luo et al., 14 May 2026). It is directly related to NormWear, a foundation model for multivariate wearable sensing that emphasized generalizable representations across heterogeneous sensor configurations and downstream transfer across health applications rather than explicit world modeling (Luo et al., 2024).

1. Conceptual scope and problem formulation

NormWear-2 addresses a gap in physiological machine learning in which most prior work emphasizes classification, short-horizon next-step prediction, event prediction, or representation learning, while long-horizon signal-level forecasting and the predictive role of interventions remain underexplored (Luo et al., 14 May 2026). In this formulation, physiological trajectories are treated as the evolution of a partially observed dynamical system, with multivariate signals $\mathbf{x}_t$ and interventions $\mathbf{u}_t$ jointly determining future behavior. The model is therefore framed as conditional forecasting,

$$p_\theta(\mathbf{x}_{t+1:t+H} \mid \mathbf{x}_{\leq t}, \mathbf{u}_{\leq t+H}),$$

rather than as static encoding or purely discriminative inference.

This design shifts the target of learning from task-specific representation quality alone to simulation-capable forecasting. In the reported formulation, interventions such as drug dosages, dialysis control parameters, meals, and lifestyle activities are not auxiliary metadata; they are incorporated as action-like drivers of temporal evolution. This allows NormWear-2 to support forecasting and simulation under heterogeneous intervention regimes spanning fitness planning, hemodialysis, diabetes management, and surgical monitoring.

Relative to the original NormWear, the distinction is structural. NormWear was introduced as a multimodal and ubiquitous foundation model for wearable sensing that accepts arbitrary combinations of channels and modalities and supports zero-shot, partial-shot, and full-shot adaptation across 18 downstream applications (Luo et al., 2024). NormWear-2 retains the channel-aware masked-autoencoding lineage, but reorients the system toward world modeling: the objective is no longer only to extract transferable embeddings, but to construct an internal latent state space whose transitions can be rolled out under intervention conditioning.

2. Backbone architecture, masking, and shared latent space

The backbone of NormWear-2 is a channel-aware masked autoencoder derived from NormWear’s encoder design (Luo et al., 14 May 2026). For an input multivariate time series $X \in \mathbb{R}^{C \times T}$, each channel is patchified into length-$P$ segments, and a Conv1D patch embedding maps each patch to a latent vector. Channel-wise, patch-wise Bernoulli masks are then applied independently, so many channel/time combinations are seen during training. The encoder comprises 12 cross-patches transformer blocks and 6 cross-channels blocks, with hidden size 768, 12 attention heads, and LayerNorm; channel attention has complexity $O(d\,C^2)$, which is more efficient than all-attention over $L \cdot C$.

The decoder uses 2 transformer blocks with hidden size 512 followed by Conv1D deconvolution to patches and a Conv1D consolidation stage. Pretraining follows an MAE-style masked reconstruction objective, with mean squared error over masked patches:

$$\mathcal{L}_{\text{recon}} = \frac{1}{|\mathcal{M}|}\sum_{(c,i)\in \mathcal{M}} \lVert x^{(c)}_i - \hat{x}^{(c)}_i \rVert_2^2.$$

This preserves the original NormWear emphasis on self-supervised reconstruction while adapting it to a broader dynamical setting.

A central architectural property is the treatment of interventions as time-aligned channels in the same multivariate sequence (Luo et al., 14 May 2026). Continuous machine parameters are included directly as additional channels. Event-based interventions such as insulin doses or meals are converted into step-function time series. Lifestyle or semantic factors are encoded via a pretrained clinical LLM into discrete embeddings, then likewise expanded into temporal step functions and included as channels. Physiological signals and interventions are thus processed by the same encoder, producing a joint latent space in which latent trajectories are explicitly action-aware.

This shared latent representation differentiates NormWear-2 from a conventional forecasting foundation model. The model does not merely ingest exogenous covariates; it embeds intervention variables into the same representational substrate as the physiological channels. A plausible implication is that this unified representation is what makes subsequent latent transition estimation compatible with heterogeneous clinical and behavioral actions without changing the backbone.

3. Intuition, insight, and latent dynamical transitions

World modeling in NormWear-2 is not implemented by retraining the backbone for multi-step rollout. Instead, forecasting is constructed on top of the pretrained encoder-decoder through a two-part mechanism described as “intuition” and “insight” (Luo et al., 14 May 2026). “Intuition” denotes the pretrained MAE backbone’s mapping between observed sequences and latent representations. “Insight” denotes a non-parametric, inference-time latent transition model estimated from the specific context window for the current subject or episode.

Latent states are obtained from encoder outputs and then clustered via K-means, with the number of clusters chosen via the elbow rule and roughly scaling with $\log(\text{\#patches})$. Each cluster $j$ is represented by a centroid $\mathbf{u}_t$0 and standard deviation $\mathbf{u}_t$1. From consecutive patch pairs in the context, empirical transition probabilities are estimated:

$\mathbf{u}_t$2

Forecasting then samples future latent states from a mixture of Gaussians over the next-state clusters.

When interventions are explicit, the same framework becomes action-conditioned. In the reported formulation,

$\mathbf{u}_t$3

and given the actual intervention $\mathbf{u}_t$4, rollout samples from the corresponding action-conditioned next-state distribution. In practice, the implementation uses either empirical counts in action-conditioned subsets or nearest-neighbor transitions in a joint state-action embedding space formed by concatenating the latent state and the action vector.

This mechanism makes personalization an inference-time property rather than a parameter-update procedure. The pretrained encoder-decoder supplies a prior latent geometry, while the transition model adapts instantly to the local context. The model variant “NormWear-2 Insight only” isolates this contribution by using only the insight stage for forecasting, and its strong performance in several datasets indicates that much of the system’s forecasting strength derives from context-specific latent transition estimation rather than from end-to-end autoregressive training.

4. Chaos-theoretic balancing and dynamical regime diversity

A defining feature of NormWear-2 is chaos-theoretic balancing during pretraining (Luo et al., 14 May 2026). The motivation is that physiological and related real-world time series can occupy qualitatively different dynamical regimes, including quasi-periodic or limit-cycle behavior, weak chaos, and strong chaos. If pretraining data are dominated by one regime type, representation learning may become brittle outside that regime.

To characterize regime diversity, NormWear-2 computes three families of nonlinear-dynamics features per time series channel: detrended fluctuation analysis exponent $\mathbf{u}_t$5, largest Lyapunov exponent $\mathbf{u}_t$6, and persistent entropy from 0D and 1D homology on Vietoris–Rips persistent homology over delay-embedded trajectories. These features are clustered, and cluster centroids are assigned composite semantic labels using threshold rules such as “Anti-corr,” “Positive-corr,” or “Non-station” for DFA; “Stable,” “Rel Chaos,” or “Rel Very Chaos” for Lyapunov exponent; and low/high connectivity or loop complexity for persistent entropy. Clusters with identical composite labels are then merged to approximate distinct dynamical regimes.

Balance is quantified over the histogram of regime memberships using an entropy-granularity score. An entropy-based balance score is defined as

$\mathbf{u}_t$7

where $\mathbf{u}_t$8 is Shannon entropy and $\mathbf{u}_t$9 measures the number of distinct occupied bins relative to the comparison set. Chaos-balance-aware sampling then iteratively retains samples that increase the balance score until a target subset size is reached.

The empirical results support the use of this criterion (Luo et al., 14 May 2026). At fixed size, more balanced subsets achieve lower generative error in zero-shot forecasting and simulation tasks. A balanced corpus of $$p_\theta(\mathbf{x}_{t+1:t+H} \mid \mathbf{x}_{\leq t}, \mathbf{u}_{\leq t+H}),$$0 samples with balance score 0.73 outperforms a larger unbalanced corpus of $$p_\theta(\mathbf{x}_{t+1:t+H} \mid \mathbf{x}_{\leq t}, \mathbf{u}_{\leq t+H}),$$1 samples with balance score 0.60 across multiple test scenarios. In a Forced Van der Pol oscillator study, the chaos-balanced pretrained model performs well in stable regions and degrades primarily after the onset of chaos while remaining reasonable, whereas the sensing-only model is poor across all regimes. This suggests that pretraining diversity in dynamical structure, rather than sample count alone, is a material determinant of latent forecast quality.

5. Evaluation protocol, datasets, and empirical performance

NormWear-2 is evaluated on five real-world physiological datasets spanning daily life, clinical biomarker tracking, and intraoperative monitoring, with records from 8,026 subjects and study durations ranging from 3.2 hours for high-resolution signals to 2.3 years for longitudinal data (Luo et al., 14 May 2026). Forecasting horizons range from approximately 31.8 seconds in VitalDB to 12 hours in PMData, 2 days in Shanghai Diabetes, and 4 dialysis sessions in KidneyDialysis.

Dataset	Subjects / duration	Signals and interventions
VitalDB	6,388 subjects, ~3.2 hours per subject	ECG, PPG, EEG, respiration; remifentanil target concentration, FiO2, gas flows, ventilation parameters
PMData	16 subjects, 5 months	heart rate, steps, distance, calories; labeled lifestyle sports
CGMacros	45 subjects, 10.9 days	glucose, heart rate, physical motion; detailed meal macronutrients
Shanghai Diabetes	125 subjects, 10.7 days	glucose, heart rate; insulin, hypoglycemic agents
KidneyDialysis	1,452 patients, average 2.3 years	heart rate, blood pressure, body temperature; blood flow rate, dialysate flow, dialysate temperature, ultrafiltration rate

The evaluation is multidimensional rather than confined to pointwise error. Time-domain metrics are MAE and Soft-DTW; frequency-domain metrics are FreqCosSim and FreqEucl; latent-domain metrics are LatentCosSim and LatentEucl. These are aggregated into a composite Final Score:

$$p_\theta(\mathbf{x}_{t+1:t+H} \mid \mathbf{x}_{\leq t}, \mathbf{u}_{\leq t+H}),$$2

Baselines include Naive, Seasonal Naive, Sundial, PandA, Chronos-2, and TiReX, all evaluated in zero-shot fashion.

Across all five datasets, NormWear-2 achieves the best composite Final Score (Luo et al., 14 May 2026). On VitalDB, its Final Score is 0.457, slightly better than NormWear-2 Insight only at 0.461, TiReX at 0.465, and Chronos-2 at 0.500. On PMData, NormWear-2 attains 0.466 versus 0.523 for TiReX and 0.541 for Chronos-2. On CGMacros, it reaches 0.474 versus 0.548 for Chronos-2. On Shanghai Diabetes, it achieves the best Final Score of 0.578. On KidneyDialysis, NormWear-2 Insight only and NormWear-2 reach 0.574 and 0.575, slightly better than Chronos-2 at 0.589. In several datasets, other models obtain the best MAE, but NormWear-2 dominates Soft-DTW, spectral metrics, and latent metrics, which indicates superior preservation of temporal shape, frequency content, and latent representation consistency rather than simple pointwise matching.

The model also maintains competitive downstream representation quality after world-model enhancement (Luo et al., 14 May 2026). In linear probing on the 18 digital-health downstream tasks introduced with NormWear, macro average improves from 84.381 for NormWear to 84.904 for NormWear-2, and micro average improves from 82.762 to 83.776. This is significant because the architecture is not limited to generative forecasting; it preserves the broad transfer behavior that motivated the original NormWear program (Luo et al., 2024).

6. Relation to prior NormWear work, naming distinctions, and limitations

NormWear-2 should be distinguished from two adjacent but different uses of the NormWear name. First, the original NormWear model is a multimodal foundation model for wearable sensing pretrained on physiological signals including PPG, ECG, EEG, GSR, and IMU, with channel-aware attention and a shared liaison [CLS] token for inter-sensor reasoning (Luo et al., 2024). Its core contribution is generalized wearable representation learning across 11 public datasets and 18 applications under zero-shot, partial-shot, and full-shot protocols. NormWear-2 inherits the encoder lineage but moves into action-conditioned world modeling, long-horizon forecasting, and chaos-aware data curation.

Second, the smartwatch gesture benchmark OpenWatch evaluates NormWear as a representative smartwatch foundation model, but it does not define a model named “NormWear-2” (Bonazzi et al., 6 May 2026). The variants in that work are NormWear-Base, in which the backbone is frozen and only a classification head is trained, and NormWear-LoRA, in which LoRA modules are inserted into the last two transformer blocks and the [CLS] attention fusion module. OpenWatch explicitly states that there is no explicit “NormWear-2” architecture or name in that paper. This distinction matters because the OpenWatch results concern wrist-based gesture recognition on commercial smartwatches, whereas NormWear-2 concerns long-horizon physiological forecasting and intervention-conditioned latent dynamics.

The limitations of NormWear-2 are also explicit (Luo et al., 14 May 2026). It does not include reinforcement learning, reward modeling, or planning. Its latent dynamics are first-order Markov in cluster space, so long-range dependencies are delegated to the backbone’s contextual representation rather than modeled directly in the transition operator. Insight uses empirical counts and nearest-neighbor transitions in joint latent-action space, which may be data-hungry in sparse contexts. The evaluated system assumes regular sampling and Z-normalization and is limited to 1D time series rather than richer multimodal combinations such as images, clinical notes, and waveforms together. Real-time deployment and low-latency streaming are not explored in depth.

These limits define the likely trajectory of subsequent work. The reported future directions include integration with reinforcement learning for clinical decision policies, extension to larger and more heterogeneous pretraining corpora, and replacement or augmentation of the non-parametric transition model with parametric alternatives such as neural ODEs, SDEs, or low-rank Koopman operators (Luo et al., 14 May 2026). A plausible implication is that NormWear-2 is best understood as a transitional architecture: it demonstrates that chaos-aware pretraining and context-specific latent transitions can yield clinically plausible, intervention-sensitive forecasting, while leaving policy optimization and richer multimodal world modeling to later systems.