acceptodds
Under review as a conference paper at ICLR 2027

Depth Beats Breadth: Scaling Wearable World Models of Human Physiology

Abstract

Most wearable foundation models operate as static encoders, summarizing windows of observed signals into representations for downstream health prediction. We develop an action-conditioned wearable that forecasts physiological trajectories from personal history and specified future behavior. How to scale such models is unclear: adding users and extending personal histories increase data volume differently, and longer histories also provide more inference context. Using a corpus of 5.4 million users and approximately 2 billion user-days, we study four scaling axes: population size, training context, model capacity, and inference context. Across the configurations studied, longer personal histories improve forecasting more than additional users, with much of the benefit coming from longer inference context. Larger models benefit more from long context, while scaling gains are most pronounced in event-anchored cohorts and composite physiological targets. We evaluate forecasting on 231 target–cohort–horizon combinations spanning 11 primary targets, six cohorts, and horizons of 1–180 days. Frozen model representations also support 15 downstream probes: 10 subject-level health targets, including type 2 diabetes and sleep apnea, and five forecasts of self-reported sickness or menstrual-cycle phase at different horizons. Together, these results provide a recipe for scaling wearable world models and a foundation for anticipating changes in health from longitudinal physiology and behavior.

Primary area:
foundation or frontier models, including LLMs
Keywords:
scaling laws, wearables, foundation models, longitudinal data, action-conditioned forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_YtwYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Problem

What is the problem/question?

How should we effectively scale action-conditioned trajectory generation models over wearable signals data, considering axes such as population size, training context (max sequence length seen during training), model capacity, and inference context (max sequence length seen during inference)?

Why is the problem/question important/impactful?

Understanding scaling laws makes it much easier to build the best possible model at a given compute and data budget, and highlight the most important directions the field should study / resources the field should curate to maximally improve model performance. This field is of increasing interest in the context of large language models, but has been less well studied in other settings.

Why is the problem/question hard?

Determining scaling laws that effectively extrapolate to unseen dataset sizes and model capacities requires nuanced understanding of the appropriate metrics to use for population size (e.g., number of unique sequences vs. number of unique patients) and model capacity (width, depth, attention heads) and are often expensive to run, as you need to fit many runs to determine their scaling relationships. You also need to use the right form of the scaling law, which can be challenging as it may change depending on the dataset and scaling regime in question.

What naive solutions have or should have been tried to solve this problem/question?

Simply run everything and fit a wide battery of laws and identify those from used in the literature to date that fit best.

Why are these naive solutions insufficient?

They aren't, usually, they just may not have been done before. Though, that is undercut by some of the novelty concerns.

Solution

How do the authors propose to solve it (in brief)

The authors perform a grid over # of patients, training context seen per patient, inference context allowed, and model capacity (on an unclear scale).

What experimental questions do the authors answer to show that they solve it?

  1. Whether their trained model shows accurate forecasting capabilities over long time horizons (up to 6 mo).
  2. Whether learned representations show utility in linear probing in addition to zero-shot forecasting capabilities.
  3. Whether increasing number of subjects vs. amount of data per subject matters more
  4. Whether higher-capacity models gain more or less from more context
  5. Whether scaling on various axes matter more for general forecasting capabilities or tasks of greater real-world impact?
  6. Whether contextual benefit matters more for inference capabilities or in training quality
What key experimental questions are missing?
  1. isoflop style analyses which ask how compute should be divided between observing more training sequences, longer sequences, or over larger model parameters at a given compute budget.
  2. What (if any) explicit scaling laws can be fit to the data that permit extrapolation of loss to larger settings.
  3. Whether training loss reliably correlates with task-specific performance.

What results are revealed from those experiments, and do they support the author's claims?

  1. The authors report results of their model's forecasting capabilities against a trailing 30d baseline. It is unclear how other kinds of models would perform on this, however, rendering it hard to make judgments of individual model performance.
  2. The authors report that forecasting is better on users with longer history when longer history is leveraged for inference, which may or may not mean that amount of data per subject is more valuable than number of subjects (see below).
  3. The authors report that larger capacity models perform better, especially when more context is leveraged.
  4. That scaling gains are larger in targeted tasks, rather than general forecasting.

Novelty

Has this been done before? If not, why not?

The authors claim that they produce the first "population-scale wearable world model for forecasting longitudinal physiological trajectories conditional on future behavior". I do not think this is true -- or that if it is true it likley has too many conditionals to be as meaningful.

There certainly are numerous generative, AR models over very large datasets of wearable data, including SensorFM and this paper (https://arxiv.org/pdf/2307.16664) that used a dataset of >2B user-days of data.

If the primary aspect of novelty here is simply that these models are not "conditioned on future behavior", I'm not sure that is as meaningful given (by the author's own statement) this model still is not a causal / RL model, and merely predicts associative effects that leverage this conditional element.

The authors also state that "Longitudinal... data introduce a further distinction: more data can mean more individuals or more history from each individual". This is actually not a distinction from natural language processing studies. More data can mean either more text sequences or longer text sequences, so I don't feel that this introduces a major novelty differential either. This question has actually been directly analyzed in the longitudinal data space as well, in "Scaling Wearable Foundation Models", which has a section entitled "Is scaling subjects or... hours per subject more helpful?"

Are there papers / references that are natural for this that the authors should highlight, and do they?

The authors do seem to cite most of the right works, but underappreciate/emphasize their releavnce, I think.

Limitations & Societal Context

Limitations

What limitations are there in this approach, and do the authors appropriately identify them?

The authors identify limitations sufficiently.

Ethical Concerns

Are there any ethical concerns with this approach, and do the authors appropriately identify them?

None.

Critical strengths

Key Strengths

These are reasons I would advocate the paper be accepted

  1. I think that scaling law research is highly impactful and helps make methods research more accessible and translatable to deployment.
  2. I think that separating out context-size, model capacity, and training dataset size is useful for furthering understanding of scaling laws in this setting.
  3. I like the specialization of generic and task-specific forecasting analysis.

Minor Strengths

These are strengths, but not sufficient alone to warrant advocating acceptance

  1. I think the paper is generally relatively clear, and I appreciate the organization of the results section around the key questions the authors evaluate.
Critical weaknesses

Key Weaknesses

These are reasons I would advocate the paper be rejected

  1. This paper tries to fit two framings at once: (1) Presenting a new "world model" over wearables data at a very large scale, capable of long-horizon, action-conditioned prediction and (2) a detailed analysis of scaling behaviors of said model / wearables models generally. This is a problem because both of these goals have different desiderata, and doing them simultaneously risks (and here may achieve) doing neither well. For (1), a key comparison would be against other existing wearables foundation models in downstream task quality, and (2) warrants studying over an established model family / architecture such that inductive insight offered by scaling law analyses can be useful for readers, and potentially a different focus or greater experimental depth than is presented here. Better specializing to one or the other (or trying to expand on and do both properly) would require both framing adjustments and experimental adjustments throughout.
  2. In the context of this as a scaling law study, it has greatly limited novelty relative to "Scaling Wearable Foundation Models". It's major differentiator against that study is (a) the question of how the amount of inference context matters relative to training contexts (which I have concerns about, mentioned below), (b) the analysis of how gains differ between general forecasting and cohort-specific forecasting, and (c) whether previous findings generalize to longer horizon models/tasks. While these are meaningful, they don't necessarily meet the novelty bar in my opinion for this venue.
  3. I do not think the inference context study is as interesting/impactful as advertised, and I believe it has some validity issues. * In terms of interestingness/impactfulness, I think generally speaking you would expect that in a large data/capacity regime, increasing sequence lengths will almost always lead to improved performance, and the choice about how much to increase sequence lengths is more a question of computational resource constraints. Similarly, if you train a model at a max sequence length of M, there is no reason beyond computational efficiency not to use an inference context length of M-1 (e.g., the full possible save the slot needed for the generated next token), so the question of how much improvement does increasing M make if you still restrict inference context to fixed k << M is somewhat strange to me. Had you done something like assessed at given total compute budgets (e.g., ISOflop curves) what max sequence length M is right, as increasing M will both increase performance and compute cost for a given dataset size and model capacity, that would be more interesting, but I don't see such a result here. * In terms of validity, critically, restricting samples to only those with an observed degree of inference context is not a random subsample of the population -- forecasting on users who have more or less total data observed may be easier or harder. Instead, you should take a population that all have the same degree of longitudinal data observed, then subsample inference context length, such that no users are being included/excluded differently at different context lengths, and the only changing variable is how much context they can use at inference time.
  4. You don't actually deliver any scaling laws (or demonstrate that they do not successfully fit in your setting) for your model type, limiting how much users can take away from this study. I'd want to see you do something like fit Chinchilla or other style of scaling law to your data, such that you can comment on extrapolation performance, how fits differ systematically across measures or settings, etc. Even if the fits don't work well, identifying that there are gaps in the scaling law literature for this style of model would be useful; without them, your results are limited in how much users can derive actionable insight from them for future studies.

Minor Weaknesses

These are weaknesses, but not sufficient alone to warrant advocating rejection

  1. I wouldn't frame your cohort specific results as showing the limitations of population level aggregation -- it is just an evaluation on different tasks. One task is forecasting from an arbitrary event, the other is a more specific prediction task. This doesn't limit the value of these observations, to be clear, I'm just suggesting a slightly different way to present them.
Questions for authors

Action Items for Rebuttal

Things that, if addressed, may change my score

If you could clarify that there is not the validity issue regarding inference context length and better identify why that arm provides material inductive value vs., e.g., presenting a unified isoflop curve view, and could further fit actual scaling laws or present real isoflop curves, that might make me improve my score, but it is unlikely to reach an accept, given the novelty concerns would remain.

Questions

Things that I want to know, but are unlikely to result in me changing my score

2 comments
AuthorsOpenReview ↗

Response to Reviewer YtwY

We thank the reviewer for a thoughtful review and for engaging with the paper. The global response covers the new baselines and how we use action conditioning. Below we address novelty relative to LSM and SensorFM and the question of isoFLOP analysis, both of which we believe stem from a framing we did not make clear in the paper.

The question we study. We ask which data to prioritize when building longitudinal health models, longer histories or additional users. In health, data is usually the binding constraint. The number of people and the length of their histories are set by who has worn a device and for how long, and more compute cannot add to either. IsoFLOP analysis answers a related and valuable question about how to spend compute on a given dataset. Our results show that breadth saturates near one million users while depth continues to help through 365 days. For example, the M model trained on 1.7M users with a full year of history performs the same as the M model trained on all 5.4M users (median R² of 0.655 for both across 308 panels), and at matched user-days the deeper model wins on 96% and 93% of panels at the two budgets. Moreover, these trends differ based on the evaluation cohort we use, where we show that scaling gains concentrate around physiological events. Hence, no single functional form can describe the scaling behavior, and we show that it should be assessed across targets, horizons, and periods of physiological change.

Novelty relative to LSM and SensorFM. These papers established scaling behavior for wearable foundation models, and we will discuss them more prominently. Neither varies how much personal history the model conditions on. LSM (Narayanswamy et al., 2025) trains on 5-hour windows from 165K people. Its subjects-versus-hours experiment holds the window at five hours, so more hours per subject adds training windows without lengthening the context, and the authors hypothesize that subject diversity may matter more as windows grow to days. SensorFM (Narayanswamy et al., 2026) samples 10 weeks per participant from 5M people, uses a 24-hour context window, and scales model and data jointly. We vary conditioning history from 7 to 365 days, separating training context from inference context, and cross both with breadth (10K to 5.4M users) and capacity (1.3M to 303M parameters), for forecasts up to 180 days. To our knowledge, this is the first test of depth versus breadth in this regime, which LSM left open, and our findings differ from the recommendation to scale both together.

Action conditioning. We agree that the model is not yet causally validated, and the paper states this. Conditioning on behavior still allows the model to predict how a person's trajectory responds to a different plan, which a static encoder cannot do. The global response gives population-level effects of alcohol and activity, along with an estimate for each user.

We will revise the introduction and related work to make these distinctions explicit and add the new results. We thank the reviewer for these comments, which have made the paper stronger.

Reviewer_YtwYOpenReview ↗

Response

I think you're being a bit too dismissive and narrow when thinking about isoFLOP analysis. IsoFLOP analysis isn't just about "how to spend compute on a given dataset" -- instead, it says "if you have a given training compute budget, how should you trade off the various choices one needs to make in model design that change how much compute a single gradient pass costs vs. how many passes you get to make. It is very natural, in fact, to incorporate context size into isoFLOP analysis. If I can spend 10^20 FLOPs training my model, do I want to spend that with more passes over shorter sequences? Or fewer passes over longer sequences? Or even fewer passes over longer sequences with a bigger model? IsoFLOP analysis identifies how to anticipate what choices from that set one should make to yield minimal loss (or any metric that can be evaluated equivalently across settings) over larger compute budgets than have been tested, such that future models can be built more efficiently. I'd strongly encourage you to take more of that perspective to heart in future revisions of the work.

Unfortunately, as I indicated in my original review, I think even with the rebuttal this is insufficient to change my score, but best of luck with your submission going forward!

Reviewer_iLiUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This manuscript studies scaling laws for an action-conditioned wearable multivariate time-series Transformer across 5.4M users, varying population size, training history length, model capacity, and inference context. It concludes that longer personal histories matter more than adding users ("depth beats breadth") and that frozen representations transfer to clinical health probes.

Critical strengths
  • Large longitudinal scale (5.4M users) with informative disaggregated evaluation across event-anchored cohorts (postpartum, GLP-1, illness, alcohol).
  • separating the effect of training history length (Q1Q_1) from inference history length (Q2Q_2).
Critical weaknesses
  1. Future-information leakage into forecast targets and asymmetric baselines
  • During hh-day forecasting, the model receives the user's realized future behaviors over days t+1…t+ht+1 \dots t+h (future sleep duration, sleep-onset deviation, active calories, and event tags), whereas the trailing 30-day mean baseline receives no future inputs.
  • This affects the discussion that scaling model capacity (1.3M to 303M parameters) yields no gain on raw physiological signals. Instead, capacity and baseline gains concentrate in proprietary composite scores (Sleep score, Readiness score, Activity/Sleep balance), which are direct algorithmic functions of the realized future sleep and activity covariates fed into the predictor.
  1. "Depth beats breadth" is an inference-window calibration effect, not a pretraining scaling law
  • Holding inference context fixed, training on longer histories (Q1Q_1, encoder quality) yields near-zero improvement
  • Nearly all of the reported depth advantage comes from supplying 30–90 days of history at test time (Q2Q_2) to anchor a user's personal baseline, which saturates by 90 days
  • In the 510M-user-day matched comparison (Table 9), the broad model (3.9M×130d3.9\text{M} \times 130\text{d}) is restricted to 30 days of inference context while the deep model (1.13M×450d1.13\text{M} \times 450\text{d}) uses 450 days, omitting the fair comparison where the broad model uses its full 130-day context.
  • Training steps, compute budgets (FLOPs), and handling of users with <365<365 days in "5.4M×365d5.4\text{M} \times 365\text{d}" cells (where only 1.7M users actually have 365 days) are not specified.
Questions for authors
  1. As I explained above, the predictor receives the user's realized future behaviors over days t+1…t+ht+1 \dots t+h. What is the model's forecasting R2R^2 and event-probe AUROC when evaluated without realized future actions (e.g., imputing trailing baseline behavior)? How does a simple regression baseline perform when given the trailing 30-day mean plus the same realized future sleep and activity covariates?
  2. In the 510M-user-day matched comparison (Table 9), the broad model is restricted to 30 days of inference context while the deep model uses 450 days. What is the difference when the broad model uses its full 130-day inference context? Additionally, how many optimizer steps were run per cell?
  3. Because input channels include both Vascular age and Age - vascular age (recovering exact chronological age), Breathing disturbance index, and CGM glucose, what AUROC does a simple logistic regression on chronological age, sex, and raw feature means achieve on the strictly held-out test split (Table 10)?

Suggestions for Future Work 4. Evaluating trajectories when future behavioral changes are unobserved or exogenously assigned would be necessary to test the model as a true counterfactual world model rather than a conditional observational interpolator.

1 comment
AuthorsOpenReview ↗

Response to Reviewer iLiU

We thank the reviewer for engaging closely with our work. The global response covers future actions as a design choice, baselines that receive the same realized future covariates, forecasts under each user's status quo in place of the realized future, and the matched comparison with the broad 3.9M-user model reading its full 130 days. Below we answer the points specific to this review.

Leakage and composite scores. We respectfully see future actions as conditioning inputs, which is different from leakage. The model takes behavior as input and predicts physiology, and that is what allows it to tell a person how a different plan would change their trajectory. Our baselines receives exactly the same realized future covariates, including the sleep and activity inputs behind the composite scores. Even so, our largest gains are on those scores. At h=30, for example, readiness reaches an R² of 0.54 with our model, compared with 0.24 for Chronos-2 and 0.32 for XGBoost. If these scores were simple functions of the covariates, models with the same covariates would recover them. Every scaling cell also sees the same actions, so the capacity, breadth, and depth comparisons are symmetric.

Matched comparison at full context. With the broad model reading its full 130 days, the 1.13M × 450-day model wins on 93% of panels (median +0.008 R²) and the 1.7M × 300-day model wins on 97% (median +0.008). The deep model is ahead on every cohort, as shown in the table in the global response.

Demographic baseline. Importantly, we do not mean to assert state-of-the-art performance on these held-out probe tasks. Instead, we aim to show the initial capability of this wearable model for tasks beyond forecasting, demonstrated by the strong AUROCs on these held-out tasks. Following past work, we consider a ridge probe on age, sex, and BMI, trained on the same users and 70/30 split as our pooled results, as a competitive baseline. Our model still outperforms it on all 12 targets. For example, AUROC is 0.89 versus 0.83 for type 2 diabetes, 0.71 versus 0.64 for atrial fibrillation, 0.83 versus 0.80 for sleep apnea, and 0.84 versus 0.80 for hypertension, and R² for systolic blood pressure is 0.19 versus 0.13. SensorFM (Narayanswamy et al., 2026) similarly found that demographics alone carry substantial signal for these tasks.

Causal interpretation. We agree that establishing the predicted effects of behavior change as causal effects requires substantially more careful validation and design, and we see this as important future work. The global response shows what the model already provides, which is population-level realistic effects of alcohol and activity along with a personalized estimate for each user.

These analyses will be added to the revision, and we thank the reviewer for pushing us toward a stronger paper.

Reviewer_3jVDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the scaling behavior of wearable world models for longitudinal physiological forecasting using a large dataset of approximately 5.4 million users. The authors systematically examine population breadth, personal history depth, model capacity, and inference context. While the large-scale experimental setting and longitudinal modeling perspective are valuable, several concerns regarding experimental fairness, methodological novelty, evaluation reliability, and practical applicability limit the strength of the conclusions.

Critical strengths
  1. The paper investigates wearable world models using approximately 5.4 million users and 2 billion user-days, providing a valuable empirical foundation for studying long-term physiological forecasting.
  2. The study examines population breadth, personal history depth, model capacity, and inference context, while evaluating performance across multiple physiological targets and event-specific cohorts. This provides useful insights into the role of longitudinal context in wearable forecasting.
Critical weaknesses
  1. The main comparison between 1.7 million users with 365 days of history (620.5 million user-days) and 5.4 million users with 30 days (162 million user-days) involves substantially different data volumes, with the former containing approximately 3.83 times more data. Although matched user-day experiments are provided in the appendix, their results are not consistently favorable across cohorts. Consequently, the evidence remains insufficient to establish that personal depth is generally more important than population breadth.
  2. The proposed architecture mainly combines established Transformer and recurrent components without introducing a substantially novel learning mechanism. Moreover, the forecasting evaluation lacks strong contemporary baselines and relies on observed future actions that may be unavailable in practical scenarios. These limitations weaken the claimed methodological contribution and practical significance.
Questions for authors
  1. Given that the main breadth versus depth comparison involves different data volumes, can the authors provide more comprehensive evidence from the matched user-day experiments to justify the central claim? In particular, how should the negative results for certain cohorts in Appendix Table 9 be interpreted?
  2. Can the authors clarify whether the reported improvements primarily arise from longer historical context rather than the proposed architecture? Comparisons with stronger forecasting baselines under identical input contexts would help establish the necessity of the architectural design.
  3. Since the forecasting evaluation uses observed future actions, how would the model perform when future behaviors are unknown or inaccurately estimated? This clarification is important for assessing the practical applicability of long-horizon physiological forecasting.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 3jVD

We thank the reviewer for the careful reading. The global response addresses the two central concerns of whether depth beats breadth at equal data volume and whether conditioning on future actions limits the evaluation. Below we answer the remaining questions.

Architecture versus context. The new baselines address this question directly. The baselines receive exactly the same 365-day history and future actions as our model, yet they trail our model at every horizon, with the widest gap around events. Since the inputs are identical, the improvement comes from the model. Moreover, the same model architecture is used throughout all of the experiments, so the scaling findings only vary context, breadth, and model size independently of architecture.

Negative entries in Table 9. Table 9 splits the deep model's advantage into a training-side part, where both models read 30 days, and an inference-side part, where the deep model reads its full history. For the illness cohorts, both parts stay below 0.011 R² in magnitude, so neither has a clear effect and small negative values can appear. With each model at its full context, the total advantage is positive on every cohort at both budgets (see the global response), including illness at 510M user-days (+0.010). We will report these totals alongside the split in the revision.

Unknown future behavior. Importantly, the model was trained to learn the relationship between actions and physiology. The ability to accurately do this can open up potential for behavior change and goal setting and is an active area of research in both robot planning and wearables, as discussed in the global response. Our evaluations show that the models have learned this well, and future work should consider the right way to serve these models to provide status-quo forecasts when future behavior is unknown. For completeness, the global response reports forecasts under each user's status quo. Around events, these forecasts stay above the trailing mean at every horizon, and on alcohol nights, at illness onset, and after childbirth they match or beat the baselines that see the real future.

We will add these analyses and clarify how users with shorter histories enter each cell. We thank the reviewer for pushing on these points, which have made the paper stronger.

The author should point out where the newly added benchmark experiments are in the revised version and mark them in a different colour. Could the author explain why the R² values for both parts are below 0.011 for the patient group, to help define the range of deep advantage?

Thank you for the follow-up. We are unable to upload a revised PDF before the deadline. However, all of the new experiments are reported in the global response above. In the final version we will add them to the paper. We are happy to clarify any of these results here.

With respect to your question, we read "the patient group" as the two illness cohorts in Table 9 and we read "both parts" as the two ways a longer history can help, which are a better-trained model and more history at prediction time. If we have misinterpreted your question, please let us know.

We hypothesize that both effects are small for illness because illness is a short-term event. To forecast the days after someone gets sick, the model mainly needs to know what that person's baseline normal looks like. Based on our results, about a month of history is enough to establish that, so going from 30 days to a full year adds little. Table 9 compares 30 days of context against a year or more, so it only measures that final stretch. The negative values for illness are all smaller than 0.01 R², and the positive ones are of the same size. We read these as no effect in either direction rather than as evidence that longer history hurts.

History still helps illness forecasts overall. Moving from 7 to 365 days of training history improves illness R² by about 0.04, more than the 0.03 gain for the general population (Table 11). However, most of that gain seems to come from the first month of history. We hypothesize that history helps more when an event recurs or when it unfolds slowly. For alcohol, a year of history contains many of a person's earlier drinking nights, which may let the model learn how that person responds (+0.015 R² at both budgets, Table 9). The start of GLP-1 treatment gains a similar amount (+0.013 to +0.018 R²), and postpartum recovery, which plays out over months, gains the most (+0.07 to +0.08 R²). We will state this directly in the revision.

The author does not seem to have clearly explained where the boundaries of the "depth advantage" lie, so I will maintain my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to all reviewers

We thank all reviewers for their careful reading. Several concerns came up across reviews, namely conditioning on future actions, baselines beyond the trailing mean, forecasting when future behavior is unknown, and depth versus breadth at equal data volume. We address them together here with new experiments and each reviewer's response covers the rest.

Future actions are part of the design. Our goal is a world model, a model of how a person's physiology responds to what they do. Such a model needs future behavior as an input to estimate how a different week of sleep, activity, or drinking would change someone's forecasts. Action-conditioned world models are an active area of research for this goal. For example, V-JEPA 2 (Assran et al., 2025) conditions video prediction on robot actions for planning, and Lee et al. (2026) recently proposed SensorWM, a wearable world model that forecasts physiological responses to lifestyle changes for health coaching. Regardless, every cell in our scaling grid receives the same future actions, so the scaling and cohort findings hold regardless of them.

Baselines with identical inputs. Our main baseline is Chronos-2 (Ansari et al., 2025), a pretrained time-series foundation model that we fine-tuned on our data. It receives exactly the same 365-day history and exactly the same future actions as our model. As an additional check, we trained XGBoost (Chen & Guestrin, 2016) on a subset of users with summary features of the same inputs. The table below reports mean R² over the 11 primary targets on held-out test users.

Cohort, horizonOursChronos-2XGBoostTrailing mean
Global, h=10.680.560.600.45
Global, h=1800.640.480.540.41
Alcohol, h=10.640.390.510.12
Severe illness onset, h=10.610.140.380.08
Postpartum, h=300.610.290.420.09

Our model outperforms both baselines in every cohort and at every horizon, especially around events. Since the baselines see identical inputs, this difference reflects the models themselves and shows that future actions themselves do not drive the performance.

Using action conditioning. On test nights with an alcohol tag, removing the tag changes the predicted next night by +2.0 bpm in lowest heart rate (standard deviation [SD] 1.2), −4.2 ms in HRV (SD 4.2), −3.9 points in readiness (SD 2.3), and −1.5 points in sleep score (SD 1.6). HRV falls on 98% of these nights but the size varies by person, with the middle 80% of effects between −9.2 and −0.6 ms. Raising a person’s active calories by one personal SD predicts an HRV increase of 4.0 ms (SD 3.9) and a 1.6 bpm drop in lowest heart rate at 90 days, rising to 6.6 ms at two SDs. The model gives population-level effects and an estimate for each individual, though these observational predictions still need causal validation.

When future actions are unknown. The model was trained to learn the relationship between actions and physiology. As discussed above, this is an active area of research for goal setting in both the video/robot space and wearables. Hence, the model should forecast best with an accurately stated plan. For completeness, we also forecast with each user's status quo in place of their realized future, using their last 5 week weekday routine for sleep and activity and their usual weekly frequency for tags. Global R² stays above the trailing mean at short horizons (0.60 versus 0.45 at h=1) and approaches it at long horizons, as behavior drifts from recent routine. Around events, it stays above the trailing mean at every horizon (0.35 versus 0.08 at h=180 postpartum). On alcohol nights, at illness onset, and postpartum, it also matches or beats baselines, even though both see the real future (0.50 versus 0.24 and 0.42 at illness onset, h=1). A recovery plan with sleep and activity one personal SD lower improves severe-illness forecasts at h=1 (0.47 to 0.48) and h=3 (0.35 to 0.37). Choosing realistic status-quo plans for serving these models is a natural direction for future work.

Depth beats breadth at equal data volume. Users with fewer than 365 days contribute all the days they have, so the 5.4M-user cells include everyone. In the matched user-day experiments, both models in each pair read their full context, so the broad models read all 30 or 130 days. The deeper model wins on every cohort at both budgets.

BudgetDeep vs broadPanels deep winsMedian ΔR²
164M450K × 365 d vs 5.4M × 30 d96%+0.018
510M1.13M × 450 d vs 3.9M × 130 d93%+0.008
510M1.7M × 300 d vs 3.9M × 130 d97%+0.008

Adding users beyond those with long histories also adds little, as the M model trained on 1.7M users with 365 days matches the M model trained on all 5.4M users (median R² of 0.655 for both across 308 panels).

We will add these analyses to the revision, and we thank the reviewers for prompting them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.