acceptodds
Under review as a conference paper at ICLR 2027

FitSim: Physiology-Grounded Persona Simulation for Long-horizon Fitness Coaching

Abstract

Personal fitness coaching is a complex process that requires general physiology knowledge alongside an understanding of a particular trainee's needs, limitations, and how they adapt to training. LLMs can give fitness advice but struggle to accumulate training context: most benchmarks evaluate only single user-agent conversations or rely on other LLMs to role-play the user for several turns. This setup drifts away from assigned user profiles across the long horizons coaching demands. We introduce FitSim, a physiology-grounded persona simulator for long-horizon fitness coaching. Unlike other environments, FitSim does not rely on agent-to-agent messaging and treats coaching as a Partially Observable Markov Decision Process. In our simulation, a coaching agent reads wearable data projected from a dynamic persona state and recommends workouts over a year of day-by-day interactions. Each day's outcome is drawn probabilistically with respect to the agent's workout recommendation, the persona's adherence traits (motivation, discipline, trust), and evolving physiology (fitness-fatigue dynamics and an injury registry). To measure agents' performance in this environment, we introduce the Coaching Gain Ratio, which estimates the relative coaching effect over an uncoached user trajectory. Additionally, we assess coaching safety by counting overtraining and injury incidents in generated trajectories. We evaluate several frontier LLMs as coaching agents and find that they recover only a fraction of the coaching gain achievable by a parameter-free linear-deload schedule and trail the oracle reference by a wide margin.

Primary area:
datasets and benchmarks
Keywords:
persona simulation, fitness coaching, agent environments

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_XV6pOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces FitSim, a 365-day closed-loop simulator for fitness coaching that combines fitness-fatigue dynamics, an injury model, and a compliance model. It introduces nine user archetypes and the Coaching Gain Ratio (CGR) to evaluate LLM coaches using synthetic wearable data. Experiments show that the LLM coaches achieve limited coaching gains and trail Linear deload in the reported results.

Critical strengths
  • FitSim enables 365-day closed-loop evaluation without LLM role-play, avoiding persona drift over long horizons.

  • The rule-based reference coaches, oracle strategies, and separate measures of fitness gain and injury days provide useful analysis of LLM coaching behavior.

Critical weaknesses
  • Although the physiological models are grounded in existing literature, the compliance coefficients and user archetypes are hand-calibrated, and FitSim is not validated against real per-user trajectories. It remains unclear how well the simulated responses to coaching recommendations represent real users.

  • Linear deload uses the simulator's intent calendar, while LLM coaches must infer training patterns from wearable data. This difference makes the reported performance gap difficult to interpret.

Questions for authors
  • What evidence supports the simulated users' responses to coaching recommendations, particularly workout compliance and fitness changes?

  • How much do the main CGR and injury-day results depend on the chosen compliance coefficients and user archetypes?

  • Would Linear deload still outperform the LLM coaches without access to the simulator's intent calendar?

2 comments

Thank you for the review! We will address the questions and concerns below.

W1 / Q1 (evidence for simulated responses).

We believe the point of the simulation is to be a good proxy for the kinds of capabilities that an agent good at coaching would need, i.e. necessary but not sufficient. As you noted, FitSim is not validated against real per-user trajectories, however, we show that it provides a realistic environment for the agent to operate in. The archetypes are calibrated to maximize the coverage of parameter space (our models have a lot of degrees of freedom, and compute is limited) including the extremities (e.g. overtrainer) -- it is extremely hard to collect real longitudinal user trajectories for such cases in an ethical manner, but exploring synthetic data could provide valuable insights about the agents under test. That said, we agree this is a limitation and discuss it as such in L465-470.

We built FitSim based on the established models in sport science literature which, in turn, were fitted based on real user data. This means, we have confidence in simulated users' responses to coaching recommendations to the same degree as we have confidence in underlying models. Compliance coefficients and archetypes are chosen to span the parameter space (App. I), not to represent a population.

Q2 (dependence on coefficients and archetypes).

Per-archetype results are in App. E. Leaving out any one archetype keeps Linear deload ahead of the best LLM. For compliance coefficients, we extended App. H to all reference coaches: under ±20% on each of ten Banister and compliance constants, the order Constant < Linear deload < partial oracle < informed oracle holds in 18 of 20 settings (in the other two, Linear deload and the partial oracle tie). We did not rerun the LLMs for every setting, but as a rough check, we fed each LLM's original prescriptions unchanged into the perturbed simulator, and the order or the models had not changed (these fixed prescriptions cannot adapt to the changed user behaviour, so this is indicative only).

W2 / Q3 (intent calendar). (copied from the response to reviewer YjTz)

Linear deload still outperforms the LLMs without the calendar. We tested several additional linear deload schedulers to achieve better parity with the LLM setting:

  • A version that prescribes every day gets the same CGR, 1.14.
  • Because the prompt says "up to N×/week", we also ran a version that never exceeds N: it issues nothing until the user logs a workout, and from then on prescribes only on weekdays on which the user has logged a workout so far. This is the history the LLM reads through list_workouts; the dosing rule is unchanged. It reaches 1.13 [1.06, 1.22] and covers 87% of training days.
  • A version that ignores the history and places its N sessions on arbitrary weekdays drops to 1.05 [0.99, 1.11]. This is a lower bound: the same dosing with knowledge of N and no use of the workout history. The LLMs score in this range although they can read the full workout history through list_workouts every day.

So the comparison holds without the privileged information as well, but we agree that it's worth reframing the scheduling claim as “inferring the user's schedule from the workout history is part of the task, and the LLMs did not do it under our default prompt”. The prompt does not name the weekdays or say that off-schedule recommendations are ignored -- both have to be inferred.

Thank you for your additional experimental results and comments. I think that additional results address most of my concerns, such as the validity of simulated users, parameter dependence, and information gap related to intent calendar. In addition, I understand the purpose of FitSim isn't to predict human responses but evaluating the ability of the coach. Accordingly, I will change my rating from 2 to 3.

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

LLMs are increasingly pitched as fitness coaches, but coaching is a long-horizon problem: choices pile up over weeks. Existing health evals are one-shot rubric tests, and simulated-user setups rely on LLM role-play, which drifts away from the assigned persona over long horizons. FitSim replaces the LLM user with a deterministic, physiology-grounded state machine: Banister fitness-fatigue dynamics, ACWR-gated injury risk, and a multinomial-logit compliance model over 9 archetypes. A coaching agent reads mock Fitbit APIs and prescribes HR-zone workouts daily for 365 days, with no chat channel. The authors propose the Coaching Gain Ratio, fitness gain relative to the same user uncoached, plus injury-day and adherence metrics

Critical strengths
  • The deterministic user model is a good design choice. Swapping LLM role-play for a seeded, physiology-grounded simulator removes persona drift and LLM-judge noise. Every metric comes from hidden state and is bit-reproducible, which is exactly what a year-long closed-loop eval needs.
  • The reference policies are well thought out. Alongside the static rules, there’s an informed oracle and a partially observed oracle that sees only the LLM’s observation surface. The partial oracle splits the gap into an observability cost (0.23 CGR) and planning slack. The health-metrics ablation is a clean negative control showing that giving the agent pre-computed trends doesn’t help.
Critical weaknesses
  • The headline LLMs trail a parameter-free linear deload comparison isn’t apples-to-apples. Constant, linear deload, and both oracles read the simulator’s intent calendar and only prescribe on intent days. The LLM is only told “up to N×/week.” The paper’s own top failure mode, prescribing on non-intent days that the persona then silently ignores, which halves the follow rate, is partly a direct consequence of this information asymmetry. Similarly, calling the LLMs overly cautious is judged against one hand-calibrated Banister/compliance setup in which hard sessions are almost always refused (P(hard) ≈ 0–2% for most archetypes).
Questions for authors
  • If the LLM gets the same intent-day information the rule-based baselines use, whether in the system prompt or through a calendar tool, how much of the follow-rate and CGR gap closes?
1 comment

Thank you for the review!

Recommendations on non-intent days are never executed (L189, App. J.2), and 42-50% of LLM recommendations fall on such days. On intent days LLM recommendations are completed as often as the baselines' (0.72-0.84 vs 0.76-0.77). A Linear deload that ignores the calendar and prescribes daily gets a follow rate of 0.38, to the LLMs. It’s worth noting that such recommendations are not harmful per se, just wasteful.

We tested several additional linear deload schedulers to achieve better parity with the LLM setting:

  • A version that prescribes every day gets the same CGR, 1.14.
  • Because the prompt says "up to N×/week", we also ran a version that never exceeds N: it issues nothing until the user logs a workout, and from then on prescribes only on weekdays on which the user has logged a workout so far. This is the history the LLM reads through list_workouts; the dosing rule is unchanged. It reaches 1.13 [1.06, 1.22] and covers 87% of training days.
  • A version that ignores the history and places its N sessions on arbitrary weekdays drops to 1.05 [0.99, 1.11]. This is a lower bound: the same dosing with knowledge of N and no use of the workout history. The LLMs score in this range although they can read the full workout history through list_workouts every day.

So the comparison holds without the privileged information as well, but we agree that it's worth reframing the scheduling claim as “inferring the user's schedule from the workout history is part of the task, and the LLMs did not do it under our default prompt”. The prompt does not name the weekdays or say that off-schedule recommendations are ignored -- both have to be inferred.

P(hard) in Table 4 is for an all-Peak session, the extreme. The oracle prescribes mostly Cardio (72%) with 17% Peak, and its sessions are completed as often as Constant's easy ones (0.77 both). Also, when users decline a recommendation and train at their own default, the session is harder than what the LLMs prescribe (TRIMP 73-83 vs 36-73, L402 and Table 3). So "cautious" holds relative to what users do on their own. Though, we agree it is a single calibration and will say so.

Reviewer_EFj7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FitSim is a physiology-grounded persona simulator for long-horizon fitness coaching. It addresses the drift of LLM-based user role-play over long horizons by modeling coaching as a Partially Observable Markov Decision Process instead of agent-to-agent conversation. A coaching agent reads wearable data projected from a dynamic persona state and recommends daily workouts over one year. Each daily outcome is sampled from the recommendation, the persona’s adherence traits (motivation, discipline, trust), and its evolving physiology (fitness-fatigue dynamics and an injury registry). Agents are scored with the Coaching Gain Ratio, which measures improvement relative to an uncoached trajectory, and with safety counts of overtraining and injury events. Frontier LLMs recover only part of the gain achieved by a parameter-free linear-deload schedule and fall well short of the oracle.

Critical strengths
  • Formulating the question into POMDP is interesting and well-motivated.
  • All metrics are computed from the simulator’s hidden state, without LLM judges, which makes evaluation reproducible
Critical weaknesses
  • The hidden state is inferable from the observations by construction, since the authors generate the sensor data from it with simple formulas. The paper does not quantify how accurately the state can be recovered without knowledge of these formulas. The state-estimation failures attributed to LLMs therefore reflect an inability to invert the simulator’s own model, and may not transfer to real wearable data.
  • All LLMs are evaluated with a single minimal prompt and one agent design, so the results give a lower bound on LLM coaching ability, not an estimate of the best achievable. The conclusions would be stronger if the authors tested other strategies, such as first 1. converting wearable data into text summaries before planning, 2. few-shot prompting, or 3. ReAct-style reasoning.
Questions for authors
  • How accurately can fitness, fatigue and ACWR be recovered from the observations alone? (by just prompting the LLM to answer questions from wearable sensing)
  • It would be interesting to see how the agents plan when given an explicit goal state, such as a target fitness level by a given date, and whether this changes the conclusions.
1 comment

Thank you for the review! We will address the questions in sections below.

W1 (hidden state inferable by construction; failures may reflect inability to invert the simulator).

The hidden state is designed to be inferable in principle but is intentionally obscured. We add a small date-deterministic noise (Appendix D, L1046) to each observation. This is needed to better reflect real-life fluctuations and sensor noise, and also prevents the model from implicitly decoding hidden state from rigid observations.

How accurately can the state be recovered? We ran the partial oracle (L328, App. C.1), which sees only the agent's observations but knows our formulas, and compared its daily state estimates with the true hidden state (45 trajectories, days 22-365). Pooled Pearson r is 0.65 for fatigue, 0.85 for the fitness delta and 0.70 for ACWR; within a single trajectory the medians are 0.53, 0.77 and 0.69. To approximate how accurately the state can be recovered without knowledge of these formulas, we also computed a simplified ACWR: daily load is the logged minutes in each HR zone, weighted 1 to 4 from Out of Range to Peak zone, and the estimate is the 7-day average load divided by the 28-day average. It reaches r = 0.62. So the load ratio that drives injury risk can be tracked from the workout log without our formulas; the task does not depend on inverting the simulator.

Does good coaching require inverting our sensor model? We conducted additional ablation and we think this is not the case under the simulator conditions. The partial oracle's advantage comes mainly from the workout history, not from the sensor readings: a variant that ignores the biometric readings scores 1.20 against 1.21 CGR, with the same correlation to the true state (fatigue r = 0.64). Removing its estimate of which intensity the user completes lowers it to 1.12. Admittedly, our partial oracle implementation does not extract a lot of additional value from sensor readings and we will explicitly mention it in the camera-ready copy. We suggest treating the partial oracle as a reference point, because the model is simple, and a better use of the sensor readings may exist. We will correct §4 and §6, which currently describe the gap as a state-estimation problem.

Learnings may not transfer to real wearable data. We agree that transferability to real data remains an open question and discuss the synthetic nature of FitSim as a limitation (L476). However, we use the established models from sport science literature to generate these data (these models that are widely used to coach real people), so we have high confidence that our synthetic results can be considered a proxy of agentic coaching quality from real data.

W2 (single prompt).

We agree that there is ample opportunity for optimizing the coaching agent from our simple setup. However, we decided to focus our work on the environment modelling side, proving its applicability and correctness. Our aim was to establish a baseline of popular LLMs' performance for agentic coaching and provide the baseline to further improve on. Thus, the prompt is kept intentionally minimal and provides a lower bound. We did test one of the strategies you list, pre-computed summaries and trends of the wearable data (App. G): CGR changed by +0.008 and the prescribed zone mix did not change.

Q1. How accurately can fitness, fatigue and ACWR be recovered from the observations alone? We have not asked the LLMs for these quantities directly, because this would have required rerunning the full trajectories with a different task. However, the observation data required to reconstruct those is available to the models via tools (see our partial oracle experiment above), but we found that it's not sufficiently used -- our main observation is that LLMs rarely attempt calculating ACWR to make informed decisions.

Q2. How the agents plan when given an explicit goal state? We agree that explicit target presents an interesting ablation opportunity, however, we decided to defer it to future work because of the content size and timing constraints. We would be happy to discuss this direction in Future work section

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.