Steering, Not Predicting: Steerable Medicine World Model

In 2026, "world model" suddenly became one of the hottest terms in AI.

From artificial general intelligence to physical world simulation, from autonomous driving to robotics, from virtual cells to drug discovery, more and more teams have started calling their next-generation models "world models." In biomedicine, companies like Bioptimus, Recursion, and GenBio AI have also released ambitious blueprints: using larger models, richer data, and more complex representations to simulate cells, diseases, drugs, and the human body.

At this juncture, I raised an inconvenient question:

> If a person stood before you, and a world model stood beside them, would you dare entrust your health to it? If not today, what conditions are still missing?

This question is not sexy. It has nothing to do with parameter counts, training corpora, GPU clusters, or how beautifully a model can generate images or how complex a molecule it can design.

But it points directly to a blind spot in the current biomedical AI narrative: we have been asking whether models can predict, but rarely asking whether models can be steered by humans.

I. We Mistook "Prediction" for the Goal

There is an almost self-evident consensus in current biomedical AI:

> The more accurate the prediction, the better the model.

This consensus is so natural that few people question it. Image recognition must be accurate, speech recognition must be accurate, machine translation must be accurate. In biomedicine, disease risk prediction must of course be accurate, drug response prediction must be accurate, survival prediction must be accurate.

But there is a hidden substitution here.

Prediction has always been a means, not an end.

The purpose of weather forecasting is not to let you appreciate the probability of a cloud producing rain, but to help you decide whether to bring an umbrella tomorrow, whether flights should be delayed, whether farmland needs irrigation. Similarly, the purpose of disease risk prediction is not to give you a probability number to three decimal places, but to help doctors and individuals make a more reliable intervention decision.

In other words, what clinical decision-making truly pursues is not "more elegant predictions," but:

> Can a living system be safely, efficiently, interpretably, and traceably steered from its current state to a better state.

I call this reliable state transition.

What is state transition?

It is not simply saying "47% risk of heart attack within three years," but asking:

- What biological state am I currently in?

- Where does this state differ from a healthy state?

- Which direction should I intervene?

- After this intervention, which modules should change first?

- If there is no change, is it the state identification that's wrong, the intervention that's wrong, or the response mechanism that's wrong?

- How should the next round be corrected?

The existing prediction paradigm has difficulty answering these questions. Not because it is not accurate enough, but because in its framework, these questions are never formally posed.

II. Rider and Horse: The Essence of a World Model is to Be Steered

Consider an ancient metaphor.

A horse walks on a mountain path. The rider does not need to control every muscle of the horse, nor calculate the angle of each hoof placement in real time. The rider only needs to give directional signals through the reins. The horse relies on its own sense of balance, perception, and steadiness to adjust its gait, adapt to terrain, and avoid falling.

> The rider is responsible for direction; the horse is responsible for steadiness.

Applying this metaphor to biomedical AI:

- The rider is the doctor or researcher. Based on clinical judgment and mechanistic understanding, they give phased directional signals: add a treatment hypothesis, modify a nutritional condition, remove a confounding factor, or change an intervention pathway.

- The horse is the world model. It must maintain inference certainty amid noise, missing data, distribution shifts, and individual differences — without crashing, running wild, or dumping all responsibility onto a black box.

- The reins are the interaction interface defined by the framework. Human intentions are translated through it into inputs the model can parse, and model feedback is translated into outputs that doctors can scrutinize, question, and correct.

We gave this framework a code name:

> SEWO — Steerable Medicine World Model.

Its core philosophy can be distilled into one sentence:

> Steering, not predicting.

This is not to say prediction is unimportant. On the contrary, without sufficient predictive capability, the model cannot even perform basic inference.

But predictive capability is merely the horse's muscle. What truly determines whether it can be used for medical decision-making is whether this horse can understand direction, walk steadily on uncertain terrain, and let the rider see where the deviation occurs when it strays.

III. This Philosophy Comes from Medicine Itself

"Steering" is not a concept invented out of thin air by the AI field. It has always existed deep within medicine.

Many truly effective medical interventions do not externally force control over the body, but give the biological system a directional signal, allowing the system's own regulatory mechanisms to restart.

Natural flavonoids may not directly "kill" cancer cells. They may target a protein interaction network, perturb a signaling subsystem, and cause the cell's own regulatory mechanisms to change. Pharmaceutical nutrients, exercise interventions, sleep reconstruction, and metabolic regulation often follow similar logic: not replacing the body's repair work, but releasing the body's own repair and remodeling capabilities.

Good doctors are the same.

A doctor does not treat the human body as a machine to be reassembled piece by piece, but rather navigates a complex, uncertain, dynamically changing living system — judging direction, setting boundaries, observing feedback, and continuously correcting.

SEWO attempts to bring this medical wisdom into AI decision-making:

> A world model should not become a lofty digital oracle.

> It should be a horse that knows the way: the doctor gives direction, and it walks steadily on uncertain terrain.

IV. Five Structural Constraints: Making "Steering" Possible

If "steering" is the goal, then a world model cannot just be a prediction engine. It must satisfy a series of structural requirements.

I start from a basic postulate:

> Life is not an accumulation of variables, but a collection of adaptive capabilities.

A person's health state is not merely whether certain indicators exceed standards, but whether multiple functional modules can still adapt to stress, repair damage, maintain homeostasis, and complete remodeling. Based on this postulate, SEWO proposes five structural constraints that form a closed loop.

CP1: State Representation — mIC Vector

First, the model must be able to articulate: what state is this person currently in.

This does not refer to a specific gene expression level, a protein concentration, or a single age number, but rather how much mobilizable intrinsic capacity remains in each functional module.

In this framework, we use the mIC vector to represent modular intrinsic capacity. It attempts to transform the life state from a single-point indicator into a multi-dimensional capability structure.

> Aging is not a number, but a vector.

CP2: Intrinsic Capacity Quantification — Capomics

Second, this capability must be measurable.

We cannot just say "immune function is okay" or "metabolic state is poor" or "inflammation is somewhat elevated." These descriptions have clinical intuition, but are difficult to enter into a world model that is inferable, comparable, and auditable.

What SEWO needs is quantitative expression on capability dimensions. For example, on a particular adaptation dimension, what is the system's capability index? How far is it from functional collapse? Has a particular intervention truly changed the capability structure, rather than just changing surface indicators?

This is precisely the role Capomics attempts to fulfill: transforming the functional capabilities of a living system into computable, comparable, and trackable structured representations.

CP3: Intervention-Response Semantics

Third, the model must be able to describe how an intervention is "perceived" by the biological system and produces a response.

The same intervention may produce completely different results under different starting states. A nutrient, a drug, a bout of exercise, a sleep restoration — for different people, these may be compensatory, stimulatory, burdensome, or even disruptive.

This is precisely the structural reason why "effective on average" does not equal "effective for this person."

Therefore, a world model cannot simply treat an intervention as an external input variable. It must understand:

- Which biological channel did the intervention enter?

- Which functional module does it act on?

- Does the current state possess the conditions for response?

- Is the response restorative, compensatory, or dysregulatory?

Without intervention-response semantics, there is no truly individualized decision-making.

CP4: Counterfactual Transition

Fourth, the model must be able to answer:

> What if this intervention were not given? What if a different direction were chosen?

Without counterfactuals, there is no causation. Without causation, we can only remain at correlational extrapolation.

In SEWO, counterfactual reasoning is not simply generating several possible futures, but reasoning along a structured causal scaffold: in the current state, if a certain intervention signal is added, how does the mIC vector change? If a certain confounding factor is removed, does the state transition path change? If the intervention order is swapped, is the final phenotype different?

What medical decision-making truly needs is often not one future, but multiple comparable futures.

CP5: Quality Control Feedback — Five-Gate Check

Fifth, and most critically: when expected changes do not appear, can the model help us diagnose which step the failure occurred at?

Was the state measured incorrectly?

Was the intervention plan wrong?

Did the response deviate from expectations?

Was the mIC transition direction reversed?

Or was the phenotype readout lagging or misjudged?

We call this process the Five-Gate Check:

> State → Intervention → Response → ΔmIC → Phenotype

Each step should be independently auditable, independently questionable, and independently falsifiable.

This is also one of the most important distinctions between SEWO and black-box prediction models. When black-box models err, they typically only tell you "the prediction was wrong." SEWO seeks to further ask: exactly which gate was wrong?

V. Why Five, Not Four or Six?

These five structural constraints are not randomly enumerated.

If there were only four, the system could easily leave an open loop. For example, with state representation, intervention semantics, and counterfactual simulation, but without quality control feedback, then once prediction fails, there is no operable correction mechanism.

If split into six, seven, or eight, it is easy to lose functional cohesion, fragmenting what should be at the same level into overly fine pieces, ultimately becoming a complex but unstable flowchart.

Five constraints happen to form a complete closed loop:

> Define state → Measure state → Design intervention → Simulate transition → Check deviation → Correct next round.

This is not to pursue formal neatness, but to give the world model the minimum structural completeness required for medical decision-making.

Many biomedical AI efforts today are essentially using increasingly large hammers to drive increasingly large nails. The hammers are increasingly refined, but few people stop to ask:

> Are we even driving nails?

SEWO asks exactly this question.

VI. The Road Ahead is Long, but the Direction is Right

I must honestly say that SEWO is still in the theoretical construction and early validation stage.

The calculation of mIC vectors depends on multiple mapping chains, each with assumptions and approximations. How Capomics indices can be stably computed across different data types, disease scenarios, and populations still requires systematic validation. If the initial declaration of the causal scaffold is wrong, how the framework can self-falsify also needs more explicit operationalization standards. Empirical research has only just begun.

But this does not affect a basic judgment:

> If the endpoint of biomedical world models is to make people dare to use them, able to use them, and able to correct when they go wrong, then "steerability" is not a nice-to-have, but a necessary condition.

Prediction is the means. Reliable state transition is the goal.

The ultimate mission of a world model is not to become a more accurate prophecy machine, but to become a horse that knows the way: it stands firm and walks far on its own, while the doctor only needs to give the correct direction.

This is what I understand as Steering, not predicting.

Discussion and criticism are welcome. Every sincere question at the preprint stage is the best peer review.

- 📄 Full Paper: https://doi.org/10.20944/preprints202605.0366.v1

- 🌐 SEWO Project: http://steerable.world