Does Your Trial Simulator Know What the Drug Does? An Intervention-Blind Test of Interventional Validity on 7,207 Registered Trials

arXiv:2611.05907 · stat.AP, cs.AI, cs.LG · Submitted 2026-11-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Does Your Trial Simulator Know What the Drug Does? An Intervention-Blind Test of Interventional Validity on 7,207 Registered Trials".

Jane: Here's the gist of the paper. Patient world models are increasingly proposed as simulators of clinical trials, and recent surveys organize their capabilities as a ladder that rises from temporal prediction, through action-conditioned prediction, to counterfactual rollouts that can support decisions on Page 1.

Tom: The paper turns the counterfactual rung into a measured quantity by constructing an intervention-blind reference and scoring simulators based on how much of their accuracy depends on knowing which intervention was tested on Page 1.

Jane: Benchmark Construction The construction involves three parts, starting with a benchmark of real counterfactuals from 30,119 completed phase 2/3 trials with results posted to ClinicalTrials.gov on Page 2.

Tom: This benchmark extracts arm-level primary-endpoint results and forms within-trial contrasts, treatment arm minus comparator arm, from 7,207 trials on Page 3.

Paper summary: Tom: We're looking at how they measured this "interventional skill" using seven thousand two hundred seventy registered trials to see if a simulator actually knows what the drug does when you swap out the treatment arm for something else.

Jane: So the authors are turning that counterfactual rollout idea into a measurable number by scoring simulators based on how much their accuracy depends on knowing which intervention was tested on page one >

Lu: They set up a benchmark using nine thousand nine hundred ninety-six within-trial contrasts from seven thousand two hundred seventy trials, focusing on treatment arm minus comparator arm results from page four >

Meng: The definition of interventional skill is that it's the error the full model removes when compared to a blind reference model, which is basically a baseline that doesn't know what treatment was applied >

Tom: So the main finding is that registry-trained simulators have zero interventional skill, meaning they aren't learning anything about the intervention itself, while a large language model reading the protocol showed an interventional skill of zero point two on page four >

Lu: But they also found that when you don't include the posted results, that interventional skill drops to a range between negative zero point one and positive zero point two on page four which is too imprecise to confirm if a model is just memorizing things.

Tom: That tells us we need this kind of measurement because just looking at accuracy isn't enough when you're trying to make decisions on page one of a trial >

Jane: It suggests that having a model that can roll out those counterfactual scenarios is necessary, but it doesn't automatically mean those scenarios are actually grounded in the physics of the drug effect >

Meng: So if you want to trust these simulators for making critical decisions, you have to test them specifically on how they handle the actual intervention, not just their general accuracy.

Meng: From a deployment standpoint, we need to know what breaks at scale, and this benchmark doesn't really simulate that kind of real traffic.

Conclusion: Tom: We're wrapping up this look at these patient world models and what they can actually do when it comes to understanding drug effects from clinical trial data on page two and page three of that paper.

Jane: Right, so we’ve seen how they measured an "interventional skill" using seven thousand two hundred seventy registered trials to see if a simulator actually knows what the drug does when you swap out the treatment arm for something else.

Lu: The authors are basically turning that counterfactual rollout idea into a measurable number by scoring simulators based on how much their accuracy depends on knowing which intervention was tested on page one >

Meng: So what this means for us is that just having a model predict an outcome is not enough; it has to be grounded in understanding the specific action taken during the trial.

Tom: Exactly. They showed that registry-trained simulators had zero interventional skill, while a large language model reading the protocol had an interventional skill of zero point two on page four >

Jane: That means if you swap arms in those tests, that large model reversed its prediction in eighty-six percent of the comparisons on page four >

Jane: It sets up the next big question: if we can measure this skill, what does it take to actually build a simulator that consistently achieves high interventional skill in the real world?

Alicem Koyun

Westcliff University

stat.AP, cs.AI, cs.LG

Submitted: 2026-11-04

Updated: 2026-11-04

Comments: Oral presentation at the NeurIPS 2026 Workshop on World Models for High-Stakes Health. 31 pages, 5 figures. Code and data: https://github.com/alicemkyn/ivbench

Code: https://github.com/alicemkyn/ivbench

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: Patient world models are increasingly proposed as simulators of clinical trials, and recent surveys organize their capabilities as a ladder that rises from temporal prediction…

Key concepts

Interventional Skill (IS)
This measures a model's ability to predict results when it knows the specific intervention tested, compared to a baseline model that doesn't know. It quantifies how much accuracy depends on knowing which treatment was applied, operationalizing a key step in testing if a simulator can handle real-world drug testing scenarios.
Intervention-Blind Reference
This is a comparison model that does not know which treatment arm was tested during the trial. It serves as a conservative baseline because it already incorporates general knowledge about the class of trials, making the measured interventional skill a lower bound on true performance.
Counterfactual Rollouts
This refers to the highest level of prediction capability being tested. It involves simulating what would happen in a trial under different hypothetical conditions or interventions that were not actually tested, which is crucial for making decisions about drug efficacy.
Registry-Trained Simulators
These are simulators trained using data from clinical trial registries like ClinicalTrials.gov. The study found these models lacked interventional skill, meaning their predictions didn't significantly improve when given specific intervention information, unlike language models.

Terminology

Summary

Patient world models are increasingly proposed as simulators of clinical trials, and recent surveys organize their capabilities as a ladder that rises from temporal prediction, through action-conditioned prediction, to counterfactual rollouts that can support decisions on Page 1. p. 0

Where the paper says it

An Intervention-Blind Test of Interventional Validity on 7,207 Registered Trials Alicem Koyun Westcliff University alicemkoyun@gmail. com Abstract Patient world models are increasingly proposed as simulators of clinical trials, and recent surveys organise their capabilities as a ladder that rises from temporal prediction, through action-conditioned prediction, to counterfactual rollouts that can support decisions.

The paper turns the counterfactual rung into a measured quantity by constructing an intervention-blind reference and scoring simulators based on how much of their accuracy depends on knowing which intervention was tested on Page 1. p. 0
Where the paper says it

gov we construct 9,996 audited within-trial contrasts (treatment-arm outcome minus comparator-arm outcome on the primary endpoint) from 7,207 trials, and we score a simulator not on how well it predicts a held-out arm but on how much of its accuracy depends on knowing which intervention was tested. The instrument is an intervention-blind reference: the same model, trained on the same trials, with every intervention-identifying field withheld.

Benchmark Construction

The construction involves three parts, starting with a benchmark of real counterfactuals from 30,119 completed phase 2/3 trials with results posted to ClinicalTrials.gov on Page 2. p. 0

Where the paper says it

From 30,119 completed phase 2/3 trials with results posted to ClinicalTrials. gov we construct 9,996 audited within-trial contrasts (treatment-arm outcome minus comparator-arm outcome on the primary endpoint) from 7,207 trials, and we score a simulator not on how well it predicts a held-out arm but on how much of its accuracy depends on knowing which intervention was tested.

This benchmark extracts arm-level primary-endpoint results and forms within-trial contrasts, treatment arm minus comparator arm, from 7,207 trials on Page 3. p. 1
Where the paper says it

gov we extract arm-level primary-endpoint results and form within-trial contrasts, treatment arm minus comparator arm, as a risk difference or a standardised mean difference. The audited release holds 9,996 contrasts from 7,207 trials, with a temporal split on the results-posting year (Section 3).

The final set consists of 9,996 contrasts from 7,207 trials on Page 4. p. 3
Where the paper says it

98]; all 125 new or re-paired test contrasts were audited individually and 8 removed by hand. The final table holds 9,996 contrasts from 7,207 trials (4,014 risk differences, 5,982 standardised differences; 7,669 inert-control, 2,327 active-comparator); v2 is kept as a sensitivity set.

Interventional Skill Measurement

Interventional skill is defined as the error the full model removes relative to an intervention-blind reference on Page 4. This skill is calculated as IS = skill(∆ˆ FULL ∆ˆ BLIND) = 1 − MSE(∆ˆ FULL) / MSE(∆ˆ BLIND) on Page 4. p. 3

Where the paper says it

Call these BLIND and FULL; BLIND− is the same reference without comparator type. Interventional skill is IS = skill(∆ˆ FULL | ∆ˆ BLIND) = 1 − MSE(∆ˆ FULL) MSE(∆ˆ BLIND) .

The authors argue that this measurement operationalizes the L3 rung of the Qazi et al. p. 1
Where the paper says it

(i) A public, audited benchmark of 9,996 real within-trial contrasts with construction code, data card, splits and every result table regenerable by one script. 1 (ii) Interventional skill, an intervention-blind measurement that operationalises the L3 rung of Qazi et al.

[2025] ladder, which is counterfactual rollouts on Page 2. p. 1
Where the paper says it

, 2025] makes this concern explicit. It organises the field on a four-rung ladder of temporal prediction (L1), action-conditioned prediction (L2), counterfactual rollouts (L3) and planning (L4), finds that electronic-health-record systems cluster at L1, and names interventional validity as the property the field cannot yet check: simulated futures “may be realistic but wrong about treatment effects”.

Model Performance and Skill Findings

The results show varying levels of skill across different systems, with registry-trained simulators having no interventional skill while a frontier language model reading the protocol has an interventional skill of 0.20 on Page 4. The authors find that perturbation tests show the large model uses the intervention as such, swapping arms reverses its prediction in 86% of contrasts on Page 4. p. 0

Where the paper says it

15], at the limit of detection. Perturbation tests show the large model uses the intervention as such: swapping the arms reverses the sign of its prediction in 86% of contrasts and substituting an unrelated drug is worse than withholding the drug.

Furthermore, where exposure to posted results can be excluded, the protocol-only interventional skill is 0. p. 0
Where the paper says it

The blind reference is itself not drug-blind: from the blind prompt a language model recovers the withheld drug class 45% of the time (chance 2%), so interventional skill is a conservative lower bound on what the intervention contributes. Where exposure to posted results can be excluded (217 contrasts completed after an open-weight model’s training cutoff) its protocol-only interventional skill is 0.

01 [−0.10, 0.12] on Page 4 and too imprecise to confirm or exclude memorization on Page 4. p. 0
Where the paper says it

Where exposure to posted results can be excluded (217 contrasts completed after an open-weight model’s training cutoff) its protocol-only interventional skill is 0. 12], too imprecise to confirm or exclude memorisation.

Reliability and Interpretation

A reliability layer shows that ranking by predicted error lowers it but that calibrated error control does not transfer under temporal shift on Page 4. The blind reference is conservative because it already encodes class-level knowledge, meaning the interventional skill is a lower bound on Page 4. p. 4

Where the paper says it

It does not isolate it completely: condition, endpoint and eligibility text imply the drug class (a biologic-naïve ulcerative-colitis trial with a remission endpoint is a trial of an advanced therapy), so the blind reference already carries class-level knowledge and IS measures the increment of the specific drug beyond what the rest of the protocol implies. 5 quantifies this and shows that IS is a conservative lower bound.

The paper concludes that interventional skill is a necessary condition for interventional validity, not a sufficient one on Page 4. p. 8
Where the paper says it

1); 7,690 result groups lack a recoverable registered arm type, and the audits were model-assisted with flagged cases reviewed by the author. Interventional skill is a necessary condition for interventional validity, not a sufficient one: positive IS could exploit a correlate of the intervention rather than its mechanism, and says nothing about individual-level counterfactuals.

Responsible Use and Limitations

The work evaluates whether models can predict randomized-trial results from protocols on Page 31. The main misuse risk is the inverse one, quoting held-out accuracy to justify a smaller trial or an external control arm on Page 31. p. 8

Where the paper says it

90 and no error control under shift. The main misuse risk is the inverse one, quoting held-out accuracy to justify a smaller trial or an external control arm; any downstream use should report interventional skill against an audited intervention-blind reference, coverage and width of intervals, and the answered fraction at a pre-fixed error target; no such model should inform a design decision without a prespecified comparison to concurrent controls [U.

Limitations include the benchmark inheriting registry failure modes, such as results being posted for a minority of trials on Page 31. p. 8
Where the paper says it

Limitations. The benchmark inherits the registry’s failure modes: results are posted for a minority of due trials, with sponsor type the strongest predictor [DeVito et al.

The exposure test is underpowered, and the findings are not evidence about the Claude models on Page 38. p. 8
Where the paper says it

The exposure test is underpowered (152 uncontaminated trials; interval ±0. 11), concerns an open-weight model at temperature 0 while the Claude runs used default sampling, and is not evidence about the Claude models, for which no postcutoff stratum exists.

Language Model Specific Gains

The language models show that their skill in predicting treatment effects is in proportion to scale on Page 4 and that the large model's skill reappears on continuous endpoints on Page 4. The authors suggest that IS be reported alongside held-out accuracy, as classifiers report a majority-class baseline on Page 4 and the plausible source of the language model’s skill is pretraining exposure to pharmacology on Page 4. p. 8

Where the paper says it

We suggest that IS be reported alongside held-out accuracy, as classifiers report a majority-class baseline, with a blindness audit of the reference. The plausible source of the language model’s skill is pretraining exposure to pharmacology and published effects of drug classes, which the registry cannot supply: it records what was tested, not what the drug does.

The paper suggests a route for registry-trained simulators (class-level and mechanism priors as inputs) on Page 4. p. 8
Where the paper says it

The plausible source of the language model’s skill is pretraining exposure to pharmacology and published effects of drug classes, which the registry cannot supply: it records what was tested, not what the drug does. That suggests a route for registry-trained simulators (class-level and mechanism priors as inputs) and a caution: real interventional skill with 0.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Measure interventional skill using an intervention-blind reference to determine how much of a simulator's accuracy depends on knowing which intervention was tested, defined as IS = 1 minus MSE(FULL) over MSE(BLIND). This quantifies interventional validity, moving beyond held-out accuracy which may be realistic but wrong about treatment effects.

  2. Develop a registry-trained gradient-boosting simulator that shows no interventional skill while the large language model achieves modest skill (0.20 for protocol-only) and smaller models sit at the limit of detection (0.08). This suggests that registry simulators are only at L1–L2, and a null cannot be attributed to noise.

  3. Implement a reliability layer that shows that ranking by predicted error lowers it but that calibrated error control does not transfer under temporal shift, and use this to select the most confident predictions based on cross-fitted conformal 90% intervals.

  4. Improve language model prompts by using the posted-record prompt, as its interventional skill is higher (0.266) than the protocol-only prompt (0.201), suggesting that the submitted version’s posted-record prompts are more informative for effect estimation on tested data.

  5. Use perturbation tests to verify if a model uses the intervention as an intervention by testing arm swap or substitution, noting that swapping the arms reverses the sign of its prediction in 86% of contrasts.

Abstract

Patient world models are increasingly proposed as simulators of clinical trials, and recent surveys rank their capabilities on a ladder from temporal prediction to counterfactual rollouts. Position on that ladder is currently assigned by reading papers rather than by measurement. We turn the counterfactual rung into a measured quantity. From 30,119 completed phase 2/3 trials with results posted to ClinicalTrials.gov we construct 9,996 audited within-trial contrasts from 7,207 trials, and we score a simulator on how much of its accuracy depends on knowing which intervention was tested. The instrument is an intervention-blind reference: the same model, trained on the same trials, with every intervention-identifying field withheld. Interventional skill is the error the full model removes relative to that reference. On a temporal split (878 test contrasts from 655 trials) a registry-trained gradient-boosting simulator removes a fifth of the error of a trivial baseline, but its interventional skill is-0.12 (95% CI-0.21 to-0.05): withholding the intervention makes it better, and no alternative featurisation, stratum, temporal cut or same-molecule retrieval yields an interval above zero. A frontier language model reading only the registered protocol has interventional skill 0.20 [0.08, 0.31]; smaller and open-weight models sit at the limit of detection. Perturbation tests show the large model uses the intervention as such: swapping the arms reverses the sign of its prediction in 86% of contrasts. The blind reference is itself not drug-blind (a language model recovers the withheld drug class 45% of the time from the blind prompt), so interventional skill is a conservative lower bound. A reliability layer shows that ranking by predicted error lowers it but that calibrated error control does not transfer under temporal shift. Benchmark, code, prompts and per-call model outputs are released.

Sources

Related papers