💻 computer science

CareLoop: Measuring the Gap Between Knowing Medicine and Practicing It in the Context of Patients' Lives

This paper introduces CareLoop, a clinical-world sandbox that evaluates AI models on their ability to practice medicine within the complex context of patients' lives by measuring their performance in discovering hidden states, adapting to constraints, and managing consequences across simulated trajectories, revealing that execution-grounded capabilities are distinct from and more challenging than simply knowing medical facts.

Original authors: Zhenhong Yang, Jintao Fei, Jun Zhao

Published 2026-09-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhenhong Yang, Jintao Fei, Jun Zhao

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Medicine is often taught as a collection of facts: a list of symptoms, a catalog of treatments, and a set of rules for diagnosis. In this view, a doctor's job is to retrieve the correct answer from memory and apply it. But in the real world, medicine is not a puzzle with a single solution waiting to be found; it is a conversation with a person whose life, beliefs, and circumstances constantly reshape the path forward. A patient might misunderstand a doctor's instructions, hide a painful detail out of shame, or simply lack the money or transportation to get a prescribed test. A medical plan that looks perfect on paper can fail if it does not account for these human realities. The challenge for artificial intelligence is not just to know the right medical facts, but to navigate the messy, unpredictable space between knowing those facts and actually helping a person.

Researchers have long tried to test whether computer programs can answer medical questions correctly. These tests usually ask a simple question: does the AI know the right diagnosis? However, a new study introduces a different kind of test, one that asks whether an AI can practice medicine in the context of a patient's life. The researchers built a simulated environment called CareLoop, designed to measure the gap between knowing medicine and practicing it. Instead of giving the AI a static list of symptoms to solve, they created a dynamic world where patients have their own memories, beliefs, and limitations. In this world, information is hidden, evidence can be delayed, and actions taken by the AI do not always lead to the expected result. The goal was to see if an AI could discover missing information, correct misunderstandings, adapt to real-world barriers, and take responsibility for a patient's care over time, rather than just offering a correct answer at the start of a conversation.

The study involved creating 120 detailed scenarios based on real, anonymized medical records. Each scenario was a contract that defined the hidden state of a patient's health, what the patient and their family believed, and what obstacles might appear. These obstacles included things like a patient misremembering their medication, a lab result arriving late, a family member refusing a treatment, or a patient being unable to afford a visit. The researchers then asked ten different AI models to act as doctors in these simulations. The models had to interact with the simulated patients and family members, who were programmed to be imperfect: they could forget instructions, lie about symptoms, or get confused. The AI had to figure out what was really happening, verify the information, and guide the patient toward a safe outcome.

The results showed a clear difference between models that simply knew medical facts and those that could navigate the complexity of a patient's life. The best-performing model, GPT-5.6 Sol, achieved the highest scores in handling real-world frictions, though its lead over the next tier of models (including GPT-5.4, DeepSeek-V4-Pro, and Qwen3.8-Max) was not statistically distinct due to overlapping confidence intervals. The study revealed that even the top models struggled with specific tasks. The biggest challenge for all of them was discovering hidden states—figuring out facts that the patient had not volunteered or that were buried in the background, which the paper identifies as the largest shared bottleneck. Another common failure was the "action-to-impact" gap: the models often took an action, like suggesting a test, but failed to ensure the test was actually done or to follow up on the results. In many cases, the AI would declare the conversation finished even though the patient's problem was not truly resolved, a mistake known as premature closure.

The study also compared how well the AI models performed against human doctors. A panel of 139 physicians reviewed the same simulated cases and ranked the AI models. While individual doctors sometimes disagreed with each other, and sometimes disagreed with the AI evaluators, the overall rankings of the models were remarkably stable. When the researchers looked at the aggregate results, the human doctors and the AI judges agreed on which models were the best and which were the worst. This suggests that the new testing method is reliable enough to distinguish between different levels of capability, even when the task is complex and the evaluation is subjective.

One of the most important findings was that finishing a conversation quickly is not the same as providing good care. Many of the AI models ended their interactions early, marking the task as complete, even though the patient still had unresolved risks or unverified information. The best models were the ones that knew when to keep an episode open, maintaining responsibility for the patient until the situation was truly safe to close. This distinction highlights a fundamental shift in how we should evaluate medical AI. It is not enough for a system to be fluent or to give a correct diagnosis in isolation. True competence in a medical setting requires the ability to manage uncertainty, verify that plans are carried out, and stay accountable for a patient's well-being even when the path is unclear.

The researchers emphasize that these results come from a simulation, not a real hospital. The study does not prove that these AI models are ready to treat patients or that they are safe for clinical use. Instead, it provides a way to measure how close they are to that goal. By exposing the specific ways in which AI fails to bridge the gap between knowledge and practice, the study offers a roadmap for improvement. It shows that the next generation of medical AI needs to be better at listening, at verifying, and at understanding that a patient's life is full of obstacles that no amount of textbook knowledge can automatically overcome. The path forward is not just about making AI smarter, but about making it more capable of walking alongside a person through the complexities of their own health journey.

Technical Summary: CareLoop – Measuring the Gap Between Knowing Medicine and Practicing It

1. Problem Statement

Current medical AI benchmarks primarily evaluate whether a model "knows" the correct medical answer (e.g., diagnosis, treatment guidelines) in static, fully observable settings. However, patient-facing care occurs in a partially observable, socio-technical world where success depends on navigating incomplete information, conflicting beliefs, real-world constraints, and delayed evidence.

The paper identifies a critical gap between knowing medicine (factual correctness) and practicing medicine (exercising agency under limited control). In real-world scenarios, a model may possess correct medical knowledge but fail to:

Uncover hidden or incorrect patient/family beliefs.
Adapt plans to feasibility constraints (cost, transport, adherence).
Verify the execution of external actions (e.g., lab results, referrals).
Maintain accountability for unresolved risks across a trajectory.
Distinguish between operational termination and safe clinical closure.

Existing agent environments often lack the specific "friction" mechanisms required to expose these latent limitations, or they focus on single-turn interactions rather than longitudinal care progression.

2. Methodology: The CareLoop Framework

CareLoop is a clinical-world sandbox designed to measure this gap through trajectory-based evaluation. It treats a clinical episode as a partially observable, multi-actor process where state is distributed across four diverging dimensions:

Underlying Clinical State ( $S_{truth}$ ): The actual medical reality.
Belief State ( $S_{belief}$ ): What the patient/family believes (which may be incorrect).
Recorded Evidence ( $S_{record}$ ): What is visible in the workspace (EHR, documents).
Executed State ( $S_{exec}$ ): What has actually been done in the world.

2.1 Case Contracts and Friction

The core artifact is the CL120 dataset: 120 frozen, de-identified EHR-derived case contracts. Before model execution, each case is authored with:

Hidden State: Facts not initially disclosed (e.g., hidden symptoms, medication errors).
Friction Strata: 10 types of real-world obstacles (e.g., misstated information, document quality issues, cost barriers, privacy concerns).
High-Order (HO) Opportunities: 480 specific response capabilities the model is expected to demonstrate (e.g., hidden-state discovery, misinformation correction, dynamic reprioritization).
Terminal Criteria: Rules for "safe bounded closure" vs. "premature closure."

2.2 Simulation Runtime

The environment simulates a multi-turn interaction involving:

Doctor Model: The AI agent providing guidance.
Patient/Family Actor: An independent, stateful agent that may misunderstand, refuse, or disclose selectively.
Care Workspace: A dynamic evidence channel where results, receipts, and external actions appear over time.
World Progression Layer: Simulates stochastic events (complications, delays) anchored in clinical causality.

The system is designed so that latent limitations (e.g., failing to verify a report) produce auditable consequences later in the trajectory (e.g., an obsolete plan leading to a safety risk).

2.3 Evaluation Metrics

The study evaluated 10 models (including GPT-5.6 Sol, DeepSeek-V4-Pro, Qwen3.8-Max) across 1,200 trajectories. Assessments were performed by:

4 Blinded LLM Judges: Generating 4,800 structured assessments.
139 Physicians: Generating 6,000 structured assessments.

Key Metrics:

Friction Capability Composite (FCC): A weighted mean (0–1) of six capability domains:
1. Information Repair
2. Constraint Navigation
3. Dynamic Reprioritization
4. Execution-Loop Repair
5. State-Sensitive Communication
6. Bounded Closure/Continuity
C-RWR (CareLoop Real-World Robustness): FCC multiplied by a safety gate (0.35 for serious errors, 0.75 for minor/moderate, 1.0 for none).
Action-to-Impact Analysis: Measures the conversion rate from "realized opportunity" $\to$ "active action" $\to$ "meaningful trajectory impact" $\to$ "effective completion."

3. Key Contributions

Patient-Centered Task Ontology: Formulates care as "agency under limited control" in a distributed, partially observable world, moving beyond static Q&A to longitudinal consequence stewardship.
Capability-Eliciting Environment: Introduces "friction" and "behavioral forks" authored before execution to force latent deficiencies to manifest as observable trajectory failures.
Trajectory-Centered Measurement System: Distinguishes between attempted action, realized consequence, and safe continuity, introducing metrics like FCC and C-RWR that penalize unsafe closures and unverified actions.
Large-Scale Empirical Study: A reproducible benchmark involving 120 cases, 1,200 trajectories, and over 10,000 structured assessments, demonstrating that stable model rankings can be recovered despite local evaluator disagreement.

4. Results

4.1 Model Performance

Ranking: GPT-5.6 Sol ranked first with an FCC of 0.824 and C-RWR of 0.753. It was followed by GPT-5.4 (0.786 FCC) and DeepSeek-V4-Pro (0.783 FCC).
Capability Gaps: The largest between-model differences occurred in Information Repair (range 0.228), Dynamic Reprioritization (0.221), and Bounded Closure (0.216). Models generally performed better in Constraint Navigation and State-Sensitive Communication.
Safety Gating: The C-RWR scores were lower than FCC for most models, indicating that clinical errors (e.g., unsafe closures, missed verification) significantly impacted robustness scores.

4.2 The Action-to-Impact Gap

Across 15,537 assessments of realized opportunities:

Models took active action in 93.5% of cases.
Actions produced meaningful trajectory impact in only 74.5%.
Effective completion occurred in 69.7%.
Hidden-state discovery was the largest shared bottleneck (effective completion: 48.4%), followed by safe bounded closure (63.6%).

4.3 Termination vs. Closure

Operational termination (e.g., reaching a turn limit or a "closed" flag) did not equate to safe clinical closure.

437 "closed" trajectories received at least one "premature closure" judgment from judges.
Models with high closure rates (e.g., Doubao-Seed-2.0-lite) often ranked lower in FCC, while top-ranked models (GPT-5.6 Sol) maintained "open" states when evidence was pending, demonstrating accountable openness.

4.4 Stability and Agreement

Ranking Stability: Despite local evaluator disagreement (single-judge ICC $\approx$ 0.25), model rankings stabilized rapidly with case coverage. Disjoint 60-case halves yielded Spearman correlations of 0.915 (FCC) and 0.927 (C-RWR).
Physician Alignment: Aggregate rankings from 139 physicians converged with the four-Judge median rankings ( $\rho = 0.988$ for FCC), validating the structured evaluation protocol.

5. Significance and Claims

The paper claims that CareLoop successfully operationalizes the evaluation of medical agents in a context that mirrors the complexity of real-world practice. Its significance lies in shifting the evaluation paradigm:

From Answers to Consequences: It demonstrates that correct medical knowledge is insufficient if the agent cannot verify execution, adapt to constraints, or manage unresolved risk.
From Terminal State to Trajectory: It argues that "safe closure" is not a binary end-state but a property of the entire trajectory, requiring accountable openness when evidence is pending.
From Control to Stewardship: It reframes medical AI not as an entity with full control, but as a steward of consequence who must navigate limited authority, independent actors, and delayed feedback.

The authors explicitly state that these findings do not establish clinical effectiveness or deployment safety. Instead, CareLoop provides a reproducible, friction-based environment to measure the "knowledge-to-practice" gap and identify specific capability deficits (particularly in information repair and dynamic updating) that must be addressed before medical agents can be trusted in open human worlds.