From Knows to Does: Why LLMs Fail the Clinical Action Test

arXiv cs.AI July 2026
Source: arXiv cs.AIArchive: July 2026
A new systematic review maps large language model performance against Miller's Pyramid, the gold standard for clinical competence. The results are stark: LLMs ace knowledge recall but collapse at the 'Shows How' and 'Does' levels, exposing a critical gap between knowing medicine and practicing it.

A comprehensive review has for the first time aligned the capabilities of large language models (LLMs) with Miller's Pyramid, the foundational framework used in medical education to assess clinical competence from 'Knows' (factual knowledge) to 'Does' (independent clinical action). The findings are a sobering reality check for the medical AI industry. While models like GPT-4 and Claude 3.5 achieve near-perfect scores on medical licensing exam questions (the 'Knows' and 'Knows How' levels), their performance plummets at the 'Shows How' level—which requires demonstrating clinical reasoning in simulated scenarios—and collapses entirely at the 'Does' level, which demands real-time decision-making with actual patients. The study identifies three core bottlenecks: an inability to handle clinical uncertainty, a lack of integration with multimodal patient data (vital signs, imaging, lab trends), and the absence of a safety-constrained decision loop that mimics the iterative judgment of a resident physician. This is not just a technical limitation; it is a product design chasm. The commercial future of medical AI will not belong to the company with the largest knowledge base, but to the one that builds a system with 'clinical reasoning resilience'—one that can recover from mistakes, ask for help, and operate within a hard safety boundary. The study's key insight is that data and compute are no longer the bottleneck; the bottleneck is aligning AI behavior with the inherently chaotic, probabilistic nature of real clinical practice.

Technical Deep Dive

The study constructs a five-level capability framework based on Miller's Pyramid: Knows (factual recall), Knows How (interpretation and explanation), Shows How (demonstration in controlled settings), Does (independent performance in real practice), and an added fifth level, Teaches (ability to mentor others). Each level imposes distinct architectural demands on an LLM.

At the Knows level, standard autoregressive transformers excel. They are trained on vast medical corpora—PubMed articles, textbooks, clinical guidelines—and can retrieve facts with high precision. Benchmarks like MedQA (USMLE Step 1) and MedMCQA show GPT-4 and Claude 3.5 achieving 86-90% accuracy, rivaling top medical students. This is a pattern-matching task: the model has seen the question or a close variant during training.

At the Knows How level, the task shifts to explanation. Models must not just answer but justify their reasoning. Chain-of-thought (CoT) prompting and retrieval-augmented generation (RAG) improve performance here. For example, using a RAG pipeline with a vector database of UpToDate content can boost explanation quality scores by 15-20% on the MedQA explanation subset. However, the model's reasoning is still linear and lacks the ability to weigh competing hypotheses.

The cliff emerges at Shows How. This level involves simulated clinical encounters—e.g., a virtual patient presenting with chest pain, where the model must order tests, interpret results, and adjust a diagnosis in real time. Current LLMs struggle with multi-turn, context-dependent reasoning. They suffer from 'diagnostic anchoring'—fixating on an initial hypothesis and failing to revise it when new contradictory data appears. A 2024 study using the Clinical Encounter Simulator (CES) benchmark found that GPT-4 changed its diagnosis only 23% of the time after receiving contradictory lab results, compared to 71% for human residents.

At the Does level, the model must integrate multimodal data streams: text (history), time-series (vital signs), images (X-rays, CT scans), and structured data (lab values). No current LLM natively handles this fusion. While multimodal models like GPT-4V can process images, they treat each modality independently and lack a unified temporal reasoning engine. For instance, a model might correctly identify pneumonia on a chest X-ray but fail to correlate it with a rising white blood cell count and a falling blood pressure trend—a combination that signals sepsis. This requires a recurrent or state-space model component that current transformer architectures lack.

Relevant Open-Source Work: The MedPrompt repository (5.2k stars) explores prompt engineering for clinical reasoning but has not solved the multi-turn issue. ClinicalBERT (3.8k stars) offers a foundation for time-series modeling but is not designed for generative tasks. BioGPT (Microsoft, 4.1k stars) excels at biomedical text generation but lacks multimodal integration. The most promising direction is Med-PaLM 2 (Google), which uses an ensemble of specialized models—one for text, one for imaging, one for structured data—with a gating network to fuse outputs. However, this architecture is proprietary and computationally expensive.

| Model | Knows (MedQA) | Knows How (Explanation Score) | Shows How (CES Accuracy) | Does (Multimodal Integration) |
|---|---|---|---|---|
| GPT-4 | 90.2% | 82.1% | 23.4% | Not evaluated |
| Claude 3.5 Sonnet | 88.7% | 79.5% | 19.8% | Not evaluated |
| Med-PaLM 2 | 86.5% | 80.3% | 31.2% | 12.7% (simulated) |
| GPT-4V | 89.1% | 81.0% | 22.1% | 9.4% (simulated) |

Data Takeaway: The 'Shows How' and 'Does' columns reveal a catastrophic drop. Even the best model (Med-PaLM 2) achieves only 31.2% on simulated clinical encounters, far below the 70% threshold considered minimally competent for a third-year medical student. The multimodal 'Does' level is essentially unsolved.

Key Players & Case Studies

Several companies are racing to bridge the 'Shows How' and 'Does' gap, each with distinct strategies.

Hippocratic AI has built a 'safety-first' LLM fine-tuned on clinical simulation data. Their approach uses a 'supervisor model' that monitors the primary LLM's outputs and flags unsafe or inconsistent decisions. In a 2024 pilot with a large hospital network, their system reduced diagnostic errors in simulated sepsis cases by 34% compared to GPT-4 alone. However, the supervisor model adds latency (average 2.3 seconds per query) and cost (3x inference cost).

Abridge focuses on the 'Does' level by integrating LLMs into the clinical workflow as a scribe and decision-support tool. Their model runs locally on a hospital's infrastructure, processing patient encounters in real time. The key innovation is a 'closed-loop' design: the model does not make autonomous decisions but instead generates structured summaries and differential diagnoses that a physician must approve. This avoids the liability issue but limits the model's utility to low-risk tasks.

Google DeepMind is pursuing the most ambitious path with Med-PaLM 2, which uses a 'Mixture of Experts' (MoE) architecture. Each expert handles a different data modality (text, imaging, genomics), and a routing network selects the appropriate expert for each query. In internal tests, Med-PaLM 2 achieved 31.2% on the CES benchmark—the highest reported—but still far from clinical readiness. The model requires 8 A100 GPUs for inference, making it impractical for deployment in resource-constrained settings.

| Company | Approach | Key Metric | Deployment Status | Cost per Query |
|---|---|---|---|---|
| Hippocratic AI | Supervisor model + fine-tuning | 34% error reduction in sepsis | Pilot with 3 hospitals | $0.12 |
| Abridge | Local, closed-loop scribe | 92% physician satisfaction | Live in 50+ clinics | $0.04 |
| Google DeepMind | MoE multimodal | 31.2% CES accuracy | Research only | $0.85 |
| OpenAI (GPT-4V) | General-purpose multimodal | 22.1% CES accuracy | API available | $0.15 |

Data Takeaway: No current product achieves clinical-grade performance at the 'Does' level. Abridge's high physician satisfaction is due to its limited scope (scribing, not decision-making). Hippocratic AI's error reduction is promising but measured in simulation, not real patients. The cost and latency of high-performance models remain prohibitive for widespread adoption.

Industry Impact & Market Dynamics

The study's findings will reshape investment and product strategy in medical AI. The global medical AI market is projected to reach $188 billion by 2030, but the current hype is concentrated in the 'Knows' and 'Knows How' levels—chatbots that answer patient questions or summarize medical records. This study reveals that the real value lies in the 'Shows How' and 'Does' levels, which remain largely untapped.

Immediate Impact: Venture capital will shift from general-purpose medical chatbots to companies that demonstrate clinical reasoning resilience. Startups that can show even a 10% improvement on the CES benchmark will attract premium valuations. Established players like Epic Systems and Cerner are already integrating LLM-based decision support, but they are cautious, limiting deployment to non-critical tasks like prior authorization and discharge summaries.

Mid-term (2-4 years): We will see a bifurcation in the market. Low-risk applications (patient education, administrative tasks) will be fully automated by LLMs. High-risk applications (diagnosis, treatment planning) will require human-in-the-loop systems with hard safety constraints. The regulatory environment will be the key determinant: the FDA has not yet approved any LLM for autonomous clinical decision-making, and this study provides evidence that such approval is years away.

Long-term (5+ years): The ultimate winner will be the company that solves the multimodal integration problem. This likely requires a new architecture—perhaps a hybrid of transformers and state-space models (like Mamba) that can handle temporal reasoning across modalities. The first company to achieve 60%+ on the CES benchmark with a deployable model will dominate the market.

| Year | Market Segment | Projected Value | Key Risk |
|---|---|---|---|
| 2025 | Low-risk LLM applications | $12B | Regulatory uncertainty |
| 2027 | Human-in-the-loop clinical support | $45B | Liability concerns |
| 2030 | Autonomous clinical decision-making | $80B | Technical unsolved (Shows How/Does) |

Data Takeaway: The market is betting on a future that the technology cannot yet deliver. The $80B autonomous segment is speculative; realistic projections should discount it by at least 50% until the 'Does' level is addressed.

Risks, Limitations & Open Questions

1. The Liability Trap: If an LLM makes a diagnostic error at the 'Does' level, who is responsible? The physician who overruled the model? The hospital that deployed it? The model developer? Current legal frameworks are unprepared for this. The study's finding that models fail at dynamic decision-making suggests that deploying them in autonomous roles would be reckless.

2. Data Bias in Clinical Simulation: The CES benchmark and similar simulations are built on curated cases that may not reflect real-world diversity. A model that performs well on CES might fail catastrophically when faced with a patient from an underrepresented demographic or a rare disease presentation. The study does not address this, but it is a critical limitation.

3. The 'Black Box' Problem: At the 'Shows How' level, models cannot explain why they changed (or failed to change) their diagnosis. This lack of interpretability is unacceptable in clinical settings where every decision must be justified. Techniques like SHAP and LIME are inadequate for multi-turn reasoning.

4. Open Question: Can LLMs Ever Reach 'Does'? The study implicitly assumes that scaling current architectures will eventually bridge the gap. But there is a growing argument that autoregressive models, by their nature, cannot handle the probabilistic, iterative nature of clinical reasoning. They may need to be replaced or augmented by reinforcement learning from human feedback (RLHF) on clinical trajectories, not just text. This is an active area of research, with the RL4Clinical repository (1.2k stars) exploring this approach.

AINews Verdict & Predictions

Verdict: This study is the most important reality check for medical AI since the release of GPT-4. It draws a clear line between what LLMs can do (ace exams) and what they cannot do (practice medicine). The industry has been conflating the two. The 'Knows' to 'Does' gap is not a minor engineering issue; it is a fundamental architectural limitation.

Prediction 1: The 'Shows How' level will be solved within 3 years. The combination of RLHF on clinical simulation data and supervisor models (like Hippocratic AI's) will push CES accuracy above 60% by 2027. This will enable deployment in controlled settings like telemedicine triage.

Prediction 2: The 'Does' level will remain unsolved for at least 7 years. Multimodal temporal reasoning is a hard AI problem. The first breakthrough will likely come from a non-transformer architecture—perhaps a state-space model or a neuro-symbolic hybrid—not from scaling existing LLMs.

Prediction 3: The commercial winners will be companies that build 'clinical reasoning resilience' into their products. This means systems that can say 'I don't know,' escalate to a human, and learn from mistakes. The companies that prioritize safety over speed will win the trust of regulators and hospitals.

What to Watch: The next major milestone will be the release of a model that achieves >40% on the CES benchmark. When that happens, expect a wave of investment and a flurry of FDA pre-submission meetings. Until then, the prudent strategy is to build for the 'Knows How' level and wait.

More from arXiv cs.AI

UntitledVectorizationLLM is not another chatbot trying to answer everything. It is a purpose-built AI tutor for CTEC 247, an engUntitledInfinity-Parser2 marks a fundamental methodological shift in document parsing. For years, the industry has been trapped UntitledThe integration of large language models into mental health support is accelerating, with platforms like Character.AI, ROpen source hub600 indexed articles from arXiv cs.AI

Archive

July 2026636 published articles

Further Reading

MedCalc-Pro: How Multi-Tool AI Agents Are Reshaping Clinical Decision-MakingMedCalc-Pro introduces a multi-tool agent architecture that allows large language models to orchestrate multiple clinicaDigital Twin & RL: How AI Simulates Treatment Trajectories for Real-Time Clinical OptimizationA novel clinical decision support framework fuses patient-specific digital twins with reinforcement learning to simulateHow Agentic AI Systems Are Building Auditable Medical Evidence Chains to Solve Healthcare's Black Box ProblemA fundamental transformation is underway in medical artificial intelligence. The field is moving beyond black-box modelsDERM-3R AI Framework Bridges Western and Traditional Medicine in DermatologyA new multimodal AI framework called DERM-3R is transforming dermatological practice by integrating Western medical diag

常见问题

这次模型发布“From Knows to Does: Why LLMs Fail the Clinical Action Test”的核心内容是什么?

A comprehensive review has for the first time aligned the capabilities of large language models (LLMs) with Miller's Pyramid, the foundational framework used in medical education t…

从“LLM Miller's Pyramid medical AI limitations”看,这个模型发布为什么重要?

The study constructs a five-level capability framework based on Miller's Pyramid: Knows (factual recall), Knows How (interpretation and explanation), Shows How (demonstration in controlled settings), Does (independent pe…

围绕“clinical decision support LLM failure modes”,这次模型更新对开发者和企业有什么影响?

开发者通常会重点关注能力提升、API 兼容性、成本变化和新场景机会,企业则会更关心可替代性、接入门槛和商业化落地空间。