Didactic knowledge or Clinical Cases? How Data Types Shape Medical Large Language Models

# Didactic knowledge or Clinical Cases? How Data Types Shape Medical Large Language Models
Source: [https://arxiv.org/html/2609.22161](https://arxiv.org/html/2609.22161)
Haochun Wang⋆Affiliation:Research Center for Social Computing and Interactive Robotics,Affiliation:Harbin Institute of Technology, Harbin 150001, ChinaSendong Zhao†Affiliation:Research Center for Social Computing and Interactive Robotics,Affiliation:E\-mail[\{hcwang,sdzhao\}@ir\.hit\.edu\.cn](mailto:{hcwang,sdzhao}@ir.hit.edu.cn)Xiao HanAffiliation:Research Center for Social Computing and Interactive Robotics,Ming MaAffiliation:Research Center for Social Computing and Interactive Robotics,Bing QinAffiliation:Research Center for Social Computing and Interactive Robotics,

###### Abstract

Medical large language models are commonly trained on mixtures of didactic data \(e\.g\., textbooks\) and clinical data \(e\.g\., patient records\), yet how these data types differentially shape model capabilities remains unclear\. We address this issue with token\-matched experiments that vary the didactic\-to\-clinical ratio and analyze how data composition affects performance, capability profiles, and error patterns across knowledge\-intensive and clinic\-oriented tasks\. We uncover an asymmetric transfer across task types: clinical data improves clinic\-oriented tasks while remaining competitive on knowledge\-intensive ones, whereas didactic data mainly improves knowledge\-intensive tasks\. Error analysis suggests a knowing–doing gap, where improvements in knowledge recall do not reliably generalize to clinical reasoning\. We further observe that modest amounts of clinical data yield most of the gains on EHR\-grounded tasks, while the optimal mixture ratio varies with the knowledge and clinical reasoning demands of downstream tasks\. These findings suggest that medical LLM data curation should be application\-driven, with higher proportions of clinical data preferred for reasoning\-intensive use cases\.

###### Keywords:

Medical LLM Training corpora composition EHR\-grounded benchmarking\.

$\\star$$\\star$footnotetext:Equal contribution\.$\\dagger$$\\dagger$footnotetext:Corresponding author\.## 1Introduction

Medical large language models \(LLMs\) are trained on heterogeneous data sources—textbooks, guidelines, exam questions, and electronic health records \(EHRs\)—often pooled into unified corpora\. This recipe has delivered strong benchmark performance\[[20](https://arxiv.org/html/2609.22161#bib.bib1)\]\. However, it implicitly treats different training data sources as interchangeable\. A fundamental question therefore remains underexplored:how do different types of medical data shape the capabilities of medical LLMs?

![Refer to caption](https://arxiv.org/html/2609.22161v1/example.png)Figure 1:Examples of didactic and clinical case\-based data across two QA formats\.Crucially, medical knowledge and practice evolve rapidly as new drugs, guidelines, and treatment protocols emerge, which motivates continual training and periodic updates to maintain currency in deployed systems\. In this setting, understanding how different data types contribute to deployment\-relevant capabilities is crucial\. Data sources differ not only in surface form, but also in the information they provide and the reasoning they elicit\. For example, a textbook passage on heart failure typically presents decontextualized mechanisms and definitions, whereas an ICU note about acute decompensation is a patient\-grounded narrative where evidence unfolds over time and decisions are made under incomplete information\. If such differences are systematic, then data composition may shape the capabilities of medical LLMs, leading to uneven gains across task types and deployment settings\. To study this effect, we categorize medical corpora into two types as shown in Figure[1](https://arxiv.org/html/2609.22161#S1.F1)\. We definedidactic dataas decontextualized medical content without patient context, andclinical dataas patient\-grounded scenarios in the training corpora of medical LLMs, inspired by the declarative and procedural knowledge in medical education\[[1](https://arxiv.org/html/2609.22161#bib.bib10),[19](https://arxiv.org/html/2609.22161#bib.bib9)\]\.

Prior work adapts LLMs to medicine via continual pretraining and fine\-tuning on mixture of didactic and clinical corpora\[[7](https://arxiv.org/html/2609.22161#bib.bib26)\]\. However, the contribution of each data type under fixed training budgets remains unclear\. To address this, we train model variants under token\-matched conditions, varying only the didactic\-to\-clinical ratio\. We evaluate on benchmarks stratified into knowledge\-intensive and clinic\-oriented tasks\. Clinic\-oriented evaluation includes reasoning\-intensive MCQs and ClinicalBench, a stagewise EHR\-grounded suite construct from MIMIC\-III\[[12](https://arxiv.org/html/2609.22161#bib.bib13)\]to assess models under realistic information constraints\.

Our experiments revealasymmetric transferacross task types: increasing the share of clinical data improves performance on both clinic\-oriented and knowledge\-intensive tasks, whereas didactic data primarily benefits knowledge\-intensive evaluation\. This transfer is even more pronounced on EHR\-grounded tasks, where didactic data yields limited gains\. We also observe that a small amount of clinical data captures most of the improvements\. To understand the mechanisms behind these patterns, we investigate three research questions \(RQ\):

- •RQ1: Do didactic and clinical data induce different reasoning directions?
- •RQ2: How does data composition interact with the knowledge and reasoning demands of downstream tasks?
- •RQ3: Do didactic and clinical data lead to different error patterns?

Our contributions can be summarized as follows:

\(1\) We introduce a token\-matched experimental framework that isolates data composition as the causal variable, enabling controlled analysis of how different medical data corpora and mixture ratios shape the capabilities of medical LLMs\.

\(2\) We construct ClinicalBench, a suite of clinical tasks derived from EHRs with temporal masking, to evaluate clinical reasoning under realistic information constraints\.

\(3\) We provide fine\-grained analysis revealing how data composition correlates with reasoning direction specialization and error distributions, offering mechanistic insights into the effects of different medical data types\.

## 2Related Works

### 2\.1Medical LLMs

Medical LLMs differ primarily in their base architectures and training data source\. Didactic\-centric models train on formalized knowledge like PubMed papers\[[16](https://arxiv.org/html/2609.22161#bib.bib2),[14](https://arxiv.org/html/2609.22161#bib.bib22)\], textbooks\[[25](https://arxiv.org/html/2609.22161#bib.bib4)\], and structured medical knowledge bases\[[23](https://arxiv.org/html/2609.22161#bib.bib5)\], while clinical\-centric models leverage real\-world clinical narratives\[[28](https://arxiv.org/html/2609.22161#bib.bib3)\]and EHRs\[[22](https://arxiv.org/html/2609.22161#bib.bib23)\]\. Recent efforts further incorporate clinical dialogues and multimodal data\[[27](https://arxiv.org/html/2609.22161#bib.bib21),[21](https://arxiv.org/html/2609.22161#bib.bib18),[15](https://arxiv.org/html/2609.22161#bib.bib19)\], with complementary approaches augmenting clinical reasoning through experience or memory\-based mechanisms\[[8](https://arxiv.org/html/2609.22161#bib.bib27)\]\. While prior works generally mix both corpora, the specific impact of each source remains underexplored\. We explicitly isolate these data types to examine their distinct effects on model capabilities\.

### 2\.2Evaluation Paradigms for Medical LLMs

Medical LLMs are commonly evaluated on exam\-style MCQ benchmarks testing knowledge recall\[[10](https://arxiv.org/html/2609.22161#bib.bib14),[5](https://arxiv.org/html/2609.22161#bib.bib25),[20](https://arxiv.org/html/2609.22161#bib.bib1)\]\. However, traditional n\-gram metrics fail to capture clinical nuance, spurring rubric\-based evaluation\[[4](https://arxiv.org/html/2609.22161#bib.bib6),[2](https://arxiv.org/html/2609.22161#bib.bib7)\]\. EHR\-grounded benchmarks\[[13](https://arxiv.org/html/2609.22161#bib.bib8),[18](https://arxiv.org/html/2609.22161#bib.bib24)\]further extend evaluation to evidence\-based QA and clinical text generation\. Overall, the evaluation of medical LLMs is shifting from multiple\-choice accuracy to multi\-dimensional frameworks spanning QA suites, multi\-turn dialogue, and EHR\-based tasks, with increasing emphasis on safety and clinical utility, including the epistemic robustness of models whose correct beliefs may collapse under clinical pressure\[[26](https://arxiv.org/html/2609.22161#bib.bib28)\]\.

## 3Study Design: Data Typology, Evaluation, and Analysis

![Refer to caption](https://arxiv.org/html/2609.22161v1/overall.png)Figure 2:Overview of the study setup, including \(a\) the training corpora, \(b\) model variants fine\-tuned on different data compositions, \(c\) evaluation benchmarks, and \(d\) comparative performance results for Qwen2\.5\-7B\.In this section, we present a comparative framework to analyze how different data types influence medical LLM capabilities\. We distinguish between didactic and clinical data, introduce a set of evaluation benchmarks, and describe analytical methods for deriving mechanistic insights into performance\. Figure[2](https://arxiv.org/html/2609.22161#S3.F2)provides an overview of this setup\.

### 3\.1Data Corpora

We categorize medical corpora into two functional types, inspired by the distinction between declarative and procedural knowledge in medical education\[[19](https://arxiv.org/html/2609.22161#bib.bib9)\]\.

##### Didactic Data

This category comprises training examples that emphasize generalizable medical knowledge without patient\-specific context, focusing onwhat is trueabout medical concepts rather thanhowto apply them in clinical practice\. We assemble didactic data from four sources: MedMCQA \(Didactic\)\[[17](https://arxiv.org/html/2609.22161#bib.bib12)\], MedQuAD\[[3](https://arxiv.org/html/2609.22161#bib.bib11)\], MedicationQA, and explanatory passages and derived QA pairs from openly licensed medical textbooks\.

##### Clinical Data

Clinical data embeds medical knowledge within clinical narratives characterized by temporal progression, incomplete information, diagnostic uncertainty, and the practical constraints of healthcare delivery\. We construct clinical data from two sources\. \(1\) MedMCQA \(Clinical\): vignette\-containing questions from MedMCQA presenting patient demographics, symptoms, or examination findings followed by diagnostic or management queries\. These questions provide examples of systematic clinical reasoning\. \(2\) MIMIC\-III with Temporal Care Simulation: We apply a temporal masking strategy to MIMIC\-III EHRs to simulate real\-world clinical dynamics\. At each care stage \(admission, hospitalization, discharge\), the model accesses only information available at that point, forcing medical LLMs to reason under uncertainty—the central challenge of clinical decision\-making\. We design seven task types covering diagnosis, treatment, and disposition prediction\.

### 3\.2Medical LLM Capabilities

We characterize medical LLM capabilities along two primary dimensions\.Knowledge Recallis the ability to retrieve and correctly apply canonical, decontextualized medical knowledge, including definitions, diagnostic criteria ,and pathophysiological mechanisms\.Clinical Reasoningis the ability to integrate patient\-specific information and reason under uncertainty, including synthesizing findings across a narrative, tracking temporal evolution, and producing context\-appropriate decisions\.

### 3\.3Evaluation for Model Variants

We organize benchmarks by capability demand into knowledge\-intensive and clinic\-oriented tasks\. Clinic\-oriented evaluation covers both MCQ\-based and EHR\-grounded settings\.

##### Knowledge\-Intensive Benchmarks

MCQ benchmarks test knowledge recall and conceptual understanding ability, including MedQA Step 1\[[10](https://arxiv.org/html/2609.22161#bib.bib14)\], MedMCQA\-Didactic, MMLU\-Medical\[[9](https://arxiv.org/html/2609.22161#bib.bib16)\], and PubMedQA\[[11](https://arxiv.org/html/2609.22161#bib.bib15)\]\.

##### Reasoning\-Intensive Benchmarks

These benchmarks require clinical reasoning and decision\-making ability over patient vignettes, including MedQA Step 2&3, MedMCQA\-Clinical, and Medbullets\[[6](https://arxiv.org/html/2609.22161#bib.bib17)\]\.

##### ClinicalBench \(EHR\-Grounded\)

We introduce ClinicalBench, constructed from MIMIC\-III under the temporal masking framework, where models generate clinical decisions from patient EHRs under realistic information constraints\. It comprises five tasks:Service Prediction\(Svc\),Primary Diagnosis\(Pdx\),Discharge Disposition\(Dd\),Discharge Medication\(Med\), andDischarge Instruction\(Di\)\.

Table 1:Performance on knowledge\-intensive and reasoning\-intensive benchmarks across model variants\.Boldindicates the best performance\.Base ModelModel VariantMedQAMedMCQAPubMedQAMMLUMedbullets\-5step1step2&3AllDidacticClinicalQwen2\.5\-7BBase50\.2250\.6750\.4355\.8059\.6053\.0070\.9843\.18Instruct\-𝒟\\mathcal\{D\}55\.6757\.0756\.3268\.6067\.2064\.8076\.8546\.10Instruct\-𝒞\\mathcal\{C\}55\.3857\.5856\.4067\.6071\.6070\.4076\.6846\.75Instruct\-ℳ1:3\\mathcal\{M\}\_\{1:3\}55\.9657\.0756\.4866\.8070\.0072\.8076\.2148\.70Instruct\-ℳ1:1\\mathcal\{M\}\_\{1:1\}57\.4456\.9057\.1969\.2069\.0073\.0076\.1247\.40Instruct\-ℳ3:1\\mathcal\{M\}\_\{3:1\}55\.8256\.9056\.3268\.8068\.8074\.4077\.1349\.03Llama3\.1\-8BBase52\.4355\.7253\.9769\.4068\.0075\.0070\.7146\.43Instruct\-𝒟\\mathcal\{D\}51\.5555\.3953\.3473\.0069\.2074\.6072\.6445\.78Instruct\-𝒞\\mathcal\{C\}51\.1058\.4254\.5272\.6072\.2071\.8071\.0744\.16Instruct\-ℳ1:3\\mathcal\{M\}\_\{1:3\}53\.4654\.3853\.8974\.6070\.6075\.2071\.4447\.08Instruct\-ℳ1:1\\mathcal\{M\}\_\{1:1\}52\.2856\.2354\.1271\.6071\.6074\.8071\.8145\.78Instruct\-ℳ3:1\\mathcal\{M\}\_\{3:1\}54\.2056\.4055\.2273\.6071\.8074\.0071\.2647\.40

Table 2:Performance on ClinicalBench across model variants\.∗indicates official instruct model\.Boldindicates the best performance per task within each base model\.Base ModelModel VariantServicePrimaryDiagnosisMedicationDischargeDispositionDischargeInstructionF1JaccardQwen2\.5\-7BBase34\.0072\.0019\.3710\.7215\.002\.14Instruct∗50\.0075\.0016\.018\.7043\.002\.51Instruct\-𝒟\\mathcal\{D\}59\.0076\.0016\.158\.7826\.002\.08Instruct\-𝒞\\mathcal\{C\}73\.0083\.0030\.1617\.7652\.002\.53Instruct\-ℳ1:3\\mathcal\{M\}\_\{1:3\}78\.0080\.0026\.9315\.5649\.002\.49Instruct\-ℳ1:1\\mathcal\{M\}\_\{1:1\}73\.0082\.0028\.1816\.4053\.002\.57Instruct\-ℳ3:1\\mathcal\{M\}\_\{3:1\}70\.0083\.0023\.0213\.0153\.002\.30Llama3\.1\-8BBase68\.0056\.0014\.257\.6716\.001\.54Instruct∗61\.0062\.0015\.048\.1344\.002\.58Instruct\-𝒟\\mathcal\{D\}77\.0055\.0017\.879\.8131\.001\.87Instruct\-𝒞\\mathcal\{C\}81\.0072\.0021\.4312\.0057\.002\.45Instruct\-ℳ1:3\\mathcal\{M\}\_\{1:3\}80\.0076\.0019\.7510\.9640\.002\.79Instruct\-ℳ1:1\\mathcal\{M\}\_\{1:1\}78\.0060\.0022\.5712\.7252\.002\.48Instruct\-ℳ3:1\\mathcal\{M\}\_\{3:1\}72\.0073\.0019\.4310\.7646\.002\.14

### 3\.4Analysis Beyond Performance

Beyond aggregate benchmark scores, we conduct three complementary analyses to investigate the mechanisms underlying performance differences between didactic and clinical training corpora: \(i\) probing directional reasoning preferences \(RQ1\), \(ii\) examining how task demands interact with data composition \(RQ2\), and \(iii\) characterizing failure patterns \(RQ3\)\.

##### Directional Reasoning

Didactic data encouragesConcept→\\rightarrowAttributesreasoning \(mapping concepts to observable properties\), while clinical data emphasizesEvidence→\\rightarrowDiagnosisinference \(from findings to conclusions\)\. We operationalize each direction with two probing tasks: Term→\\rightarrowSymptoms and Term→\\rightarrowDefinition for the former; Symptoms→\\rightarrowDiagnosis and EHR→\\rightarrowDiagnosis \(Pdx\) for the latter\. To quantify directional preference, we define the Direction Bias Index \(DBI\):

DBI=S¯C→A−S¯E→DS¯all\\text\{DBI\}=\\frac\{\\bar\{S\}\_\{\\text\{C\}\\rightarrow\\text\{A\}\}\-\\bar\{S\}\_\{\\text\{E\}\\rightarrow\\text\{D\}\}\}\{\\bar\{S\}\_\{\\text\{all\}\}\}\(1\)whereS¯C→A\\bar\{S\}\_\{\\text\{C\}\\rightarrow\\text\{A\}\}andS¯E→D\\bar\{S\}\_\{\\text\{E\}\\rightarrow\\text\{D\}\}are mean scores for each direction andS¯all\\bar\{S\}\_\{\\text\{all\}\}is the overall mean\. Positive DBI indicates concept\-oriented bias; negative indicates evidence\-to\-diagnosis preference\.

##### Demand Decomposition

Aggregate benchmark scores quantify overall performance, but they do not reveal which demand profiles are most sensitive to corpus composition\. To localize these effects, we decompose each MCQ item along two axes: knowledge demand \(KK\) and reasoning demand \(RR\)\. Then, we partition the evaluation set into four quadrants \(K±​R±K^\{\\pm\}R^\{\\pm\}\)\. We report accuracy for each quadrant and assess \(i\) which data composition results in the largest improvements within each quadrant, and \(ii\) whether mixed corpora exhibit complementarity beyond a ratio\-weighted linear expectation\.

##### Failure Analysis

Accuracy indicates whether a model answers correctly, but it does not explain what types of error are made\. To test whether corpus composition is associated with systematic differences in failure patterns, we annotate each incorrectly answered question with a single primary error cause\. We use four categories: knowledge errors \(EKnow\.E\_\{\\text\{Know\.\}\}\), reasoning errors \(EReas\.E\_\{\\text\{Reas\.\}\}\), context errors \(ECtx\.E\_\{\\text\{Ctx\.\}\}\), and format errors \(EFmt\.E\_\{\\text\{Fmt\.\}\}\)\.

## 4Experiments

### 4\.1Experimental Setup

##### Base Models

We conduct experiments on Qwen2\.5\-7B\-Base and Llama3\.1\-8B\-Base\. We choose base models rather than instruction\-tuned variants to ensure a clean experimental setup\. All observed capabilities can be attributed to our controlled supervision\.

##### Training Variants

We train five variants for each base model: didactic\-only \(𝒟\\mathcal\{D\}\), clinical\-only \(𝒞\\mathcal\{C\}\), and three mixtures at ratios 1:3, 1:1, and 3:1 \(ℳ1:3\\mathcal\{M\}\_\{1:3\},ℳ1:1\\mathcal\{M\}\_\{1:1\},ℳ3:1\\mathcal\{M\}\_\{3:1\}\), all under the same token budget, schedule, and hyperparameters\.

##### Evaluation Metrics

For MCQs and ClinicalBench classification tasks \(Svc,Pdx,Dd\), we report accuracy\. For taskMed, we use Jaccard and micro\-F1; for taskDi, we adopt multi\-dimensional GPT\-based evaluation\[[24](https://arxiv.org/html/2609.22161#bib.bib20)\]assessing accuracy, clarity, faithfulness, and completeness\.

### 4\.2Overall Performance

Tables[1](https://arxiv.org/html/2609.22161#S3.T1)and[2](https://arxiv.org/html/2609.22161#S3.T2)present results across medical QA benchmarks and ClinicalBench\. The two corpora yield asymmetric transfer across evaluation settings\. On knowledge\-intensive QA, Instruct\-𝒟\\mathcal\{D\}holds a slight advantage on some benchmarks \(e\.g\., MMLU\-Medical\), while Instruct\-𝒞\\mathcal\{C\}performs better on reasoning\-intensive MCQs and dominates across all ClinicalBench tasks\. Instruct\-𝒟\\mathcal\{D\}yields limited and unstable gains on EHR\-grounded settings, whereas Instruct\-𝒞\\mathcal\{C\}remains competitive on knowledge\-intensive QA while excelling in EHR contexts\. This reveals an*asymmetric transfer*: clinical data supports both task types, while didactic data primarily benefits knowledge\-intensive evaluation—suggesting clinical data reinforces factual knowledge through application, whereas didactic data does not cultivate the procedural competencies required for EHR\-grounded reasoning\. Mixed corpora provide complementary benefits, with mixture variants outperforming both single\-source counterparts on several benchmarks\. On ClinicalBench, a small clinical fraction captures most of the improvement over Instruct\-𝒟\\mathcal\{D\}, with diminishing returns from further increases\. The optimal ratio is task\-dependent: Instruct\-ℳ1:3\\mathcal\{M\}\_\{1:3\}favors EHR\-grounded tasks, while Instruct\-ℳ3:1\\mathcal\{M\}\_\{3:1\}is competitive on knowledge\-intensive QA—suggesting corpus composition should be selected based on the target use case\.

### 4\.3Directional Reasoning Analysis \(RQ1\)

To avoid confounding data\-type effects with coverage differences, we restrict this analysis to diseases appearing in both corpora\. We evaluate Symptoms→\\rightarrowDiagnosis on both typical and atypical symptom sets, using the atypical subset to probe robustness to clinical presentations\. As shown in Table[3](https://arxiv.org/html/2609.22161#S4.T3), results confirm our hypothesis: Instruct\-𝒟\\mathcal\{D\}outperforms Instruct\-𝒞\\mathcal\{C\}onConcept→\\rightarrowAttributestasks, while Instruct\-𝒞\\mathcal\{C\}excels onEvidence→\\rightarrowDiagnosistasks across both typical and atypical presentations\. The DBI captures this shift: starting from a near\-balanced base \(DBI≈0\\approx 0\), Instruct\-𝒟\\mathcal\{D\}yields a positive DBI and Instruct\-𝒞\\mathcal\{C\}a negative DBI, while Instruct\-ℳ1:1\\mathcal\{M\}\_\{1:1\}remains near zero\. This directional specialization explains the asymmetric transfer in Section[4\.2](https://arxiv.org/html/2609.22161#S4.SS2): clinical data implicitly exercises concept knowledge through diagnostic application, enabling transfer to knowledge\-intensive tasks; didactic data lacking such inferential practice does not develop the synthesis capabilities required for clinical reasoning\.

Table 3:Performance on each directional mapping task on Qwen2\.5\-7B\. A positive DBI indicates concept\-oriented bias; a negative indicates evidence\-to\-diagnosis preference\. Sym\., Def\., and Diag\. stand for Symptoms, Definition, and Diagnosis respectively\.Boldindicates the best performance across model variants\.Model VariantTerm→\\rightarrowSym\.Term→\\rightarrowDef\.Sym\.→\\rightarrowDiag\.EHR→\\rightarrowDiag\.DBITypicalAtypicalBase53\.1841\.9826\.8317\.5672\.00\+0\.01Instruct\-𝒟\\mathcal\{D\}64\.7048\.4030\.2424\.3976\.00\+0\.09Instruct\-𝒞\\mathcal\{C\}57\.7247\.0035\.1228\.2983\.00\-0\.09Instruct\-ℳ1:1\\mathcal\{M\}\_\{1:1\}62\.8046\.7234\.6324\.3982\.00\-0\.02
### 4\.4Quadrant\-Wise Analysis \(RQ2\)

##### Quadrant\-Wise Performance

Table[4](https://arxiv.org/html/2609.22161#S4.T4)suggests that different quadrants favor different data compositions\. In the reasoning\-dominant quadrant \(K−​R\+K^\{\-\}R^\{\+\}\), Instruct\-ℳ1:3\\mathcal\{M\}\_\{1:3\}achieves the highest accuracy, which is consistent with clinical data being more helpful when inference is required despite limited explicit knowledge in the prompt\. In the knowledge\-dominant quadrant \(K\+​R−K^\{\+\}R^\{\-\}\), Instruct\-ℳ1:1\\mathcal\{M\}\_\{1:1\}performs best, indicating that combining the two corpora can be beneficial even when the question format is not strongly reasoning\-intensive\. In the dual\-demand quadrant \(K\+​R\+K^\{\+\}R^\{\+\}\), Instruct\-𝒞\\mathcal\{C\}yields the strongest result, suggesting that evidence\-driven inference can be a limiting factor when both knowledge and reasoning requirements are high\. Finally, in the low\-demand quadrant \(K−​R−K^\{\-\}R^\{\-\}\), Instruct\-ℳ3:1\\mathcal\{M\}\_\{3:1\}performs best, indicating that when questions require limited multi\-step inference, additional didactic exposure is sufficient and may even be advantageous\.

Table 4:Performance across knowledge\-reasoning quadrants on Qwen2\.5\-7B\.Boldindicates the best performance per quadrant\.Model VariantK−​R−K^\{\-\}R^\{\-\}K−​R\+K^\{\-\}R^\{\+\}K\+​R−K^\{\+\}R^\{\-\}K\+​R\+K^\{\+\}R^\{\+\}Base72\.5073\.0158\.1044\.00Instruct\-𝒟\\mathcal\{D\}83\.8573\.0169\.8348\.33Instruct\-𝒞\\mathcal\{C\}82\.1374\.6070\.9453\.00Instruct\-ℳ1:3\\mathcal\{M\}\_\{1:3\}82\.1376\.1967\.6050\.33Instruct\-ℳ1:1\\mathcal\{M\}\_\{1:1\}82\.8274\.6073\.1851\.66Instruct\-ℳ3:1\\mathcal\{M\}\_\{3:1\}84\.8773\.0171\.5050\.00
##### Synergy Analysis

We define the Synergy Index \(SI\) as the deviation of observed mixture performance from a ratio\-weighted linear interpolation of the two single\-source variants:

SI=Accℳ−\(α⋅Acc𝒟\+\(1−α\)⋅Acc𝒞\)\\text\{SI\}=\\text\{Acc\}\_\{\\mathcal\{M\}\}\-\\left\(\\alpha\\cdot\\text\{Acc\}\_\{\\mathcal\{D\}\}\+\(1\-\\alpha\)\\cdot\\text\{Acc\}\_\{\\mathcal\{C\}\}\\right\)\(2\)whereα\\alphais the didactic proportion\. As shown in Table[5](https://arxiv.org/html/2609.22161#S4.T5), Instruct\-ℳ1:3\\mathcal\{M\}\_\{1:3\}exceeds the linear expectation inK−​R\+K^\{\-\}R^\{\+\}but falls below it inK\+​R−K^\{\+\}R^\{\-\}, with Instruct\-ℳ3:1\\mathcal\{M\}\_\{3:1\}showing the opposite pattern\. These results indicate that mixture optimization is not about finding a universally superior ratio, but about matching composition to task demands: misaligned mixtures underperform linear expectations, while well\-matched ones unlock complementary benefits\.

Table 5:Synergy analysis with ratio\-weighted linear expectation\. Positive values indicate performance exceeding the mixture\-proportion linear baseline\.Model VariantK−​R−K^\{\-\}R^\{\-\}K−​R\+K^\{\-\}R^\{\+\}K\+​R−K^\{\+\}R^\{\-\}K\+​R\+K^\{\+\}R^\{\+\}Instruct\-ℳ1:3\\mathcal\{M\}\_\{1:3\}\-0\.43\+1\.99\-3\.06\-1\.50Instruct\-ℳ1:1\\mathcal\{M\}\_\{1:1\}\-0\.17\+0\.80\+2\.80\+1\.00Instruct\-ℳ3:1\\mathcal\{M\}\_\{3:1\}\+1\.45\-0\.40\+1\.39\+0\.50

### 4\.5Failure Mode Analysis \(RQ3\)

Table[6](https://arxiv.org/html/2609.22161#S4.T6)shows that knowledge errors constitute the majority of failures across all variants, but error profiles differ by data composition\. Instruct\-𝒟\\mathcal\{D\}shows a higher share of reasoning errors and lower share of context errors than the base model, suggesting aknowing\-doing gap—knowledge is acquired but not effectively applied\. Instruct\-𝒞\\mathcal\{C\}reduces both total errors and the share of knowledge errors, reinforcing that clinical contexts strengthen knowledge recall through application\. Instruct\-ℳ1:1\\mathcal\{M\}\_\{1:1\}falls between the two\. This pattern is most pronounced in theK\+​R\+K^\{\+\}R^\{\+\}quadrant \(Figure[3](https://arxiv.org/html/2609.22161#S4.F3)\), where Instruct\-𝒟\\mathcal\{D\}yields the highest reasoning error rate and Instruct\-𝒞\\mathcal\{C\}the lowest knowledge error rate\. Overall, the results suggest that clinical data improves robustness on errors tied to evidence\-driven inference, while didactic data more often struggles when questions require synthesizing multiple cues, even when factual knowledge is available\.

Figure 3:Error distribution in theK\+​R\+K^\{\+\}R^\{\+\}quadrant\. Bars show the proportion of each error type within the total errors for each variant in this quadrant\.Table 6:Error type distribution across data compositions, whereEKnow\.E\_\{\\text\{Know\.\}\},EReas\.E\_\{\\text\{Reas\.\}\},ECtx\.E\_\{\\text\{Ctx\.\}\}, andEFmt\.E\_\{\\text\{Fmt\.\}\}denote knowledge, reasoning, context, and format errors, respectively\. Percentages indicate the proportion within total errors for each model\.Model VariantEKnow\.E\_\{\\text\{Know\.\}\}EReas\.E\_\{\\text\{Reas\.\}\}ECtx\.E\_\{\\text\{Ctx\.\}\}EFmt\.E\_\{\\text\{Fmt\.\}\}Base229\(67\.4%\)38\(11\.2%\)60\(17\.6%\)13\(3\.8%\)Instruct\-𝒟\\mathcal\{D\}195\(71\.4%\)51\(18\.7%\)24\(8\.8%\)3\(1\.1%\)Instruct\-𝒞\\mathcal\{C\}182\(69\.7%\)38\(14\.6%\)34\(13\.0%\)7\(2\.7%\)Instruct\-ℳ1:1\\mathcal\{M\}\_\{1:1\}185\(70\.6%\)43\(16\.4%\)30\(11\.5%\)4\(1\.5%\)

## 5Conclusion

In this study, we examine how didactic and clinical corpora shape medical LLM capabilities under token\-matched experiments\. Clinical data performs strongly on clinic\-oriented tasks while remaining competitive on knowledge\-intensive ones, whereas didactic data yields limited gains on clinic\-oriented evaluation—revealing an asymmetric transfer consistent with aknowing–doing gap\. The optimal composition varies with task demands and a modest clinical fraction captures most gains on EHR\-grounded tasks\. These findings highlight the importance of composition\-aware, application\-driven data curation for medical LLMs\.

#### Acknowledgements

This work was supported in part by the National Natural Science Foundation of China \(Grant No\. 62576126\) and the Key R&D Program of Heilongjiang Province \(Grant No\. 2023ZX01A11\)\.

## References

- \[1\]J\. R\. AndersonThe architecture of cognition\.Psychology Press,London, England\.Cited by:[§1](https://arxiv.org/html/2609.22161#S1.p2.1)\.
- \[2\]R\. K\. Arora, J\. Wei, R\. S\. Hicks, P\. Bowman, J\. Quiñonero\-Candela, F\. Tsimpourlas, M\. Sharman, M\. Shah, A\. Vallone, A\. Beutel, J\. Heidecke, and K\. Singhal\(2025\)HealthBench: evaluating large language models towards improved human health\.External Links:2505\.08775,[Link](https://arxiv.org/abs/2505.08775)Cited by:[§2\.2](https://arxiv.org/html/2609.22161#S2.SS2.p1.1)\.
- \[3\]A\. Ben Abacha and D\. Demner\-Fushman\(2019\)A question\-entailment approach to question answering\.BMC bioinformatics20\(1\),pp\. 511\.Cited by:[§3\.1](https://arxiv.org/html/2609.22161#S3.SS1.SSS0.Px1.p1.1)\.
- \[4\]A\. Ben Abacha, W\. Yim, G\. Michalopoulos, and T\. Lin\(2023\)An investigation of evaluation methods in automatic medical note generation\.InFindings of the Association for Computational Linguistics: ACL 2023,A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 2575–2588\.Cited by:[§2\.2](https://arxiv.org/html/2609.22161#S2.SS2.p1.1)\.
- \[5\]Y\. Cai, L\. Wang, Y\. Wang, G\. de Melo, Y\. Zhang, Y\. Wang, and L\. He\(2024\)Medbench: a large\-scale chinese benchmark for evaluating medical large language models\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.38,pp\. 17709–17717\.Cited by:[§2\.2](https://arxiv.org/html/2609.22161#S2.SS2.p1.1)\.
- \[6\]H\. Chen, Z\. Fang, Y\. Singla, and M\. Dredze\(2025\)Benchmarking large language models on answering and explaining challenging medical questions\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),L\. Chiruzzo, A\. Ritter, and L\. Wang \(Eds\.\),Albuquerque, New Mexico,pp\. 3563–3599\.External Links:ISBN 979\-8\-89176\-189\-6Cited by:[§3\.3](https://arxiv.org/html/2609.22161#S3.SS3.SSS0.Px2.p1.1)\.
- \[7\]C\. Christophe, T\. Raha, S\. Maslenkova, M\. U\. Salman, P\. Kanithi, M\. A\. Pimentel, and S\. Khan\(2024\)Beyond fine\-tuning: unleashing the potential of continuous pretraining for clinical LLMs\.\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 10549–10561\.Cited by:[§1](https://arxiv.org/html/2609.22161#S1.p3.1)\.
- \[8\]X\. Han, Y\. Fan, S\. Zhao, H\. Wang, and B\. Qin\(2026\)Gsem: graph\-based self\-evolving memory for experience augmented clinical reasoning\.arXiv preprint arXiv:2603\.22096\.Cited by:[§2\.1](https://arxiv.org/html/2609.22161#S2.SS1.p1.1)\.
- \[9\]D\. Hendrycks, C\. Burns, S\. Basart, A\. Zou, M\. Mazeika, D\. Song, and J\. Steinhardt\(2021\)Measuring massive multitask language understanding\.External Links:2009\.03300Cited by:[§3\.3](https://arxiv.org/html/2609.22161#S3.SS3.SSS0.Px1.p1.1)\.
- \[10\]D\. Jin, E\. Pan, N\. Oufattole, W\. Weng, H\. Fang, and P\. Szolovits\(2021\)What disease does this patient have? a large\-scale open domain question answering dataset from medical exams\.Appl\. Sci\. \(Basel\)11\(14\),pp\. 6421\(en\)\.Cited by:[§2\.2](https://arxiv.org/html/2609.22161#S2.SS2.p1.1),[§3\.3](https://arxiv.org/html/2609.22161#S3.SS3.SSS0.Px1.p1.1)\.
- \[11\]Q\. Jin, B\. Dhingra, Z\. Liu, W\. Cohen, and X\. Lu\(2019\)PubMedQA: a dataset for biomedical research question answering\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),K\. Inui, J\. Jiang, V\. Ng, and X\. Wan \(Eds\.\),Hong Kong, China,pp\. 2567–2577\.Cited by:[§3\.3](https://arxiv.org/html/2609.22161#S3.SS3.SSS0.Px1.p1.1)\.
- \[12\]A\. Johnson, T\. Pollard, and R\. Mark\(2023\)MIMIC\-III clinical database\.PhysioNet\.Cited by:[§1](https://arxiv.org/html/2609.22161#S1.p3.1)\.
- \[13\]S\. Kweon, J\. Kim, H\. Kwak, D\. Cha, H\. Yoon, K\. Kim, J\. Yang, S\. Won, and E\. Choi\(2024\)EHRNoteQA: an llm benchmark for real\-world clinical practice using discharge summaries\.InAdvances in Neural Information Processing Systems,A\. Globerson, L\. Mackey, D\. Belgrave, A\. Fan, U\. Paquet, J\. Tomczak, and C\. Zhang \(Eds\.\),Vol\.37,pp\. 124575–124611\.Cited by:[§2\.2](https://arxiv.org/html/2609.22161#S2.SS2.p1.1)\.
- \[14\]Y\. Labrak, A\. Bazoge, E\. Morin, P\. Gourraud, M\. Rouvier, and R\. Dufour\(2024\)BioMistral: a collection of open\-source pretrained large language models for medical domains\.InFindings of the Association for Computational Linguistics: ACL 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 5848–5864\.Cited by:[§2\.1](https://arxiv.org/html/2609.22161#S2.SS1.p1.1)\.
- \[15\]C\. Li, C\. Wong, S\. Zhang, N\. Usuyama, H\. Liu, J\. Yang, T\. Naumann, H\. Poon, and J\. Gao\(2023\)LLaVA\-med: training a large language\-and\-vision assistant for biomedicine in one day\.InAdvances in Neural Information Processing Systems,A\. Oh, T\. Naumann, A\. Globerson, K\. Saenko, M\. Hardt, and S\. Levine \(Eds\.\),Vol\.36,pp\. 28541–28564\.Cited by:[§2\.1](https://arxiv.org/html/2609.22161#S2.SS1.p1.1)\.
- \[16\]R\. Luo, L\. Sun, Y\. Xia, T\. Qin, S\. Zhang, H\. Poon, and T\. Liu\(2022\)BioGPT: generative pre\-trained transformer for biomedical text generation and mining\.Briefings in Bioinformatics23\(6\),pp\. bbac409\.External Links:ISSN 1477\-4054Cited by:[§2\.1](https://arxiv.org/html/2609.22161#S2.SS1.p1.1)\.
- \[17\]A\. Pal, L\. K\. Umapathi, and M\. Sankarasubbu\(2022\)MedMCQA: a large\-scale multi\-subject multi\-choice dataset for medical domain question answering\.InProceedings of the Conference on Health, Inference, and Learning,G\. Flores, G\. H\. Chen, T\. Pollard, J\. C\. Ho, and T\. Naumann \(Eds\.\),Proceedings of Machine Learning Research, Vol\.174,pp\. 248–260\.Cited by:[§3\.1](https://arxiv.org/html/2609.22161#S3.SS1.SSS0.Px1.p1.1)\.
- \[18\]A\. Pampari, P\. Raghavan, J\. Liang, and J\. Peng\(2018\)EmrQA: a large corpus for question answering on electronic medical records\.InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing,E\. Riloff, D\. Chiang, J\. Hockenmaier, and J\. Tsujii \(Eds\.\),Brussels, Belgium,pp\. 2357–2368\.Cited by:[§2\.2](https://arxiv.org/html/2609.22161#S2.SS2.p1.1)\.
- \[19\]H\. G\. Schmidt and R\. M\. J\. P\. RikersHow expertise develops in medicine: knowledge encapsulation and illness script formation\.Med\. Educ\.41\(12\),pp\. 1133–1139\(en\)\.Cited by:[§1](https://arxiv.org/html/2609.22161#S1.p2.1),[§3\.1](https://arxiv.org/html/2609.22161#S3.SS1.p1.1)\.
- \[20\]K\. Singhal, S\. Azizi, T\. Tu, S\. S\. Mahdavi, J\. Wei, H\. W\. Chung, N\. Scales, A\. Tanwani, H\. Cole\-Lewis, S\. Pfohl, P\. Payne, M\. Seneviratne, P\. Gamble, C\. Kelly, A\. Babiker, N\. Schärli, A\. Chowdhery, P\. Mansfield, D\. Demner\-Fushman, B\. Agüera Y Arcas, D\. Webster, G\. S\. Corrado, Y\. Matias, K\. Chou, J\. Gottweis, N\. Tomasev, Y\. Liu, A\. Rajkomar, J\. Barral, C\. Semturs, A\. Karthikesalingam, and V\. Natarajan\(2023\)Large language models encode clinical knowledge\.Nature620\(7972\),pp\. 172–180\(en\)\.Cited by:[§1](https://arxiv.org/html/2609.22161#S1.p1.1),[§2\.2](https://arxiv.org/html/2609.22161#S2.SS2.p1.1)\.
- \[21\]T\. Tu, M\. Schaekermann, A\. Palepu, K\. Saab, J\. Freyberg, R\. Tanno, A\. Wang, B\. Li, M\. Amin, Y\. Cheng, E\. Vedadi, N\. Tomasev, S\. Azizi, K\. Singhal, L\. Hou, A\. Webson, K\. Kulkarni, S\. S\. Mahdavi, C\. Semturs, J\. Gottweis, J\. Barral, K\. Chou, G\. S\. Corrado, Y\. Matias, A\. Karthikesalingam, and V\. Natarajan\(2025\)Towards conversational diagnostic artificial intelligence\.Nature642\(8067\),pp\. 442–450\(en\)\.Cited by:[§2\.1](https://arxiv.org/html/2609.22161#S2.SS1.p1.1)\.
- \[22\]G\. Wang, G\. Yang, Z\. Du, L\. Fan, and X\. Li\(2023\)ClinicalGPT: large language models finetuned with diverse medical data and comprehensive evaluation\.External Links:2306\.09968Cited by:[§2\.1](https://arxiv.org/html/2609.22161#S2.SS1.p1.1)\.
- \[23\]H\. Wang, S\. Zhao, Z\. Qiang, Z\. Li, C\. Liu, N\. Xi, Y\. Du, B\. Qin, and T\. Liu\(2025\)Knowledge\-tuning large language models with structured medical knowledge bases for trustworthy response generation in chinese\.ACM Transactions on Knowledge Discovery from Data19\(2\),pp\. 1–17\.Cited by:[§2\.1](https://arxiv.org/html/2609.22161#S2.SS1.p1.1)\.
- \[24\]X\. Wang, G\. Chen, S\. Dingjie, Z\. Zhiyi, Z\. Chen, Q\. Xiao, J\. Chen, F\. Jiang, J\. Li, X\. Wan, B\. Wang, and H\. Li\(2024\)CMB: a comprehensive medical benchmark in Chinese\.InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),K\. Duh, H\. Gomez, and S\. Bethard \(Eds\.\),Mexico City, Mexico,pp\. 6184–6205\.Cited by:[§4\.1](https://arxiv.org/html/2609.22161#S4.SS1.SSS0.Px3.p1.1)\.
- \[25\]C\. Wu, W\. Lin, X\. Zhang, Y\. Zhang, W\. Xie, and Y\. Wang\(2024\)PMC\-llama: toward building open\-source language models for medicine\.Journal of the American Medical Informatics Association31\(9\),pp\. 1833–1843\.External Links:ISSN 1527\-974XCited by:[§2\.1](https://arxiv.org/html/2609.22161#S2.SS1.p1.1)\.
- \[26\]B\. Xiao, X\. Tian, X\. Song, H\. Wang, G\. Song, S\. Zhao, and B\. Qin\(2026\)When correct beliefs collapse: epistemic resilience of llms under clinical pressure\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 8720–8764\.Cited by:[§2\.2](https://arxiv.org/html/2609.22161#S2.SS2.p1.1)\.
- \[27\]Q\. Xie, Q\. Chen, A\. Chen, C\. Peng, Y\. Hu, F\. Lin, X\. Peng, J\. Huang, J\. Zhang, V\. Keloth, X\. Zhou, L\. Qian, H\. He, D\. Shung, L\. Ohno\-Machado, Y\. Wu, H\. Xu, and J\. Bian\(2024\)Me llama: foundation large language models for medical applications\.External Links:2402\.12749,[Link](https://arxiv.org/abs/2402.12749)Cited by:[§2\.1](https://arxiv.org/html/2609.22161#S2.SS1.p1.1)\.
- \[28\]X\. Yang, A\. Chen, N\. PourNejatian, H\. C\. Shin, K\. E\. Smith, C\. Parisien, C\. Compas, C\. Martin, M\. G\. Flores, Y\. Zhang, T\. Magoc, C\. A\. Harle, G\. Lipori, D\. A\. Mitchell, W\. R\. Hogan, E\. A\. Shenkman, J\. Bian, and Y\. Wu\(2022\)GatorTron: a large clinical language model to unlock patient information from unstructured electronic health records\.External Links:2203\.03540,[Link](https://arxiv.org/abs/2203.03540)Cited by:[§2\.1](https://arxiv.org/html/2609.22161#S2.SS1.p1.1)\.