Paper deep dive

SAGE: Semantic Anchor-Guided Evolution for Grounded Medical QA Data Synthesis

Chuan Li, Chengyu Wang, Cen Chen, Ye Lyu, Mingyuan Fan, Ming Gao

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 82

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Developing reliable models for clinical tasks, such as Medical Question Answering (QA), is severely constrained by the limited availability of high-quality, expert-annotated training data. This challenge is exacerbated by stringent privacy requirements and the impracticality of utilizing large open-source corpora or proprietary cloud APIs within resource-limited clinical settings. To address these obstacles, we introduce SAGE (\textit{Semantic Anchor-Guided Evolution}), a novel data synthesis framework that enables small, locally deployed models to generate high-quality medical training data. SAGE leverages lightweight, publicly available taxonomies such as MeSH as semantic anchors, imposing a structured prior to effectively guide and ground the data generation process. At its core, SAGE iteratively interleaves atomic (individual concept-based) and associative (relation-based) synthesis, bootstrapping training data from minimal seeds. This approach eliminates the need for large collections of medical documents or reliance on external APIs, providing a practical solution for on-premises data creation. Extensive experiments across multiple medical question-answering benchmarks demonstrate that models fine-tuned with SAGE-synthesized data consistently outperform those trained using self-derived or conventional document-based paradigms, highlighting tangible improvements in data efficiency and resource utilization for medical LLM development. Code is available at this https URL.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

78,211 characters extracted from source content.

Expand or collapse full text

SAGE: Semantic Anchor-Guided Evolution for Grounded Medical QA Data Synthesis Chuan Li 1 , Chengyu Wang 2 , Cen Chen 1 * , Ye Lyu 1 , Mingyuan Fan 1 , Ming Gao 1 1 East China Normal University, Shanghai, China 2 Alibaba Group, Hangzhou, China lichuan, yelyu, mingyuan_fmy@stu.ecnu.edu.cn chengyu.wcy@alibaba-inc.com, cenchen,mgao@dase.ecnu.edu.cn Abstract Developing reliable models for clinical tasks, such as Medical Question Answering (QA), is severely constrained by the limited availabil- ity of high-quality, expert-annotated training data. This challenge is exacerbated by strin- gent privacy requirements and the impractical- ity of utilizing large open-source corpora or proprietary cloud APIs within resource-limited clinical settings. To address these obstacles, we introduce SAGE (Semantic Anchor-Guided Evolution), a novel data synthesis framework that enables small, locally deployed models to generate high-quality medical training data. SAGE leverages lightweight, publicly avail- able taxonomies such as MeSH as semantic anchors, imposing a structured prior to effec- tively guide and ground the data generation pro- cess. At its core, SAGE iteratively interleaves atomic (individual concept-based) and asso- ciative (relation-based) synthesis, bootstrap- ping training data from minimal seeds. This approach eliminates the need for large col- lections of medical documents or reliance on external APIs, providing a practical solution for on-premises data creation. Extensive ex- periments across multiple medical question- answering benchmarks demonstrate that mod- els fine-tuned with SAGE-synthesized data consistently outperform those trained using self-derived or conventional document-based paradigms, highlighting tangible improvements in data efficiency and resource utilization for medical LLM development. Code is available at https://github.com/DIaacKr/SAGE. 1 Introduction Large language models (LLMs) have demon- strated remarkable capabilities, prompting signif- icant interest in their deployment within high- stakes, knowledge-intensive domains such as medicine (Zhang et al., 2024; Guo et al., 2025; Yang et al., 2024a, 2025; Xu et al., 2025; Huang * Corresponding author Self- derived Document- based Ours Derive Append Synthesize Extract Express Express Derived Anchor Corpus Seed Derived Example Data Data Seed Anchor Seed Extract Data What is insulin? What is the function of insulin? How many amino acids constitute the chemical essence of insulin? C fibers are unmyelinated and are responsible for transmitting slow, dull, aching pain, which is consistent with the sharp, localized pain described. FAKE EASY Figure 1: Comparison of data synthesis paradigms. Con- ventional methods either derive data using implicit para- metric knowledge or synthesize data from document corpus. In contrast, SAGE synthesizes high-quality medical data via semantic anchor. et al., 2025a; Yu et al., 2025a; Wang et al., 2025; Chen et al., 2025b). However, further progress is hampered by the scarcity of high-quality, expert- annotated datasets. In medicine, assembling large- scale training corpora is not only expensive and time-consuming, but also subject to stringent pri- vacy and ethical constraints (Price and Cohen, 2019; Kaissis et al., 2020; Rajpurkar et al., 2022; Thirunavukarasu et al., 2023).This poses a fundamental bottleneck for transforming general- purpose LLMs into reliable medical experts. Synthetic data generation has emerged as a promising direction; yet, existing approaches predominantly rely on proprietary, large-scale LLMs as generative engines (Wang et al., 2023b; Gudibande et al., 2023). Such dependence intro- duces challenges regarding data privacy, opera- tional costs, and practical feasibility in resource- constrained environments, including hospitals (Car- lini et al., 2021). Addressing this underexplored issue requires a practical and resource-efficient method for synthesizing high-quality training data using strictly limited, local seed resources, thereby eliminating dependency on external APIs or mas- sive proprietary corpora (Labrak et al., 2024). Prior work in data synthesis primarily falls into two paradigms, each with clear limitations, as arXiv:2610.08093v1 [cs.CL] 6 Oct 2026 shown in Figure 1. The self-derived paradigm lever- ages implicit parametric knowledge to efficiently generate diverse data (Wang et al., 2023b; Xu et al., 2024; Yu et al., 2025b); yet, the absence of external grounding can result in unanchored outputs and hal- lucinations, which is an unacceptable risk for clini- cal applications. Conversely, the document-based paradigm grounds generation in authoritative texts to promote factual accuracy (Maini et al., 2024; Miao et al., 2025). However, document-based methods require extensive preprocessing of large corpora, which are often unavailable or unstruc- tured in clinical contexts. Furthermore, document- based approaches tend to encourage mere “mem- orization” of text, rather than the “integration” of complex clinical reasoning, constraining model ex- pressivity and applicability. We contend that structured medical taxonomies, Medical Subject Headings (MeSH) 1 , represent a widely accessible yet underutilized resource. Un- like lengthy documents or corpora, MeSH serves as a standardized dictionary, which is a topolog- ical scaffold of medical concepts, that is both lightweight and authoritative. In this work, we introduce SAGE (Semantic Anchor-Guided Evo- lution), a grounded QA data synthesis framework specifically designed for resource-constrained envi- ronments. SAGE treats MeSH concepts as seman- tic anchors to guide and ground the generative pro- cess. This neuro-symbolic approach enables small, locally deployed models to produce logically rig- orous content, adhering to a structured prior while reducing both computational and data overhead. SAGE orchestrates data synthesis leveraging se- mantic anchors through a dual-track, parallel, it- erative process that balances exploitation and ex- ploration, as shown in Figure 1. To ensure fac- tual precision, SAGE employs atomic generation, targeting individual semantic anchors via ontol- ogy definitions to produce definition-centric ques- tions and mitigate hallucination risk. Meanwhile, SAGE utilizes associative generation, interweav- ing multiple concepts as anchor to construct coher- ent medical scenarios requiring multi-step reason- ing. Crucially, these tracks operate concurrently, maximizing the coverage of the concept space of core concepts; they are refined through ongoing it- erative calibration that broadens the semantic scope of the synthesized dataset. This evolutionary loop enables bootstrapping from a minimal seed set, or- 1 https://w.ncbi.nlm.nih.gov/mesh/ ganically expanding from atomic facts to complex associative reasoning. As a result, SAGE cultivates comprehensive and valuable training corpora from compact, structured sources. Our method achieves state-of-the-art perfor- mance across a broad range of medical QA bench- marks, with an average accuracy of 68.12% on five widely used medical QA datasets, outperforming other self-derived and document-based methods. Notably, it also exceeds the results achieved by sev- eral high-quality open-source datasets and models trained on massive data. In summary, our main contributions are: •We introduce SAGE, a Semantic Anchor- Guided Evolution framework for synthesizing high-quality, grounded medical QA data using small models and public taxonomies, without reliance on large corpora or external APIs. • We propose a parallel, iterative synthesis strat- egy that integrates atomic generation for fac- tual coverage and associative generation for reasoning complexity, dynamically balancing exploitation and exploration. • SAGE-grounded data substantially improve factual reliability and consistency. Models fine-tuned on SAGE data consistently outper- form strong baselines on medical QA bench- marks in resource-constrained settings. 2 Related Work Self-Derived Data Synthesis. Self-bootstrapping data synthesis methods, which autonomously gen- erate large-scale training data, have attracted in- creasing research interest as alternatives to manual annotation (Li et al., 2023; Yu et al., 2023; Luo et al., 2023; Xu et al., 2024; Tang et al., 2024; Li et al., 2024; Zeng et al., 2024; Huang et al., 2025b; Liu et al., 2025; Riaz et al., 2025; Zhao et al., 2025b,a). A representative example is Self- Instruct (Wang et al., 2023b), which prompts LLMs to produce diverse instructions and corresponding responses, expanding training data from a small seed set. Subsequent variants, including Chain-of- Thought Self-Instruct (Yu et al., 2025b), introduce “think-then-generate” pipelines with automatic fil- tering for higher-quality examples. However, these approaches lack rigorous mechanisms for factual control: filtering criteria typically prioritize seman- tic richness or complexity rather than factuality. In high-stakes fields such as medicine, this limitation becomes critical, leading to severe factual halluci- nations and adversely affecting model fine-tuning. Document-Based Data Synthesis. Document- based approaches utilize structured or unstruc- tured corpora to synthesize training data for LLMs (Zhang and Yang, 2023; Yehudai et al., 2024; Köksal et al., 2024; Maini et al., 2024; Yang et al., 2024b; Dedhia et al., 2025; Chen et al., 2025c). For instance, Easy Datasets (Miao et al., 2025) guides models to create question-answer pairs from the same source text by assigning varied “genre- audience” roles, thereby enhancing the diversity and scale of synthetic data for domain adaptation. Similarly, Wrap (Maini et al., 2024) leverages off- the-shelf LLMs to rephrase noisy web text into higher-quality synthetic formats, improving both compute efficiency and downstream performance. While these methods ensure factual grounding through anchoring to source documents, their suc- cess heavily relies on access to large, high-quality document collections. In low-resource scenarios where such corpora are unavailable, the practical utility of document-based synthesis is limited. 3 Methodology We present Semantic Anchor-Guided Evolution (SAGE), a medical QA synthesis framework that alternates between neural generation in text space and symbolic projection onto a fixed medical taxon- omy. Unlike unconstrained self-generation, SAGE uses MeSH concepts as semantic anchors to orga- nize the generation trajectory while retaining the flexibility of a language model to express relations that are not explicitly encoded as taxonomy edges. 3.1 Problem Formulation and Neural-Symbolic Evolution Our objective is to synthesize a medical question- answering dataset, denoted byD syn , using a lo- cally deployed language modelMand a small seed datasetD seed . We use an authoritative medical tax- onomy, such as MeSH, as an external structured prior. Let T = (V,E )(1) denote the fixed taxonomy, whereVis the set of medical concept nodes andE represents their hier- archical relations. SAGE maintains an evolving population of ac- tive anchor configurations S (t) ⊆ 2 V \∅(2) at epocht. Each anchor configurationA ∈ S (t) contains one or more MeSH concepts and specifies the semantic scope of a generation step. Given an anchor configurationA, the language model generates a candidate QA instancex = (q x ,r x ,y x ) , whereq x is the question,r x is the gen- erated reasoning, andy x is the answer label. Af- ter candidate selection,q x is mapped back to the taxonomy and may contribute an updated anchor configuration to the next epoch. This process can be viewed as neural-symbolic evolution over a fixed ontological scaffold. The language model serves as a neural variation op- erator that explores possible medical expressions and relations, whereas the taxonomy provides a discrete concept space for anchoring, projection, and population update. SAGE does not perform a hand-designed traversal of taxonomy edges; in- stead, semantic exploration is carried out by the language model in text space and is iteratively con- solidated in the MeSH concept space. Accordingly, the taxonomy itself remains fixed, while the active populationS (t) evolves across epochs. 3.2 MeSH-Structured Semantic Anchors We define a taxonomy-aware mapping function φ T :X q → 2 V ,(3) which maps a medical questionq ∈X q to a set of MeSH concepts. In implementation,φ T performs synonym-expanded terminology matching and re- tains concepts according to predefined MeSH cate- gory and hierarchy-depth constraints. Additional details and evaluations of the mapping function are provided in Appendix H. The setφ T (q x )is referred to as the semantic anchor configuration of instancex. These anchors do not prescribe a complete reasoning path. Instead, they provide a structured semantic scaffold that specifies which medical concepts should organize the generation. This allows the model to draw on its parametric knowledge while keeping the iterative synthesis trajectory connected to a curated medical concept space. The initial anchor population is constructed as S (0) =φ T (q x )| x∈D seed , φ T (q x )̸=∅. (4) 3.3 Dual-Track Neural Generation To support both concept-focused grounding and relation-oriented exploration, as illustrated in Fig- ure 2, we route each anchor configuration accord- ing to its cardinality. Singleton configurations Q Q Concept Anchor Question Reasoning Mutate A Gout Aspirin Cyclosporins Renal Insufficiency, Chronic Kidney Failure, Chronic Lead Edge MeSHTaxonomy LLM Exemplar #1 Anchor #1 Exemplar #n Anchor #n Seed Anchors A Consistency Check Diversity Check Extract Express Pool Update Q A Q A Explanation: Conditions in which the KIDNEYS perform below the normal level for more than three months... Anchor: Renal Insufficiency, Chronic. Perspective: physiopathology. Used with organs and diseases for disordered function in disease states. Sample Atomic Associative LLM AQ Figure 2: Overview of the SAGE framework. MeSH anchors guide dual-track generation, while consistency and diversity checks select QA instances whose projected concepts update the anchor population across epochs. are assigned to atomic generation, whereas multi- concept configurations are assigned to associative generation. LetA∈S (t) be the target anchor configuration for the current generation step. The generation of a candidate instance x is defined as follows: x∼ ( P atomic (x|c,K c ) ifA =c, P assoc (x|A,E (t) few )if|A| > 1 (5) whereP atomic governs single-concept synthesis and P assoc governs multi-concept synthesis. Atomic Generation. This track processes single- ton anchor configurations (|A| = 1), including newly surfaced concepts introduced during anchor evolution. Letcdenote the sole concept inA. To provide explicit taxonomy-derived grounding with- out relying on few-shot demonstrations, we con- struct a taxonomy-derived contextK c . Specifically, we retrieve the authoritative definitionδ(c)and the set of allowed qualifiersQ(c)(e.g., pathology, ad- verse effects) fromT. We defineK c =δ(c),u i , whereu i ∼U (Q(c))is a clinical qualifier sampled for the present generation. The distribution over candidate instances is then given by: P atomic (x|c,K c ) = P M (x| Prompt(c,δ(c),u i )) (6) To introduce controlled variation, we generateN candidate instances in parallel using different qual- ifiers fromQ(c)when available. This qualifier- conditioned sampling provides controlled variation around concept c. Associative Generation. This track processes multi-concept anchor configurations (|A| > 1) and encourages relation-oriented generation among the anchored concepts. For each synthesis step, we construct a contextE (t) few composed ofn few exam- ples drawn from a growing pool that contains both the initial seed instances and previously accepted samples. The few-shot pool and sampled context are defined as: F (t) =D seed ∪ t−1 [ i=0 D (i) valid , E (t) few ⊆F (t) . (7) The generative modelMis then conditioned on AandE (t) few to synthesize a candidatexthat is en- couraged to connect the concepts in the anchor configuration (e.g., generating a clinical scenario linkingA =“Aspirin”, “Chronic Pain”). 3.4Symbolic Selection and Anchor Evolution Each epoch comprises answer-consistency selec- tion, symbolic projection, and semantic mutation. Calibration via Consistency and Diversity Checking.Given the computational over- head of large-scale reward models in our resource-constrained setting, we combine answer- consistency checking with anchor-level diver- sity control. First, we adopt Self-Consistency (SC) (Wang et al., 2023a; Yu et al., 2025b) as the candidate-selection criterion. For each candidate instancex = (q x ,r x ,y x ), we generateKauxiliary answersy ′ j K j=1 conditioned onq x . We define the answer-consistency indicator as I(x) = 1 y x = MajorVote y ′ 1 ,...,y ′ K . (8) Candidates satisfying the self-consistency criterion are incorporated into the accepted setD (t) valid . Second,SAGE maintains concept- and configuration-level occurrence counts to control diversity across epochs. Anchor configurations whose expansion budgets have been exhausted are excluded from further sampling. Symbolic Projection and Semantic Mutation. Transitioning from epochttot+1, we construct the next active anchor populationS (t+1) by combin- ing retained output configurations with singleton anchors instantiated from newly surfaced concepts. First, we collect the projected anchor configura- tions of the accepted samples: S (t) proj = n φ T (q x )| x∈D (t) valid , φ T (q x )̸=∅ o . (9) Second, during generation (e.g., discussing As- pirin for chronic pain), the model may introduce new, contextually relevant medical concepts absent from the input anchors configuration (e.g., natu- rally surfacing the side effect“Gastric Ulcer”). Let A in denote the input anchor configuration used to generate candidate instancex, and define the newly discovered concepts ∆(x) as: ∆(x) = φ T (q x ) in .(10) These newly surfaced concepts are recognized through projection onto the MeSH concept space. Each new concept is additionally instantiated as a singleton anchor configuration for the subse- quent epoch, providing an opportunity for dedi- cated Atomic Generation. The resulting singleton mutation configurations are collected as: S (t) mut =      c| c∈ [ x∈D (t) valid ∆(x)      (11) Finally, the active anchor population for the next epoch is obtained by combining retained projected configurations with singleton mutation configura- tions:S (t+1) =S (t) proj ∪S (t) mut . Newly surfaced con- cepts are additionally instantiated as singleton an- chor configurations, allowing them to receive dedi- cated taxonomy-derived grounding through Atomic Generation in subsequent epochs. Meanwhile, ac- cepted output configurations are propagated to sup- port continued associative exploration. This update expands the explored region of the MeSH concept space while keeping the generation trajectory con- nected to the fixed MeSH scaffold. 4 Experiments 4.1 Experimental Setup To reflect resource-constrained on-premise settings, we use Qwen3-4B (Yang et al., 2025) with thinking disabled as the common data-synthesis backbone for all methods. To assess the quality of each result- ing dataset, we fine-tune a fresh Qwen3-4B-Base model using an identical supervised fine-tuning protocol. Unless otherwise specified, all trainee models are fine-tuned for 5 epochs with a learn- ing rate of1× 10 −5 and a global batch size of 16. Further configurations are provided in Appendix B. 4.2 Baselines and Benchmarks We compare SAGE against three categories of baselines: (1) Self-derived synthesis methods: Self-Instruct (Wang et al., 2023b), CoT-Self- Instruct (Yu et al., 2025b), ScaleQuest (Ding et al., 2025), KPDDS (Huang et al., 2025b), and Math- Scale (Tang et al., 2024). To ensure fairness, we generate 10k samples for all methods using the same seed set. (2) Document-based methods: Easy-Dataset (Miao et al., 2025), Genie (Yehudai et al., 2024), Self-QA (Zhang and Yang, 2023), and Wrap (Maini et al., 2024). We use MeSH ontology scope notes as the unified reference corpus for these baselines. (3) High-quality open-source datasets: m1k, m23k (Huang et al., 2025a), Huatuogpt-o1- SFT (Chen et al., 2025b), and MedReason33k (Wu et al., 2025). For evaluation, we utilize a suite of ten medical question-answering benchmarks that probe differ- ent aspects of medical reasoning and generalization. Following Chen et al. (2025b), our suite includes MedMCQA (Pal et al., 2022), MedQA (Jin et al., 2021), and PubMedQA (Jin et al., 2019), which fo- cus on domains well-aligned with our seed data. To measure cross-domain generalization, we employ the medical subset of GPQA (Rein et al., 2024) and MMLU-Pro (Wang et al., 2024). For robustness and advanced reasoning, we use high-difficulty benchmarks: Lancet 2 , NEJM 3 , MedBullets (Chen et al., 2025a), and MedXpertQA (Zuo et al., 2025). Accuracy is adopted as the evaluation metric for all tasks following previous works (Chen et al., 2025b; Huang et al., 2025a; Yu et al., 2025a; Wang et al., 2025). 2 https://w.thelancet.com/ 3 https://w.nejm.org/ ModelMethod/DatasetScaleAVGMedMCQAMedQAPubMedQAMMLU-ProGPQA (med)(med) Qwen3-4B-BaseFew-shot PromptingNA26.1529.1429.6935.2014.6622.05 Qwen3-4B NoThinkingNANA62.5456.7863.6371.0064.6356.67 Qwen3-4B ThinkingNANA66.1859.8172.0372.6071.6054.87 Open Source Datasets Qwen3-4B-Basem1k (2025a)1k63.4059.2672.1974.2067.4943.85 Qwen3-4B-Basem23k (2025a)23k67.3964.1976.2076.0068.7951.79 Qwen3-4B-BaseHuatuo-SFT (2025b)20k52.8054.0860.8056.7053.6838.72 Qwen3-4B-BaseMedReason (2025)33k58.6053.6060.4176.9063.1338.97 Self-derived Method Qwen3-4B-BaseSelf-Instruct (2023b)10k53.6652.3550.1269.7053.2942.82 Qwen3-4B-BaseScaleQuest (2025)10k62.9157.4267.1676.8064.9548.21 Qwen3-4B-BaseKPDDS (2025b)10k58.9952.5261.7471.6062.1546.92 Qwen3-4B-BaseMathScale (2024)10k57.8555.0158.9273.7060.8540.77 Qwen3-4B-BaseCoT-Self-Instruct (2025b)10k63.71 62.0867.4070.0061.8957.18 Document-based Method Qwen3-4B-BaseEasy-Dataset (2025)28k49.1146.9044.0770.3048.4035.90 Qwen3-4B-BaseGenie (2024)30k49.9849.3251.7763.9048.9935.90 Qwen3-4B-BaseSelf-QA (2023)230k46.3448.9851.0648.9047.1035.64 Qwen3-4B-BaseWrap (2024)98K50.20 47.5347.1369.5047.1039.74 Ours Qwen3-4B-BaseOurs, w/o Iteration10k66.4564.2468.2680.2060.3359.23 Qwen3-4B-BaseOurs, w/o Parallel10k66.8164.0467.9582.8060.7858.46 Qwen3-4B-BaseOurs, w/o Atomic gen.10k66.2165.2966.6977.2061.8960.00 Qwen3-4B-BaseOurs, w/o Associative gen.10k66.1365.2469.8477.9062.0255.64 Qwen3-4B-BaseOurs (Full)10k68.1264.9870.0780.8063.1961.54 Table 1: Performance of models trained on datasets generated or collected by various paradigms, evaluated on mainstream medical QA benchmarks. The highest average score within each paradigm isunderlined. The performance of our method is highlighted in bold. Units are in percentage (%). 4.3 Main Results As shown in Table 1, SAGE consistently outper- forms baseline methods. Compared to self-derived approaches, SAGE achieves substantial gains by anchoring generation in factual knowledge. It also surpasses document-based methods, suggesting that our approach effectively integrates atomic facts into coherent scenarios rather than merely extract- ing them; we provide concrete case studies illustrat- ing this distinction in Appendix J. Notably, SAGE matches or exceeds the performance of large-scale datasets and models fine-tuned on massive CoT data. This efficiency highlights the value of our evolutionary strategy, which maximizes informa- tion density by balancing concept mastery with scenario-based reasoning. 4.4 Ablation Study Contribution of Evolutionary Components. We conducted comprehensive ablation studies on all four components of our proposed method; the results are summarized in Table 1, alongside ex- tended validations of our mapping function, corpus selection, and factual filtering strategies (Appen- dices H, G, I). MedMCQA MedQA PubMedQA MMLU-Pro (med) GPQA(med) Lancet MedBullets (op4) MedBullets (op5) MedXpertQA NEJM 20 40 60 80 64.98% 70.07% 80.80% 63.19% 61.54% 66.02% 59.42% 46.43% 22.71% 62.69% Wrap Self-Instruct CoT-Self-Instruct Ours Figure 3: Comparison between our method and base- lines across diverse medical QA benchmarks, including out-of-distribution tasks. Specifically, for the w/o Iteration variant, the tar- get dataset size was achieved by increasing multi- sample generation within a single iteration, without leveraging the parametric knowledge of the LLM to evolve the concept graph across successive itera- tions. For the w/o Parallel variant, the target sample size was obtained by relaxing the criteria for se- mantic anchor deduplication in subsequent itera- tions. Furthermore, versions lacking either of the 135710 Training Data Size (k) 62 63 64 65 66 67 68 Accuracy (%) 61.90% 66.05% 63.86% 66.82% 63.23% 67.16% 63.67% 67.68% 63.71% 68.12% CoT-Self-Instruct ours Figure 4: Performance scaling analysis comparing our method with CoT-Self-Instruct across different training data sizes from 1k to 10k samples. two question types consistently exhibited marked performance degradation, underscoring the signifi- cance of maintaining diversity in question formats. Scalability and Efficiency. Figure 4 illustrates our method’s high data ef- ficiency, maintaining strong performance across varying scales and rivaling much larger parameter models (Appendix E). Token usage analysis (Ap- pendix L) reveals it consumes significantly fewer tokens than baselines, confirming SAGE’s supe- rior information density. Additionally, our offline semantic mapping ensures extreme computational efficiency compared to standard online API tools (Appendix H). Generalization and Robustness. As shown in Fig- ure 3, our method consistently outperforms base- lines on out-of-distribution (OOD) QA benchmarks (Appendix K). Crucially, this strong generaliza- tion extends beyond QA to other clinical tasks like MedNLI and MTS-Dialog (Appendix D). Fur- thermore, ablation studies across various generator models and architectures (Appendix C) confirm the robustness of our approach irrespective of the underlying model. 5 Discussion 5.1 Qualitative Case: Mitigating Hallucination and Enhancing Reasoning. While document-based data synthesis methods can generate factually correct content, they often pro- duce questions that assess only simple concept re- call, rather than complex reasoning or the appli- cation of knowledge to specific clinical scenarios. This limitation significantly diminishes the learning value of the resulting training corpus. In contrast, self-derived synthesis methods can create corpora that emphasize reasoning, but typically lack mech- anisms for factual verification or refinement. This often results in severe hallucinations, which are un- acceptable in high-risk fields such as medicine. A detailed qualitative comparison of these limitations is provided in Appendix J. We begin with a single medical concept serving as a semantic anchor, accompanied by explanatory text for knowledge grounding. Using a specific examination perspective, we perform atomic gener- ation to produce question-answer pairs that consoli- date the model’s factual understanding of the given concept. Subsequently, leveraging the LLM’s ca- pacity for self-association, we extract multiple con- cepts involved in the generated questions. These ex- tracted concepts then serve as semantic anchors for the next phase of associative generation, enabling the creation of training materials that strengthen the model’s ability to apply concepts in complex medi- cal scenarios. A concrete example is illustrated in Figure 5. As shown in the figure, starting from the core concept “Chronic Pain” with “pathology” as the examination perspective, the LLM generates ques- tions that naturally introduce related concepts such as “Spinal Nerve Roots” and “Ganglia, Spinal”. In the subsequent associative generation phase, these concepts are used as scaffolds to construct appro- priate clinical scenarios, design educationally valu- able questions, and provide responses that integrate accurate knowledge with advanced reasoning. Un- like the baselines, this two-stage process effectively mitigates the hallucination risks of self-derived methods while overcoming the shallow reasoning limitations inherent in document-based approaches. 5.2 Quantitative Assessment: Factuality, Reasoning, and Utility. We evaluated 1,000 generated samples using DeepSeek-V3.2 (Guo et al., 2025) across med- ical factuality, reasoning, and utility (Figure 6). The document-based Wrap (Maini et al., 2024) achieves high factuality but lacks reasoning depth (consistent with Figure 10), limiting its utility. Con- versely, the self-derived ScaleQuest (Ding et al., 2025) shows moderate reasoning but suffers from factual hallucinations. SAGE achieves a superior balance of both. To ensure our factuality claims extend beyond LLM-as-a-judge, we further vali- dated SAGE on standard hallucination benchmarks (TruthfulQA (Lin et al., 2022) and MedHALT (Pal et al., 2023)), where it consistently minimized con- Which of the following best describes a pathological change associated with chronic pain? A. Increased expression of transient receptor potential vanilloid 1 receptors in dorsal root ganglionneurons B. Decreased activity of opioid receptors in the central nervous system C. Loss of myelination in peripheral nerves due to autoimmune destruction D. Increased production of endogenous opioids in the brain during acute pain A 42-year-old female presents with a 2-week history of dull, aching pain in her right forearm. She reports that the pain worsens with activity and is not responsive to over-the-counter analgesics. On examination, a firm, non-tender, smooth nodule is noted on the posterior aspect of her wrist. She denies any history of trauma. Which of the following is the most likely diagnosis? A. Ganglion cystB. Opioid-induced neuropathy C. Spinal nerve rootcompression D. Chronic pain due to opioid use The patient presents with a chronic, dull pain in the forearm, which is not relieved by over-the-counter analgesics, suggesting a non-inflammatory or non- metabolic cause. The presence of a firm, non-tender nodule on the wrist is a classic sign of a ganglion cyst, which is a common benign tumor arising from the ganglia of the tendon sheaths. While spinal nerve root compression or opioid-induced neuropathy could cause chronic pain, the presence of a ganglion cyst makes this the most likely diagnosis. Pathology Chronic Pain Perspective: Anchor： Atomic Generation Option A is correct because... Anchors: Associative Generation Chronic Pain Ganglion Cysts Receptors, Opioid Analgesics, Opioid Spinal Nerve Roots Ganglia, Spinal Explanation：... Figure 5: A detailed case on dual-track evolutionary synthesis. Utility FactualityReasoning 2.01 3.01 3.38 1.93 4.59 1.77 3.54 4.01 4.19 ScaleQuest Wrap Ours Figure 6: Comparison of different methods across factu- ality, reasoning, and utility. fabulations compared to all baselines (Appendix F). 5.3 Combinatorial Validity and Diversity. We conducted a detailed investigation of two self- derived methods that also generate data based on concepts. Specifically, the concepts autonomously generated by the LLM in KPDDS and MathScale were mapped to specific concepts within the MeSH ontology using a medical embedding model (Deka et al., 2022). For each method, we generated a total of 10,000 data points and calculated two metrics for every batch of 1,000 samples: the average semantic depth of the concepts used and the diversity of concept combinations, as shown in Figures 7 and 8. As illustrated, with increasing data scale, our method consistently maintains high levels of both semantic depth and combinatorial diversity. In 200040006000800010000 Number of Generated Samples 3.0 3.5 4.0 4.5 5.0 Depth 3.25 3.23 3.27 3.21 3.25 3.25 3.37 3.35 3.36 3.31 3.31 3.36 4.78 4.91 4.86 4.85 4.84 4.86 KPDDS MathScale Ours Figure 7: Comparison of semantic depth (MeSH Tree) of generated concepts during the generation process. contrast, the baseline methods, which rely on the LLM’s self-directed concept generation, tend to produce broad, shallow, and frequently recurring topical concepts. Subsequent data regeneration based on statistical co-occurrence patterns within these limited initial batches fails to overcome com- mon concept pairings, resulting in a rapid collapse in diversity. 6 Conclusion We introduced SAGE, a Semantic Anchor-Guided Evolution framework that mitigates the data scarcity bottleneck for medical LLMs under strict privacy constraints. By grounding atomic and as- sociative generation in public taxonomies, SAGE enables small local models to autonomously pro- duce pedagogically structured data. Results show it outperforms prevalent paradigms without relying on massive corpora, establishing a practical path- way for developing reliable medical AI assistants directly from a compact knowledge base. 200040006000800010000 Number of Generated Samples 0.0 0.2 0.4 0.6 0.8 1.0 Diversity 0.38 0.040 0.015 0.0070.007 0.003 0.38 0.042 0.024 0.011 0.006 0.004 0.71 0.68 0.66 0.67 0.67 0.66 KPDDS MathScale Ours Figure 8: Comparison of dynamic diversity with in- creasing generated samples. Diversity is measured by the discovery rateR = N new B , whereN new denotes the number of unique unseen concept combinations, andB is the batch size. Limitations While SAGE provides a resource-efficient founda- tion for grounded QA data synthesis, its current de- sign still leaves several limitations for future work. The effectiveness of the framework is partly tied to the scope and organization of external taxonomies such as MeSH, and applying SAGE to domains without mature semantic resources may require constructing task-specific knowledge schemas. Al- though structured anchoring helps reduce hallu- cination, the generated scenarios are still shaped by the combinatorial coverage of anchor concepts and may not fully reflect the nuanced, open-ended reasoning patterns present in authentic clinical nar- ratives. In addition, due to the scale of the syn- thesized data, we did not conduct comprehensive clinician double-checking for all generated sam- ples, which may leave some subtle clinical inaccu- racies unexamined. Our evaluation also primarily focuses on question-answering benchmarks, while broader validation on interactive clinical tasks, such as patient dialogue simulation or diagnostic deci- sion support, would further clarify the practical utility of the framework. Future directions include dynamically expanding semantic anchors, integrat- ing multimodal medical knowledge such as imag- ing data, and adapting SAGE to other high-stakes, knowledge-intensive domains beyond medicine. Ethics Statement This research utilizes only publicly available, de- identified resources: the structured MeSH taxon- omy and open-source medical QA datasets are used exclusively for research purposes. All training data are synthetically generated, circumventing privacy and confidentiality concerns associated with real patient information. No human subjects, personal health data, or protected clinical records were in- volved in this study. Nonetheless, we acknowledge that models trained on synthetic data may repro- duce or amplify biases present in the underlying taxonomies or generation processes, and their out- puts should not be interpreted as clinical advice. Any real-world deployment of this methodology or its derivatives must undergo rigorous clinical validation, adhere to all relevant regulatory stan- dards, and be implemented in human-supervised healthcare workflows. Acknowledgements This work was supported by the Guizhou Provincial Program on Commercialization of Scientific and Technological Achievements (Qiankehezhongyindi [2025] No. 006). References Asma Ben Abacha, Wen-wai Yim, Yadan Fan, and Thomas Lin. 2023. An empirical study of clinical note generation from doctor-patient encounters. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Lin- guistics, pages 2291–2302. Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. 2021. Extracting training data from large language models. In 30th USENIX security symposium (USENIX Security 21), pages 2633–2650. Hanjie Chen, Zhouxiang Fang, Yash Singla, and Mark Dredze. 2025a. Benchmarking large language mod- els on answering and explaining challenging medical questions. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Asso- ciation for Computational Linguistics: Human Lan- guage Technologies (Volume 1: Long Papers), pages 3563–3599. Junying Chen, Zhenyang Cai, Ke Ji, Xidong Wang, Wanlong Liu, Rongsheng Wang, and Benyou Wang. 2025b. Towards medical complex reasoning with llms through medical verifiable problems. In Find- ings of the Association for Computational Linguistics: ACL 2025, pages 14552–14573. Zihong Chen, Wanli Jiang, Jinzhe Li, Zhonghang Yuan, Huanjun Kong, Wanli Ouyang, and Nanqing Dong. 2025c. Graphgen: Enhancing supervised fine-tuning for llms with knowledge-driven synthetic data gener- ation. arXiv preprint arXiv:2505.20416. Bhishma Dedhia, Yuval Kansal, and Niraj K Jha. 2025. Bottom-up domain-specific superintelligence: A re- liable knowledge graph is what we need. arXiv preprint arXiv:2507.13966. Pritam Deka, Anna Jurek-Loughrey, and P Deepak. 2022.Improved methods to aid unsupervised evidence-based fact checking for online health news. Journal of Data Intelligence, 3(4):474–504. Yuyang Ding, Xinyu Shi, Xiaobo Liang, Juntao Li, Zhaopeng Tu, Qiaoming Zhu, and Min Zhang. 2025. Unleashing llm reasoning capability via scalable question synthesis from scratch. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13414–13438. Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, and Dawn Song. 2023. The false promise of imitating proprietary llms. arXiv preprint arXiv:2305.15717. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. 2025. Deepseek-r1 in- centivizes reasoning in llms through reinforcement learning. Nature, 645(8081):633–638. Xiaoke Huang, Juncheng Wu, Hui Liu, Xianfeng Tang, and Yuyin Zhou. 2025a. m1: Unleash the potential of test-time scaling for medical reasoning with large language models. arXiv preprint arXiv:2504.00869. Yiming Huang, Xiao Liu, Yeyun Gong, Zhibin Gou, Yelong Shen, Nan Duan, and Weizhu Chen. 2025b. Key-point-driven data synthesis with its enhance- ment on mathematical reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 24176–24184. Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2021. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Ap- plied Sciences, 11(14):6421. Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. 2019. PubMedQA: A dataset for biomedical research question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Lan- guage Processing (EMNLP-IJCNLP), pages 2567– 2577, Hong Kong, China. Association for Computa- tional Linguistics. Georgios A Kaissis, Marcus R Makowski, Daniel Rück- ert, and Rickmer F Braren. 2020. Secure, privacy- preserving and federated machine learning in medi- cal imaging. Nature Machine Intelligence, 2(6):305– 311. Abdullatif Köksal, Timo Schick, Anna Korhonen, and Hinrich Schütze. 2024. Longform: Effective instruc- tion tuning with reverse instructions. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7056–7078. Yanis Labrak, Adrien Bazoge, Emmanuel Morin, Pierre- Antoine Gourraud, Mickaël Rouvier, and Richard Dufour. 2024. Biomistral: A collection of open- source pretrained large language models for medical domains. In Findings of the Association for Compu- tational Linguistics: ACL 2024, pages 5848–5864. Chen Li, Weiqi Wang, Jingcheng Hu, Yixuan Wei, Nan- ning Zheng, Han Hu, Zheng Zhang, and Houwen Peng. 2024. Common 7b language models already possess strong math capabilities. arXiv preprint arXiv:2403.04706. Rumeng Li, Xun Wang, and Hong Yu. 2023. Two direc- tions for clinical data generation with large language models: data-to-label and label-to-data. In Proceed- ings of the Conference on Empirical Methods in Nat- ural Language Processing. Conference on Empirical Methods in Natural Language Processing, volume 2023, page 7129. Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. Truthfulqa: Measuring how models mimic human falsehoods. In Proceedings of the 60th annual meet- ing of the association for computational linguistics (volume 1: long papers), pages 3214–3252. Haoxiong Liu, Yifan Zhang, Yifan Luo, and Andrew C Yao. 2025. Augmenting math word problems via iterative question composing. In Proceedings of the AAAI Conference on Artificial Intelligence, vol- ume 39, pages 24605–24613. Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jian- guang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. 2023. Wiz- ardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583. Pratyush Maini, Skyler Seto, Richard Bai, David Grang- ier, Yizhe Zhang, and Navdeep Jaitly. 2024. Rephras- ing the web: A recipe for compute and data-efficient language modeling. In Proceedings of the 62nd An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14044– 14072. Ziyang Miao, Qiyu Sun, Jingyuan Wang, Yuchen Gong, Yaowei Zheng, Shiqi Li, and Richong Zhang. 2025. Easy dataset: A unified and extensible framework for synthesizing llm fine-tuning data from unstructured documents. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Process- ing: System Demonstrations, pages 960–968. Ankit Pal, Logesh Kumar Umapathi, and Malaikan- nan Sankarasubbu. 2022. Medmcqa: A large-scale multi-subject multi-choice dataset for medical do- main question answering. In Conference on health, inference, and learning, pages 248–260. PMLR. Ankit Pal, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. 2023. Med-halt: Medical domain hallucination test for large language models. In Pro- ceedings of the 27th Conference on Computational Natural Language Learning (CoNLL), pages 314– 334. W Nicholson Price and I Glenn Cohen. 2019. Privacy in the age of medical big data. Nature medicine, 25(1):37–43. Pranav Rajpurkar, Emma Chen, Oishi Banerjee, and Eric J Topol. 2022. Ai in health and medicine. Na- ture medicine, 28(1):31–38. David Rein, Betty Li Hou, Asa Cooper Stickland, Jack- son Petty, Richard Yuanzhe Pang, Julien Dirani, Ju- lian Michael, and Samuel R Bowman. 2024. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling. Haris Riaz, Sourav Sanjukta Bhabesh, Vinayak Aran- nil, Miguel Ballesteros, and Graham Horwood. 2025. Metasynth: Meta-prompting-driven agentic scaffolds for diverse synthetic data generation. In Findings of the Association for Computational Linguistics: ACL 2025, pages 18770–18803. Alexey Romanov and Chaitanya Shivade. 2018. Lessons from natural language inference in the clini- cal domain. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Process- ing, pages 1586–1596, Brussels, Belgium. Associa- tion for Computational Linguistics. Zhengyang Tang, Xingxing Zhang, Benyou Wang, and Furu Wei. 2024. Mathscale: Scaling instruction tun- ing for mathematical reasoning. In Forty-first Inter- national Conference on Machine Learning. Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. 2023. Large language models in medicine. Nature medicine, 29(8):1930– 1940. Bingning Wang, Haizhou Zhao, Huozhi Zhou, Liang Song, Mingyu Xu, Wei Cheng, Xiangrong Zeng, Yupeng Zhang, Yuqi Huo, Zecheng Wang, et al. 2025. Baichuan-m1: Pushing the medical capa- bility of large language models.arXiv preprint arXiv:2502.12671. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023a. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations. Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023b. Self-instruct: Aligning language models with self-generated instructions. In Proceed- ings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers), pages 13484–13508. Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. 2024. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. In The Thirty- eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track. Juncheng Wu, Wenlong Deng, Xingxuan Li, Sheng Liu, Taomian Mi, Yifan Peng, Ziyang Xu, Yi Liu, Hyunjin Cho, Chang-In Choi, et al. 2025. Medreason: Elicit- ing factual medical reasoning steps in llms via knowl- edge graphs. arXiv preprint arXiv:2504.00993. Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, and Daxin Jiang. 2024. Wizardlm: Empowering large pre-trained language models to follow complex instructions. In The Twelfth International Conference on Learning Representations. Weiwen Xu, Hou Pong Chan, Long Li, Mahani Alju- nied, Ruifeng Yuan, Jianyu Wang, Chenghao Xiao, Guizhen Chen, Chaoqun Liu, Zhaodonghui Li, et al. 2025. Lingshu: A generalist foundation model for unified multimodal medical understanding and rea- soning. arXiv preprint arXiv:2506.07044. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024a. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115. Zitong Yang, Neil Band, Shuangping Li, Emmanuel Candes, and Tatsunori Hashimoto. 2024b.Syn- thetic continued pretraining.arXiv preprint arXiv:2409.07431. Asaf Yehudai, Boaz Carmeli, Yosi Mass, Ofir Arviv, Nathaniel Mills, Eyal Shnarch, and Leshem Choshen. 2024. Achieving human parity in content-grounded datasets generation. In The Twelfth International Conference on Learning Representations. Hongzhou Yu, Tianhao Cheng, Yingwen Wang, Wen He, Qing Wang, Ying Cheng, Yuejie Zhang, Rui Feng, and Xiaobo Zhang. 2025a. FinemedLM-o1: Enhanc- ing medical knowledge reasoning ability of LLM from supervised fine-tuning to test-time training. In Second Conference on Language Modeling. Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhen- guo Li, Adrian Weller, and Weiyang Liu. 2023. Metamath: Bootstrap your own mathematical ques- tions for large language models. arXiv preprint arXiv:2309.12284. Ping Yu, Jack Lanchantin, Tianlu Wang, Weizhe Yuan, Olga Golovneva, Ilia Kulikov, Sainbayar Sukhbaatar, Jason Weston, and Jing Xu. 2025b. Cot-self-instruct: Building high-quality synthetic prompts for rea- soning and non-reasoning tasks.arXiv preprint arXiv:2507.23751. Weihao Zeng, Can Xu, Yingxiu Zhao, Jian-Guang Lou, and Weizhu Chen. 2024.Automatic instruction evolving for large language models. arXiv preprint arXiv:2406.00770. Kaiyan Zhang, Sihang Zeng, Ermo Hua, Ning Ding, Zhang-Ren Chen, Zhiyuan Ma, Haoxin Li, Ganqu Cui, Biqing Qi, Xuekai Zhu, et al. 2024. Ultramedi- cal: Building specialized generalists in biomedicine. Advances in Neural Information Processing Systems, 37:26045–26081. Xuanyu Zhang and Qing Yang. 2023. Self-qa: Unsuper- vised knowledge guided language model alignment. arXiv preprint arXiv:2305.11952. Xueliang Zhao, Wei Wu, Jian Guan, Zhuocheng Gong, and Lingpeng Kong. 2025a. Promptcot 2.0: Scaling prompt synthesis for large language model reasoning. arXiv preprint arXiv:2509.19894. Xueliang Zhao, Wei Wu, Jian Guan, and Lingpeng Kong. 2025b. Promptcot: Synthesizing olympiad- level problems for mathematical reasoning in large language models. arXiv preprint arXiv:2503.02324. Yuxin Zuo, Shang Qu, Yifei Li, Zhang-Ren Chen, Xuekai Zhu, Ermo Hua, Kaiyan Zhang, Ning Ding, and Bowen Zhou. 2025. MedxpertQA: Benchmark- ing expert-level medical reasoning and understand- ing. In Forty-second International Conference on Machine Learning. A Data and Knowledge Base A.1 Knowledge Base Description Medical Subject Headings (MeSH) is a comprehen- sive controlled vocabulary developed by the Na- tional Library of Medicine (NLM) 4 , primarily used for indexing, cataloging, and searching biomedical and health-related information. As the authorita- tive thesaurus for the MEDLINE 5 and PubMed 6 databases, MeSH enforces uniform terminology and consistency in the retrieval of biomedical lit- erature. Its hierarchical tree structure organizes descriptors from broad, general categories down to specific concepts; each descriptor is assigned one or more “Tree Numbers” that indicate its precise location within this hierarchy. This organization robustly encodes semantic relationships between medical concepts, making MeSH an essential foun- dation for constructing domain-specific knowledge graphs and ontologies. We leverage the built-in synonym information within the MeSH ontology to implement the mapping functionφ, as described in Section 3, using a lightweight and efficient string matching approach. To focus on medically relevant content, we retain only mid-to-deep level terms from core medical categories as candidate semantic anchors(including Anatomy; Diseases; Chemicals and Drugs; Analytical, Diagnostic and Therapeutic Techniques, and Equipment; Psychiatry and Psy- chology; and Phenomena and Processes). Further- more, when selecting a single anchor and its associ- ated inquiry perspective, we consider only the core medical perspectives pertinent to that concept, such as clinical diagnosis and treatment, pharmacology, pathophysiology, and mechanism-based inquiry. A.2 Seed Dataset Overview We utilize the m1k (Huang et al., 2025a) dataset as the seed for our entire data synthesis pipeline. It provides concepts distributions and few-shot exam- ples that serve as initial scaffolds for guiding small- scale models in complex data generation tasks. The m1k dataset represents a curated medical QA benchmark, distilled from an initial pool of 196K public samples across MedMCQA, MedQA- USMLE, HeadQA, and PubMedQA. The construc- tion process involved strict difficulty filtering, iso- lating 37K complex questions that Qwen2.5-7B 4 https://w.nlm.nih.gov/ 5 https://w.nlm.nih.gov/medline/medline_home. html 6 https://pubmed.ncbi.nlm.nih.gov/ and Qwen2.5-32B failed to solve. These were subsequently augmented with verified Chain-of- Thought (CoT) solutions from DeepSeek-R1. Ul- timately, through diversity sampling stratified by MeSH categories and balanced by source, a core training set of 1,000 instances was established. A.3 Benchmark Datasets To comprehensively assess the efficacy of our pro- posed method, we utilize a suite of established medical QA benchmarks: • MedMCQA (Pal et al., 2022): A massive dataset of over 194,000 multiple-choice ques- tions covering 21 medical disciplines and 2,400 topics. Originated from real-world In- dian entrance exams (AIIMS and NEET-PG), we utilize the specific 4,183-question test set adopted by (Chen et al., 2025b) (Apache-2.0). •MedQA (Jin et al., 2021): Derived from USMLE practice materials, this corpus con- tains 12,723 questions compiled from 18 clin- ical textbooks. Solving these requires expert- level reasoning and the ability to synthe- size evidence from multiple documents. We employ the 1,273-sample test split follow- ing (Chen et al., 2025b) (C-BY-4.0). • PubMedQA (Jin et al., 2019): A biomedical QA task grounded in PubMed abstracts. It features 1,000 expert-labeled multiple-choice questions drawn from a broader collection of 211k articles (MIT). • MMLU-Pro (Medical) (Wang et al., 2024): The Professional Medicine segment of the Massive Multitask Language Understanding benchmark, targeting advanced clinical knowl- edge. We adopt the test splits used in (Chen et al., 2025b) (MIT). • GPQA (Medical) (Rein et al., 2024): Rep- resenting the biomedical portion of the Graduate-Level QA dataset, these questions are expert-written and designed to be Google- proof, resisting simple retrieval strategies. We strictly use the biology/medical subset and splits from (Chen et al., 2025b) (C-BY-4.0). •Lancet & NEJM: Two compact datasets con- structed from clinical case reports published in The Lancet and the New England Jour- nal of Medicine (NEJM), curated according to (Huang et al., 2025a). •MedBullets (Chen et al., 2025a): A collec- tion of practice items from the MedBullets platform. Following (Huang et al., 2025a), we focus on the high-difficulty subsets (Levels 4 and 5), labeled as MedBullets_Op4 and Med- Bullets_Op5, each containing approximately 100 rigorous questions. • MedXpertQA (Zuo et al., 2025): A set of 50 expert-crafted questions requiring multi- step reasoning, employed here for qualitative analysis (MIT). B Implementation Details and Hyperparameters Candidate instances are sampled with a tempera- ture of0.7, top-pof0.95, and a maximum genera- tion length of 4,096 tokens. For each eligible an- chor configuration, we generateN = 2candidates. Associative generation usesn few = 2examples sampled from the growing few-shot pool, whereas self-consistency checking usesK = 3auxiliary answers. C Training on Weaker Models To demonstrate the robustness of our method in real-world scenarios, we evaluated data generated from Qwen3-4B on weaker models. As shown in Figure 9, our method achieved significant advan- tages across different model architectures, high- lighting its strong adaptability. DGeneralization on non-QA Benchmark To demonstrate SAGE’s generalization on non-QA benchmark, we further evaluated it on MedNLI (Romanov and Shivade, 2018) (Natural Language Inference) and MTS-Dialog (Abacha et al., 2023) (Medical Summarization). As shown in Table 2, SAGE consistently outperforms synthesis baselines on both tasks. SAGE provides the strongest gains among all synthesis paradigms. E Performance Ceiling and Cross-Scale Comparison To provide context on the difficulty of the clini- cal benchmarks evaluated in this study and offer insight for non-specialist readers, we present a per- formance ceiling and a cross-scale comparison. Ta- ble 3 compares our model against recent state-of- the-art (SOTA) open-source medical LLMs, heavy- weight general models, and flagship reasoning mod- els. MedMCQAMedQAPubMedQA MMLU-Pro(med)GPQA(med) 0 20 40 60 80 Accuracy (%) 49.3 50.6 66.2 50.9 38.5 57.9 58.2 79.8 50.6 52.3 59.0 59.9 83.7 53.7 53.3 Qwen2.5-3B-Instruct MedMCQAMedQAPubMedQA MMLU-Pro(med)GPQA(med) 50.6 49.2 73.3 49.3 39.5 56.8 59.7 82.8 46.6 49.0 60.3 61.3 85.4 50.9 60.0 Llama3.2-3B-Instruct BaselineCoT-Self-Instructours Figure 9: Performance on the benchmark after training Qwen2.5-3B-Instruct and Llama3.2-3B-Instruct. StrategyMedNLIMTS Wrap72.0883.62 CoT-Self-Instruct84.6083.16 SAGE87.2785.02 Table 2: Generalization performance of SAGE com- pared to baselines. All trainee models are based on Qwen3-4B-Base. We report Accuracy for MedNLI and BERTScore for MTS-Dialog. All results are scaled to percentages (%). As the table illustrates, these clinical bench- marks are challenging. Even parameter-heavy mod- els like UltraMedical-8B struggle to surpass a 63% average. By providing this ceiling, it becomes ev- ident that pushing a 4B model to a 68.12% aver- age, matching the 8B SOTA and approaching the performance of a 72B general model, is a highly non-trivial achievement. This underscores the effi- ciency and high quality of the SAGE synthetic data paradigm. F Quantitative Assessment of Hallucinations Beyond LLM-as-a-Judge To provide a more comprehensive assessment and move beyond relying solely on LLM-judges, we conducted an additional rigorous hallucination anal- ysis using two complementary, established bench- marks: •TruthfulQA (Medical subset) (Lin et al., 2022): A gold standard for evaluating general truthfulness and the model’s susceptibility to common misconceptions. •MedHALT (Fake Question subset) (Pal et al., 2023): A specialized medical bench- mark designed specifically to test “confabu- lation”, the tendency of models to generate realistic-looking answers for non-existent or fabricated medical queries. We report the Hallucination Rate (where lower is better, denoted by↓) in Table 4. SAGE consistently maintains the lowest hallucination rate across both benchmarks. This confirms that our method effec- tively calibrates the model’s factual boundaries by grounding generation in MeSH taxonomies, rather than relying on unconstrained self-generation. G In-depth Comparison with Document-Based Methods Our problem setting is specifically designed to simulate a practical, resource-constrained scenario (e.g., a small-scale specialized clinic), where mas- sive, clean medical corpora are often inaccessible due to privacy or computing limitations. In this context, the MeSH taxonomy serves as a founda- tional medical dictionary, while the 1K seed Q&A samples represent years of accumulated patient case records. Our initial baseline comparisons uti- lized MeSH Scope Notes as the primary corpus for document-based methods (e.g., Genie) specifically to ensure strict fairness in knowledge sources, pre- venting any unintended advantage from external data. To comprehensively assess whether the perfor- mance gap between SAGE and document-based methods stems from the brevity of scope notes rather than the synthesis paradigm itself, we con- ducted two rigorous ablation experiments: 1.Data Augmentation: We augmented the document-based baseline with the identical 1K seed Q&A cases used by SAGE. ModelMedMCQAMedQAPubMedQAMMLU-ProGPQAAVG (med)(med) SAGE-4B64.9870.0780.8063.1961.5468.12 ∼8B Scale UltraMedical-8B (Zhang et al., 2024)59.1771.6470.3061.4350.2662.56 m1-7B (Huang et al., 2025a)61.4971.9674.0063.6546.6763.55 MedReason-8B (Wu et al., 2025)60.1770.7878.5065.0852.5665.42 HuatuoGPT-o1-8B (Chen et al., 2025b)63.7875.0280.5064.2357.1868.14 Above 70B Scale Qwen2.5-72B-Instruct66.6074.5570.8966.0662.0568.03 UltraMedical-70B (Zhang et al., 2024)72.9483.9080.0073.9458.7273.90 Ceiling Reference Deepseek-V3-032475.3088.6669.3384.6766.0076.79 Deepseek-R1-052880.2093.0768.3290.1073.2780.99 Table 3: Cross-scale comparison and performance ceiling on clinical benchmarks. Results are reported in percentages (%). SAGE-4B matches the performance of SOTA 8B medical models and approaches the 72B general model. StrategyTruthfulQA↓MedHALT↓ Wrap32.6690.84 CoT-SI22.6191.59 SAGE20.6089.28 Table 4: Hallucination rates on TruthfulQA (Medical subset) and MedHALT (Fake Question subset). Lower values indicate better factuality. All trainee models are based on Qwen3-4B-Base. “CoT-SI” stands for CoT- Self-Instruct. Results are reported in percentages (%). 2.Corpus Substitution: We replaced the scope notes entirely with a traditional medical cor- pus (PubMed abstracts), extracting an equiv- alent volume of documents to strictly match SAGE’s 3.51M token budget. As shown in Table 5, simply scaling the token budget or switching to traditional medical docu- ments yields only marginal improvements for the document-based baseline, which remains signifi- cantly inferior to SAGE. This demonstrates that SAGE’s superiority does not stem from simply ex- posing the model to more text. Instead, its effec- tiveness lies in its unique ability to distill static, dictionary-style facts into dense, high-quality clin- ical reasoning instructions, which is far more ef- fective for fine-tuning small LLMs in data-scarce environments. H Robustness and Evaluation of the Mapping Function The mapping functionφis fundamentally grounded in string matching, which guarantees its lightweight nature and extremely low computa- tional overhead. However, to ensure high clinical specificity and robustness against extraction errors, φextends significantly beyond naive string match- ing by employing a synonym-expanded matching and structural filtering pipeline. Specifically, we incorporate the comprehensive MeSH synonym dictionary to capture semantic variations during the matching process. Furthermore, to eliminate shallow, noisy vocabulary (e.g., broad terms like “Disease”), we strictly filter out terms with a MeSH tree depth< 4and restrict anchors exclusively to core medical categories (e.g., Anatomy, Diseases). This design ensures that all extracted semantic an- chors maintain high clinical specificity without sac- rificing processing speed. Quantitative Error AnalysisTo provide a quan- titative evaluation of the concept extraction, we an- alyzed 1,000 seed samples using the official MeSH on Demand (MoD) tool, a sophisticated online en- gine utilizing MetaMap and PubMed k-N map- ping which can be used as the pseudo-gold stan- dard. As shown in Table 6, while naive string matching achieves a higher recall, it suffers from remarkably low precision (36.52%). This indicates that while it successfully extracts many terms, the anchor set StrategyData SourceTokensMedMCQAMedQAPubMedQAMMLU-ProGPQAAVG GenieScope Notes1.69M49.3251.7763.9048.9935.9049.98 GenieNotes + Distilled Q&A3.51M46.8655.1575.3052.3135.1352.95 GeniePubMed Abstracts3.51M52.3848.6361.0051.2140.0050.64 SAGENotes + Distilled Q&A3.51M64.9870.0780.8063.1961.5468.12 Table 5: Ablation study on data sources and token budgets. All trainee models are Qwen3-4B-Base. MMLU-Pro and GPQA refer to their respective medical subsets. Results are reported in percentages (%). Extraction MethodPrecisionRecall Naive String Matching36.5255.30 SAGE Mapping (Ours)55.6812.21 Table 6: Quantitative evaluation of extraction methods against the MeSH on Demand pseudo-gold standard. Results are reported in percentages (%). is heavily flooded with clinically uninformative vo- cabulary. SAGE intentionally trades off the recall of these broad terms (12.21%) to maximize the clin- ical density and precision (55.68%) of the extracted anchors. In the context of data synthesis, feeding small LLMs with a few highly precise, deep seman- tic anchors is vastly superior to flooding them with numerous shallow, noisy concepts. Thus, the high precision of ourφfunction safeguards the struc- tural integrity and clinical density of the generation trajectories. Downstream Performance and Computational Overhead To definitively validate the robust- ness ofφ, we compared the downstream QA per- formance, privacy constraints, and computational overhead of these extraction strategies. All ex- periments utilized Qwen3-4B-Base as the trainee model. As Table 7 illustrates, naive string matching per- forms poorly (64.02% average) due to the inter- ference of noisy anchors. Conversely, while the official MoD tool yields strong downstream results, it requires continuous external network requests, violating the strict data privacy setting of practi- cal clinical environments, and incurs an immense time overhead (∼40×slower), making it unscal- able for large-scale data synthesis. Our SAGE map- ping achieves virtually identical downstream per- formance (68.12% vs. 68.14%) compared to the heavy online expert tool, but executes securely of- fline at a fraction of the computational cost. This confirms that our framework is highly robust to ex- traction noise and optimally designed for resource- constrained, privacy-sensitive scenarios. I Ablation on Factual Grounding and Filtering Strategies While self-consistency voting is a widely adopted proxy for uncertainty estimation in resource- constrained settings (e.g., in CoT-Self-Instruct), it serves primarily as a confidence filter to prune un- stable reasoning paths and does not necessarily guarantee factual accuracy. If a model exhibits sys- tematic knowledge gaps, repeated reasoning paths may converge on the same incorrect answer, poten- tially allowing consistently erroneous samples to pass the filtering process. To prevent these systematic knowledge gaps, SAGE does not rely solely on post-hoc filtering. Through Atomic Generation, structured factual in- formation (MeSH definitions) is explicitly incorpo- rated into the generation trajectory. This ensures the synthesis is grounded in authoritative knowl- edge from the start, mitigating the risk of the model consistently repeating its internal flawed guesses. To validate this design choice, we tested an alter- native approach using contradiction filtering, where MeSH Scope Notes were used as a reference for the small model to post-verify its outputs. As shown in Table 8, this approach yielded inferior results. In resource-constrained settings, small models lack the discriminative capacity to accurately detect sub- tle semantic contradictions. Consequently, it is more effective to embed factual knowledge directly into the training data generation phase and use self- consistency to reinforce the model’s confidence. J Qualitative Analysis of Baseline Methods In this section, we present a detailed case study illustrating the limitations of existing synthesis paradigms discussed in Section 5.1. As illus- trated in Figure 10, the Wrap method, representing document-based synthesis, generates a simplistic question about distinguishing “acute pain” from Extraction MethodAvg Time / SamplePrivacyMedMCQAMedQAPubMedQAMMLU-ProGPQAAVG Naive Matching1.15sOffline63.8169.6070.4061.9554.3664.02 MeSH on Demand48.24sOnline API66.0867.1683.2062.7461.5468.14 SAGE Mapping1.31sOffline64.9870.0780.8063.1961.5468.12 Table 7: Comparison of extraction methods regarding downstream QA performance, computational overhead, and privacy. MMLU-Pro and GPQA refer to their respective medical subsets. Downstream metrics are reported in percentages (%). Filtering StrategyMedMCQAMedQAPubMedQAMMLU-ProGPQAAVG Contradiction Filtering64.6768.0370.8060.7258.4664.54 SAGE (Consistency + Atomic Gen.)64.9870.0780.8063.1961.5468.12 Table 8: Performance comparison between post-hoc contradiction filtering and the SAGE approach. MMLU-Pro and GPQA refer to their respective medical subsets. Results are reported in percentages (%). How is acute pain different from other types of pain? A 32-year-old man presents with sudden onset of severe pain and numbness in his right hand after sleeping with his hand in a flexed position for several hours. He reports that the pain is sharp and localized, and he cannot move his fingers without significant discomfort. Which type of nerve fiber is most likely involved in this condition? A. A-alpha fibers B. C fibers C. A-gamma fibers D. B fibers The patient's symptoms include sharp, localized pain and numbness in the right hand, which occurs after prolonged flexion of the hand. This is consistent with a condition involving small-diameter nerve fibers, particularly those responsible for transmitting pain signals. C fibers are unmyelinated and are responsible for transmitting slow, dull, aching pain, which is consistent with the sharp, localized pain described. However, the numbness suggests a possible issue with sensory nerve function, which might involve other fibers. But in the context of the question, the primary symptom of sharp pain points towards C fibers. Therefore, the correct answer is B. Document-based Synthesis It is distinct in its intensity, distress, and clear onset and duration. Self-derived Synthesis Figure 10: A detailed case on self-derived and document-based synthesis. other types of pain. The corresponding response, while factually inoffensive, is superficial and of limited educational utility. On the other hand, CoT- Self-Instruct, representing self-derived synthesis, produces a question about the functions and clini- cal localization of different nerve fiber types. How- ever, due to insufficient factual grounding, it fails to include A-delta fibers as a valid option within the appropriate clinical context, and incorrectly at- tributes sharp pain to C fibers. This constitutes a serious hallucination that contradicts fundamental neurophysiological consensus. K Evaluation on Extended Benchmarks We evaluated all baseline methods on the extended benchmarks and compiled their performance. The results, presented in Figure 9, demonstrate that our approach consistently achieves leading perfor- mance. 010002000300040005000600070008000 Number of Generated Tokens 0 100 200 300 400 500 600 700 800 Count / Density Mean: 2655.2 Median: 2033.0 Figure 11: Distribution chart of token lengths in the test set on Qwen3-4B with thinking mode. L Token Efficiency Analysis To demonstrate the token efficiency of our method, we compared the token lengths of inference re- sponses across multiple baselines on ten bench- mark test sets (results shown in Figures 11, 12, 13, and 14). Our method uses the non-thinking mode ModelMethod/DatasetScaleAVG-5AVG-10LancetMedBul_op4MedBul_op5MedXpertNEJM Qwen3-4B-BaseFew-shot PromptingNA26.1524.8830.5829.5524.0311.1822.72 Qwen3-4B NoThinkingNANA62.5454.1559.4750.3248.0513.0457.88 Qwen3-4B ThinkingNANA66.1857.9862.8659.0950.9713.7362.19 Open Source Datasets Qwen3-4B-Basem1k1k63.4056.9861.6559.7454.8717.0559.54 Qwen3-4B-Basem23k23k 67.3961.4465.0567.5361.0420.5063.35 Qwen3-4B-BaseHuatuoGPT-o1-SFT20k52.8048.2256.3151.6243.5115.7351.08 Qwen3-4B-BaseMedReason33k58.6051.7651.9454.2249.3516.8452.24 Self-derived Method Qwen3-4B-Baseself-instruct10k53.6647.2650.7347.0841.2313.1852.07 Qwen3-4B-Basescalequest10k62.9154.1758.2551.3047.4013.6656.55 Qwen3-4B-BaseKPDDS10k 58.9950.2653.8844.1641.5613.8754.23 Qwen3-4B-BaseMathScale10k57.8550.3759.2247.7339.2914.6353.57 Qwen3-4B-BaseCot-Self-Instruct10k63.7156.8064.3257.1448.7019.1260.20 Document-based Method Qwen3-4B-BaseEasy-Dataset28k49.1141.4044.1737.3434.0911.3241.46 Qwen3-4B-BaseGenie30k49.9844.8754.8544.1637.9914.2247.60 Qwen3-4B-BaseSelf-QA230k 46.3442.6250.7342.2139.2913.6648.59 Qwen3-4B-BaseWrap98K50.2045.4353.4044.4842.2115.1148.09 Ours Qwen3-4B-Basew/o Iteration10k66.4558.5163.8358.1250.3221.3959.20 Qwen3-4B-Basew/o Parallel10k66.8158.3362.3857.1448.3823.1258.21 Qwen3-4B-Basew/o Atomic gen.10k66.2158.1762.1457.4751.3022.1557.55 Qwen3-4B-Basew/o Associative gen.10k 66.1359.6268.2059.0952.2723.6062.35 Qwen3-4B-BaseOurs(Full)10k68.1259.7966.0259.4246.4322.7162.69 Table 9: Performance comparison on additional datasets. Since the first five medical datasets are detailed in the main text, this table focuses on the remaining datasets and the overall AVG-10 scores (%). of Qwen3-4B for data synthesis, mainly consider- ing the efficiency and practicality of offline data generation under resource-constrained settings. We further compare thinking and non-thinking modes for data synthesis using both SAGE and CoT-Self- Instruct, with each setting repeated over three runs. As shown in Table 10, using thinking mode dur- ing data synthesis does not bring consistent im- provements. As illustrated, our approach not only achieves state-of-the-art performance as shown in Section 4.3, but also maintains significantly lower token lengths compared to all baseline methods. Specifically, our method yields an average token length of only230.3, which is notably lower than the2655.2tokens of Qwen3-4B’s thinking mode and the2947.3tokens of the M23K dataset distilled from R1-style responses. Furthermore, the Wrap method exhibits excessively short token lengths, averaging only57.5, which aligns with the analysis presented in Appendix J. Therefore, our method achieves a favorable balance between downstream effectiveness and inference-time token efficiency. M Stability Across Independent Runs To further examine the stability of SAGE, we con- duct three independent runs for SAGE and two primary synthetic-data baselines, CSI (CoT-Self- Instruct) and WRAP, using the same experimental setting. Table 11 reports the mean accuracy and standard deviation across five medical QA bench- marks. As shown in the table, SAGE consistently achieves the best average performance across the evaluated benchmarks across runs. This suggests that the observed improvements are not driven by a single favorable run, but reflect stable gains from the proposed grounded synthesis framework. N Comparison with Fine-Tuning on Original Training Sets We also compare SAGE with supervised fine- tuning on the original training sets of datasets that provide dedicated training splits. Specifically, we fine-tune the same base LLM on the training sets of MedMCQA, MedQA, and PubMedQA, which con- tain approximately 10k, 100k, and 200k samples, respectively. Table 12 reports the mean accuracy and standard deviation across multiple runs. Despite using ex- isting human-curated training data, fine-tuning on the original dataset training splits yields substan- tially lower performance than SAGE. One possible reason is that these datasets were not primarily de- signed for instruction tuning of LLMs: their target MethodAVG-5MedMCQAMedQAPubMedQAMMLU-Pro (med)GPQA (med) SAGE (thinking) 67.85± 0.26 65.11± 0.89 67.61± 0.39 82.27± 0.81 62.48± 0.75 61.80± 1.18 SAGE (non-thinking) 68.05± 0.22 65.26± 0.98 68.24± 1.85 82.03± 0.98 62.21± 1.02 62.50± 2.31 CoT-Self-Instruct (thinking) 63.57± 0.50 62.34± 0.13 68.34± 0.95 69.63± 0.23 60.98± 0.55 56.58± 0.97 CoT-Self-Instruct (non-thinking) 63.75± 0.10 61.47± 0.92 67.11± 0.94 74.30± 4.07 62.26± 0.44 53.59± 3.20 Table 10: Comparison between thinking and non-thinking modes for data synthesis. All results are reported as mean ± standard deviation over three runs. AVG-5 denotes the average score over the five medical benchmarks (%). MethodAVG-5MedMCQAMedQAPubMedQAMMLU-Pro (med)GPQA (med) SAGE 68.05± 0.22 65.26± 0.98 68.24± 1.85 82.03± 0.98 62.21± 1.02 62.50± 2.31 CSI 63.75± 0.10 61.47± 0.92 67.11± 0.94 74.30± 4.07 62.26± 0.44 53.59± 3.20 WRAP 49.18± 0.95 47.81± 0.32 47.03± 0.79 66.37± 3.10 46.40± 0.69 38.29± 1.41 Table 11: Mean accuracy and standard deviation over three independent runs. SAGE shows consistently strong performance with low variance across five medical QA benchmarks (%). DatasetTrain-SetSAGE MedMCQA 57.42± 0.76 65.26± 0.98 MedQA 56.79± 1.85 68.24± 1.85 PubMedQA 50.60± 1.12 82.03± 0.98 Table 12: Comparison between fine-tuning on origi- nal dataset training splits and fine-tuning on SAGE- generated data. Results are reported as mean accuracy and standard deviation across multiple runs (%). 05001000150020002500300035004000 Number of Generated Tokens 0 1000 2000 3000 4000 5000 6000 7000 Count / Density Mean: 2947.3 Median: 3998.0 Figure 12: Distribution chart of token lengths in the test set after training with m23k dataset on Qwen3-4B-Base. outputs are often short answer-only labels, pro- viding limited supervision for learning reasoning- oriented response patterns. In contrast, SAGE con- structs explanation-rich instruction data grounded in semantic anchors, which provides stronger su- pervision for medical question answering. O Prompt Templates To diversify our iterative data generation process, we employed two distinct question types: atomic generation and associative generation. The specific prompt templates used are provided in Figure 15 and Figure 16, respectively. 05001000150020002500300035004000 Number of Generated Tokens 0 2000 4000 6000 8000 10000 Count / Density Mean: 57.5 Median: 14.0 Figure 13: Distribution chart of token lengths in the test set after training the data synthesized by the Wrap method on Qwen3-4B-Base. 05001000150020002500300035004000 Number of Generated Tokens 0 1000 2000 3000 4000 5000 Count / Density Mean: 230.3 Median: 200.0 Figure 14: Distribution chart of token lengths in the test set after training the data synthesized by our method on Qwen3-4B-Base. Prompt Template for Associative Generation You are a medical expert creating high-quality multiple-choice questions for medical education. Few-shot examples: Example 1 (Concepts: terms_str): q_text Example 2 (Concepts: terms_str): q_text Task: Generate a NEW medical multiple-choice question that meaningfully incorporates the concepts from these MeSH terms: target_terms Requirements: 1. Create a medically relevant, challenging question 2. Include 3-6 plausible options labeled A, B, C, D, ... 3. Only ONE option should be correct 4. The question should require medical reasoning, not just recall 5. Do NOT copy from the examples Output format STRICTLY: [Question Begin] your generated question with options [Question End] [Reasoning Begin] step-by-step medical reasoning [Reasoning End] [Answer Begin] correct option letter[Answer End] Figure 15: The prompt template used for associative generation. Placeholders denote: terms_str (concepts associated with the few-shot examples), q_text (the content of the few-shot example questions), and tar- get_terms (the set of MeSH terms to be incorporated into the new question). Prompt Template for Atomic Generation You are a medical expert. Target Concept: term Definition: scope_note Relevant Aspects: qual_str Task: Create a detailed multiple-choice question specifically focused on the term. Output Format: [Question Begin]...[Question End] [Reasoning Begin]...[Reasoning End] [Answer Begin] A[Answer End] Figure 16: The prompt template used for atomic generation. Placeholders denote: term (the target medical concept), scope_note (the official defini- tion/explanation of the term), and qual_str (relevant sub-aspects or qualifiers associated with the term).

Instant research discovery