SEMA-RAG: A Self-Evolving Multi-Agent Retrieval-Augmented Generation Framework for Medical Reasoning

# SEMA-RAG: A Self-Evolving Multi-Agent Retrieval-Augmented Generation Framework for Medical Reasoning
Source: [https://arxiv.org/html/2605.17101](https://arxiv.org/html/2605.17101)
Yongfeng Huang1Ruiying Chen211footnotemark:1James Cheng1 1CSE, The Chinese University of Hong Kong2Wuhan University of Technology \{yfhuang22,jcheng\}@cse\.cuhk\.edu\.hk355227@whut\.edu\.cn

###### Abstract

Retrieval\-Augmented Generation \(RAG\) is widely employed to mitigate risks such as hallucinations and knowledge obsolescence in medical question answering, yet its predominantly single\-round, static retrieval paradigm misaligns with the multi\-stage process of clinical reasoning\. This compressed workflow induces two structural deficiencies: question\-to\-query translation often lacks clinically grounded semantic interpretation, and retrieval lacks iterative sufficiency feedback, making it difficult to form reliable evidence chains\. We argue that both issues stem from a deeper cause—overloading a single reasoning chain with heterogeneous tasks of interpretation, exploration, and adjudication—and that the remedy is to reconstruct the workflow via task decoupling and dynamic multi\-round exploration\. To this end, we proposeSEMA\-RAG, aSelf\-EvolvingMulti\-Agent RAG framework for medical question answering, which assigns these roles to three specialist agents: theInterpreter Agentfor clinical schema interpretation, theExplorer Agentfor sufficiency\-driven self\-evolving retrieval, and theArbiter Agentfor evidence adjudication and answer selection\. Across five benchmarks and five LLM backbones, SEMA\-RAG improves the strongest baseline by\+6\.46accuracy points on average, measured per backbone\.

SEMA\-RAG: A Self\-Evolving Multi\-Agent Retrieval\-Augmented Generation Framework for Medical Reasoning

Yongfeng Huang1††thanks:These authors contributed equally to this work\.Ruiying Chen211footnotemark:1James Cheng1††thanks:Corresponding author\.1CSE, The Chinese University of Hong Kong2Wuhan University of Technology\{yfhuang22,jcheng\}@cse\.cuhk\.edu\.hk355227@whut\.edu\.cn

## 1Introduction

In recent years, large language models \(LLMs\) have shown specific capabilities in understanding and reasoning about medical knowledge when applied in healthcareKunget al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib2)\); Omaret al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib1)\)\. However, they remain prone to hallucinations and outdated information in high\-stakes clinical settingsOmiyeet al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib3)\); Roustan and Bastardot \([2025](https://arxiv.org/html/2605.17101#bib.bib4)\)\. Retrieval\-Augmented Generation \(RAG\), which incorporates external authoritative evidence to support the generation process, has been widely adopted to mitigate these risksLewiset al\.\([2020](https://arxiv.org/html/2605.17101#bib.bib5)\)\.

\(%\)84\.384\.377\.077\.064\.564\.550\.850\.878\.378\.386\.586\.583\.383\.370\.370\.356\.356\.383\.183\.1MMLUMedQA\-USMedMCQAPubMedQA\*BioASQ50506060707080809090Average Best BaselineSEMA\-RAGFigure 1:Benchmark\-level accuracy averaged over five LLM backbones\.However, standard RAG frameworks typically treat retrieval as a static, single\-round auxiliary step, misaligning with the multi\-stage process of clinical reasoning: clinicians often first interpret patient narratives as searchable clinical questions, then progressively gather and verify information to address evidence gaps, and weigh and integrate redundant or contradictory evidence to ultimately form judgments based on relatively robust evidenceLinnet al\.\([2012](https://arxiv.org/html/2605.17101#bib.bib6)\); Yazdaniet al\.\([2017](https://arxiv.org/html/2605.17101#bib.bib7)\)\. In contrast, single\-round static RAG compresses this process into a single retrieval and generation step\. This is akin to requiring clinicians to simultaneously analyze, retrieve, evaluate, and diagnose immediately upon receiving initial medical records, without adjusting their reasoning as new evidence emerges\. This typically leads to two structural flaws:*\(i\)*the translation from question to query lacks clinical semantic interpretation, making implicit constraints difficult to articulate explicitlySoldainiet al\.\([2017](https://arxiv.org/html/2605.17101#bib.bib8)\); and*\(ii\)*the retrieval process lacks mechanisms for sufficiency assessment and feedback, hindering self\-evolving iterative convergence under insufficient evidence and thus weakening reliable evidence\-chain formationMallenet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib10)\); Shiet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib9)\)\.

These shortcomings, we argue, are not independent issues but rather symptoms of a deeper problem:overloading heterogeneous tasks into a single reasoning chain\. When question interpretation, evidence exploration, and answer adjudication are tightly coupled, the cognitive load increases and the steps become interdependent, making it hard for the model to promptly adjust retrieval and reasoning when evidence is insufficient or conflictingWanget al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib11)\); Liuet al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib12)\)\. Thus, the key is not to intensify single\-round reasoning, but to restructure RAG to better match the phased clinical workflow by extending single\-round queries into multi\-round iterative exploration\. After each retrieval round, the system evaluates whether the evidence covers key constraints and then chooses the next action: terminate exploration and proceed to decision integration if sufficient, or generate targeted follow\-up queries to fill gaps if insufficient\. This mechanism, which continuously updates the direction of queries and retrieval based on the evaluation results of each round, enables the system to adjust and converge as evidence accumulates progressively\. In this sense, the process constitutes a form ofintra\-test\-time self\-evolution, in which the system adaptively updates its query and retrieval trajectory during task execution while remaining tightly coupled to the current problem instanceGaoet al\.\([2026](https://arxiv.org/html/2605.17101#bib.bib69)\)\. For simplicity, we use the term "self\-evolving" in the remainder of the paper to refer to this intra\-test\-time setting\.

To this end, we proposeSEMA\-RAG\(Self\-EvolvingMulti\-Agent RAG\)\. This framework simulates clinical workflows through task decoupling and role specialization, decomposing complex clinical reasoning into three collaborative modules:Interpreter Agent \(I\-Agent\)maps unstructured inputs to structured clinical semantics;Explorer Agent \(E\-Agent\)implements self\-evolving, evidence\-sufficiency\-driven retrieval for convergent exploration;Arbiter Agent \(A\-Agent\)performs comprehensive adjudication based on closed\-loop evidence\.

We evaluate SEMA\-RAG on five medical question\-answering benchmarks\. As shown in Figure[1](https://arxiv.org/html/2605.17101#S1.F1), SEMA\-RAG consistently outperforms representative baselines in terms of average accuracy on each benchmark when averaged over multiple underlying LLMs\. Across five benchmarks and five LLM backbones, it improves the strongest baseline by an average of\+6\.46accuracy points, validating convergent evidence\-chain construction via task decoupling, role specialization, and evidence sufficiency\-driven self\-evolving retrieval\. Our main contributions are as follows:

- •We proposeSEMA\-RAG, a multi\-agent RAG framework for medical question answering, which models clinical reasoning processes via role division and collaboration\.
- •We develop a self\-evolvingExplorer Agentthat updates queries based on evidence gaps, steering retrieval toward medical reasoning objectives\.
- •We validate SEMA\-RAG on five medical Q&A benchmarks across multiple underlying LLMs, achieving consistent improvements over baselines\.

![Refer to caption](https://arxiv.org/html/2605.17101v1/figure/method.png)Figure 2:Overview ofSEMA\-RAG: \(i\)I\-Agentstructures the input questionQQinto a clinical schema tupleQ′Q^\{\\prime\}for retrieval; \(ii\)E\-Agentconducts sufficiency\-driven self\-evolving multi\-round retrieval to obtain a converged evidence setC∗C^\{\*\}; \(iii\)A\-Agentadjudicates evidence into a traceable reportRRand selects the final answer grounded inRR\.
## 2Preliminaries

### 2\.1Task Formulation of Medical RAG

Given a medical questionQQ, the system selects the final answery~\\tilde\{y\}from the discrete candidate set𝒴\\mathcal\{Y\}\(y∈𝒴y\\in\\mathcal\{Y\}\)\. Under question\-only retrieval conditions, the system may only retrieve evidence from the medical corpus𝒞\\mathcal\{C\}\. The core RAG consists of a retrieval operatorRet​\(⋅\)\\mathrm\{Ret\}\(\\cdot\)and a generation operator:C=Ret​\(Q\)C=\\mathrm\{Ret\}\(Q\), and predicts accordingly:

y~=arg⁡maxy∈𝒴⁡p​\(y∣Q,C\)\.\\tilde\{y\}=\\arg\\max\_\{y\\in\\mathcal\{Y\}\}p\(y\\mid Q,C\)\.

### 2\.2Multi\-Agent Roles and Abstraction

We employ role\-based division of labor, with three agents collaborating to complete the medical question answering process:I\-Agenthandles question interpretation,E\-Agentmanages evidence exploration, andA\-Agentoversees answer adjudication\.

Three agents share the same underlying language model, differentiated solely by role\-specific prompts\. We denote the output of the shared LLM conditioned on a role promptP​m​trPmt\_\{r\}and an inputXXasAgentr​\(P​m​tr,X\)\\mathrm\{Agent\}\_\{r\}\(Pmt\_\{r\},X\), whereXXmay be a set containing multiple elements\.

## 3Method

Figure[2](https://arxiv.org/html/2605.17101#S1.F2)illustrates the overall SEMA\-RAG framework, which comprises three role\-based agents with responsibilities described below\.

### 3\.1I\-Agent as a Question Interpreter

I\-Agent does not merely rephrase the input medical questionQQ; instead, it semantically structuresQQand projects it onto an explicitClinical Schema\. This process externalizes latent clinical intent and key constraints, providing stable anchors for subsequent retrieval and reasoning\.

Specifically, I\-Agent produces a clinical schema tupleQ′Q^\{\\prime\}with four components:*\(i\)*clinical intentointo\_\{\\mathrm\{int\}\}, describing the implied task type \(e\.g\., diagnosis, treatment, dosage\);*\(ii\)*medical entitiesoento\_\{\\mathrm\{ent\}\}, identifying core medical objects \(e\.g\., diseases, drugs\);*\(iii\)*clinical constraintsoconso\_\{\\mathrm\{cons\}\}, specifying applicability conditions \(e\.g\., pregnant, renal impairment, adult\); and*\(iv\)*an initial retrieval queryqinitq\_\{\\mathrm\{init\}\}, a concise, search\-oriented question distilled from the above schema\.

Formally, I\-Agent mapsQQto the schema tuple:

Q′\\displaystyle Q^\{\\prime\}=⟨oint,oent,ocons,qinit⟩\\displaystyle=\\langle o\_\{\\mathrm\{int\}\},\\,o\_\{\\mathrm\{ent\}\},\\,o\_\{\\mathrm\{cons\}\},\\,q\_\{\\mathrm\{init\}\}\\rangle=AgentI​\(P​m​tI,Q\)\.\\displaystyle=\\mathrm\{Agent\}\_\{\\mathrm\{I\}\}\(Pmt\_\{\\mathrm\{I\}\},\\,Q\)\.
To make the schema tuple usable by the dense retriever, we further linearizeQ′Q^\{\\prime\}into a retrieval\-ready query string:

q^init=Linearize​\(Q′\)\\hat\{q\}\_\{\\mathrm\{init\}\}=\\mathrm\{Linearize\}\(Q^\{\\prime\}\)=Concat​\(qinit,⊕,oint,⊕,oent,⊕,ocons\),=\\mathrm\{Concat\}\(q\_\{\\mathrm\{init\}\},\\oplus,o\_\{\\mathrm\{int\}\},\\oplus,o\_\{\\mathrm\{ent\}\},\\oplus,o\_\{\\mathrm\{cons\}\}\),where⊕\\oplusdenotes a semicolon separator\. Here,Linearize​\(⋅\)\\mathrm\{Linearize\}\(\\cdot\)is a parameter\-free function without field\-specific weights or additional control tokens\. It preservesqinitq\_\{\\mathrm\{init\}\}as the core query while explicitly incorporatingoint,oento\_\{\\mathrm\{int\}\},o\_\{\\mathrm\{ent\}\}, andoconso\_\{\\mathrm\{cons\}\}, making clinically important but implicit constraints more visible to the retriever and reducing semantic drift in the initial retrieval stage\. The resulting queryq^init\\hat\{q\}\_\{\\mathrm\{init\}\}is used to initialize E\-Agent, whileQ′Q^\{\\prime\}remains the clinical anchor for subsequent coordination\.

### 3\.2E\-Agent as a Knowledge Explorer

E\-Agent begins with the linearized schema queryq^init\\hat\{q\}\_\{\\mathrm\{init\}\}generated by I\-Agent and progressively completes the evidence through a self\-evolving iterative retrieval process, ultimately constructing the final evidence setC∗C^\{\*\}\.

##### Retrieval Space Initialization

We construct a dense vector retrieval space based on the medical corpus𝒞\\mathcal\{C\}\. Using a parameter\-frozen medical dual encoder, we map queries and documents to the same vector space, whereEqry​\(⋅\)E\_\{\\mathrm\{qry\}\}\(\\cdot\)andEdoc​\(⋅\)E\_\{\\mathrm\{doc\}\}\(\\cdot\)denote the query and document encoders, respectively\. Given a queryqq, its Top\-kkcandidate documents \(passages/chunks\) are retrieved based on vector similarity as follows:

TopK​\(q\)=Top​\-​kD∈𝒞⁡⟨Eqry​\(q\),Edoc​\(D\)⟩\.\\mathrm\{TopK\}\(q\)=\\operatorname\*\{Top\\text\{\-\}k\}\_\{D\\in\\mathcal\{C\}\}\\left\\langle E\_\{\\mathrm\{qry\}\}\(q\),\\,E\_\{\\mathrm\{doc\}\}\(D\)\\right\\rangle\.whereTop​\-​k\\operatorname\*\{Top\\text\{\-\}k\}returns thekkdocuments with the largest similarity scores\.

##### Self\-Evolving Evidence Retrieval Loop

Using the linearized schema queryq^init\\hat\{q\}\_\{\\mathrm\{init\}\}as the initial query, we set𝒬1=\{q^init\}\\mathcal\{Q\}\_\{1\}=\\\{\\hat\{q\}\_\{\\mathrm\{init\}\}\\\}andC0=∅C\_\{0\}=\\varnothing, where𝒬t\\mathcal\{Q\}\_\{t\}is the query set for thettth retrieval round andCtC\_\{t\}is the accumulated evidence set after roundtt\. Each retrieved documentDiD\_\{i\}is associated with a deterministic document identifierID​\(Di\)\\mathrm\{ID\}\(D\_\{i\}\), which is retained throughout the pipeline for exact deduplication and source tracing\. At roundtt, E\-Agent performs retrieval for each query in𝒬t\\mathcal\{Q\}\_\{t\}and updates the evidence set:

𝒟t=⋃q∈𝒬tTopK​\(q\),\\mathcal\{D\}\_\{t\}=\\bigcup\_\{q\\in\\mathcal\{Q\}\_\{t\}\}\\mathrm\{TopK\}\(q\),Ct=Ct−1∪\{Di∈𝒟t:ID​\(Di\)∉IDs​\(Ct−1\)\}\.C\_\{t\}=C\_\{t\-1\}\\cup\\\{D\_\{i\}\\in\\mathcal\{D\}\_\{t\}:\\mathrm\{ID\}\(D\_\{i\}\)\\notin\\mathrm\{IDs\}\(C\_\{t\-1\}\)\\\}\.
Conditioned on clinical anchorsQ′Q^\{\\prime\}, the current textual query set𝒬t\\mathcal\{Q\}\_\{t\}, and the evidence setCtC\_\{t\}, E\-Agent predicts a sufficiency flagsts\_\{t\}, a gap descriptiongtg\_\{t\}, and the next query set𝒬t\+1\\mathcal\{Q\}\_\{t\+1\}:

=AgentE​\(P​m​tE,\[Q′,𝒬t,Ct\]\),\\displaystyle=\\mathrm\{Agent\}\_\{\\mathrm\{E\}\}\\\!\\left\(Pmt\_\{\\mathrm\{E\}\},\\,\[Q^\{\\prime\},\\,\\mathcal\{Q\}\_\{t\},\\,C\_\{t\}\]\\right\),𝒬t\+1\\displaystyle\\mathcal\{Q\}\_\{t\+1\}=\{qt\+1⟨1⟩,…,qt\+1⟨m⟩\}\.\\displaystyle=\\left\\\{q\_\{t\+1\}^\{\\langle 1\\rangle\},\\ldots,q\_\{t\+1\}^\{\\langle m\\rangle\}\\right\\\}\.wherest∈\{0,1\}s\_\{t\}\\in\\\{0,1\\\}indicates evidence sufficiency: ifst=1s\_\{t\}=1, the evidence is sufficient and we set𝒬t\+1=∅\\mathcal\{Q\}\_\{t\+1\}=\\varnothing; otherwise \(st=0s\_\{t\}=0\), evidence gaps remain, andgtg\_\{t\}identifies missing conditions or reasoning steps, from which𝒬t\+1\\mathcal\{Q\}\_\{t\+1\}generatesmmcandidate follow\-up queries targeting these gaps\.

Whenst=0s\_\{t\}=0, the generated𝒬t\+1\\mathcal\{Q\}\_\{t\+1\}is issued in the next round to retrieve additional evidence, and the results are incorporated into the update ofCt\+1C\_\{t\+1\}\.

Iteration terminates whenst=1s\_\{t\}=1,t=Tmaxt=T\_\{\\max\}, or stagnation occurs \(i\.e\.,𝒬t\+1=∅\\mathcal\{Q\}\_\{t\+1\}=\\varnothing\)\. Upon termination, we obtain the closed evidence set, record the actual number of iterationsT≤TmaxT\\leq T\_\{\\max\}, and store the self\-evolving trajectory:

C∗=CT,τ=\{\[𝒬1,C1\],…,\[𝒬T,CT\]\}\.C^\{\*\}=C\_\{T\},\\;\\tau=\\left\\\{\[\\mathcal\{Q\}\_\{1\},C\_\{1\}\],\\ldots,\[\\mathcal\{Q\}\_\{T\},C\_\{T\}\]\\right\\\}\.

### 3\.3A\-Agent as an Evidence Arbiter

A\-Agent adjudicates evidence by organizing the converged setC∗C^\{\*\}into a traceable evidence report and generating a discrete answer from it\.

##### Evidence Adjudication and Report Construction

Given redundant and potentially conflicting evidence, A\-Agent first adjudicatesC∗C^\{\*\}by removing irrelevant or duplicated content, identifying consistencies and conflicts, and organizing supporting and refuting clues into a structured evidence reportRR\. For traceability, we retain the original document identifier for each retrieved documentDi∈C∗D\_\{i\}\\in C^\{\*\}, forming the source set

𝒮∗=\{\(ID​\(Di\),Di\)∣Di∈C∗\}\.\\mathcal\{S\}^\{\*\}=\\\{\(\\mathrm\{ID\}\(D\_\{i\}\),D\_\{i\}\)\\mid D\_\{i\}\\in C^\{\*\}\\\}\.A\-Agent then generates the evidence report

R=AgentA​\(P​m​tadj,\[Q,C∗,𝒮∗\]\),R=\\mathrm\{Agent\}\_\{\\mathrm\{A\}\}\(Pmt\_\{\\mathrm\{adj\}\},\[Q,C^\{\*\},\\mathcal\{S\}^\{\*\}\]\),whereRRexplicitly organizes key conclusions relevant to the question along with their source indices, provides a reconciled synthesis of conflicting evidence, and offers a stable basis for final answer selection\.

##### Evidence\-Grounded Answering

Upon obtaining the evidence reportRR, A\-Agent performs discrete answer selection over the candidate answer set𝒴\\mathcal\{Y\}:

y~=AgentA​\(P​m​tans,\[Q,R\]\),\\tilde\{y\}=\\mathrm\{Agent\}\_\{\\mathrm\{A\}\}\(Pmt\_\{\\mathrm\{ans\}\},\[Q,R\]\),wherey~\\tilde\{y\}is the final predicted answer\.

## 4Experiments

### 4\.1Experimental Setup

ModelMMLU\-MedMedQA\-USMedMCQAPubMedQA\*BioASQ\-Y/NAveragedeepseek\-v3\.1DeepSeek\-AIet al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib33)\)\+ CoTWeiet al\.\([2022](https://arxiv.org/html/2605.17101#bib.bib32)\)88\.1577\.5371\.6938\.4080\.1071\.17\+ MedCPTJinet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib14)\)85\.1273\.8462\.6643\.2076\.3868\.24\+ MedRAGXionget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib13)\)88\.6177\.1467\.9944\.6078\.4871\.36\+i\-MedRAGXionget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib27)\)85\.8674\.7865\.6550\.6080\.5871\.49\+ SEMA\-RAG \(Ours\)91\.4689\.9575\.0959\.2082\.8579\.71kimi\-k2Teamet al\.\([2025b](https://arxiv.org/html/2605.17101#bib.bib34)\)\+ CoTWeiet al\.\([2022](https://arxiv.org/html/2605.17101#bib.bib32)\)84\.3977\.8572\.0853\.6085\.7674\.74\+ MedCPTJinet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib14)\)89\.8180\.6873\.8550\.2081\.3975\.19\+ MedRAGXionget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib13)\)91\.3781\.5473\.2052\.6085\.6076\.86\+i\-MedRAGXionget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib27)\)91\.2881\.7874\.1354\.6083\.1776\.99\+ SEMA\-RAG \(Ours\)91\.4686\.4176\.0755\.8088\.6779\.68qwen3\-coder\-plusYanget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib35)\)\+ CoTWeiet al\.\([2022](https://arxiv.org/html/2605.17101#bib.bib32)\)89\.2676\.9073\.0647\.2081\.7273\.63\+ MedCPTJinet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib14)\)87\.4275\.2667\.4446\.6075\.8970\.52\+ MedRAGXionget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib13)\)89\.4481\.5469\.2649\.2072\.3372\.35\+i\-MedRAGXionget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib27)\)89\.2677\.3870\.2648\.6082\.5273\.60\+ SEMA\-RAG \(Ours\)92\.1086\.1774\.2356\.0083\.0178\.30gemini\-2\.0\-flashGoogle \([2025](https://arxiv.org/html/2605.17101#bib.bib37)\)\+ CoTWeiet al\.\([2022](https://arxiv.org/html/2605.17101#bib.bib32)\)58\.2265\.1241\.3340\.2068\.4554\.66\+ MedCPTJinet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib14)\)62\.3570\.5444\.9042\.8070\.0658\.13\+ MedRAGXionget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib13)\)74\.2983\.1950\.8744\.2072\.6565\.04\+i\-MedRAGXionget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib27)\)65\.4777\.6946\.7851\.2077\.9963\.83\+ SEMA\-RAG \(Ours\)80\.9990\.4271\.6059\.2088\.1978\.08glm\-4\.0\-flashGLMet al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib36)\)\+ CoTWeiet al\.\([2022](https://arxiv.org/html/2605.17101#bib.bib32)\)68\.1450\.4348\.2240\.0058\.9053\.14\+ MedCPTJinet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib14)\)73\.0058\.2150\.6342\.2060\.8456\.98\+ MedRAGXionget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib13)\)77\.5960\.8852\.5946\.8062\.7860\.13\+i\-MedRAGXionget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib27)\)73\.2853\.5751\.4748\.4064\.7258\.29\+ SEMA\-RAG \(Ours\)76\.6863\.7954\.3451\.2072\.9863\.80

Table 1:Accuracy \(%\) comparison ofSEMA\-RAGand baselines on five medical QA benchmarks across different LLMs\. Bold indicates the best result within each model block and underline indicates the second\-best\.#### 4\.1\.1Evaluation Benchmarks

To systematically evaluate SEMA\-RAG’s performance and generalisation capabilities across diverse medical question\-answering scenarios, we select five widely used datasets from the MIRAGE benchmarkXionget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib13)\): three medical examination datasets \(MMLU\-MedHendryckset al\.\([2021](https://arxiv.org/html/2605.17101#bib.bib18)\), MedQA\-USJinet al\.\([2020](https://arxiv.org/html/2605.17101#bib.bib19)\), MedMCQAPalet al\.\([2022](https://arxiv.org/html/2605.17101#bib.bib20)\)\) and two biomedical research QA datasets \(PubMedQA\*Jinet al\.\([2019](https://arxiv.org/html/2605.17101#bib.bib21)\)and BioASQ\-Y/NTsatsaroniset al\.\([2015](https://arxiv.org/html/2605.17101#bib.bib22)\); Kritharaet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib23)\)\)\. Together, they cover general medical knowledge, clinical examinations, and biomedical literature inference\. Following MIRAGE’s filtering and preprocessing pipeline, we retain only discrete biomedical classification questions, use PubMedQA\* \(with the original evidence context removed\), and apply question\-only retrieval for all tasks\.

#### 4\.1\.2Models and Baselines

To assess SEMA\-RAG’s robustness across backbones, we instantiate the framework on five publicly accessible LLMs:deepseek\-v3\.1DeepSeek\-AIet al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib33)\),kimi\-k2Teamet al\.\([2025b](https://arxiv.org/html/2605.17101#bib.bib34)\),qwen3\-coder\-plusYanget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib35)\),gemini\-2\.0\-flashGoogle \([2025](https://arxiv.org/html/2605.17101#bib.bib37)\), andglm\-4\.0\-flashGLMet al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib36)\)\. These models originate from different providers, encompass diverse pretraining configurations and capability focuses\.

We selected three representative methods for comparison to characterize performance differences across no retrieval, single\-round retrieval, and iterative retrieval\. The no\-retrieval setting employsCoTWeiet al\.\([2022](https://arxiv.org/html/2605.17101#bib.bib32)\), relying solely on model\-internal knowledge for chain\-of\-reasoning\. The single\-round retrieval setting employsMedCPTJinet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib14)\)as the medical domain retriever, further contrasting it withMedRAGXionget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib13)\)’s retrieval\-fusion framework\. The iterative retrieval setting utilizesi\-MedRAGXionget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib27)\), which generates subsequent queries to drive multi\-round retrieval and accumulate evidence\.

#### 4\.1\.3Implementation Details

Followingi\-MedRAGXionget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib27)\)and MedRAGXionget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib13)\), we retrieve from TextbooksJinet al\.\([2021](https://arxiv.org/html/2605.17101#bib.bib38)\)and StatPearlsStatPearls Publishing \([2025](https://arxiv.org/html/2605.17101#bib.bib39)\)on all benchmarks\. We use MedCPTJinet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib14)\)as the dense retriever and perform FAISS\-based retrieval over these corpora\.

All methods are evaluated in a zero\-shot setting\. Unless stated otherwise, SEMA\-RAG usesTmax=2T\_\{\\max\}=2,k=16k=16, andm=3m=3\. We set temperature to 1\.0 for I/E\-Agent and 0\.0 for A\-Agent\. Other baselines follow their official settings\.

### 4\.2Main Results: Consistent and Significant Improvements

Table[1](https://arxiv.org/html/2605.17101#S4.T1)reports results on five medical QA benchmarks across five underlying LLMs\. Overall,SEMA\-RAG achieves the best average accuracy within every model block, indicating that the improvement is backbone\-agnostic rather than tied to a particular LLM\.

A consistent pattern is that the gains are larger on benchmarks where premature commitment under incomplete evidence is costly\. This matches SEMA\-RAG’s self\-evolving exploration: it checks evidence sufficiency, performs targeted follow\-up retrieval when needed, and only then moves to adjudication, leading to more reliable evidence chains\.

Overall, this consistent and substantial advantage is not accidental\. It directly stems from SEMA\-RAG’s successful simulation of expert clinical reasoning by decoupling interpretation, exploration, and adjudication, thereby alleviating the cognitive overload bottleneck diagnosed in the Introduction\. To unpack the source of these gains, we next present a core component analysis\.

### 4\.3Core Component Analysis

In this section, we quantitatively evaluate the contributions of the three roles in SEMA\-RAG via role\-wise removal analysis\. Specifically, we compare the full framework with three variants: \(i\)w/o I\-Agent, which removes the question interpretation module and directly uses the raw question as the initial retrieval query; \(ii\)w/o E\-Agent, which removes the self\-evolving retrieval mechanism and performs only a single round of static retrieval based on the linearized schema query generated by I\-Agent; and \(iii\)w/o A\-Agent, which removes the answer adjudication module and directly generates answers from the final evidence set\. The results are summarized in Table[2](https://arxiv.org/html/2605.17101#S4.T2)\.

I\-AgentE\-AgentA\-AgentMedQA\-USPubMedQA\*✗✓✓85\.4754\.20✓✗✓83\.5850\.80✓✓✗86\.4953\.60✓✓✓89\.9559\.20Table 2:Role\-wise removal results of SEMA\-RAG on MedQA\-US and PubMedQA\* \(deepseek\-v3\.1\)\.#### 4\.3\.1Resolving Ambiguous Queries with I\-Agent

Table[2](https://arxiv.org/html/2605.17101#S4.T2)shows that removing the I\-Agent consistently degrades performance, indicating that question\-to\-query translation benefits from clinically grounded semantic interpretation\. Without this step, queries often remain underspecified and fail to surface implicit constraints, which increases the chance that retrieval drifts toward generic, symptom\-level evidence rather than the decision\-critical clinical setting\.

As shown in Table[4\.5](https://arxiv.org/html/2605.17101#S4.SS5)\(Step 1\), I\-Agent makes*hospital day 7*retrieval\-actionable, anchoring evidence in the late\-onset inpatient context and enabling option\-level discrimination forS\. aureus\. By extracting a structured clinical schema and making such key conditions retrievable, I\-Agent keeps retrieval aligned with the intended clinical scenario and provides a cleaner substrate for subsequent exploration and adjudication\.

#### 4\.3\.2Achieving Dynamic Reasoning with E\-Agent

Table[2](https://arxiv.org/html/2605.17101#S4.T2)shows that removing E\-Agent causes the largest performance drop, highlighting that the core gain comes from a self\-evolving, sufficiency\-driven closed\-loop retrieval rather than static retrieval\. Without E\-Agent, the system loses explicit sufficiency feedback and is more prone to stop with partially covered evidence, leaving key constraints unresolved\.

To configure this loop, we vary the maximum exploration depthTmaxT\_\{\\max\}withm=3m=3\. Figure[3](https://arxiv.org/html/2605.17101#S4.F3)suggests that most of the benefit is captured within two rounds, with performance peaking aroundTmax∈\{2,3\}T\_\{\\max\}\\in\\\{2,3\\\}; beyond that, deeper exploration saturates and can introduce noise\. Notably, the two\-round setting already outperforms our reproducedi\-MedRAG baseline \(which uses three fixed retrieval rounds\), suggesting that the gain comes from self\-evolving, sufficiency\-driven closed\-loop exploration rather than simply increasing the number of iterations\.

11223355779986868787888889899090919186\.5786\.5789\.9589\.9590\.1090\.1089\.8789\.8788\.4588\.4587\.6787\.67Max Iterations \(TmaxT\_\{\\max\}\)Accuracy \(%\)SEMA\-RAG \(m=3m=3\)Figure 3:Impact of max iterationsTmaxT\_\{\\max\}\(fixm=3m=3\) on MedQA\-US \(deepseek\-v3\.1\)\.

### 4\.4Further Analysis

#### 4\.4\.1Synergy of the Multi\-Agent Architecture

Table[2](https://arxiv.org/html/2605.17101#S4.T2)shows that the full SEMA\-RAG consistently performs best across both MedQA\-US and PubMedQA\*\. Ablating any single agent leads to a clear drop, suggesting that the gains do not come from one isolated module but from role\-specialized collaboration that mirrors the staged clinical workflow: I\-Agent anchors clinically grounded interpretation, E\-Agent drives sufficiency\-driven evidence completion, and A\-Agent consolidates and adjudicates evidence for option selection\.

#### 4\.4\.2Impact of Query Breadth

Building on the depth analysis in Figure[3](https://arxiv.org/html/2605.17101#S4.F3), we examine how the per\-round query breadthmmaffects E\-Agent’s exploration\. WithTmax=2T\_\{\\max\}=2fixed, Table[3](https://arxiv.org/html/2605.17101#S4.T3)shows a clear monotonic trend: increasingmmimproves accuracy, but the gains quickly taper\. This suggests that expanding the query set helps cover complementary evidence gaps in early exploration, while additional branches beyond a moderate breadth tend to introduce overlapping or low\-yield retrievals\. We therefore useTmax=2T\_\{\\max\}=2andm=3m=3as the default setting in subsequent experiments, balancing coverage and efficiency\.

VariantMedQA\-US \(%\)SEMA\-m1 \(m=1m=1\)86\.72SEMA\-m2 \(m=2m=2\)89\.00SEMA\-m3 \(m=3m=3\)89\.95Table 3:Effect of query breadthmm\(fixTmax=2T\_\{\\max\}=2\) on MedQA\-US \(deepseek\-v3\.1\)\.

### 4\.5Qualitative Case Study: A Head\-to\-Head Comparison

Table[4\.5](https://arxiv.org/html/2605.17101#S4.SS5)shows a representative MedQA\-US case comparing MedRAG and SEMA\-RAG, where the decisive cue is the temporal constraint \(hospital day 7\), pointing to a hospital\-acquired rather than community\-acquired etiology\.

MedQA\-US Question 0024Context: A 62\-year\-old patient has been hospitalized for a week due to a stroke\. On hospital day 7, he develops a fever \(38\.4°C\) and purulent cough\. Vitals: HR 88, RR 20, BP 110/85\. Physical exam: right basal crackles\. Chest X\-ray: new right\-sided consolidation\.Labs \(Selected\): WBC 8,900/mm³ \(Neutrophils 72%, Bands 4%\), Hb 16 g/dL, Platelets 280,000/mm³\.Question: What is the most likely causal microorganism?A\.Streptococcus pneumoniaeB\.Mycobacterium tuberculosisC\.Haemophilus influenzaeD\.Staphylococcus aureus\\rowcolor\[RGB\]242,242,242Baseline: MedRAG\\rowcolor\[RGB\]250,250,250Retrieved Evidence:"Community\-acquired pneumonia \(CAP\) is most frequently caused byStreptococcus pneumoniae…" ; "Stroke patients are at risk for aspiration pneumonia…"\\rowcolor\[RGB\]250,250,250Analysis & Answer:The patient presents with typical signs of pneumonia\.Given thatS\. pneumoniaeis the most common cause, it is the likely pathogen…Prediction: A✗\\rowcolor\[RGB\]225,235,245Ours: SEMA\-RAG\\rowcolor\[RGB\]240,248,2551\. I\-Agent \(Interpretation\)\\rowcolor\[RGB\]240,248,255Structured Schema:Clinical Intent: Infectious etiology & pathogen identificationMedical Entities: stroke; HAP/aspiration pneumonia; right\-basal crackles; new right consolidation; fever \+ purulent coughClinical Constraints: 62y;hospital day 7; post\-stroke aspiration risk; neutrophil predominanceInitial Query: "hospital day 7 post\-stroke pneumonia right consolidation most likely causative organism"\\rowcolor\[RGB\]240,248,2552\. E\-Agent \(Exploration\)\\rowcolor\[RGB\]240,248,255Iteration 1:Evidence: "…Post\-stroke patients have high risk of aspiration… pneumonia causes include anaerobes and streptococci…"Gap:Evidence does not distinguish pathogens based on hospitalization duration \(day 7\)\.Sufficiency:s1=0s\_\{1\}=0\(Insufficient\)→\\rightarrowNext Query: "most likely pathogen hospital\-acquired pneumonia vs community\-acquired"\\rowcolor\[RGB\]240,248,255Iteration 2:Evidence: "…Hospital\-acquired pneumonia \(HAP\) is defined as pneumonia≥\\geq48h after admission…"Key Find: "For late\-onset HAP \(≥\\geq5 days\),common pathogens includeStaphylococcus aureus\(MRSA\) andPseudomonas…"Sufficiency:s2=1s\_\{2\}=1\(Sufficient\)\\rowcolor\[RGB\]240,248,2553\. A\-Agent \(Adjudication\)\\rowcolor\[RGB\]240,248,255Report: … Hospital day 7 indicates HAP rather than CAP under standard definitions; combined with the candidate set, the evidence most consistently supportsS\. aureus\.Among the provided options,S\. aureusis the only matching HAP pathogen\.\\rowcolor\[RGB\]240,248,255Prediction: D✓

Table 4:A case of how SEMA\-RAG helps deepseek\-v3\.1 find the correct answer on MedQA\-US \(Question 0024\) by making thekey clinical constraintexplicit and retrievingdecision\-critical evidence, while MedRAG’s single\-round retrieval leads to amisleading rationale\.In theBaselineblock, MedRAG relies on a single round of static retrieval\. Its evidence remains centered on generic pneumonia cues and aspiration risk, without explicitly anchoring retrieval to the hospital day 7 condition\. This shifts the evidence toward typical community\-acquired pathogens and leads to an incorrect selection of Streptococcus pneumoniae \(Option A\), which mismatches the inpatient, late\-onset setting in the question\.

In contrast,SEMA\-RAGmakes the temporal constraint retrieval\-actionable and carries it through to option selection\. TheI\-Agentsurfaces hospital day 7 as a key constraint, theE\-Agentdetects the missing distinction between community\- and hospital\-acquired spectra and performs targeted follow\-up retrieval, and theA\-Agentconsolidates the resulting evidence and maps it to the candidate set, yielding the correct choice Staphylococcus aureus \(Option D\)\. Overall, the case shows how task decoupling with sufficiency\-driven self\-evolving retrieval helps form a more reliable evidence chain and prevents premature decisions under insufficient evidence\.

### 4\.6Cost and Efficiency Analysis

Table[5](https://arxiv.org/html/2605.17101#S4.T5)compares methods in terms of both accuracy and inference cost\. Calls denotes the number of LLM invocations per question, while Retr\. counts the number of vector retrieval operations; Time reports the average end\-to\-end latency per question; Tok\./Q denotes the average total token consumption per question\. Because SEMA\-RAG employs sufficiency\-driven early stopping, we report per\-question averages over the MedQA\-US set\.

MethodCallsRetr\.TimeAcc\.Tok\./Q\(\#\)\(\#\)\(s\)\(%\)\(\#\)CoT1\.00\.02\.577\.53713\.7MedRAG1\.01\.03\.277\.142264\.9i\-MedRAG3\.09\.08\.874\.7821516\.6SEMA\-RAG4\.83\.49\.589\.9519488\.4Table 5:Efficiency comparison on MedQA\-US \(deepseek\-v3\.1\)\.The results reveal three main patterns\. First, SEMA\-RAG consistently improves decision quality over single\-round baselines\. Second, these gains come with a moderate overhead relative to single\-pass methods, as expected from multi\-agent, multi\-round inference\. Third, compared with the iterative baselinei\-MedRAG, SEMA\-RAG achieves a more favorable accuracy–efficiency trade\-off, indicating that sufficiency\-driven early stopping allocates additional computation more effectively than fixed\-step iteration\. Taken together, these results suggest that SEMA\-RAG improves decision quality in a practically affordable regime, making the added overhead worthwhile for high\-stakes medical QA\.

## 5Related Work

### 5\.1Retrieval\-Augmented Medical Reasoning

Early medical LLMs systems largely depended on parametric knowledge or single\-round retrievalLuoet al\.\([2022](https://arxiv.org/html/2605.17101#bib.bib41)\); Singhalet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib40)\)\. Methods such as MedRAGXionget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib13)\)and MedCPTJinet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib14)\)improve domain retrieval with medical dual encoders and RAG pipelines, yet still follow a retrieve\-once\-then\-answer pattern, which often falls short on questions requiring multi\-hop evidence integration and clinically constrained reasoningJianget al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib42)\); Trivediet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib43)\)\.

Recent work has shifted from one\-shot retrieval to iterative RAGGaoet al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib44)\); Zhaoet al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib45)\)\. In general domains, Self\-RAGAsaiet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib24)\)and CRAGYanet al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib25)\)use self\-reflection to trigger re\-retrieval, while in healthcarei\-MedRAGXionget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib27)\)iteratively refines queries via follow\-up questions\. However, without explicit clinical intent and constraints modeling, iterations may devolve into shallow rewrites, yielding inefficient retrieval, drift, and uncontrolled expansionZhaoet al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib45)\); Zhuet al\.\([2026](https://arxiv.org/html/2605.17101#bib.bib46)\)\.

### 5\.2Agentic Collaboration in Medicine

Multi\-agent systems enhance complex\-task solving through role specialization and coordinationGuoet al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib48)\); Zonget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib62)\); Tranet al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib47)\)\(e\.g\., CAMELLiet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib26)\), MetaGPTHonget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib28)\)\)\. ReActYaoet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib29)\)further couples reasoning with tool use, allowing agents to act and retrieve information during inference\. In healthcare, MedAgentsTanget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib30)\)and Agent\-HospitalLiet al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib31)\)similarly show that multi\-role clinical collaboration improves diagnosis and decision quality\.

However, prior healthcare multi\-agent work largely centers on deliberation under the assumption that key evidence is already in\-context, leaving evidence acquisition unsystematicChenet al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib64)\); Gorenshteinet al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib50)\); Wanget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib49)\)\. In particular, gap identification, sufficiency\-driven termination, and evidence adjudication/integration are often missing, weakening reliable external evidence grounding and closed evidence\-chain formation in real clinical tasksLi \([2025](https://arxiv.org/html/2605.17101#bib.bib61)\); Amugongoet al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib63)\)\.

## 6Conclusion

We proposeSEMA\-RAG, a self\-evolving multi\-agent framework for medical question answering that restructures retrieval\-augmented generation according to the staged process of clinical reasoning, with theI\-Agentfor clinical schema interpretation, theE\-Agentfor sufficiency\-driven evidence exploration, and theA\-Agentfor evidence adjudication and final answer selection\. Across five medical QA benchmarks and five LLM backbones, SEMA\-RAG consistently outperforms strong baselines, improving the strongest baseline by an average of\+6\.46accuracy points, while ablations verify the necessity of the interpret–explore–adjudicate loop for reliable evidence\-chain construction\. Additional experiments further support its robustness across retrievers, smaller models, and more open\-ended interactive settings\. These findings suggest that medical RAG should move beyond static single\-round retrieval toward more adaptive and reliable evidence construction\.

## Limitations

Although we extend the evaluation beyond discrete\-choice medical QA with additional open\-ended and multi\-turn benchmarks, the current study is still limited to benchmark\-based settings rather than realistic clinical workflows such as longitudinal EHR reasoning or record\-grounded decision support\.

Our framework also depends on the quality and coverage of the retrieval corpus\. If critical evidence is missing, outdated, or only partially retrieved, the self\-evolving loop may still converge to incomplete grounding\. In addition, the current sufficiency criterion is not explicitly designed for option\-level separability or generative completeness\.

Finally, SEMA\-RAG introduces additional inference cost due to role specialization and multi\-round exploration\. Although this overhead is more efficient than fixed\-step iterative baselines, it remains higher than single\-round methods\. The current design also lacks explicit relevance filtering during evidence accumulation, making performance sensitive to stopping criteria and exploration hyperparameters\.

## References

- L\. M\. Amugongo, P\. Mascheroni, S\. Brooks, S\. Doering, and J\. Seidel \(2025\)Retrieval augmented generation for large language models in healthcare: A systematic review\.4\(6\),pp\. e0000877\(en\)\.External Links:ISSN 2767\-3170,[Link](https://dx.plos.org/10.1371/journal.pdig.0000877),[Document](https://dx.doi.org/10.1371/journal.pdig.0000877)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p2.1)\.
- 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\.arXiv\.Note:arXiv:2505\.08775 \[cs\]External Links:[Link](http://arxiv.org/abs/2505.08775),[Document](https://dx.doi.org/10.48550/arXiv.2505.08775)Cited by:[§F\.3](https://arxiv.org/html/2605.17101#A6.SS3.SSS0.Px1.p1.1)\.
- A\. Asai, Z\. Wu, Y\. Wang, A\. Sil, and H\. Hajishirzi \(2023\)Self\-RAG: Learning to Retrieve, Generate, and Critique through Self\-Reflection\.arXiv\.Note:arXiv:2310\.11511 \[cs\]External Links:[Link](http://arxiv.org/abs/2310.11511),[Document](https://dx.doi.org/10.48550/arXiv.2310.11511)Cited by:[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p2.1)\.
- X\. Chen, J\. Xiang, S\. Lu, Y\. Liu, M\. He, and D\. Shi \(2025\)Evaluating large language models and agents in healthcare: key challenges in clinical applications\.5\(2\),pp\. 151–163\.External Links:ISSN 2667\-1026,[Link](https://www.sciencedirect.com/science/article/pii/S2667102625000294),[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.imed.2025.03.002)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p2.1)\.
- DeepSeek\-AI, A\. Liu, B\. Feng, B\. Xue, B\. Wang, B\. Wu, C\. Lu, C\. Zhao, C\. Deng, C\. Zhang, C\. Ruan, D\. Dai, D\. Guo, D\. Yang, D\. Chen, D\. Ji, E\. Li, F\. Lin, F\. Dai, F\. Luo, G\. Hao, G\. Chen, G\. Li, H\. Zhang, H\. Bao, H\. Xu, H\. Wang, H\. Zhang, H\. Ding, H\. Xin, H\. Gao, H\. Li, H\. Qu, J\. L\. Cai, J\. Liang, J\. Guo, J\. Ni, J\. Li, J\. Wang, J\. Chen, J\. Chen, J\. Yuan, J\. Qiu, J\. Li, J\. Song, K\. Dong, K\. Hu, K\. Gao, K\. Guan, K\. Huang, K\. Yu, L\. Wang, L\. Zhang, L\. Xu, L\. Xia, L\. Zhao, L\. Wang, L\. Zhang, M\. Li, M\. Wang, M\. Zhang, M\. Zhang, M\. Tang, M\. Li, N\. Tian, P\. Huang, P\. Wang, P\. Zhang, Q\. Wang, Q\. Zhu, Q\. Chen, Q\. Du, R\. J\. Chen, R\. L\. Jin, R\. Ge, R\. Zhang, R\. Pan, R\. Wang, R\. Xu, R\. Zhang, R\. Chen, S\. S\. Li, S\. Lu, S\. Zhou, S\. Chen, S\. Wu, S\. Ye, S\. Ye, S\. Ma, S\. Wang, S\. Zhou, S\. Yu, S\. Zhou, S\. Pan, T\. Wang, T\. Yun, T\. Pei, T\. Sun, W\. L\. Xiao, W\. Zeng, W\. Zhao, W\. An, W\. Liu, W\. Liang, W\. Gao, W\. Yu, W\. Zhang, X\. Q\. Li, X\. Jin, X\. Wang, X\. Bi, X\. Liu, X\. Wang, X\. Shen, X\. Chen, X\. Zhang, X\. Chen, X\. Nie, X\. Sun, X\. Wang, X\. Cheng, X\. Liu, X\. Xie, X\. Liu, X\. Yu, X\. Song, X\. Shan, X\. Zhou, X\. Yang, X\. Li, X\. Su, X\. Lin, Y\. K\. Li, Y\. Q\. Wang, Y\. X\. Wei, Y\. X\. Zhu, Y\. Zhang, Y\. Xu, Y\. Xu, Y\. Huang, Y\. Li, Y\. Zhao, Y\. Sun, Y\. Li, Y\. Wang, Y\. Yu, Y\. Zheng, Y\. Zhang, Y\. Shi, Y\. Xiong, Y\. He, Y\. Tang, Y\. Piao, Y\. Wang, Y\. Tan, Y\. Ma, Y\. Liu, Y\. Guo, Y\. Wu, Y\. Ou, Y\. Zhu, Y\. Wang, Y\. Gong, Y\. Zou, Y\. He, Y\. Zha, Y\. Xiong, Y\. Ma, Y\. Yan, Y\. Luo, Y\. You, Y\. Liu, Y\. Zhou, Z\. F\. Wu, Z\. Z\. Ren, Z\. Ren, Z\. Sha, Z\. Fu, Z\. Xu, Z\. Huang, Z\. Zhang, Z\. Xie, Z\. Zhang, Z\. Hao, Z\. Gou, Z\. Ma, Z\. Yan, Z\. Shao, Z\. Xu, Z\. Wu, Z\. Zhang, Z\. Li, Z\. Gu, Z\. Zhu, Z\. Liu, Z\. Li, Z\. Xie, Z\. Song, Z\. Gao, and Z\. Pan \(2025\)DeepSeek\-V3 Technical Report\.arXiv\.Note:arXiv:2412\.19437 \[cs\]External Links:[Link](http://arxiv.org/abs/2412.19437),[Document](https://dx.doi.org/10.48550/arXiv.2412.19437)Cited by:[§4\.1\.2](https://arxiv.org/html/2605.17101#S4.SS1.SSS2.p1.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.2.1)\.
- H\. Gao, J\. Geng, W\. Hua, M\. Hu, X\. Juan, H\. Liu, S\. Liu, J\. Qiu, X\. Qi, Y\. Wu, H\. Wang, H\. Xiao, Y\. Zhou, S\. Zhang, J\. Zhang, J\. Xiang, Y\. Fang, Q\. Zhao, D\. Liu, Q\. Ren, C\. Qian, Z\. Wang, M\. Hu, H\. Wang, Q\. Wu, H\. Ji, and M\. Wang \(2026\)A Survey of Self\-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence\.arXiv\.Note:arXiv:2507\.21046 \[cs\]External Links:[Link](http://arxiv.org/abs/2507.21046),[Document](https://dx.doi.org/10.48550/arXiv.2507.21046)Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p3.1)\.
- Y\. Gao, Y\. Xiong, X\. Gao, K\. Jia, J\. Pan, Y\. Bi, Y\. Dai, J\. Sun, M\. Wang, and H\. Wang \(2024\)Retrieval\-Augmented Generation for Large Language Models: A Survey\.arXiv\.Note:arXiv:2312\.10997 \[cs\]External Links:[Link](http://arxiv.org/abs/2312.10997),[Document](https://dx.doi.org/10.48550/arXiv.2312.10997)Cited by:[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p2.1)\.
- T\. GLM, A\. Zeng, B\. Xu, B\. Wang, C\. Zhang, D\. Yin, D\. Zhang, D\. Rojas, G\. Feng, H\. Zhao, H\. Lai, H\. Yu, H\. Wang, J\. Sun, J\. Zhang, J\. Cheng, J\. Gui, J\. Tang, J\. Zhang, J\. Sun, J\. Li, L\. Zhao, L\. Wu, L\. Zhong, M\. Liu, M\. Huang, P\. Zhang, Q\. Zheng, R\. Lu, S\. Duan, S\. Zhang, S\. Cao, S\. Yang, W\. L\. Tam, W\. Zhao, X\. Liu, X\. Xia, X\. Zhang, X\. Gu, X\. Lv, X\. Liu, X\. Liu, X\. Yang, X\. Song, X\. Zhang, Y\. An, Y\. Xu, Y\. Niu, Y\. Yang, Y\. Li, Y\. Bai, Y\. Dong, Z\. Qi, Z\. Wang, Z\. Yang, Z\. Du, Z\. Hou, and Z\. Wang \(2024\)ChatGLM: A Family of Large Language Models from GLM\-130B to GLM\-4 All Tools\.arXiv\.Note:arXiv:2406\.12793 \[cs\]External Links:[Link](http://arxiv.org/abs/2406.12793),[Document](https://dx.doi.org/10.48550/arXiv.2406.12793)Cited by:[§4\.1\.2](https://arxiv.org/html/2605.17101#S4.SS1.SSS2.p1.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.26.1)\.
- L\. Gong, A\. Wang, Y\. Lai, W\. Ma, and Y\. Liu \(2025\)The Dialogue That Heals: A Comprehensive Evaluation of Doctor Agents’ Inquiry Capability\.arXiv\.Note:arXiv:2509\.24958 \[cs\]External Links:[Link](http://arxiv.org/abs/2509.24958),[Document](https://dx.doi.org/10.48550/arXiv.2509.24958)Cited by:[§F\.3](https://arxiv.org/html/2605.17101#A6.SS3.SSS0.Px2.p1.1)\.
- Google \(2025\)Gemini 2\.0 flash: model card\.Note:Accessed 2025\-12\-28External Links:[Link](https://modelcards.withgoogle.com/assets/documents/gemini-2-flash.pdf)Cited by:[§4\.1\.2](https://arxiv.org/html/2605.17101#S4.SS1.SSS2.p1.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.20.1)\.
- A\. Gorenshtein, M\. Omar, B\. S\. Glicksberg, G\. N\. Nadkarni, and E\. Klang \(2025\)AI Agents in Clinical Medicine: A Systematic Review\.Health Informatics\(en\)\.External Links:[Link](http://medrxiv.org/lookup/doi/10.1101/2025.08.22.25334232),[Document](https://dx.doi.org/10.1101/2025.08.22.25334232)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p2.1)\.
- T\. Guo, X\. Chen, Y\. Wang, R\. Chang, S\. Pei, N\. V\. Chawla, O\. Wiest, and X\. Zhang \(2024\)Large language model based multi\-agents: a survey of progress and challenges\.InProceedings of the Thirty\-Third International Joint Conference on Artificial Intelligence,IJCAI ’24\.External Links:ISBN 978\-1\-956792\-04\-1,[Link](https://doi.org/10.24963/ijcai.2024/890),[Document](https://dx.doi.org/10.24963/ijcai.2024/890)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p1.1)\.
- D\. Hendrycks, C\. Burns, S\. Basart, A\. Zou, M\. Mazeika, D\. Song, and J\. Steinhardt \(2021\)Measuring Massive Multitask Language Understanding\.arXiv\.Note:arXiv:2009\.03300 \[cs\]External Links:[Link](http://arxiv.org/abs/2009.03300),[Document](https://dx.doi.org/10.48550/arXiv.2009.03300)Cited by:[§B\.1](https://arxiv.org/html/2605.17101#A2.SS1.p1.1),[§4\.1\.1](https://arxiv.org/html/2605.17101#S4.SS1.SSS1.p1.1)\.
- S\. Hong, M\. Zhuge, J\. Chen, X\. Zheng, Y\. Cheng, C\. Zhang, J\. Wang, Z\. Wang, S\. K\. S\. Yau, Z\. Lin, L\. Zhou, C\. Ran, L\. Xiao, C\. Wu, and J\. Schmidhuber \(2024\)MetaGPT: Meta Programming for A Multi\-Agent Collaborative Framework\.arXiv\.Note:arXiv:2308\.00352 \[cs\]External Links:[Link](http://arxiv.org/abs/2308.00352),[Document](https://dx.doi.org/10.48550/arXiv.2308.00352)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p1.1)\.
- Z\. Jiang, F\. Xu, L\. Gao, Z\. Sun, Q\. Liu, J\. Dwivedi\-Yu, Y\. Yang, J\. Callan, and G\. Neubig \(2023\)Active retrieval augmented generation\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 7969–7992\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.495/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.495)Cited by:[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p1.1)\.
- D\. Jin, E\. Pan, N\. Oufattole, W\. Weng, H\. Fang, and P\. Szolovits \(2020\)What Disease does this Patient Have? A Large\-scale Open Domain Question Answering Dataset from Medical Exams\.arXiv\.Note:arXiv:2009\.13081 \[cs\]External Links:[Link](http://arxiv.org/abs/2009.13081),[Document](https://dx.doi.org/10.48550/arXiv.2009.13081)Cited by:[§B\.2](https://arxiv.org/html/2605.17101#A2.SS2.p1.1),[§4\.1\.1](https://arxiv.org/html/2605.17101#S4.SS1.SSS1.p1.1)\.
- 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\.Applied SciencesNatureBriefings in BioinformaticsACM Transactions on Information SystemsPLOS Digital HealthIntelligent Medicine11\(14\)\.External Links:[Link](https://www.mdpi.com/2076-3417/11/14/6421),ISSN 2076\-3417Cited by:[item Textbooks](https://arxiv.org/html/2605.17101#A3.I1.ix1.p1.1),[§4\.1\.3](https://arxiv.org/html/2605.17101#S4.SS1.SSS3.p1.1)\.
- Q\. Jin, B\. Dhingra, Z\. Liu, W\. W\. Cohen, and X\. Lu \(2019\)PubMedQA: A Dataset for Biomedical Research Question Answering\.arXiv\.Note:arXiv:1909\.06146 \[cs\]External Links:[Link](http://arxiv.org/abs/1909.06146),[Document](https://dx.doi.org/10.48550/arXiv.1909.06146)Cited by:[§B\.4](https://arxiv.org/html/2605.17101#A2.SS4.p1.1),[§4\.1\.1](https://arxiv.org/html/2605.17101#S4.SS1.SSS1.p1.1)\.
- Q\. Jin, W\. Kim, Q\. Chen, D\. C\. Comeau, L\. Yeganova, W\. J\. Wilbur, and Z\. Lu \(2023\)MedCPT: Contrastive Pre\-trained Transformers with large\-scale PubMed search logs for zero\-shot biomedical information retrieval\.Bioinformatics39\(11\),pp\. btad651\(en\)\.External Links:ISSN 1367\-4811,[Link](https://academic.oup.com/bioinformatics/article/doi/10.1093/bioinformatics/btad651/7335842),[Document](https://dx.doi.org/10.1093/bioinformatics/btad651)Cited by:[§4\.1\.2](https://arxiv.org/html/2605.17101#S4.SS1.SSS2.p2.1),[§4\.1\.3](https://arxiv.org/html/2605.17101#S4.SS1.SSS3.p1.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.10.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.16.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.22.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.28.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.4.1),[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p1.1)\.
- A\. Krithara, A\. Nentidis, K\. Bougiatiotis, and G\. Paliouras \(2023\)BioASQ\-QA: A manually curated corpus for Biomedical Question Answering\.Scientific Data10\(1\),pp\. 170\(en\)\.External Links:ISSN 2052\-4463,[Link](https://www.nature.com/articles/s41597-023-02068-4),[Document](https://dx.doi.org/10.1038/s41597-023-02068-4)Cited by:[§B\.5](https://arxiv.org/html/2605.17101#A2.SS5.p1.1),[§4\.1\.1](https://arxiv.org/html/2605.17101#S4.SS1.SSS1.p1.1)\.
- T\. H\. Kung, M\. Cheatham, A\. Medenilla, C\. Sillos, L\. De Leon, C\. Elepaño, M\. Madriaga, R\. Aggabao, G\. Diaz\-Candido, J\. Maningo, and V\. Tseng \(2023\)Performance of ChatGPT on USMLE: Potential for AI\-assisted medical education using large language models\.PLOS Digital Health2\(2\),pp\. e0000198\(en\)\.External Links:ISSN 2767\-3170,[Link](https://dx.plos.org/10.1371/journal.pdig.0000198),[Document](https://dx.doi.org/10.1371/journal.pdig.0000198)Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p1.1)\.
- P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. Kiela \(2020\)Retrieval\-augmented generation for knowledge\-intensive nlp tasks\.InProceedings of the 34th International Conference on Neural Information Processing Systems,NIPS ’20,Red Hook, NY, USA\.External Links:ISBN 9781713829546Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p1.1)\.
- G\. Li, H\. Hammoud, H\. Itani, D\. Khizbullin, and B\. Ghanem \(2023\)CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society\.InAdvances in Neural Information Processing Systems,A\. Oh, T\. Naumann, A\. Globerson, K\. Saenko, M\. Hardt, and S\. Levine \(Eds\.\),Vol\.36,pp\. 51991–52008\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2023/file/a3621ee907def47c1b952ade25c67698-Paper-Conference.pdf)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p1.1)\.
- J\. Li, Y\. Lai, W\. Li, J\. Ren, M\. Zhang, X\. Kang, S\. Wang, P\. Li, Y\. Zhang, W\. Ma, and Y\. Liu \(2025\)Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents\.arXiv\.Note:arXiv:2405\.02957 \[cs\]External Links:[Link](http://arxiv.org/abs/2405.02957),[Document](https://dx.doi.org/10.48550/arXiv.2405.02957)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p1.1)\.
- X\. Li \(2025\)A review of prominent paradigms for LLM\-based agents: tool use, planning \(including RAG\), and feedback learning\.InProceedings of the 31st International Conference on Computational Linguistics,O\. Rambow, L\. Wanner, M\. Apidianaki, H\. Al\-Khalifa, B\. D\. Eugenio, and S\. Schockaert \(Eds\.\),Abu Dhabi, UAE,pp\. 9760–9779\.External Links:[Link](https://aclanthology.org/2025.coling-main.652/)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p2.1)\.
- A\. Linn, C\. Khaw, H\. Kildea, and A\. Tonkin \(2012\)Clinical reasoning: a guide to improving teaching and practice\.Australian Family Physician41\(1\-2\),pp\. 18–20\.Note:January/February issueExternal Links:[Link](https://www.racgp.org.au/afp/2012/january-february/clinical-reasoning)Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p2.1)\.
- N\. F\. Liu, K\. Lin, J\. Hewitt, A\. Paranjape, M\. Bevilacqua, F\. Petroni, and P\. Liang \(2024\)Lost in the middle: how language models use long contexts\.Transactions of the Association for Computational Linguistics12,pp\. 157–173\.External Links:[Link](https://aclanthology.org/2024.tacl-1.9/),[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00638)Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p3.1)\.
- 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\.23\(6\),pp\. bbac409\.Note:arXiv:2210\.10341 \[cs\]External Links:ISSN 1467\-5463, 1477\-4054,[Link](http://arxiv.org/abs/2210.10341),[Document](https://dx.doi.org/10.1093/bib/bbac409)Cited by:[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p1.1)\.
- A\. Mallen, A\. Asai, V\. Zhong, R\. Das, D\. Khashabi, and H\. Hajishirzi \(2023\)When not to trust language models: investigating effectiveness of parametric and non\-parametric memories\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 9802–9822\.External Links:[Link](https://aclanthology.org/2023.acl-long.546/),[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.546)Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p2.1)\.
- M\. Omar, G\. N\. Nadkarni, E\. Klang, and B\. S\. Glicksberg \(2024\)Large language models in medicine: A review of current clinical trials across healthcare applications\.PLOS Digital Health3\(11\),pp\. e0000662\(en\)\.External Links:ISSN 2767\-3170,[Link](https://dx.plos.org/10.1371/journal.pdig.0000662),[Document](https://dx.doi.org/10.1371/journal.pdig.0000662)Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p1.1)\.
- J\. A\. Omiye, H\. Gui, S\. J\. Rezaei, J\. Zou, and R\. Daneshjou \(2024\)Large Language Models in Medicine: The Potentials and Pitfalls: A Narrative Review\.Annals of Internal Medicine177\(2\),pp\. 210–220\(en\)\.External Links:ISSN 0003\-4819, 1539\-3704,[Link](https://www.acpjournals.org/doi/10.7326/M23-2772),[Document](https://dx.doi.org/10.7326/M23-2772)Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p1.1)\.
- A\. Pal, L\. K\. Umapathi, and M\. Sankarasubbu \(2022\)MedMCQA : A Large\-scale Multi\-Subject Multi\-Choice Dataset for Medical domain Question Answering\.arXiv\.Note:arXiv:2203\.14371 \[cs\]External Links:[Link](http://arxiv.org/abs/2203.14371),[Document](https://dx.doi.org/10.48550/arXiv.2203.14371)Cited by:[§B\.3](https://arxiv.org/html/2605.17101#A2.SS3.p1.1),[§4\.1\.1](https://arxiv.org/html/2605.17101#S4.SS1.SSS1.p1.1)\.
- D\. Roustan and F\. Bastardot \(2025\)The Clinicians’ Guide to Large Language Models: A General Perspective With a Focus on Hallucinations\.Interactive Journal of Medical Research14,pp\. e59823\(en\)\.External Links:ISSN 1929\-073X,[Link](https://www.i-jmr.org/2025/1/e59823),[Document](https://dx.doi.org/10.2196/59823)Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p1.1)\.
- F\. Shi, X\. Chen, K\. Misra, N\. Scales, D\. Dohan, E\. Chi, N\. Schärli, and D\. Zhou \(2023\)Large language models can be easily distracted by irrelevant context\.InProceedings of the 40th International Conference on Machine Learning,ICML’23\.Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p2.1)\.
- 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\.620\(7972\),pp\. 172–180\.External Links:ISSN 1476\-4687,[Link](https://doi.org/10.1038/s41586-023-06291-2),[Document](https://dx.doi.org/10.1038/s41586-023-06291-2)Cited by:[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p1.1)\.
- L\. Soldaini, A\. Yates, and N\. Goharian \(2017\)Learning to reformulate long queries for clinical decision support\.J\. Assoc\. Inf\. Sci\. Technol\.68\(11\),pp\. 2602–2619\.External Links:ISSN 2330\-1635,[Link](https://doi.org/10.1002/asi.23924),[Document](https://dx.doi.org/10.1002/asi.23924)Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p2.1)\.
- StatPearls Publishing \(2025\)StatPearls \[internet\]\.StatPearls Publishing,Treasure Island \(FL\)\.Note:NCBI BookshelfNCBI Bookshelf ID: NBK430685\. Accessed 2026\-01\-04External Links:[Link](https://www.ncbi.nlm.nih.gov/books/NBK430685/)Cited by:[item StatPearls](https://arxiv.org/html/2605.17101#A3.I1.ix2.p1.1),[§4\.1\.3](https://arxiv.org/html/2605.17101#S4.SS1.SSS3.p1.1)\.
- X\. Tang, A\. Zou, Z\. Zhang, Z\. Li, Y\. Zhao, X\. Zhang, A\. Cohan, and M\. Gerstein \(2024\)MedAgents: large language models as collaborators for zero\-shot medical reasoning\.InFindings of the Association for Computational Linguistics: ACL 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 599–621\.External Links:[Link](https://aclanthology.org/2024.findings-acl.33/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.33)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p1.1)\.
- G\. Team, A\. Kamath, J\. Ferret, S\. Pathak, N\. Vieillard, R\. Merhej, S\. Perrin, T\. Matejovicova, A\. Ramé, M\. Rivière, L\. Rouillard, T\. Mesnard, G\. Cideron, J\. Grill, S\. Ramos, E\. Yvinec, M\. Casbon, E\. Pot, I\. Penchev, G\. Liu, F\. Visin, K\. Kenealy, L\. Beyer, X\. Zhai, A\. Tsitsulin, R\. Busa\-Fekete, A\. Feng, N\. Sachdeva, B\. Coleman, Y\. Gao, B\. Mustafa, I\. Barr, E\. Parisotto, D\. Tian, M\. Eyal, C\. Cherry, J\. Peter, D\. Sinopalnikov, S\. Bhupatiraju, R\. Agarwal, M\. Kazemi, D\. Malkin, R\. Kumar, D\. Vilar, I\. Brusilovsky, J\. Luo, A\. Steiner, A\. Friesen, A\. Sharma, A\. Sharma, A\. M\. Gilady, A\. Goedeckemeyer, A\. Saade, A\. Feng, A\. Kolesnikov, A\. Bendebury, A\. Abdagic, A\. Vadi, A\. György, A\. S\. Pinto, A\. Das, A\. Bapna, A\. Miech, A\. Yang, A\. Paterson, A\. Shenoy, A\. Chakrabarti, B\. Piot, B\. Wu, B\. Shahriari, B\. Petrini, C\. Chen, C\. L\. Lan, C\. A\. Choquette\-Choo, C\. J\. Carey, C\. Brick, D\. Deutsch, D\. Eisenbud, D\. Cattle, D\. Cheng, D\. Paparas, D\. S\. Sreepathihalli, D\. Reid, D\. Tran, D\. Zelle, E\. Noland, E\. Huizenga, E\. Kharitonov, F\. Liu, G\. Amirkhanyan, G\. Cameron, H\. Hashemi, H\. Klimczak\-Plucińska, H\. Singh, H\. Mehta, H\. T\. Lehri, H\. Hazimeh, I\. Ballantyne, I\. Szpektor, I\. Nardini, J\. Pouget\-Abadie, J\. Chan, J\. Stanton, J\. Wieting, J\. Lai, J\. Orbay, J\. Fernandez, J\. Newlan, J\. Ji, J\. Singh, K\. Black, K\. Yu, K\. Hui, K\. Vodrahalli, K\. Greff, L\. Qiu, M\. Valentine, M\. Coelho, M\. Ritter, M\. Hoffman, M\. Watson, M\. Chaturvedi, M\. Moynihan, M\. Ma, N\. Babar, N\. Noy, N\. Byrd, N\. Roy, N\. Momchev, N\. Chauhan, N\. Sachdeva, O\. Bunyan, P\. Botarda, P\. Caron, P\. K\. Rubenstein, P\. Culliton, P\. Schmid, P\. G\. Sessa, P\. Xu, P\. Stanczyk, P\. Tafti, R\. Shivanna, R\. Wu, R\. Pan, R\. Rokni, R\. Willoughby, R\. Vallu, R\. Mullins, S\. Jerome, S\. Smoot, S\. Girgin, S\. Iqbal, S\. Reddy, S\. Sheth, S\. Põder, S\. Bhatnagar, S\. R\. Panyam, S\. Eiger, S\. Zhang, T\. Liu, T\. Yacovone, T\. Liechty, U\. Kalra, U\. Evci, V\. Misra, V\. Roseberry, V\. Feinberg, V\. Kolesnikov, W\. Han, W\. Kwon, X\. Chen, Y\. Chow, Y\. Zhu, Z\. Wei, Z\. Egyed, V\. Cotruta, M\. Giang, P\. Kirk, A\. Rao, K\. Black, N\. Babar, J\. Lo, E\. Moreira, L\. G\. Martins, O\. Sanseviero, L\. Gonzalez, Z\. Gleicher, T\. Warkentin, V\. Mirrokni, E\. Senter, E\. Collins, J\. Barral, Z\. Ghahramani, R\. Hadsell, Y\. Matias, D\. Sculley, S\. Petrov, N\. Fiedel, N\. Shazeer, O\. Vinyals, J\. Dean, D\. Hassabis, K\. Kavukcuoglu, C\. Farabet, E\. Buchatskaya, J\. Alayrac, R\. Anil, Dmitry, Lepikhin, S\. Borgeaud, O\. Bachem, A\. Joulin, A\. Andreev, C\. Hardin, R\. Dadashi, and L\. Hussenot \(2025a\)Gemma 3 Technical Report\.arXiv\.Note:arXiv:2503\.19786 \[cs\]External Links:[Link](http://arxiv.org/abs/2503.19786),[Document](https://dx.doi.org/10.48550/arXiv.2503.19786)Cited by:[§F\.2](https://arxiv.org/html/2605.17101#A6.SS2.p1.1)\.
- K\. Team, Y\. Bai, Y\. Bao, G\. Chen, J\. Chen, N\. Chen, R\. Chen, Y\. Chen, Y\. Chen, Y\. Chen, Z\. Chen, J\. Cui, H\. Ding, M\. Dong, A\. Du, C\. Du, D\. Du, Y\. Du, Y\. Fan, Y\. Feng, K\. Fu, B\. Gao, H\. Gao, P\. Gao, T\. Gao, X\. Gu, L\. Guan, H\. Guo, J\. Guo, H\. Hu, X\. Hao, T\. He, W\. He, W\. He, C\. Hong, Y\. Hu, Z\. Hu, W\. Huang, Z\. Huang, Z\. Huang, T\. Jiang, Z\. Jiang, X\. Jin, Y\. Kang, G\. Lai, C\. Li, F\. Li, H\. Li, M\. Li, W\. Li, Y\. Li, Y\. Li, Z\. Li, Z\. Li, H\. Lin, X\. Lin, Z\. Lin, C\. Liu, C\. Liu, H\. Liu, J\. Liu, J\. Liu, L\. Liu, S\. Liu, T\. Y\. Liu, T\. Liu, W\. Liu, Y\. Liu, Y\. Liu, Y\. Liu, Y\. Liu, Z\. Liu, E\. Lu, L\. Lu, S\. Ma, X\. Ma, Y\. Ma, S\. Mao, J\. Mei, X\. Men, Y\. Miao, S\. Pan, Y\. Peng, R\. Qin, B\. Qu, Z\. Shang, L\. Shi, S\. Shi, F\. Song, J\. Su, Z\. Su, X\. Sun, F\. Sung, H\. Tang, J\. Tao, Q\. Teng, C\. Wang, D\. Wang, F\. Wang, H\. Wang, J\. Wang, J\. Wang, J\. Wang, S\. Wang, S\. Wang, Y\. Wang, Y\. Wang, Y\. Wang, Y\. Wang, Y\. Wang, Z\. Wang, Z\. Wang, Z\. Wang, C\. Wei, Q\. Wei, W\. Wu, X\. Wu, Y\. Wu, C\. Xiao, X\. Xie, W\. Xiong, B\. Xu, J\. Xu, J\. Xu, L\. H\. Xu, L\. Xu, S\. Xu, W\. Xu, X\. Xu, Y\. Xu, Z\. Xu, J\. Yan, Y\. Yan, X\. Yang, Y\. Yang, Z\. Yang, Z\. Yang, Z\. Yang, H\. Yao, X\. Yao, W\. Ye, Z\. Ye, B\. Yin, L\. Yu, E\. Yuan, H\. Yuan, M\. Yuan, H\. Zhan, D\. Zhang, H\. Zhang, W\. Zhang, X\. Zhang, Y\. Zhang, Y\. Zhang, Y\. Zhang, Y\. Zhang, Y\. Zhang, Y\. Zhang, Z\. Zhang, H\. Zhao, Y\. Zhao, H\. Zheng, S\. Zheng, J\. Zhou, X\. Zhou, Z\. Zhou, Z\. Zhu, W\. Zhuang, and X\. Zu \(2025b\)Kimi K2: Open Agentic Intelligence\.arXiv\.Note:arXiv:2507\.20534 \[cs\]External Links:[Link](http://arxiv.org/abs/2507.20534),[Document](https://dx.doi.org/10.48550/arXiv.2507.20534)Cited by:[§4\.1\.2](https://arxiv.org/html/2605.17101#S4.SS1.SSS2.p1.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.8.1)\.
- K\. Tran, D\. Dao, M\. Nguyen, Q\. Pham, B\. O’Sullivan, and H\. D\. Nguyen \(2025\)Multi\-Agent Collaboration Mechanisms: A Survey of LLMs\.arXiv\.Note:arXiv:2501\.06322 \[cs\]External Links:[Link](http://arxiv.org/abs/2501.06322),[Document](https://dx.doi.org/10.48550/arXiv.2501.06322)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p1.1)\.
- H\. Trivedi, N\. Balasubramanian, T\. Khot, and A\. Sabharwal \(2023\)Interleaving retrieval with chain\-of\-thought reasoning for knowledge\-intensive multi\-step questions\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 10014–10037\.External Links:[Link](https://aclanthology.org/2023.acl-long.557/),[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.557)Cited by:[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p1.1)\.
- G\. Tsatsaronis, G\. Balikas, P\. Malakasiotis, I\. Partalas, M\. Zschunke, M\. R\. Alvers, D\. Weissenborn, A\. Krithara, S\. Petridis, D\. Polychronopoulos, Y\. Almirantis, J\. Pavlopoulos, N\. Baskiotis, P\. Gallinari, T\. Artiéres, A\. N\. Ngomo, N\. Heino, E\. Gaussier, L\. Barrio\-Alvers, M\. Schroeder, I\. Androutsopoulos, and G\. Paliouras \(2015\)An overview of the BIOASQ large\-scale biomedical semantic indexing and question answering competition\.BMC Bioinformatics16\(1\),pp\. 138\(en\)\.External Links:ISSN 1471\-2105,[Link](https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-015-0564-6),[Document](https://dx.doi.org/10.1186/s12859-015-0564-6)Cited by:[§B\.5](https://arxiv.org/html/2605.17101#A2.SS5.p1.1),[§4\.1\.1](https://arxiv.org/html/2605.17101#S4.SS1.SSS1.p1.1)\.
- L\. Wang, W\. Xu, Y\. Lan, Z\. Hu, Y\. Lan, R\. K\. Lee, and E\. Lim \(2023\)Plan\-and\-solve prompting: improving zero\-shot chain\-of\-thought reasoning by large language models\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 2609–2634\.External Links:[Link](https://aclanthology.org/2023.acl-long.147/),[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.147)Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p3.1)\.
- W\. Wang, Z\. Ma, Z\. Wang, C\. Wu, J\. Ji, W\. Chen, X\. Li, and Y\. Yuan \(2025\)A survey of LLM\-based agents in medicine: how far are we from baymax?\.InFindings of the Association for Computational Linguistics: ACL 2025,W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 10345–10359\.External Links:[Link](https://aclanthology.org/2025.findings-acl.539/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-acl.539),ISBN 979\-8\-89176\-256\-5Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p2.1)\.
- J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, B\. Ichter, F\. Xia, E\. H\. Chi, Q\. V\. Le, and D\. Zhou \(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.InProceedings of the 36th International Conference on Neural Information Processing Systems,NIPS ’22,Red Hook, NY, USA\.External Links:ISBN 9781713871088Cited by:[§4\.1\.2](https://arxiv.org/html/2605.17101#S4.SS1.SSS2.p2.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.15.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.21.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.27.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.3.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.9.1)\.
- G\. Xiong, Q\. Jin, Z\. Lu, and A\. Zhang \(2024\)Benchmarking retrieval\-augmented generation for medicine\.InFindings of the Association for Computational Linguistics: ACL 2024,L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 6233–6251\.External Links:[Link](https://aclanthology.org/2024.findings-acl.372/),[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.372)Cited by:[Appendix C](https://arxiv.org/html/2605.17101#A3.p1.1),[§4\.1\.1](https://arxiv.org/html/2605.17101#S4.SS1.SSS1.p1.1),[§4\.1\.2](https://arxiv.org/html/2605.17101#S4.SS1.SSS2.p2.1),[§4\.1\.3](https://arxiv.org/html/2605.17101#S4.SS1.SSS3.p1.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.11.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.17.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.23.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.29.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.5.1),[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p1.1)\.
- G\. Xiong, Q\. Jin, X\. Wang, M\. Zhang, Z\. Lu, and A\. Zhang \(2025\)Improving Retrieval\-Augmented Generation in Medicine with Iterative Follow\-up Questions\.Pacific Symposium on Biocomputing\. Pacific Symposium on Biocomputing30,pp\. 199–214\(eng\)\.External Links:ISSN 2335\-6936,[Document](https://dx.doi.org/10.1142/9789819807024%5F0015)Cited by:[Appendix C](https://arxiv.org/html/2605.17101#A3.p1.1),[§4\.1\.2](https://arxiv.org/html/2605.17101#S4.SS1.SSS2.p2.1),[§4\.1\.3](https://arxiv.org/html/2605.17101#S4.SS1.SSS3.p1.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.12.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.18.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.24.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.30.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.6.1),[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p2.1)\.
- S\. Yan, J\. Gu, Y\. Zhu, and Z\. Ling \(2024\)Corrective Retrieval Augmented Generation\.arXiv\.Note:arXiv:2401\.15884 \[cs\]External Links:[Link](http://arxiv.org/abs/2401.15884),[Document](https://dx.doi.org/10.48550/arXiv.2401.15884)Cited by:[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p2.1)\.
- A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv, C\. Zheng, D\. Liu, F\. Zhou, F\. Huang, F\. Hu, H\. Ge, H\. Wei, H\. Lin, J\. Tang, J\. Yang, J\. Tu, J\. Zhang, J\. Yang, J\. Yang, J\. Zhou, J\. Zhou, J\. Lin, K\. Dang, K\. Bao, K\. Yang, L\. Yu, L\. Deng, M\. Li, M\. Xue, M\. Li, P\. Zhang, P\. Wang, Q\. Zhu, R\. Men, R\. Gao, S\. Liu, S\. Luo, T\. Li, T\. Tang, W\. Yin, X\. Ren, X\. Wang, X\. Zhang, X\. Ren, Y\. Fan, Y\. Su, Y\. Zhang, Y\. Zhang, Y\. Wan, Y\. Liu, Z\. Wang, Z\. Cui, Z\. Zhang, Z\. Zhou, and Z\. Qiu \(2025\)Qwen3 Technical Report\.arXiv\.Note:arXiv:2505\.09388 \[cs\]External Links:[Link](http://arxiv.org/abs/2505.09388),[Document](https://dx.doi.org/10.48550/arXiv.2505.09388)Cited by:[§4\.1\.2](https://arxiv.org/html/2605.17101#S4.SS1.SSS2.p1.1),[Table 1](https://arxiv.org/html/2605.17101#S4.T1.1.1.14.1)\.
- S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao \(2023\)ReAct: Synergizing Reasoning and Acting in Language Models\.arXiv\.Note:arXiv:2210\.03629 \[cs\]External Links:[Link](http://arxiv.org/abs/2210.03629),[Document](https://dx.doi.org/10.48550/arXiv.2210.03629)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p1.1)\.
- S\. Yazdani, M\. Hosseinzadeh, and F\. Hosseini \(2017\)Models of clinical reasoning with a focus on general practice: A critical review\.Journal of Advances in Medical Education & Professionalism5\(4\),pp\. 177–184\(eng\)\.External Links:ISSN 2322\-2220Cited by:[§1](https://arxiv.org/html/2605.17101#S1.p2.1)\.
- Y\. Zhang, M\. Li, D\. Long, X\. Zhang, H\. Lin, B\. Yang, P\. Xie, A\. Yang, D\. Liu, J\. Lin, F\. Huang, and J\. Zhou \(2025\)Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models\.arXiv\.Note:arXiv:2506\.05176 \[cs\]External Links:[Link](http://arxiv.org/abs/2506.05176),[Document](https://dx.doi.org/10.48550/arXiv.2506.05176)Cited by:[§F\.1](https://arxiv.org/html/2605.17101#A6.SS1.p1.1)\.
- S\. Zhao, Y\. Yang, Z\. Wang, Z\. He, L\. K\. Qiu, and L\. Qiu \(2024\)Retrieval Augmented Generation \(RAG\) and Beyond: A Comprehensive Survey on How to Make your LLMs use External Data More Wisely\.arXiv\.Note:arXiv:2409\.14924 \[cs\]External Links:[Link](http://arxiv.org/abs/2409.14924),[Document](https://dx.doi.org/10.48550/arXiv.2409.14924)Cited by:[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p2.1)\.
- Y\. Zhu, H\. Yuan, S\. Wang, J\. Liu, W\. Liu, C\. Deng, H\. Chen, Z\. Liu, Z\. Dou, and J\. Wen \(2026\)Large Language Models for Information Retrieval: A Survey\.44\(1\),pp\. 1–54\.Note:arXiv:2308\.07107 \[cs\]External Links:ISSN 1046\-8188, 1558\-2868,[Link](http://arxiv.org/abs/2308.07107),[Document](https://dx.doi.org/10.1145/3748304)Cited by:[§5\.1](https://arxiv.org/html/2605.17101#S5.SS1.p2.1)\.
- C\. Zong, Y\. Yan, W\. Lu, J\. Shao, E\. Huang, H\. Chang, and Y\. Zhuang \(2024\)Triad: A Framework Leveraging a Multi\-Role LLM\-based Agent to Solve Knowledge Base Question Answering\.arXiv\.Note:arXiv:2402\.14320 \[cs\]External Links:[Link](http://arxiv.org/abs/2402.14320),[Document](https://dx.doi.org/10.48550/arXiv.2402.14320)Cited by:[§5\.2](https://arxiv.org/html/2605.17101#S5.SS2.p1.1)\.

## Appendix

## Appendix AAlgorithm Pseudocode

Algorithm[1](https://arxiv.org/html/2605.17101#alg1)formally describes the inference flow of SEMA\-RAG, highlighting the interaction between agents\.

Algorithm 1Inference Process of SEMA\-RAG0:Question

QQ, corpus

𝒞\\mathcal\{C\}, max iterations

TmaxT\_\{\\max\}
0:Final answer

y~\\tilde\{y\}
1:Stage 1: I\-Agent

2:

Q′=⟨oint,oent,ocons,qinit⟩←AgentI​\(P​m​tI,Q\)Q^\{\\prime\}=\\langle o\_\{\\mathrm\{int\}\},o\_\{\\mathrm\{ent\}\},o\_\{\\mathrm\{cons\}\},q\_\{\\mathrm\{init\}\}\\rangle\\leftarrow\\mathrm\{Agent\}\_\{\\mathrm\{I\}\}\(Pmt\_\{\\mathrm\{I\}\},Q\)
3:

q^init←Linearize​\(Q′\)\\hat\{q\}\_\{\\mathrm\{init\}\}\\leftarrow\\mathrm\{Linearize\}\(Q^\{\\prime\}\)
4:

C0←∅C\_\{0\}\\leftarrow\\varnothing
5:

𝒬1←\{q^init\}\\mathcal\{Q\}\_\{1\}\\leftarrow\\\{\\hat\{q\}\_\{\\mathrm\{init\}\}\\\}
6:Stage 2: E\-Agent

7:for

t=1t=1to

TmaxT\_\{\\max\}do

8:

𝒟t←⋃q∈𝒬tTopK​\(q\)\\mathcal\{D\}\_\{t\}\\leftarrow\\bigcup\_\{q\\in\\mathcal\{Q\}\_\{t\}\}\\mathrm\{TopK\}\(q\)
9:

Ct←Dedup​\(Ct−1∪𝒟t\)C\_\{t\}\\leftarrow\\mathrm\{Dedup\}\(C\_\{t\-1\}\\cup\\mathcal\{D\}\_\{t\}\)
10:

\[st,gt,𝒬t\+1\]←AgentE​\(P​m​tE,\[Q′,𝒬t,Ct\]\)\[s\_\{t\},g\_\{t\},\\mathcal\{Q\}\_\{t\+1\}\]\\leftarrow\\mathrm\{Agent\}\_\{\\mathrm\{E\}\}\(Pmt\_\{\\mathrm\{E\}\},\[Q^\{\\prime\},\\mathcal\{Q\}\_\{t\},C\_\{t\}\]\)
11:if

st=1s\_\{t\}=1then

12:break

13:endif

14:if

𝒬t\+1=∅\\mathcal\{Q\}\_\{t\+1\}=\\varnothingthen

15:break

16:endif

17:endfor

18:

T←tT\\leftarrow t
19:

C∗←CtC^\{\*\}\\leftarrow C\_\{t\}
20:Stage 3: A\-Agent

21:

𝒮∗←\{\(ID​\(Di\),Di\)∣Di∈C∗\}\\mathcal\{S\}^\{\*\}\\leftarrow\\\{\(\\mathrm\{ID\}\(D\_\{i\}\),D\_\{i\}\)\\mid D\_\{i\}\\in C^\{\*\}\\\}
22:

R←AgentA​\(P​m​tadj,\[Q,C∗,𝒮∗\]\)R\\leftarrow\\mathrm\{Agent\}\_\{\\mathrm\{A\}\}\(Pmt\_\{\\mathrm\{adj\}\},\[Q,C^\{\*\},\\mathcal\{S\}^\{\*\}\]\)
23:

y~←AgentA​\(P​m​tans,\[Q,R\]\)\\tilde\{y\}\\leftarrow\\mathrm\{Agent\}\_\{\\mathrm\{A\}\}\(Pmt\_\{\\mathrm\{ans\}\},\[Q,R\]\)
24:return

y~\\tilde\{y\}

## Appendix BDataset Details

We evaluate SEMA\-RAG on five standard medical question\-answering benchmarks from the MIRAGE suite: MMLU\-Med, MedQA\-US, MedMCQA, PubMedQA\*, and BioASQ\-Y/N\. Table[6](https://arxiv.org/html/2605.17101#A2.T6)summarizes the dataset sizes and answer formats used in our experiments, and we briefly describe each benchmark below\.

### B\.1MMLU\-Med

The MMLU\-Med dataset is a subset of the Massive Multitask Language Understanding \(MMLU\) benchmarkHendryckset al\.\([2021](https://arxiv.org/html/2605.17101#bib.bib18)\)\. It covers six distinct subtasks:Clinical Knowledge,Medical Genetics,Anatomy,Professional Medicine,College Biology, andCollege Medicine\. The questions are designed to measure knowledge acquired during preclinical and clinical medical training, formatted as 4\-choice multiple\-choice questions\.

### B\.2MedQA\-US

MedQA\-USJinet al\.\([2020](https://arxiv.org/html/2605.17101#bib.bib19)\)is derived from the United States Medical Licensing Examination \(USMLE\)\. It represents highly complex clinical case studies that require multi\-hop reasoning and domain\-specific knowledge to solve\. Following the standard setting in MIRAGE, we use the 4\-option English version of the dataset\. The questions typically present a patient vignette followed by a query about diagnosis, prognosis, or pharmacology\.

### B\.3MedMCQA

MedMCQAPalet al\.\([2022](https://arxiv.org/html/2605.17101#bib.bib20)\)is a large\-scale dataset collected from Indian medical entrance examinations \(AIIMS and NEET\-PG\)\. It covers a wide range of 21 medical subjects, including surgery, pediatrics, and pharmacology\. The questions vary significantly in difficulty and length, testing both memorization of medical facts and application of concepts in clinical scenarios\.

### B\.4PubMedQA\*

PubMedQAJinet al\.\([2019](https://arxiv.org/html/2605.17101#bib.bib21)\)is a biomedical research question\-answering dataset\. The task requires answering "Yes", "No", or "Maybe" to a research question based on a provided abstract\.PubMedQA\*refers to the setting where the original context \(abstract\) is removed, forcing the model to retrieve external evidence to answer the question\. This setting tests the system’s ability to find relevant biomedical literature to support a scientific conclusion\.

### B\.5BioASQ\-Y/N

BioASQ\-Y/N is a subset of the BioASQ Task B benchmarkTsatsaroniset al\.\([2015](https://arxiv.org/html/2605.17101#bib.bib22)\); Kritharaet al\.\([2023](https://arxiv.org/html/2605.17101#bib.bib23)\)\. It consists of biomedical questions that require a strict "Yes" or "No" answer\. These questions are expert\-constructed and reflect real\-world information needs of biomedical researchers\. The task is challenging because it often involves specific gene\-disease associations or protein interactions that require precise fact\-checking\.

Dataset\#SamplesTaskMMLU\-Med10894\-choice MCQMedQA\-US12734\-choice MCQMedMCQA41834\-choice MCQPubMedQA\*5003\-choice Y/N/MBioASQ\-Y/N6182\-choice Y/NTable 6:Statistics of the medical QA datasets from MIRAGE used in our experiments\.

## Appendix CRetrieval Corpus Details

Followingi\-MedRAGXionget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib27)\)and MedRAGXionget al\.\([2024](https://arxiv.org/html/2605.17101#bib.bib13)\), we employ a hybrid retrieval corpus that combines medical textbooks with point\-of\-care clinical summaries, covering both foundational concepts and practical clinical knowledge\.

TextbooksThis component is the released medical textbook collection used in prior medical QA benchmarksJinet al\.\([2021](https://arxiv.org/html/2605.17101#bib.bib38)\)\. It contains widely used reference textbooks spanning core biomedical sciences and clinical specialties, and is particularly helpful for queries requiring standard definitions, canonical mechanisms, and established medical principles\.

StatPearlsStatPearlsStatPearls Publishing \([2025](https://arxiv.org/html/2605.17101#bib.bib39)\)is a point\-of\-care clinical review resource that provides high\-yield summaries across diseases, diagnostics, and treatments\. In our setup, we use the publicly available StatPearls articles \(e\.g\., via the NCBI Bookshelf releases\) as in prior work, which complements textbooks with concise, practice\-oriented evidence for retrieval\.

## Appendix DError Analysis

To probe where SEMA\-RAG can still fail, we analyze a representative MedQA\-US error \(Question 0060\) in Table[7](https://arxiv.org/html/2605.17101#A4.T7)\. This case is informative because the system retrieves the correct biochemical cue at the*class*level, yet still makes an incorrect discrete choice among the remaining candidates\.

Error Case \(MedQA\-US Q0060\)Question \(brief\): Septic shock with pelvic infectious focus; phenol/90∘C assay indicates a Lipid A–like motif⇒\\RightarrowGram\-negative signal\.Options:A\.Coagulase\-positive, Gram\-positive cocciB\.Encapsulated, Gram\-negative coccobacilliC\.Spore\-forming, Gram\-positive bacilliD\.Lactose\-fermenting, Gram\-negative rods\\rowcolor\[RGB\]240,248,2551\) I\-Agent: Extracts anchors \(pelvic source, shock/DIC\-like labs\) and treats the biochemical clue as the key discriminator\.\\rowcolor\[RGB\]240,248,2552\) E\-Agent: Retrieves evidence consistent with an LPS/Lipid A signature and correctly narrows the class toGram\-negative, ruling outA/C…\\rowcolor\[RGB\]255,245,2453\) A\-Agent: Commits among the remaining candidates without enforcing*option\-separating*evidence \(rod vs\. coccobacillus; lactose fermentation\) …Prediction: B✗Ground Truth: DTable 7:A representative failure where retrieval supports only*class\-level*elimination, while*option\-level*discrimination remains under\-supportedHere,I\-Agentidentifies the decision anchors from both presentation and assay\. The clinical picture indicates severe sepsis with a pelvic infectious focus, accompanied by DIC\-like abnormalities\. The phenol\-heating assay reveals a phosphorylatedN\-acetylglucosamine dimer with multiple fatty acids, which strongly suggests a Lipid A–type structure\. Guided by these anchors,E\-Agentretrieves evidence linking Lipid A to LPS and therefore to Gram\-negative organisms, which is sufficient to eliminate the Gram\-positive distractors\.

The failure arises when moving from class identification to option selection\. The retrieved evidence supports ruling out the Gram\-positive options, but it does not provide*option\-separating*signals within the remaining Gram\-negative candidates\. When the evidence report lacks an explicit bridge from the clinical setting to the discriminative phenotype expected in blood culture,A\-Agentcan be pulled toward salient surface descriptors and commit to an unsupported candidate\.

This exposes a limitation of the current sufficiency and adjudication design\. The loop may stop once it reaches a correct coarse conclusion, even though an additional round is still needed to uniquely determine the answer option\. A practical implication is that sufficiency should be judged against*option\-level separability*, not only against class\-level plausibility\. A simple fix is to tighten E\-Agent’s stopping rule by requiring evidence that supports one remaining option while directly excluding the other plausible candidates\. If this condition is not met, the system should issue a final follow\-up query targeting discriminative attributes and then re\-adjudicate\. We leave such option\-aware sufficiency calibration to future work\.

## Appendix EAdditional Implementation Details

##### Computational Resources

All experiments were executed in an API\-based inference setting\. The dense retrieval index was constructed and queried locally, while all LLM calls were served by the corresponding model providers\.

##### Retriever Settings

We use the MedCPT Query Encoder and Article Encoder to embed queries and passages, respectively\. We then perform FAISS\-based dense retrieval over Textbooks and StatPearls and globally rank the retrieved candidates\.

## Appendix FAdditional Robustness and Transfer Experiments

### F\.1Retriever Robustness

To examine whether SEMA\-RAG depends strongly on a specific retriever, we further compare the domain\-specific retriever MedCPT with the general\-purpose retrieverqwen3\-embedding\-4bZhanget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib65)\)on MedQA\-US, using deepseek\-v3\.1 as the backbone\. Table[8](https://arxiv.org/html/2605.17101#A6.T8)shows that SEMA\-RAG consistently improves over the corresponding single\-round retrieval baseline under both retrievers\. Meanwhile, MedCPT remains the stronger default choice in this clinical setting, suggesting that domain\-specific retrieval is still advantageous for medical QA\.

Methodqwen3\-embedding\-4bMedCPTCoT77\.5377\.53MedRAG76\.4377\.14SEMA\-RAG87\.4389\.95Table 8:Accuracy \(%\) on MedQA\-US with different retrievers \(deepseek\-v3\.1\)\.
### F\.2Smaller\-Model Robustness

To assess whether the gains of SEMA\-RAG rely mainly on strong large\-scale backbones, we further evaluate the framework on MedQA\-US using a much smaller open\-source model,gemma3:4bTeamet al\.\([2025a](https://arxiv.org/html/2605.17101#bib.bib66)\)\. All other settings remain the same as in the main experiments\. Table[9](https://arxiv.org/html/2605.17101#A6.T9)shows that although the absolute performance of all methods drops on the smaller model, SEMA\-RAG still maintains a clear advantage over the strongest single\-round baseline\. This result suggests that task decoupling remains beneficial even when the base model has weaker instruction\-following ability\.

MethodAcc\. \(%\)CoT51\.77MedRAG56\.01SEMA\-RAG60\.41Table 9:Results on MedQA\-US using gemma3:4b as the backbone\.
### F\.3Beyond Discrete QA: Open\-Ended and Interactive Settings

To examine whether SEMA\-RAG generalizes beyond discrete\-choice medical QA, we further evaluate it on two benchmarks covering open\-ended generation and multi\-turn medical dialogue\. In these experiments, we keep the I\-Agent and E\-Agent unchanged, and only adapt the final stage of the A\-Agent to generate free\-text outputs instead of selecting a discrete option\. This setting isolates the contribution of the evidence loop under more open\-ended output formats\.

##### HealthBench\.

HealthBench is an open\-ended health\-domain benchmark designed to assess response quality under clinician\-authored rubricsAroraet al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib67)\)\. To test whether SEMA\-RAG can transfer to grounded free\-text generation with minimal modification, we evaluate deepseek\-v3\.1 on 500 randomly sampled English questions from HealthBench main using the official scoring script, and report the overall score together with three rubric dimensions: accuracy, completeness, and instruction following\.

As shown in Table[10](https://arxiv.org/html/2605.17101#A6.T10), SEMA\-RAG consistently outperforms MedRAG across all reported metrics\. This suggests that the evidence completion loop remains effective when the target output shifts from option selection to grounded free\-text generation, providing a stronger factual basis for open\-ended responses\.

MethodAvg\.Acc\.Comp\.Instr\.Follow\.MedRAG26\.8731\.3430\.1741\.66SEMA\-RAG33\.6437\.2238\.3047\.58Table 10:Results on HealthBench for grounded open\-ended response generation using deepseek\-v3\.1\.
##### MAQuE\.

MAQuE is a multi\-turn medical dialogue benchmark that evaluates response quality in simulated clinical communication settingsGonget al\.\([2025](https://arxiv.org/html/2605.17101#bib.bib68)\)\. Compared with single\-turn QA, it places greater emphasis on maintaining robust, relevant, and contextually appropriate responses across iterative interactions\.

To test whether SEMA\-RAG remains effective in interactive settings, we evaluate deepseek\-v3\.1 on 200 randomly sampled MAQuE test cases using the official evaluation script, and report four communication\-oriented metrics: Accuracy, Robustness, Relevance, and Empathy\. Table[11](https://arxiv.org/html/2605.17101#A6.T11)shows that SEMA\-RAG maintains clear gains over MedRAG on all four metrics\. This improvement suggests that sufficiency\-driven exploration remains beneficial in interactive clinical scenarios, where the system must sustain grounded response generation over multiple turns\.

MethodAcc\.Rob\.Rel\.Emp\.MedRAG52\.5064\.8674\.0066\.40SEMA\-RAG61\.5075\.3882\.0072\.00Table 11:Results on MAQuE for multi\-turn clinical response generation using deepseek\-v3\.1\.Overall, these results suggest that the SEMA\-RAG framework generalizes beyond discrete medical QA to more open\-ended and interactive forms of clinical response generation\.

## Appendix GPrompt Templates

To facilitate reproducibility, we provide the system instructions used for the three agents in SEMA\-RAG\. Note that the A\-Agent uses two prompts for evidence adjudication and final answer selection, respectively\.

I\-Agent Prompt \(Clinical Schema Interpreter\)\.Role: You are an expert clinician\.

Goal: Given an unstructured medical question, extract an explicit Clinical Schema that makes the implied intent and constraints searchable\. Focus on what must be retrieved and do not answer the question itself\.

Input: Medical Question: \{research\_topic\}

Task: Identify: 1\. clinical intent \(task type\), 2\. core medical entities \(salient concepts from the question\), 3\. key constraints \(time course, demographics, setting, comorbidities, severity, contraindications, risk factors, anatomical or functional qualifiers\), 4\. a concise retrieval query aligned with the schema \(q\_init\)\.

Key Instructions: \- Entities should be small, focused, and primarily grounded in the question itself\. \- Include only the most retrieval\-relevant concepts; avoid broad, redundant, or unnecessary enumeration\. \- Merge obvious synonyms, near\-duplicates, or simple morphological variants into one canonical medical expression when possible\. \- Prefer the main clinical concept, condition, mechanism, finding, test, treatment, population, anatomical target, or other decision\-critical concept that is necessary for retrieval\. \- If the question provides candidate answers or options, do not mechanically include all of them as entities; include an option only if needed for retrieval or candidate discrimination\. \- For multiple\-choice, judgment, or open\-ended questions, center the schema on the stem and its decision\-critical medical concepts rather than listing answer choices\. \- Constraints should capture only decision\-relevant qualifiers explicitly stated or strongly implied by the question\. \- Preserve key medical relations when they are essential for retrieval, such as derivation, origin, cause, association, indication, or contraindication\. \- q\_init should retrieve the knowledge needed to answer the question, remain neutral, and avoid prematurely inferring a conclusion\. \- q\_init should be short, medically precise, and should not simply concatenate all entities or options\. \- Use precise medical terminology\. \- Do not add explanations, rationale, or extra keys\.

Output JSON: \{ "intent": "<short clinical task type\>", "entities": \["<entity1\>", "<entity2\>"\], "constraints": \["<constraint1\>", "<constraint2\>"\], "q\_init": "<one concise neutral search\-style query\>" \}

Figure 4:Prompt template for the I\-Agent clinical schema interpreter\.E\-Agent Prompt \(Self\-Evolving Explorer\)\.Role: You are an evidence sufficiency auditor and query refiner for medical question answering\.

Goal: Determine whether the current retrieved evidence is sufficient to answer the medical question under the given Clinical Schema\. Do not answer the question itself\.

Input: Clinical Schema: \{clinical\_schema\} Current Query Set: \{query\_list\} Retrieved Evidence Summaries: \{summaries\}

Key Instructions: \- Assess whether the current evidence sufficiently covers the key intent, entities, and constraints in the Clinical Schema\. \- Judge sufficiency based on whether the evidence is enough to support final answer selection, or to distinguish among competing candidate answers when relevant\. \- Evidence may be relevant yet still insufficient; do not mark sufficiency = 1 unless the evidence is adequate for confident answer selection\. \- If the evidence is insufficient, identify the single most important missing fact, missing distinction, or unresolved clinical criterion\. \- Generate 1 to 3 follow\-up queries that directly target this gap\. \- Follow\-up queries must be specific, self\-contained, non\-redundant, and explicitly grounded in the Clinical Schema\. \- For questions with candidate answers, prioritize queries that help distinguish among candidates rather than broad background expansion\. \- Prefer targeted refinement over broad exploratory expansion\. \- Do not repeat an existing query unless revision is necessary\. \- If the current evidence is already sufficient, return no follow\-up queries\.

Rules: \- If sufficiency = 1, set "gap" to "N/A" and "queries" to \[\]\. \- If sufficiency = 0, "gap" must be specific, concrete, and decision\-relevant rather than generic\. \- Queries should target missing clinical distinctions, time conditions, population constraints, contraindications, severity, mechanisms, diagnostic criteria, or option\-level discrimination when relevant\. \- Return JSON only\.

Output JSON: \{ "sufficiency": 0 or 1, "gap": "<short concrete description of the most important missing evidence\>", "queries": \["<query1\>", "<query2\>", "<query3\>"\] \}

Figure 5:Prompt template for the E\-Agent self\-evolving explorer\.A\-Agent Prompt \(Phase 1: Evidence Adjudicator\)\.Role: You are a medical evidence adjudicator\.

Goal: Synthesize the final retrieved evidence into a concise, traceable report that can support final answer selection\. Do not directly answer the question\. Only organize, adjudicate, and summarize the evidence\.

Input: Medical Question: \{research\_topic\} Clinical Schema: \{clinical\_schema\} Final Query Set: \{query\_list\} Retrieved Evidence Summaries: \{summaries\}

Key Instructions: \- Review the retrieved evidence in light of the medical question and Clinical Schema\. \- Focus on the most decision\-relevant evidence and remove redundancy\. \- Identify which evidence directly supports a candidate conclusion, which evidence conflicts with it, and which evidence is only background, indirect, or weakly relevant\. \- When multiple pieces of evidence overlap, merge them into one concise statement\. \- When evidence is incomplete, uncertain, indirect, or conflicting, make that explicit rather than resolving it prematurely\. \- Preserve traceability by attaching source identifiers or summary indices whenever available\. \- Every claim in the report must be supported by the provided summaries; do not infer unsupported medical facts\. \- Do not introduce external medical knowledge\. \- Do not perform final answer selection\.

Figure 6:Prompt template for the A\-Agent evidence adjudicator\.A\-Agent Prompt \(Phase 1: Evidence Adjudicator\) \(continued\)\.Rules: \- Keep the report concise, traceable, and decision\-oriented\. \- Prefer evidence that is directly relevant to the question over general background knowledge\. \- If there is no real conflicting evidence, return an empty list for "key\_conflicting\_or\_limiting\_evidence"\. \- If source identifiers are unavailable, use summary indices or short summary labels consistently\. \- Do not repeat the same evidence across multiple fields unless necessary\. \- Return JSON only\.

Output JSON: \{ "question\_focus": "<one short sentence stating what must be decided\>", "key\_supporting\_evidence": \[ \{ "claim": "<concise evidence\-supported statement\>", "source\_ids": \["<source1\>", "<source2\>"\] \} \], "key\_conflicting\_or\_limiting\_evidence": \[ \{ "claim": "<concise conflicting, uncertain, or limiting statement\>", "source\_ids": \["<source1\>", "<source2\>"\] \} \], "evidence\_synthesis": "<short integrated synthesis of what the evidence supports, what remains uncertain, and what distinction matters most for final answer selection\>" \}

Figure 7:Prompt template for the A\-Agent evidence adjudicator\.A\-Agent Prompt \(Phase 2: Evidence\-Grounded Answering\)\.Role: You are a medical AI assistant\.

Goal: Answer the multiple\-choice medical question using the provided evidence adjudication report\.

Input: Medical Question: \{research\_topic\} Evidence Adjudication Report: \{adjudication\_report\}

Key Instructions: \- Select exactly one final answer: A, B, C, or D\. \- First rely on the evidence adjudication report\. \- If the report contains relevant evidence, choose the option best supported by that evidence\. \- If the report is incomplete, weak, or lacks directly relevant evidence, use medical knowledge to reason and choose the most appropriate answer\. \- Do not output reasoning, JSON, code blocks, or any extra text\.

Output Format: Final Answer: \[A/B/C/D\]

Figure 8:Prompt template for the A\-Agent evidence\-grounded answerer\.