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

Conference: ACL 2026 Findings
arXiv: 2605.17101
Code: None
Area: Medical NLP
Keywords: Medical QA, Multi-agent RAG, Self-evolving retrieval, Evidence chain construction, Clinical reasoning

TL;DR

The authors propose SEMA-RAG, a self-evolving multi-agent Retrival-Augmented Generation framework. By simulating phased clinical reasoning via three specialized agents (Interpreter, Explorer, Arbiter), it outperforms the strongest baselines by an average of +6.46 accuracy points across 5 medical QA benchmarks.

Background & Motivation

Background: RAG is widely used to mitigate LLM hallucinations and outdated knowledge in medical QA; however, existing RAG methods primarily adopt a single-turn static retrieval paradigm.

Limitations of Prior Work: (1) The transformation from question to query lacks clinical semantic interpretation, making it difficult to explicitize implicit constraints; (2) Retrieval lacks a sufficiency feedback mechanism, hindering the formation of reliable evidence chains; (3) Coupling three heterogeneous tasks—interpretation, exploration, and adjudication—within a single reasoning chain imposes excessive cognitive load.

Key Challenge: Single-turn static RAG effectively asks a clinician to simultaneously analyze, retrieve, evaluate, and diagnose immediately upon receiving an initial case file, failing to adjust reasoning based on new evidence—a significant mismatch with the multi-stage clinical reasoning process.

Goal: Reconstruct the RAG workflow to match phased clinical reasoning: expanding single-turn queries into multi-turn iterative exploration, evaluating evidence sufficiency after each retrieval turn to decide the next action.

Key Insight: Task decoupling + role specialization—assigning interpretation, exploration, and adjudication to three collaborating specialized agents.

Core Idea: A three-agent division of labor (I-Agent interprets \(\rightarrow\) E-Agent performs sufficiency-driven self-evolving retrieval \(\rightarrow\) A-Agent performs evidence arbitration), enhancing the reliability of medical RAG through closed-loop evidence chain construction.

Method

Overall Architecture

SEMA-RAG consists of three agent roles sharing the same underlying LLM, distinguished solely by role prompts: (1) I-Agent maps the raw question to a structured clinical schema; (2) E-Agent accumulates evidence turn-by-turn through a sufficiency-driven self-evolving retrieval loop; (3) A-Agent arbitrates the converged evidence set and outputs the final answer.

graph TD
    A["Medical Question"] --> B["I-Agent Question Interpreter<br/>Maps to clinical schema Q′<br/>Intent / Entities / Constraints / Initial Query"]
    B --> C["E-Agent Knowledge Explorer<br/>MedCPT Dense Retrieval (Turn t)"]
    C --> D{"Evidence Sufficiency Flag s_t"}
    D -->|"s_t=0 & T < T_max: Locate gap g_t, generate m follow-up queries"| C
    D -->|"s_t=1 or T = T_max"| E["Closed Evidence Set C*"]
    E --> F["A-Agent Evidence Arbiter<br/>Denoising/Hedging, organize structured report R"]
    F --> G["Final Answer"]

Key Designs

1. I-Agent (Question Interpreter): Mapping questions to structured clinical schemas prior to retrieval

Critical constraints in medical questions are often implicit—"7th day of hospitalization" implies a nosocomial infection. Using the raw sentence for retrieval often misses this semantic layer. The I-Agent maps unstructured questions into a schema tuple \(Q' = \langle o_{\text{int}}, o_{\text{ent}}, o_{\text{cons}}, q_{\text{init}} \rangle\), representing clinical intent, medical entities, clinical constraints, and initial retrieval queries. By explicitizing implicit constraints, subsequent retrieval turns have clear alignment targets.

2. E-Agent (Knowledge Explorer): Multi-turn self-evolving retrieval driven by evidence sufficiency

Single-turn static retrieval cannot guarantee coverage of all key constraints. E-Agent treats retrieval as a closed loop: after each turn (using MedCPT for dense retrieval), it evaluates a sufficiency flag \(s_t \in \{0,1\}\). If \(s_t=0\) (insufficient), it identifies the evidence gap \(g_t\) and generates \(m\) targeted follow-up queries \(\mathcal{Q}_{t+1}\). The process terminates when \(s_t=1\) or \(T_{\max}\) is reached, yielding the closed evidence set \(C^*\).

This "gap identification \(\rightarrow\) targeted follow-up" early-stopping mechanism is the core performance driver—removing E-Agent leads to a 6.37 point drop on MedQA-US, the largest among the three agents. It is also more efficient than fixed-turn iterations, gathering sufficient evidence with fewer tokens rather than exhausting a budget and potentially introducing noise.

3. A-Agent (Evidence Arbiter): Denoising and hedging converged evidence for traceable judgment

Evidence accumulated via multi-turn retrieval is often redundant or contradictory. The A-Agent serves as an arbitrator: it deduplicates and denoises the evidence set, identifies consistencies and conflicts, and organizes supporting/refuting clues into a structured report \(R\). Finally, it performs discrete answer selection: \(\tilde{y} = \text{Agent}_A(\text{Pmt}_{\text{ans}}, [Q, R])\). Decoupling evidence integration into a separate role provides the model with a stable basis for judgment.

Case Example: Workflow for a "Fever on Hospitalization Day 7" question

For a question with implicit nosocomial infection clues: The I-Agent first parses it into a schema—Intent: "Differential Diagnosis," Entities: "Fever + Day 7 Hospitalization," Constraints explicitly labeled: "nosocomial." Based on this, it generates \(q_{\text{init}}\). In Turn 1, the E-Agent determines \(s_1=0\) because retrieved evidence only covers community-acquired infections. The gap \(g_1\) is identified as "nosocomial pathogen spectrum and catheter-related infections," prompting \(m=3\) follow-up queries. Turn 2 achieves \(s_2=1\) as targeted evidence is found. The A-Agent then denoises this set, hedges against community-acquired distractors, organizes report \(R\), and selects the answer. The process averages 4.8 LLM calls, 3.4 retrievals, and 9.5s latency, consuming 19,488 tokens—more efficient and 15.17% more accurate than i-MedRAG's fixed 3-turn approach.

Loss & Training

• Training-free: The three agents share the same base LLM and are distinguished via role prompting.
• Default Hyperparameters: \(T_{\max}=2\), \(k=16\) (Top-k retrieval), \(m=3\) (follow-up queries per turn).
• Temperature: 1.0 for I/E-Agent, 0.0 for A-Agent (deterministic output).

Key Experimental Results

Main Results (5 Benchmarks × 5 LLM Backbones, Accuracy %)

Method	MMLU-Med	MedQA-US	MedMCQA	PubMedQA*	BioASQ	Average
deepseek-v3.1 + CoT	88.15	77.53	71.69	38.40	80.10	71.17
deepseek-v3.1 + MedRAG	88.61	77.14	67.99	44.60	78.48	71.36
deepseek-v3.1 + i-MedRAG	85.86	74.78	65.65	50.60	80.58	71.49
deepseek-v3.1 + SEMA-RAG	91.46	89.95	75.09	59.20	82.85	79.71
gemini-2.0-flash + CoT	58.22	65.12	41.33	40.20	68.45	54.66
gemini-2.0-flash + SEMA-RAG	80.99	90.42	71.60	59.20	88.19	78.08

Ablation Study (deepseek-v3.1, MedQA-US / PubMedQA*)

Configuration	MedQA-US	PubMedQA*
w/o I-Agent	85.47	54.20
w/o E-Agent	83.58	50.80
w/o A-Agent	86.49	53.60
Full SEMA-RAG	89.95	59.20

Key Findings

• Removing the E-Agent leads to the most significant performance degradation (-6.37 on MedQA-US), confirming self-evolving retrieval as the core benefit.
• Query width \(m\): \(m=1\) (86.72%), \(m=2\) (89.00%), \(m=3\) (89.95%); gains show diminishing returns.
• Exploration depth \(T_{\max}\): Optimal at 2-3 turns; exceeding this may introduce noise.
• Efficiency: SEMA-RAG averages 4.8 LLM calls / 3.4 retrievals / 9.5s latency and 19,488 tokens—fewer tokens than i-MedRAG (21,517) with a +15.17% accuracy improvement.

Highlights & Insights

• The multi-agent architecture precisely simulates the phased workflow of clinical reasoning (Interpretation → Exploration → Adjudication), demonstrating the universality of task decoupling.
• The sufficiency-driven early stopping mechanism is more efficient than fixed-turn iterations, achieving higher accuracy with fewer tokens.
• Significant improvements on Gemini-2.0-flash (+23.42 average) suggest the framework provides stronger enhancement for relatively weaker models.
• Case studies vividly demonstrate how the "structured interpretation → gap identification → targeted retrieval" sequence forms a reliable evidence chain.

Limitations & Future Work

• Evaluation is limited to benchmark environments and lacks validation in real clinical workflows (e.g., longitudinal EHR reasoning).
• The framework depends on the quality and coverage of the retrieval corpus; self-evolution may converge on incomplete evidence if key knowledge is missing.
• Sufficiency judgment criteria are not yet optimized for option-level separability or generative completeness.
• Additional overhead from multi-turn reasoning is superior to fixed-step baselines but remains higher than single-turn methods.

Related Work & Insights

• MedRAG / MedCPT provide the retrieval foundation for medical domains, while SEMA-RAG builds a multi-turn closed loop upon them.
• SEMA-RAG improves upon i-MedRAG's iterative approach by introducing a sufficiency feedback mechanism.
• Multi-agent collaboration (CAMEL / MetaGPT / MedAgents) principles can be extended to other high-stakes domains requiring multi-stage reasoning.
• The "gap detection + targeted follow-up" pattern in self-evolving retrieval can inspire complex RAG system designs outside of medicine.

Rating

• Novelty: ⭐⭐⭐⭐ Task decoupling + sufficiency-driven self-evolving retrieval are clear innovations.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ 5 benchmarks × 5 LLM backbones + complete ablation + efficiency analysis + case studies.
• Writing Quality: ⭐⭐⭐⭐ Rigorous formalization and intuitive clinical analogies.
• Value: ⭐⭐⭐⭐⭐ Significant and consistent improvements in medical QA; framework logic is broadly applicable.

Related Papers

• [ACL 2026] HeteroRAG: A Heterogeneous Retrieval-Augmented Generation Framework for Medical Vision Language Tasks
• [ACL 2026] MultiDx: A Multi-Source Knowledge Integration Framework towards Diagnostic Reasoning
• [ACL 2025] Towards Omni-RAG: Comprehensive Retrieval-Augmented Generation for Large Language Models in Medical Applications
• [ACL 2026] MARCH: Multi-Agent Radiology Clinical Hierarchy for CT Report Generation
• [ACL 2026] RA-RRG: Multimodal Retrieval-Augmented Radiology Report Generation with Key Phrase Extraction