DiagAgent: Autonomous Diagnostic Agents

Updated 3 July 2026

DiagAgent are a class of autonomous diagnostic agents that perform complex medical reasoning and multi-turn clinical dialogues using LLMs and reinforcement learning paradigms.
They integrate techniques such as Bayesian inference, mutual information heuristics, and multimodal data fusion to dynamically refine diagnostic hypotheses and improve decision accuracy.
Their design prioritizes interpretability and traceability, enabling robust failure diagnosis, specialist collaboration, and continuous system adaptation in clinical settings.

DiagAgent refers to a class of autonomous diagnostic agents designed to conduct complex medical reasoning, multi-stage hypothesis refinement, and evidence-driven decision-making, typically leveraging LLMs or vision-LLMs (VLMs) as core policy engines. DiagAgent frameworks have been developed in both classic dialogue-based settings and modern multimodal, multi-agent, and reinforcement learning (RL) paradigms. They aim to emulate physician-like diagnostic strategies, manage multi-turn clinical dialogues, interactively query and incorporate diverse data modalities, and—crucially—offer interpretable, auditable rationales and root-cause diagnostics. Below is a comprehensive technical synopsis of the most prominent DiagAgent designs, methodological innovations, representative benchmarks, and directions for future development.

1. Core Architectures and Functional Paradigms

DiagAgent implementations fall broadly into three categories: interpretable dialogue managers with doctor-inspired inquiry logic, interactive RL-based agents trained in virtual clinical environments, and tool-augmented or multimodal diagnostic systems that orchestrate external knowledge, domain-specific tools, and sub-agents.

Dialogue Manager-Based: Systems such as BR-Agent formalize the diagnostic process as a Markov Decision Process (MDP) where each state encodes observed symptom vectors and each action is either a targeted inquiry (ask symptom) or a commit action (diagnose disease). Probabilistic inference (Bayesian networks), mutual-information heuristics, and logic-switching actuate human-like inquiry strategies, with clear two-stage flows separating symptom gathering from hypothesis refinement (Liu et al., 2022).
Reinforcement Learning and Virtual Environments: RL-based DiagAgents (e.g., DoctorAgent-RL, DiagAgent trained in DiagGym) model consultations as episodic MDPs over sequences of patient-agent exchanges. The learning objective maximizes long-horizon reward, decomposed into diagnostic accuracy, information efficiency, and protocol compliance, with policies learned via advanced methods like Group Relative Policy Optimization and validated in high-fidelity simulators or large multi-turn datasets (Feng et al., 26 May 2025, Qiu et al., 28 Oct 2025).
Tool-Augmented and Modular Multi-Agent Systems: Modern DiagAgents integrate LLM controllers with clinically validated diagnostic tools (e.g., CheXStruct for chest X-ray quantification; specialist sub-agents for cardiology) and orchestrate specialist or generalist reasoning chains. Modular systems such as HeartAgent sequence multiple sub-agents for specialist differential diagnosis, reviewer cross-validation, and reference verification, assembling structured rationales and explicit evidence chains for each output (Zhou et al., 11 Mar 2026, Lee et al., 26 Feb 2026).

2. Decision Logic, Reasoning, and Interpretability

A cornerstone of DiagAgent research is transparent, auditable reasoning, ensuring that each decision or recommendation can be decomposed, interrogated, and traced. Key features include:

Interpretable Scoring and Decision Rules: Bayesian posteriors over diseases, mutual information matrices for symptom value, and logic-switcher mechanisms (shallow MLPs outputting a logic-mixture scalar) yield conceptually transparent next-action strategies. RL-based agents map explicit reward functions to learning targets, while chain-of-thought (CoT) prompting and summary-generation enforce stepwise evidence logging (Liu et al., 2022, Qiu et al., 28 Oct 2025).
Multi-Agent Collaboration and Division of Reasoning Labor: Systems such as MeDxAgent and HeartAgent coordinate specialist and generalist sub-agents, utilize knowledge-graph-based candidate expansion, and consolidate outputs via selector modules—emulating interdisciplinary team-based care (Zhou et al., 11 Mar 2026, Sanghvi et al., 2 Jun 2026).
Traceability and Reference Verification: Diagnostic outputs are coupled with reasoning trajectories (minichains-of-thought), pointers to prior cases (episodic memory), explicit procedural heuristics, and, in some cases, automated passage retrieval and citation (e.g., BM25 + MedCPT reranking followed by LLM-based factual check) (Zhou et al., 11 Mar 2026).

3. Learning Algorithms and Data Regimens

DiagAgents leverage diverse training regimens depending on system complexity and available data.

Supervised and Policy-Gradient RL: Classic agents use advantage actor-critic or PPO/GRPO to optimize policy networks, often initialized by supervised fine-tuning on synthetic or expert-annotated dialogues and then further refined through self-play or simulated patient interaction (Feng et al., 26 May 2025, Qiu et al., 28 Oct 2025).
Interactive Simulated Environments: DiagGym (for DiagAgent) and MTMedDialog (for DoctorAgent-RL) provide patient simulators that yield examination outcomes and test trajectories, permitting large-scale RL without human-in-the-loop cost (Qiu et al., 28 Oct 2025, Feng et al., 26 May 2025).
Reflection, Memory, and Self-Evolution: Evo-MedAgent introduces a three-store memory (episodic, procedural, tool reliability), updated at test time via retrieval, LLM-based reflection, and post-hoc scoring. No additional fine-tuning is needed; policy refinement is performed online via prompt augmentation (Shen et al., 15 Apr 2026).
Dataset Expansion and Benchmarking: GMD-12, MTMedDialog, MeDxBench, DiagBench, and ChestAgentBench are representative large-scale diagnostic benchmarks, each with different specialties, domain coverage, and annotation detail (Liu et al., 2022, Feng et al., 26 May 2025, Qiu et al., 28 Oct 2025, Sanghvi et al., 2 Jun 2026, Shen et al., 15 Apr 2026).

4. Empirical Performance and Comparative Results

DiagAgents consistently outperform static, single-shot, and non-interactive LLM baselines on both traditional and multi-turn benchmarks. The following table summarizes representative quantitative results from recent studies (absolute top-1 or composite accuracy, unless otherwise stated):

System	Benchmark	DiagAgent (%)	Best Baseline (%)	ΔAccuracy
Evo-MedAgent	ChestAgentBench MCQ	79	68	+11
DoctorAgent-RL	MTMedDialog	58.9	52.6	+6.3
DiagAgent	DiagBench E2E	61.3	47.1	+14.2
HeartAgent	MIMIC (top-3)	60.0	44.0	+16
MeDxAgent	MeDxBench (top-1)	57.4	47.1	+10.3
CXReasonAgent	CXReasonDial Faith	99.2	43–58 (LVLM)	+41–56

Quiz-style metrics (MCQ, rubric adherence), chain-of-thought faithfulness, strict dialogue success, and rationale quality also see consistent gains over instruction-tuned and prompt-engineered LLMs (Shen et al., 15 Apr 2026, Feng et al., 26 May 2025, Qiu et al., 28 Oct 2025, Zhou et al., 11 Mar 2026, Sanghvi et al., 2 Jun 2026, Lee et al., 26 Feb 2026).

5. Auditability, Failure Diagnosis, and Safety

Auditability and robust root-cause analysis are fundamental requirements for clinical and agentic deployment.

Constraint-Based Failure Attribution: AgentRx automatically diagnoses failed agent trajectories by synthesizing global and dynamic constraints, generating detailed violation logs, and using an LLM-based judge to localize both step-index and failure category from a nine-way taxonomy (e.g., invention of new information, plan-adherence failure, misinterpretation of tool output) (Barke et al., 2 Feb 2026).
Safety Guardrails and Provenance: AgentDoG introduces a three-dimensional risk taxonomy (source, failure mode, harm) and diagnostic guardrail models, achieving over 92% trajectory-level safety accuracy and providing stepwise and sentence-level attribution scores for each decision (Liu et al., 26 Jan 2026).
Transparent Rationales and Reference Verification: Sub-agent outputs, explicit scoring matrices, and chain-of-thoughts support precise auditing, while reference verification agents attach evidence to each diagnostic justification (Zhou et al., 11 Mar 2026).

6. Adaptation, Generalization, and Modular Expansion

DiagAgents are designed for extensibility:

Rapid Tooling and Specialty Expansion: Modular tool interfaces (e.g., CheXStruct, knowledge bases, imaging analyzers) permit seamless addition of new diagnostic tasks or integration of multimodal inputs without retraining the core LLM (Lee et al., 26 Feb 2026, Zhou et al., 11 Mar 2026).
Interactive Adaptation: Memory-augmented agents and reflection/heuristics update mechanisms support adaptation to evolving clinical distributions and continual improvement without dataset retraining (Shen et al., 15 Apr 2026).
Multi-Agent Orchestration: Flows such as MeDxAgent’s division of labor (summarizer, demographic collector, candidate generator, evidence-gap agent) enable hybridization of specialist and generalist strategies and evidence-driven information gathering (Sanghvi et al., 2 Jun 2026).

7. Limitations and Future Directions

Remaining limitations include:

Patient Simulation Fidelity: Most systems simulate patient responses; real-world performance may differ absent live trials (Qiu et al., 28 Oct 2025, Feng et al., 26 May 2025).
Data Modalities: Many DiagAgents are text- or image-centric; structured labs, EHR, and cross-modality fusion remain underexplored.
Failure Category Coverage: Failure diagnosis taxonomies, while robust, may require expansion for new domains (e.g., pediatrics, open-world, robotics) (Barke et al., 2 Feb 2026, Liu et al., 26 Jan 2026).
Interpretability in High-Complexity Trajectories: As system complexity and length scale, maintaining actionable interpretability and auditability poses ongoing challenges.

Promising directions include integrating online patient interaction, multimodal fusion, recurrent memory traces, formal logic constraints, and specialized safety/guardrail models for “DiagAgent” safety, transparency, and regulatory compliance.

For further detail, see (Liu et al., 2022, Feng et al., 26 May 2025, Qiu et al., 28 Oct 2025, Barke et al., 2 Feb 2026, Zhou et al., 11 Mar 2026, Shen et al., 15 Apr 2026, Liu et al., 26 Jan 2026, Lee et al., 26 Feb 2026), and (Sanghvi et al., 2 Jun 2026).