💬 NLP

ClinicalAgents: Multi-Agent Orchestration for Clinical Decision Making with Dual-Memory

The paper introduces ClinicalAgents, a novel multi-agent framework that leverages a Dual-Memory architecture and Monte Carlo Tree Search-based dynamic orchestration to simulate expert clinicians' iterative reasoning, thereby significantly improving diagnostic accuracy and explainability over existing static methods.

Original authors: Zhuohan Ge, Haoyang Li, Yubo Wang, Nicole Hu, Chen Jason Zhang, Qing Li

Published 2026-05-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhuohan Ge, Haoyang Li, Yubo Wang, Nicole Hu, Chen Jason Zhang, Qing Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a very tricky mystery, like diagnosing a patient with a complex set of symptoms. In the real world, a human doctor doesn't just look at the symptoms and guess the answer immediately. Instead, they act like detectives: they form a theory, look for clues to prove or disprove it, realize they are missing a piece of evidence, go back and ask for a new test, and then revise their theory.

The paper introduces ClinicalAgents, a new computer system designed to mimic this exact human detective work, rather than just acting like a super-fast encyclopedia.

Here is how it works, broken down into simple concepts:

1. The Problem: The "One-Note" vs. The "Detective"

Most current AI models (Large Language Models) are like one-note geniuses. If you give them a medical case, they try to read the whole story and spit out an answer in one go. If they miss a tiny detail or get stuck on a wrong idea, they often keep going down that wrong path, leading to a wrong diagnosis. They are rigid; they don't know how to say, "Wait, I need more info," and go back.

ClinicalAgents is different. It acts like a team of specialized detectives working together.

2. The Team: A "Conductor" and Specialists

Instead of one AI doing everything, this system uses a Multi-Agent approach. Think of it as a hospital ward with different specialists:

One agent looks at medical images (like an X-ray).
One agent suggests which tests to order.
One agent looks at the patient's history.
One agent formulates the final diagnosis.

But a team needs a leader. This is where the Clinical Orchestrator comes in. It's the Conductor of the orchestra. It doesn't just tell everyone what to do in a fixed order (like a script). Instead, it listens to the team, decides who needs to speak next, and changes the plan on the fly.

3. The Secret Sauce: "Dual Memory"

To keep the team from getting confused or forgetting things, the system uses two types of memory, like a detective's notebook and a library:

Working Memory (The Active Notebook): This is a mutable, changing record of the current patient. It holds the symptoms, the tests done so far, and the current theories. As the investigation progresses, this notebook gets updated in real-time so every agent knows exactly where they stand.
Experience Memory (The Library): This is a static library of medical guidelines and past cases. When the team is stuck, the Conductor can pull a relevant rulebook or a similar past case from this library to help guide the investigation.

4. The Magic Trick: "Backtracking" (The "Wait, Let's Go Back" Button)

This is the most important part. In many AI systems, if you make a mistake early on, the whole process fails. ClinicalAgents has a Backtracking Mechanism.

Imagine the Conductor is leading the team. At any point, if the team realizes, "Hey, we are missing a crucial clue (like a specific blood test result) to prove our theory," the system doesn't force a guess. Instead, it hits the Backtrack button.

It stops the current flow.
It goes back to the "Perceive" stage.
It asks for the missing information.
Once the new info arrives, it updates the "Notebook" and starts the reasoning process again with the new facts.

The paper describes this decision-making process using a method called Monte Carlo Tree Search (MCTS). Think of this as the Conductor mentally simulating different paths: "If I ask for Test A, what happens? If I ask for Test B, what happens?" It chooses the path that is most likely to lead to the correct answer, rather than just following a pre-written script.

5. The Results: Better than the Rest

The researchers tested this system against other AI models (both single AI brains and other AI teams) using a standard medical benchmark called MedChain.

The Outcome: ClinicalAgents performed better than all the others.
Why? It was better at ordering the right tests, interpreting images, and finally diagnosing the patient correctly.
Efficiency: Interestingly, it didn't just win by using more computer power or reading more text. It won because it was smarter about how it used its resources. It knew exactly when to stop and ask for more info, saving time and money compared to other methods that just kept guessing.

Summary

In short, ClinicalAgents is a smart, team-based AI system that doesn't just "guess" a medical diagnosis. It acts like a human doctor: it forms a hypothesis, checks the evidence, realizes when it's missing a piece of the puzzle, goes back to get that piece, and then updates its conclusion. By using a "Conductor" to manage the team and two types of memory to stay organized, it solves medical mysteries more accurately and reliably than previous AI methods.

Technical Summary: ClinicalAgents

Problem Statement

Clinical decision-making (CDM) is a complex, non-linear process requiring physicians to iteratively collect evidence, formulate hypotheses, and refine diagnoses based on evolving patient data. While Large Language Models (LLMs) have shown promise in healthcare, they often struggle with the rigorous, hypothesis-driven reasoning required for accurate diagnosis. Existing approaches typically rely on static, linear mappings from symptoms to diagnoses or pre-scripted multi-agent workflows. These methods fail to capture the iterative nature of clinical reasoning, lack mechanisms for dynamic replanning or backtracking when evidence is insufficient, and often treat memory as a window-bounded dialogue context, leading to information fragmentation and inconsistent intermediate states. Consequently, current systems may settle on premature diagnoses even when critical evidence is missing, increasing the risk of misdiagnosis.

Methodology

The paper proposes ClinicalAgents, a novel multi-agent framework designed to simulate the cognitive workflow of expert clinicians. The system is built upon three core components:

1. Dual-Memory Architecture

To ensure consistent state tracking and context-aware reasoning across multi-round interactions, ClinicalAgents employs a dual-memory system:

Working Memory ( $M^{work}_t$ ): A mutable state board that maintains the evolving patient context, including the accumulated evidence set ( $E_t$ ), the set of diagnostic hypotheses ( $H_t$ ), and the action trajectory ( $\tau_t$ ). This ensures all agents share a synchronized view of the diagnostic progress.
Experience Memory ( $M^{exp}$ ): A static external store containing structured clinical guidelines ( $D_{guide}$ $D_{g u i d e}$ ) and a historical case database ( $D_{cdc}$ $D_{c d c}$ ). This memory supports the system via a hybrid Retrieval-Augmented Generation (RAG) mechanism:
- Passive Augmentation: Retrieving relevant guidelines to inform reasoning.
- Active Evidence-Seeking: Using historical cases to identify potentially missing but clinically significant evidence ( $E^p_t$ ) that the current diagnostic context may have overlooked.

2. Clinical Orchestrator with MCTS

The core innovation is a Clinical Orchestrator that manages agent interactions not through a fixed sequence, but via a Monte Carlo Tree Search (MCTS) process.

Action Space: The orchestrator selects from a dynamic action space $A$ $A$ comprising:
- $A_{agent}$ : Activating specialized agents (e.g., imaging, treatment, diagnosis).
- $a_{rag}$ : Retrieving knowledge from Experience Memory.
- $a_{back}$ : Triggering backtracking to a previous state in the workflow.
- $a_{term}$ : Terminating the process upon reaching clinical closure.
Hypothesis-Driven Expansion: The orchestrator iteratively generates hypotheses and verifies them against evidence. If verification reveals missing critical evidence ( $E^m_t \neq \emptyset$ ), the system triggers the backtracking mechanism.
Backtracking Mechanism: Formulated as a decision function $\Phi$ , this mechanism allows the system to revert to earlier stages (e.g., from diagnosis back to information gathering) to request or retrieve missing data, rather than proceeding with incomplete information.
Reward Function: The MCTS is guided by a reward function $R_t$ that maximizes information gain, specifically rewarding actions that reduce the set of missing evidence ( $\Delta E^m_t > 0$ ) or increase diagnostic confidence ( $\Delta c_t > 0$ ), while penalizing uninformative actions.

3. Workflow

The system operates in a Perceive $\rightarrow$ Hypothesize $\rightarrow$ Verify $\rightarrow$ Update loop. The orchestrator dynamically selects agents, retrieves necessary knowledge, and if evidence is insufficient, initiates backtracking to refine the hypothesis or gather missing data before proceeding.

Key Contributions

Hypothesis-Guided Framework: The introduction of ClinicalAgents, which explicitly operationalizes the hypothetico-deductive reasoning process of clinicians, moving beyond static generation to an iterative, evidence-seeking loop.
MCTS-Based Orchestration: A novel orchestration mechanism that uses MCTS to adaptively select, reorder, and iterate agent actions. This overcomes the rigidity of pre-scripted workflows and enables evidence-triggered backtracking.
Dual-Memory Integration: A design combining a mutable working memory for state synchronization with a static experience memory for knowledge retrieval. This reduces information loss and redundancy while improving reasoning consistency.
Empirical Validation: Comprehensive experiments demonstrating state-of-the-art performance on multi-stage clinical benchmarks, with ablation studies confirming the complementary value of the orchestrator and dual-memory components.

Experimental Results

The authors evaluated ClinicalAgents on the MedChain benchmark, a five-stage clinical workflow (Specialty Referral, Test Ordering, Examination, Diagnosis, Treatment), and compared it against strong baselines including single LLMs, single-agent systems (e.g., ReAct, RAG), and existing multi-agent frameworks (e.g., MedAgents, MedChain-Agents).

Overall Performance: ClinicalAgents achieved the highest average score (0.5107) among all evaluated methods, outperforming the best multi-agent baseline (MedChain-Agents, 0.4880) by +4.7% and the strongest single-agent baseline (ReAct, 0.4696) by +8.8%.
Task-Specific Gains: The system showed particularly strong improvements in Test Ordering (IoU 0.4820), attributed to the verification-and-backtracking mechanism that iteratively refines examination plans by adding critical tests and removing redundant ones.
Ablation Studies: Removing the Dual-Memory or the Orchestrator resulted in significant performance drops, confirming that both components are necessary and complementary. The full system outperformed the backbone model (GPT-5.2) by +13.0%.
Backbone Sensitivity: The framework demonstrated robustness across different backbone LLMs (including GPT-5.2, DeepSeek-R1, Qwen3-VL, and HuaTuoGPT), consistently improving performance regardless of the underlying model's capability.
Efficiency: ClinicalAgents achieved superior performance with lower token consumption and latency compared to Self-Consistency (SC) scaling and other multi-agent baselines, indicating that gains stem from effective orchestration rather than increased compute budgets.
Reliability: In cases where the initial hypothesis was incorrect, the backtracking mechanism successfully triggered in 83% of instances and recovered the correct diagnosis in 64%.

Significance and Claims

The paper claims that ClinicalAgents represents a significant step toward more transparent and dependable agentic systems for clinical decision support. By bridging the gap between static AI models and the dynamic, iterative nature of human clinical reasoning, the framework addresses critical limitations in current AI-driven CDM, such as rigidity, information fragmentation, and premature convergence. The authors posit that their approach provides a practical pathway for AI systems to act not just as passive predictors, but as active, evidence-seeking collaborators that can rigorously verify hypotheses and adapt to new information, thereby enhancing both diagnostic accuracy and explainability. The work emphasizes that the system is designed to function as a supportive tool to augment, rather than replace, human clinical judgment.