Complex clinical reasoning requires models to update diagnostic hypotheses as new evidence emerges and to coordinate different medical specialities under limited consultation resources. Existing LLM-based clinical reasoning systems typically perform single-pass prediction or rely on fixed multi-agent workflows, making expert participation either static or unnecessarily exhaustive. We propose Sparse Multi-Stage Expert-Agent Routing, a language-based clinical reasoning framework that models diagnosis as a stage-wise routing process. Given progressively available clinical evidence derived from multiple modalities, the framework maintains an evolving case state and adaptively activates a sparse set of medical expert agents, supported by expert-specific memory across stages. To evaluate free-text diagnostic conclusions beyond surface similarity, we further introduce ClinFEScore, a fact-aware semantic evaluation protocol for clinical reasoning outputs. On reconstructed multi-stage cases from MAC and AgentClinic-NEJM, our framework reduces the average number of activated experts from 17.0 to 3.0 whilst maintaining strong fact-level diagnostic quality. On 200 real-world hospital MDT cases, ClinFEScore correlates strongly with clinician judgements (Spearman's ρ=0.81; Pearson's r=0.87), whilst our method achieves 91.5% clinician-verified diagnostic accuracy with approximately five expert-agent/LLM calls per case. These results support sparse stage-wise coordination as an efficient and clinically relevant approach to LLM-based clinical reasoning.
Figures & tables
Figure 1: Illustration of the evidence-driven activation process for an individual expert agent within the sparse routing framework. Clinical evidence derived from multiple modalities is transformed into a unified case representation and an expert-agent activation score. When the score exceeds the adaptive threshold, the corresponding expert agent is invoked to produce stage-wise reasoning output.
Figure 2: Overview of the proposed sparse multi-stage expert routing framework for complex clinical reasoning. Multimodal clinical evidence is organised into stage-wise evidence representations and routed through the activation mechanism in Fig. 1 . The router selectively activates medical expert agents across stages, whose outputs are aggregated into a final clinical report. The clinical content shown above is for illustrative purposes only.
SA
ST
MT
S
M
ClinFEScore ↑
Average Activated Experts ↓
Energy ↓
Fact Similarity
Fact Precision
Fact Recall
Ground Truth
0.9972 ±0.0049
0.9904 ±0.0079
0.9887 ±0.0076
–
–
✓
–
–
–
–
0.9434 ±0.0050
0.9590 ±0.0066
0.9556 ±0.0068
1.0000 ±0.0000
1.2000 ±0.0000
–
✓
–
–
–
0.9537 ±0.0046
0.9691 ±0.0083
0.9676 ±0.0073
9.0000 ±0.0000
17.1000 ±0.0000
–
✓
✓
–
–
0.9845 ±0.0043
0.9706 ±0.0071
0.9690 ±0.0066
17.0000 ±0.0000
34.2000 ±0.0000
–
✓
✓
✓
–
0.9936 ±0.0050
0.9788 ±0.0068
0.9768 ±0.0056
3.0119 ±0.2437
7.1488 ±0.8047
Table 1: Comparison of different system configurations, showing the effects of single-agent inference (SA), single-stage full expert-agent consultation (ST), multi-stage full expert-agent consultation (MT), sparse expert-agent routing (S) and memory augmentation (M). Ground Truth is reported as an upper-bound reference, and the best model results excluding Ground Truth are shown in bold . gemini-3.1-pro-preview is used as the backbone model.
Backbone Model
ClinFEScore ↑
Fact Similarity
Fact Precision
Fact Recall
Grok-4.20
0.9822 ±0.0078
0.9649 ±0.0061
0.9623 ±0.0058
Claude-Opus-4.7
0.9829 ±0.0079
0.9650 ±0.0061
0.9640 ±0.0059
GPT-5.4
0.9839 ±0.0080
0.9741 ±0.0070
0.9724 ±0.0068
Gemini-3.1-Pro-Preview
0.9940 ±0.0043
0.9792 ±0.0063
0.9780 ±0.0062
Hulu-Med-4B
0.9939 ±0.0049
0.9795 ±0.0069
0.9783 ±0.0062
Table 2: Comparison of different backbone models under the same routing and memory configuration. The upper block lists general-purpose LLMs, whilst the lower block lists medical-domain LLMs. The best results are shown in bold and the second-best results are underlined .
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Expert memory structure. The memory module is organised as a triangular interaction among expert memory, short-term memory and long-term memory, supporting stable yet adaptive reasoning across stages.
Routing Layers
ClinFEScore ↑
Average Activated Experts ↓
Energy ↓
Fact Similarity
Fact Precision
Fact Recall
2
0.9941 ±0.0050
0.9798 ±0.0070
0.9785 ±0.0066
3.0119 ±0.2437
7.1488 ±0.8047
4
0.9940 ±0.0049
0.9792 ±0.0070
0.9780 ±0.0065
4.0000 ±0.0000
9.4000 ±0.0000
Appendix
Table 3: Comparison between two-layer and four-layer routing structures under the same backbone model and evaluation setting. The best results are shown in bold . Gemini-3.1-Pro-Preview is used as the backbone model.
While Large Language Models (LLMs) have demonstrated potential in healthcare, they often struggle with the complex, non-linear reasoning required for accurate clinical diagnosis. Existing methods typically rely on static, linear mappings from symptoms to diagnoses, failing to capture the iterative, hypothesis-driven reasoning inherent in human clinicians. To bridge this gap, we introduce ClinicalAgents, a novel multi-agent framework designed to simulate the cognitive workflow of expert clinicians. Unlike rigid sequential chains, ClinicalAgents employs a dynamic orchestration mechanism modeled as a Monte Carlo Tree Search (MCTS) process. This allows an orchestrator to iteratively generate hypotheses, actively verify evidence, and trigger backtracking when critical information is missing. The foundation of this framework is a Dual-Memory architecture: a mutable working memory that maintains the evolving patient state for context-aware reasoning, and a static experience memory that retrieves clinical guidelines and historical cases via an active feedback loop. Extensive experiments demonstrate that ClinicalAgents achieves the best performance among evaluated baselines, significantly enhancing both diagnostic accuracy and explainability compared to strong single-agent and multi-agent baselines. Our code is released at https://github.com/ZhuohanGe/ClinicalAgents-Code.
Zhuohan Ge, Haoyang Li, Yubo Wang +3
The Hong Kong Polytechnic University, Hong Kong SAR, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China
We present Multi-Agent Reasoning and Coordination (MARC), an open-source framework that replaces monolithic LLM prompting with deterministic multi-agent orchestration for clinical reasoning. MARC coordinates role-specialized agents for extraction, reasoning, answer generation, and evaluation, with explicit context passing and traceable intermediate outputs, enabling stage-wise failure attribution. We additionally introduce a Decomposer module that generates task-specific agent prompts from a plain-language description, eliminating manual prompt engineering. The framework supports both API-based and local CPU-compatible deployments and is entirely configurable via YAML, without code modifications. MARC is designed to be model-agnostic, interpretable, and accessible to clinical domain experts without programming expertise. The full framework is available at https://github.com/Penn-RAIL/MARC-v1.
Saisha Shetty, Satvik Tripathi, Austin Lin +6
1. College of Engineering, University of California, Davis · 2. Perelman School of Medicine, University of Pennsylvania · 3. School of Engineering and Applied Science, University of Pennsylvania +2
Large language models applied to clinical prediction exhibit case-level heterogeneity: simple cases yield consistent outputs, while complex cases produce divergent predictions under minor prompt changes. Existing single-agent strategies sample from one role-conditioned distribution, and multi-agent frameworks use fixed roles with flat majority voting, discarding the diagnostic signal in disagreement. We propose CAMP (Case-Adaptive Multi-agent Panel), where an attending-physician agent dynamically assembles a specialist panel tailored to each case's diagnostic uncertainty. Each specialist evaluates candidates via three-valued voting (KEEP/REFUSE/NEUTRAL), enabling principled abstention outside one's expertise. A hybrid router directs each diagnosis through strong consensus, fallback to the attending physician's judgment, or evidence-based arbitration that weighs argument quality over vote counts. On diagnostic prediction and brief hospital course generation from MIMIC-IV across four LLM backbones, CAMP consistently outperforms strong baselines while consuming fewer tokens than most competing multi-agent methods, with voting records and arbitration traces offering transparent decision audits.
Yuxing Lu, Yushuhong Lin, Jason Zhang
Georgia Institute of Technology · Peking University