Beyond Federated Learning: Decentralized Agentic Systems for Self-Evolving Clinical Intelligence

Haoyuan Li, Mathias Funk, Aaqib SaeedarXiv 2026 · arXiv

This perspective paper, recently published by researchers at the Eindhoven University of Technology, proposes a paradigm shift in medical AI. Rather than focusing on aggregating raw data or model weights, the authors introduce Decentralized Agentic Systems (DAS)—a framework where AI agents across different hospitals share their "reasoning trajectories" to build a collective, self-evolving clinical intelligence.

TL;DR

The paper argues for a move from "data aggregation" to "experience aggregation" in healthcare AI. By sharing abstracted reasoning paths (semantic trajectories) instead of raw data or model gradients, Decentralized Agentic Systems (DAS) enable heterogeneous AI agents to learn from each other's clinical successes and failures while preserving institutional sovereignty and patient privacy.

What Problem Does This Paper Address?

Clinical expertise is inherently distributed. A rural doctor develops unique insights into resource-constrained care that an academic specialist might never see. However, current medical AI is "stubbornly centralized" or relies on Federated Learning (FL).

The authors identify three core problems with the status quo:

  1. Expertise Fragmentation: Real clinical mastery resides in the reasoning process and outcomes, not just the static records in an Electronic Health Record (EHR).
  2. Scaling Limits of Static Knowledge: LLMs trained on static datasets quickly become outdated and lack the "procedural knowledge" of how to apply facts under clinical pressure.
  3. Sovereignty-Utility Tension: Hospitals are often unwilling to share raw data or institutional protocols (their "intellectual assets") with centralized systems.

While Federated Learning (FL) helps with privacy, it requires all participating models to have the same architecture and only shares numerical gradients, which are uninterpretable and don't capture the "why" behind a clinical decision.

Methods

The authors propose the DAS framework, which consists of three architectural layers:

  1. Distributed Agentic Nodes: Local, autonomous AI agents (e.g., using models like BioMistral or Med-PaLM 2) tailored to specific institutional needs. They operate in a perception-reasoning-action cycle, storing their experiences locally.
  2. Semantic Abstraction Layer: This is the "secret sauce." It involves two steps:
  3. Experience Distillation: Filtering raw logs to keep only those with high "epistemic value" (e.g., a successful diagnosis of a rare condition).
  4. Trajectory Standardization: Mapping local reasoning into a "semantic trajectory"—a structured, human-readable record of observations, hypotheses, tool calls, and outcomes using standardized concepts (like SNOMED-CT).
  5. Collective Intelligence Layer: A federated repository where these standardized trajectories are aggregated. Local nodes can then "retrieve" relevant trajectories from this global pool to update their own reasoning backbones via in-context learning or fine-tuning.
The DAS architecture showing the flow from local experience to collective intelligence.

Datasets

Not discussed in the paper. As a "Perspective" piece, the authors focus on formalizing the architectural framework and theoretical foundations rather than benchmarking on specific datasets. They do, however, provide illustrative clinical scenarios involving heart failure management and rare disease diagnosis (hATTR).

Key Insights and Ideas

  • Experience > Data: The authors distinguish between data (static records) and experience (the dynamic process of perceiving signals and observing outcomes). DAS treats experience as the primary unit of learning.
  • Model Heterogeneity: Unlike standard FL, DAS allows a hospital using a small, local model to learn from a trajectory generated by a state-of-the-art model at a major research center. The "knowledge" is in the reasoning path, not the model weights.
  • Interpretability and Auditability: Because the shared unit is a "semantic trajectory," clinicians can actually read and audit the knowledge being shared. This is a massive leap over the "black box" gradients of Federated Learning.
A comparison of DAS against traditional distributed learning paradigms.

Key Results

The paper provides a qualitative comparison (Table 1) showing that DAS is the only paradigm that simultaneously supports model heterogeneity, captures procedural reasoning, and provides human-readable shared knowledge.

The authors illustrate the potential through two scenarios:

  • Rare Disease Diagnosis: A cardiology agent at one hospital learns to identify a rare amyloidosis (hATTR) by retrieving a reasoning trajectory from a neurology agent at another hospital who successfully connected peripheral neuropathy to cardiac biomarkers.
  • Heart Failure Management: A home monitoring system refines its "early warning" triggers by assimilating longitudinal trajectories from both rural primary care and urban cardiology practices.

Limitations and Future Work

The authors are candid about the significant hurdles ahead:

  • Hallucination Propagation: If one agent generates a "plausible but false" reasoning path, it could corrupt the collective knowledge base (a version of the Byzantine failure problem).
  • Semantic Privacy Leakage: Even without names, a specific sequence of rare disease queries could act as a "behavioral fingerprint," potentially identifying a patient.
  • Evaluation Metrics: We lack standard ways to measure "collective intelligence." The authors suggest new metrics like "marginal improvement from collective knowledge."
  • Incentives: How do you prevent "free-rider" institutions that consume collective knowledge without contributing their own high-quality trajectories?

The paper concludes that while the technical and governance challenges are steep, the shift toward decentralized, experience-sharing agents is a necessary evolution for AI to truly master the complexities of longitudinal patient care.

Summary updated

Abstract

Clinical expertise is inherently distributed: it resides not in any single dataset or institution but in the accumulated reasoning experiences of thousands of clinicians navigating diverse patient populations, resource constraints, and evolving treatment protocols. Yet current approaches to scaling medical AI, from centralized large language models to federated learning, remain anchored to a data-centric paradigm that treats knowledge as static parameters to be aggregated. In this Perspective, we argue for a fundamental shift from data aggregation to experience aggregation. We propose Decentralized Agentic Systems (DAS), a framework in which distributed AI agents operating across healthcare institutions synthesize collective clinical intelligence by sharing abstracted reasoning trajectories rather than raw data or model parameters. Unlike federated learning, which averages numerical gradients across homogeneous model architectures, DAS exchanges structured records of clinical reasoning, specifically the sequence of observations, hypotheses, tool invocations, and outcome feedback that constitute diagnostic experience. We examine the architectural principles, deployment considerations, governance requirements, and critical open challenges of this framework, laying the groundwork for a collaborative medical AI ecosystem where distributed expertise autonomously adapts to the complexities of longitudinal patient care.
BibTeX
@misc{li2026beyond,
  author = {Haoyuan Li and Mathias Funk and Aaqib Saeed},
  title = {{Beyond Federated Learning: Decentralized Agentic Systems for Self-Evolving Clinical Intelligence}},
  year = {2026},
  eprint = {2610.35453},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.35453},
}
Figure