MedLatentDx: Latent Multi-Agent Communication for Cross-Hospital Rare-Disease Diagnosis

A new multi-agent framework called MedLatentDx enables hospitals to collaborate on rare-disease diagnosis without exposing identifiable patient data, according to a paper submitted in 2026. The system transmits compact latent representations instead of clinical text, addressing a core tension between data utility and privacy regulations that restrict cross-institutional information sharing. [1] Rare diseases collectively affect over 300 million patients across more than 7,000 conditions, yet no single hospital accumulates enough cases of any one condition to develop reliable diagnostic expertise. [1] Cross-hospital collaboration could pool distributed, case-specific evidence, but privacy rules generally prohibit the transmission of identifiable clinical text across institutional boundaries. [1] Existing medical agent systems often depend on textual evidence exchange, while raw latent states such as hidden states and KV caches can still leak prompt-derived clinical content. [1] MedLatentDx addresses this by keeping private clinical records and retrieved cases local to each hospital agent. The agents send compact latent KV blocks to a host agent that performs the diagnosis. [1] The framework supports two deployment configurations: when all hospitals use the same large language model backbone, it employs latent KV distillation; when hospitals operate different LLM backbones, it uses cross-family latent alignment. [1] The approach was evaluated on CrossRare-Bench, a self-built large-scale rare-disease benchmark with hospital-level partitions, where it improved cross-hospital diagnostic performance while reducing reconstructable clinical content compared with raw-latent communication baselines. [1] The work builds on a growing body of research into agentic systems for rare disease. DeepRare, another framework, uses a central host agent with a memory bank and multiple specialized agent servers that manage over 40 tools for tasks such as phenotype extraction and variant prioritization. [3] RareAgents takes a different approach, assembling a multi-disciplinary team of specialist agents that engage in iterative discussions and maintain dynamic long-term memories to reach consensus on diagnosis and treatment plans. [4] More broadly, the concept of pure latent collaboration among LLM agents has been explored in LatentMAS, a training-free framework in which agents generate auto-regressive latent thoughts and exchange information through shared latent working memory stored in layer-wise KV caches. [5] Privacy-preserving collaboration remains a central challenge. The MedLatentDx authors note that raw latent states may still reveal clinical content, and their framework explicitly measures and reduces reconstructable clinical information relative to baselines. [1] The benchmark, CrossRare-Bench, introduces hospital-level data partitions to simulate realistic institutional silos, a design choice that distinguishes it from prior rare-disease datasets. [1] The paper was submitted to arXiv in June 2026 through arXivLabs, a framework that allows collaborators to develop and share new features on the platform. [1]