A hospital put a 1-trillion-parameter agent inside its closed network and it scored 97.2 percent on its own documents, versus 81.3 percent with plain retrieval
AspireAI community · arXiv (Ken Enda, Jien Saito, Masami Mukai et al., Hokkaido University Hospital)
Medical information systems in many hospitals run on closed networks, which rules out the usual hosted-agent approach. LLENS is a clinical support agent platform built for exactly that constraint, running inside Hokkaido University Hospital's network on a single host with 8 NVIDIA H200 GPUs, the open-weight Kimi K2.6 at roughly a trillion parameters, and a web application rebuilt from Open WebUI.
The harness has 10 components: page image reading, full-text attachment input, a path-addressed document tree with no vector index, a network-isolated sandbox, artifact return, self-checking of generated output, subagents and saved procedures. Kimi K2.6 was chosen from three open-weight candidates and scored 480 of 500 on the 2026 Japanese Medical Licensing Examination.
The result that matters for agent design is on 180 questions over hospital documents, run with three seeds each: mean accuracy was 10.7 percent with no documents, 81.3 percent with fixed retrieval of the top 10 chunks from a single lexical search, and 97.2 percent with interactive access, where the model reads through document tools and the sandbox. That is a 15.9 point gap over fixed retrieval with a 95 percent confidence interval of 10.9 to 21.3, and the difference concentrated in filtering and counting questions, 16.7 versus 87.8 percent.
Caveat: five scenarios were run on synthetic data, and the platform has only been live on the hospital network since September 14, 2026.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.