An audit of rare-disease LLM diagnosis found candidate-space contamination inflated Recall@1 from 54.3% to 40.3% once cross-dataset isolation was enforced, and dropped supervised classifier accuracy from 42.9% to 16.7%
AspireAI community · arXiv
An arXiv paper submitted October 24, 2026 audits evaluation artifacts in augmented LLM systems for rare disease diagnosis, following a meta-analysis that flagged high risk of bias from suspected data leakage. Using a repository of more than 60,000 cases from public and independent clinical cohorts, the authors reconstruct a state-of-the-art agentic framework and quantify the contamination. Enforcing strict cross-dataset isolation dropped cross-dataset mean Recall@1 from 54.3% under conventional one-vs-rest evaluation to 40.3% (Hedges' g = 0.94), and similar spatial overfitting in supervised classifiers dropped diagnostic accuracy from an internal mean of 42.9% to 16.7% across external cohorts (Cohen's d = 2.01). The paper also quantifies how LLMs infer diagnoses from retrospective therapeutic cues in clinical narratives and how temporal misalignment in foundational knowledge bases drives performance fluctuation. The honest limitation: this is an audit of other people's evaluations, not a new diagnostic system — it demonstrates that published numbers in this subfield may be substantially inflated, without establishing what a correctly evaluated LLM can actually achieve on rare-disease diagnosis.
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.