An interpretable arbitrator resolved 93.5 percent of deceased-donor kidney placement decisions and deferred the rest to a human, instead of making every call itself
AspireAI community · arXiv
Nearly 30 percent of kidneys recovered from deceased donors in the United States are never transplanted because they are hard to place through the standard waitlist sequence, and flagging kidneys at risk of delayed or failed placement for out-of-sequence allocation comes at the cost of bypassing candidates higher on the waitlist. The Arbitrator framework combines interpretable models and acts only when prediction uncertainty cannot change the preferred action, using Organ Procurement and Transplantation Network match-run data from 2016 to 2019 and 2023 to 2025 excluding the COVID-disrupted years. It resolves 93.5 percent of placement decisions using explicit criteria learned from data and defers the remaining 6.5 percent to a human or a more complex model, correctly identifying 70.1 percent of historically discarded low-KDPI kidneys as out-of-sequence candidates and projecting that up to 43.9 percent of historically discarded kidneys could be flagged for expedited placement. The authors also exploit the interpretability to build an agentic audit comparing learned rules against published clinical literature to flag contentious patterns. The limitation is that the projected transplant gains are a modelling projection on retrospective data, not observed outcomes, and clinical adoption would face the constraint that out-of-sequence allocation bypasses higher-priority candidates.
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.