Reduce supervision requirements for Med-TrustRouter

Reduce the labeled-training requirement of the Med-TrustRouter action-value heads through few-shot or self-supervised learning so that the router can be adapted with less outcome supervision in a new domain.

Background

Med-TrustRouter is trained on a labeled learn split whose gold answers are used to fit per-action value heads and estimate the expected correctness of each knowledge-use action. Consequently, the router cannot be deployed zero-shot in a new medical domain unless outcomes are first collected for supervision.

The paper explicitly identifies reducing this supervision burden through few-shot or self-supervised action-value heads as unresolved. Solving this problem would improve the practicality and portability of regression-safe routing in domains where labeled clinical outcomes are expensive or unavailable.

References

The router is trained: it requires a labeled learn split to fit the action-value heads, so it cannot be deployed zero-shot in a new domain without first collecting outcomes; reducing this supervision through few-shot or self-supervised heads remains open.

— Learning When to Trust Medical Knowledge: Regression-Safe Routing for Clinical Reasoning  (2610.19663 - Zhang et al., 14 Oct 2026) in Section Discussion and Limitations, paragraph (3)