Referring-expression models like ScanRefer often face a spatial disambiguation problem: when several instances of the same object class appear in a scene, the model must decide which one a query refers to. Confidence in that decision is typically derived from per-query softmax logits.

Those logits, however, are not always stored during training. The new arXiv paper (2610.14210) examines a common substitute: cross-epoch agreement, or how consistently the model picks the same instance across training epochs.

The abstract frames agreement as a stand-in for softmax confidence under spatial disambiguation, but stops before presenting results. Based on the excerpt, the paper's contribution is to investigate whether this substitution is valid and to lay out the conditions under which the two signals might diverge.