Improving JEV Reranking via Span Readouts and Selective Adaptation

The study improves JEV decision models by replacing pointer-based heads with mean-span pooling and applying selective backbone adaptation. This approach uses low-rank updates in attention and gated DeltaNet projections to enhance reranking accuracy without significant parameter overhead.