Establish executable-candidate coverage gains from TR-GRPO

Establish whether the complete TR-GRPO training intervention increases executable-candidate coverage relative to the DualVLN baseline.

Background

The paper separates candidate support from candidate selection and reports historical probe statistics showing that selection is generally effective when an executable candidate is present. However, the logged subset is historical rather than a final-checkpoint counterfactual comparison.

Consequently, the available evidence does not show whether TR-GRPO itself increases the frequency with which the candidate set contains an executable option. Establishing this comparison would directly test the proposed mechanism by which TR-GRPO reshapes trajectory-generation support.

References

This diagnostic is historical rather than a final-checkpoint counterfactual evaluation. It therefore does not establish that TR-GRPO increased executable-candidate coverage relative to the baseline.

— Revealing and Bridging the Physical Executability Gap in Vision-Language Navigation  (2610.25687 - He et al., 19 Oct 2026) in Appendix, Section A9, “Historical Candidate Support and Selection”