Cause of the NFCorpus null retrieval association

Determine whether restricted variation in NFCorpus query density explains why the association between query density and retrieval degradation under embedding truncation does not replicate at 32 dimensions.

Background

The paper finds that, on SciFact using MiniLM-L6, denser queries tend to experience less retrieval degradation under PCA truncation, whereas the corresponding 32-dimensional association is not observed consistently on NFCorpus using GTE-small. NFCorpus has substantially less within-dataset variation in query density than SciFact, but the benchmarks also differ in dataset and encoder, preventing attribution of the discrepancy to density variation alone.

The authors test whether restricting SciFact queries to narrower density bands reproduces the NFCorpus null pattern. The point estimates attenuate as the density range narrows, but the confidence interval for the attenuation includes zero, so the evidence does not establish whether restricted density variation explains the NFCorpus result.

References

Restricting SciFact to narrower density bands attenuates the point estimates monotonically, but the attenuation confidence interval includes zero, leaving range restriction unresolved.

— Geometric Structure Predicts Per-Example Neighborhood Loss Under PCA Truncation  (2610.17742 - Loganathan et al., 13 Oct 2026) in Section "Downstream Retrieval Consequences"