Work

Seeing Both Levels: Evidence-Weighted Paper Scores That Reveal Passage-Level Retrieval Degradation

Year: 2026
Type: preprint
Abstract: A retrieval evaluation should reward finding both the right paper and its supporting passages. Companion work showed that a paper-level score can remain exactly unchanged while passage retrieval deteriorates. We test what should replace that score. Two simple candidates fixed before confirmation are an evidence-weighted paper score, assigning half its credit to paper rank and half to within-paper evidence quality, and the mean of paper and flat passage nDCG. Both met the pre-set confirmation rule on 424 fresh TRAIN-derived QASPER questions with a GPT reranker and on 848 DEV questions with a different reranker, although Mean failed the pair-discrimination criterion. Natural differences between five real systems also expose the gap: paper scoring misses 72.4% of passage-level differences, versus 19.0% and 4.0% for the two fixes. Improvements have limits. Weight tuning did not help: interior weights formed a near-flat region, and the selected weight failed confirmation. In the fault answer check, both fixed scores track answer quality better than the paper score, but not better than plain passage scoring: all three have answer-harm AUC 0.630. The fixes miss fewer answer harms but fall more often when answers do not worsen. A grouped-score hypothesis fails; label-free detectors fail the pre-set 0.70 AUC bar with both cross-encoder and local language-model judges. We recommend a fixed combined score with both raw scores retained, and provide a reproducible protocol that separates discovery, confirmation, answer validity and failures.
Language: English
Open Access status: green
OpenAlex ID: W7236638384
Cites:
Cited by:
Related to: 10