Auditing Whether Grounded Language Models Preserve the Selected Object

Summary

Grounded language-model pipelines first select an object, retrieve passages associated with it, and then use the evidence to answer a question. This study audits whether the selected object survives that handoff, noting that conventional benchmark recall may track a dataset-linked object rather than the object actually selected by the pipeline. The authors examine 600 HybridQA questions and 1,463 resolvable records in which the selected object matches the passage traced by the dataset. When every retrieval rule receives the same decoded selected title, exact key lookup and exact title matching return the target on every record. Body-only BM25 at a top-five cutoff omits the target on 389 records, or 26.6%, whereas hybrid retrieval with reranking omits it on 14 records, or 1.0%. The selected and dataset-traced identities differ on 329 of 1,792 resolvable records, and top-five checks based on original-question rankings disagree on 106 records, or 5.9%. Frozen-reader comparisons associate the aligned object’s presence with a 28.6-to-31.0-point improvement in exact match. In a deliberately selected 64-item cohort, removing the aligned passage sharply reduces exact match, while removing a similar-length comparison passage does not reproduce the decline. The authors release the Returned-Object Profile, an executable record containing the target, returned-ID field, cutoff, membership rule, and complete expected population, together with data and an offline replay.