The Copy Ceiling: An Input-Exposure Control for Ontology-Grounded Generation over Curated Corpora
Grounded models scored high because the context often already contained the answers.
The paper proposes “exposure accounting” to separate items exposed in retrieved context from items recovered beyond it. Across ten models, grounded recall averaged 0.92, but every model performed below a verbatim-copy baseline. Only three unexposed credited items appeared in 11,360 gold-item observations, and all three failed the relational audit. The author frames the method as a control for corpus-derived evaluations, not as proof that reasoning did or did not occur. ArXiv · AI/CL/LG's note
The paper proposes “exposure accounting” to separate items exposed in retrieved context from items recovered beyond it. Across ten models, grounded recall averaged 0.92, but every model performed below a verbatim-copy baseline. Only three unexposed credited items appeared in 11,360 gold-item observations, and all three failed the relational audit. The author frames the method as a control for corpus-derived evaluations, not as proof that reasoning did or did not occur. ArXiv · AI/CL/LG's note
score 5