Signal The Copy Ceiling: An Input-Exposure Control for Ontology-Grounded Generation over Curated Corpora

Summary

The paper argues that when a language model answers questions from a curated corpus using graph-based retrieval, a large improvement in apparent grounding does not by itself establish that the model is reasoning over the retrieved structure, because the provided context may already expose the correct answer. To address this, the authors propose exposure accounting, a method that classifies each gold answer item by whether the shown context already reveals it and whether the model's answer actually recovers it. Its reference point, called the copy ceiling, is the recall achievable simply by copying the context verbatim, and the signed gain over this baseline allows model performance to be measured against a deterministic standard without relying on a judge. Across ten models tested, unaided recall averaged 0.26 while grounded recall averaged 0.92, yet the gain over the copy baseline was uniformly negative for every model. This pattern suggests that apparent performance gains do not necessarily reflect genuine reasoning ability. The authors present exposure accounting as a standing control that can be applied to evaluations built on curated corpora.