Language model recall limited by copying from retrieval context
The Copy Ceiling: An Input-Exposure Control for Ontology-Grounded Generation over Curated Corpora
Computation and LanguageComputers and Society
Summary
This paper finds that when language models answer questions using information retrieved from a curated database, they mostly copy the exact answers already shown to them rather than reasoning beyond what they see. The researchers measure how much models just repeat exposed information versus genuinely infer unseen answers. Most of the credit for correct answers comes from the information already exposed in the context, with very little true reasoning beyond it. Their method helps distinguish when models are just copying versus when they might be reasoning, but it doesn't prove reasoning itself.
What this means in practice
- •For natural language processing engineers: Evaluate language model outputs for true reasoning by identifying when answers copy versus extend from retrieved database context.
- •For chatbot developers: Improve response validation by measuring whether chatbot answers are verbatim recalls from their source data or demonstrate inference beyond exposed content.
Authors
John J. O'Hare
Abstract
When a language model answers from a curated corpus via graph-based retrieval, a large grounding uplift does not establish reasoning over the retrieved structure: the context may already expose the gold answers. We propose exposure accounting, which classifies each gold item by whether the shown context exposes it and whether the answer recovers it. Its scalar reference is the copy ceiling, the recall a verbatim copy of the context achieves; signed gain over copy measures the model's recall relative to this deterministic, judge-free baseline. Across ten models, unaided recall averages 0.26 and grounded recall 0.92, yet gain over copy is uniformly negative (-0.067 to -0.022). Of 11,360 gold-item observations, representing 1,136 target instances evaluated under ten models, only three unexposed items receive lexical credit. A stratified model-judged audit of 423 observations, with a symmetric quotation-verification policy, estimates that 97.1% of credited items assert the requested relation; all three unexposed credits fail relational adjudication. On targets the scaffold does not expose, lexical recovery falls from 0.121 unaided to 0.004 grounded; adjudication validates 71 of the 92 unaided credits and none of the three grounded credits, without establishing full-frame relational recovery rates. Rephrasing questions outside the graph's title vocabulary reduces exposure from 0.964 to 0.328, while an absence-triggered fallback activates on only 2 of 506 questions. A paired production study improves judged quality by +0.27 pooled, but negative controls do not establish content specificity beyond a well-formed on-corpus block. These results support exposure accounting as a standing control for corpus-derived evaluations. The accounting distinguishes exposed-item omissions from beyond-exposure recoveries; it does not determine whether reasoning occurred.