Retrieval-grounded models show negative reasoning gain over simple copying
Researchers introduce exposure accounting, a diagnostic framework that reveals a critical blind spot in retrieval-augmented generation systems. When language models answer questions grounded in curated knowledge graphs, their apparent performance gains vanish once you subtract what a simple verbatim copy of the context would achieve. Testing ten models across 1,136 instances shows unaided recall of 0.26 jumping to 0.92 with grounding, yet actual reasoning gain stays negative. This work matters because it exposes how retrieval systems can inflate benchmark scores through context leakage rather than genuine structural reasoning, forcing the field to rethink how we measure grounding quality.
Modelwire context
ExplainerThe paper's real contribution isn't just measuring context leakage, it's formalizing a method to subtract it from performance metrics. This lets researchers isolate whether a model is reasoning over structure or simply regurgitating nearby text, a distinction most benchmarks don't attempt to measure.
This connects directly to the broader reckoning around evaluation rigor visible across recent work. OSWorld-Pro decomposed agent failures into subgoals to enable root-cause diagnosis; SLITE made entailment reasoning explicit rather than opaque; and onPanda improved annotation granularity to catch token-level errors. The Copy Ceiling paper follows the same pattern: it exposes a blind spot in how we measure performance by introducing a diagnostic layer beneath the headline number. All four papers reflect pressure to move beyond coarse accuracy metrics toward frameworks that reveal what's actually happening inside the system.
If the exposure accounting method gets adopted in the next round of retrieval-augmented generation benchmarks (expect new KBQA or knowledge-grounded QA leaderboards by Q1 2027), that signals the field is taking context leakage seriously. If major RAG papers continue reporting headline recall without subtracting baseline copy performance, the finding will have failed to shift practice.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “The Copy Ceiling: An Input-Exposure Control for Ontology-Grounded Generation over Curated Corpora”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.