Abstract
Retrieval-augmented generation treats the document store as a frozen input, and the systems that instead let an agent curate one never measure what curation does to the store. We invert the framing: the knowledge base is the model. A training agent answers a supervised question against the current store, is shown the gold, then edits the store; an unchanged reader is later examined on a frozen snapshot under a fixed action budget. Where offline graph construction is unsupervised, (question, answer) pairs are our labels -- and that supervision is what makes the structure cheap. Per point of corpus indexed it returns 1.6x the action saving and 1.8x the accuracy of an unsupervised entity index covering everything, using 1,913 links against its 196,112. On questions the store trained on, an unchanged reader spends 31% fewer actions at higher accuracy, and the result reproduces on an official PhantomWiki generation whose questions we did not write. To measure how far this reaches we introduce a key-coverage gradient, a probe varying how much of a question the training set touched, replacing a train/test split's pass/fail with a decay curve. Generalization proves endpoint-dependent: accuracy carries to unseen questions (+0.167 F1 where both of a question's keys were indexed, +0.100 where one was, zero where neither) while the action saving stays on trained questions. Because that decay is indexed by coverage rather than by novelty, more training extends it -- and the store is undertrained, not saturated: coverage grows linearly in new questions and stops the moment training repeats them, so a hundred questions reach a quarter of the corpus and four times as many would close the gap.
Abstract (translated)
URL
https://arxiv.org/abs/2608.21829
