acceptodds
Desk rejected at ICLR 2027

Revocation Before Rebuild: Semantic Closure for Dense Retrieval Under Tight Handling Windows

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

An authenticated revocation takes effect before a 10M-chunk dense index can be rebuilt: the handling window in our study is one hour, whereas Full Rebuild has a 6.8 h p95 on matched hardware. During this interval, deleting named records leaves answer-bearing mirrors, quotations, migrations, and paraphrases retrievable. We formalize the required serving state as a candidate-bounded semantic revocation closure and instantiate it in RevocEdge-10M, which spans five revocation operators, a 10.2M-chunk Wikipedia stream, and a temporally distinct technical-document stream. PRAU estimates event-time closure from provenance and embedding neighborhoods, applies operator-conditioned admissibility gating, and edits only local document representations in the frozen retrieval space. Ordinary Wikipedia churn favors Partial Rebuild by 1.0 Exposure@20 point, but mirror-heavy, bursty, and authority-skewed events reverse the ordering: PRAU lowers exposure by 2.7, 6.6, and 4.2 points while using rather than GPU-hours. The technical stream reproduces the reversal, with PRAU moving from 0.8 point behind Partial Rebuild on standard updates to 1.6–4.0 points ahead on stress slices; against SISA-Retrieval, the stress-slice advantage reaches 14.8 points on Wikipedia and 14.6 points on technical documents. PRAU publishes in 46.7 min p95 with % one-hour completion and reduces hard-slice answer disclosure to % while retaining % lawful-answer support. These results identify semantic reach as the retrieval-maintenance object that separates efficient updates before rebuild.

Primary area:
foundation or frontier models, including LLMs
Keywords:
retrieval rag and knowledge dynamics, Revocation Before Rebuild

Related papers

Discussion (0)

Sign in to comment.