inversion-and-drop-as-dual-boundary
OUT derived (depth 7)
Created 2026-08-25T03:50:14+00:00 · Reviewed 2026-08-25T04:02:18+00:00
Kandpal's retrieval inversion (parametric accuracy *decreases* with document count while contextual accuracy *increases*) and ROME's multi-hop drop (40.5%→7.4% after editing) are the same geometric boundary observed from the read and write sides respectively: the rank-one addressable subspace fills the head directions parametrically but is silent on the tail, and the context window fills exactly those missing tail directions.
Justifications
SL — The Kandpal inversion shows the read-side boundary (parametric channel fades, contextual channel strengthens as a function of document count), while the ROME multi-hop drop shows the write-side boundary (rank-one editing cannot extend to composed/multi-hop directions). Together they bracket the same subspace: the set of directions the parametric channel can address.
Antecedents (all must be IN):
- OUT retrieval-inversion-as-geometric-signature — The inversion of accuracy trends—parametric accuracy *decreases* with document count while contextual accuracy *increases*—is the operational signature of the geometric addressability boundary: the inversion occurs precisely at the subspace boundary (where the fact direction exits the rank-one addressable space), not at a frequency threshold, making the boundary a geometric rather than statistical object.
- OUT parametric-write-subspace-boundary — The operational boundary between parametric recall and contextual retrieval is precisely the geometric boundary of the rank-one addressable subspace: facts whose subject-key projection aligns with the locally-stored key covariance are parametrically editable, while facts outside this subspace must be externally supplied.
- IN rome-gptj-mquake-cf-multi-hop-drop — ROME-edited GPT-J answers only 7.4% of MQuAKE-CF multi-hop questions, down from 40.5% before editing