acceptodds
Desk rejected at ICLR 2027

Revision Collisions in Sequential Model Editing: Family-Aligned Write Paths Reduce Factual Interference

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

Abstract

Sequential corrections can remain individually successful while collectively corrupting unrelated facts: after 100 Llama-3.1-8B revisions, a single low-rank write path sustains 95.8% edit success yet adds 32.4 percentage points of net exact-match error on fixed, unedited factual probes. Write-path decomposition exposes the source of this interference by fixing MEMIT targets, edited layers, request order, summed read semantics, and a 2,752,512-parameter online budget while varying writable organization; among partitioned policies, online destination assignment is the experimental variable. RevSep routes temporal, relational, and attribute revisions to family-specialized paths with shared overflow, while random, label-shuffled, route-blind, and shared controls separate semantic placement from generic dispersion. At matched partition geometry, RevSep reduces edit-100 probe error from 30.6 to 27.1 points versus random assignment, a paired mean shift of points (95% bootstrap interval ), and from 31.3 to 27.1 versus label-shuffled assignment ( []), while retaining 96.2% edit success. Relative to the shared path, the reduction is 5.3 points; the shared-to-random-to-aligned ordering also appears with zero subject–relation–template overlap, on Mistral-7B-v0.3, and in mixed-family histories, whereas route corruption and 91% occupancy progressively compress the gain. Write destination is therefore a controllable source of lifelong-edit retention: matching updates to specialized writable state protects factual behavior beyond equal-budget load dispersion.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
memory editing and lifelong learning, Revision Collisions in Sequential Model Editing

Related papers

Discussion (0)

Sign in to comment.