Learning Edit-Adaptable Parameters for Multi-Hop Knowledge Propagation
LEAP is a MAML-based meta-training framework that learns edit-adaptable LoRA parameters so that ordinary single-hop knowledge edits also propagate to their multi-hop consequences, without changing the test-time editing procedure itself.
meta-training/— meta-trains the edit-adaptable LoRA adapter (meta_adapter).evaluation/— evaluates ameta_adapteron MQuAKE multi-hop edit propagation.analysis/— interpretability analysis (logit lens, activation patching) of how LEAP changes edit propagation internally.
Code is released under the MIT License (see LICENSE). Parts of evaluation/ and analysis/ are adapted from CaKE (MIT License, Copyright (c) 2025 ZJUNLP) — see THIRD_PARTY_LICENSES/ and each subdirectory's README for details.