Saved in:
Bibliographic Details
Main Authors: Li, Ruilin, Wang, Yibin, Zhu, Wenhong, Li, Chenglin, Zhang, Jinghao, Li, Chenliang, Yan, Junchi, Wang, Jiaqi
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2512.04753
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911410324242432
author Li, Ruilin
Wang, Yibin
Zhu, Wenhong
Li, Chenglin
Zhang, Jinghao
Li, Chenliang
Yan, Junchi
Wang, Jiaqi
author_facet Li, Ruilin
Wang, Yibin
Zhu, Wenhong
Li, Chenglin
Zhang, Jinghao
Li, Chenliang
Yan, Junchi
Wang, Jiaqi
contents Knowledge editing aims to update specific facts in large language models (LLMs) without full retraining. Prior efforts sought to tune the knowledge layers of LLMs, achieving improved performance in controlled, teacher-forced evaluations. However, they still encounter challenges in real-world autoregressive generation scenarios, which greatly limit their practical applicability. Our empirical analysis reveals two issues: (1) Most methods degrade pre-trained capabilities after injecting new knowledge; (2) They may exhibit a discrepancy between stored parametric knowledge and inference-time autoregressive generation behavior. To this end, we propose EtCon, an edit-then-consolidate paradigm that couples targeted edits with post-edit consolidation. Specifically, our framework comprises two stages: (1) Targeted Proximal Supervised Fine-Tuning (TPSFT) performs a constrained targeted edit to update parametric knowledge while controlling policy drift. (2) Group Relative Policy Optimization (GRPO) consolidates the edit by aligning autoregressive trajectories with the intended fact. Extensive experiments demonstrate that our EtCon improves editing reliability and real-world generalization, while better preserving pre-trained capabilities.
format Preprint
id arxiv_https___arxiv_org_abs_2512_04753
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle EtCon: Edit-then-Consolidate for Reliable Knowledge Editing
Li, Ruilin
Wang, Yibin
Zhu, Wenhong
Li, Chenglin
Zhang, Jinghao
Li, Chenliang
Yan, Junchi
Wang, Jiaqi
Computation and Language
Knowledge editing aims to update specific facts in large language models (LLMs) without full retraining. Prior efforts sought to tune the knowledge layers of LLMs, achieving improved performance in controlled, teacher-forced evaluations. However, they still encounter challenges in real-world autoregressive generation scenarios, which greatly limit their practical applicability. Our empirical analysis reveals two issues: (1) Most methods degrade pre-trained capabilities after injecting new knowledge; (2) They may exhibit a discrepancy between stored parametric knowledge and inference-time autoregressive generation behavior. To this end, we propose EtCon, an edit-then-consolidate paradigm that couples targeted edits with post-edit consolidation. Specifically, our framework comprises two stages: (1) Targeted Proximal Supervised Fine-Tuning (TPSFT) performs a constrained targeted edit to update parametric knowledge while controlling policy drift. (2) Group Relative Policy Optimization (GRPO) consolidates the edit by aligning autoregressive trajectories with the intended fact. Extensive experiments demonstrate that our EtCon improves editing reliability and real-world generalization, while better preserving pre-trained capabilities.
title EtCon: Edit-then-Consolidate for Reliable Knowledge Editing
topic Computation and Language
url https://arxiv.org/abs/2512.04753