TY - GEN
T1 - Multi-Hop Knowledge Editing via Critic-Guided Multi-Agent Reasoning
AU - Li, Xudong
AU - Tian, Yuhang
AU - Song, Dandan
AU - Wu, Zhijing
AU - Zhang, Shuhao
AU - Yang, Jun
AU - Huo, Yongyu
AU - Zhou, Changzhi
AU - Zhang, Xinyu
AU - Li, Chenhao
AU - Ma, Huipeng
AU - Zhang, Luan
AU - Xu, Yan
AU - Liu, Qian
N1 - Publisher Copyright:
© 2026 Association for Computational Linguistics.
PY - 2026
Y1 - 2026
N2 - Knowledge within large language models (LLMs) inevitably lags behind an evolving world, motivating knowledge editing methods that update facts without expensive retraining. In multi-hop knowledge editing, models must not only recall updated facts but also correctly propagate them through multi-hop reasoning chains. However, most existing approaches rely on unidirectional, feed-forward pipelines, decomposing questions and retrieving edited facts in a rigid hop-wise sequence. This design is brittle: a minor retrieval error or logical mismatch at an early hop can become a silent failure that cascades to the final answer without an explicit recovery mechanism. To address this limitation, we propose Critic-Guided Multi-Agent Reasoning for Knowledge Editing (CARE), a framework for closed-loop post-edit reasoning. A CRITIC agent performs chain-level verification by checking both global coherence and step-wise correctness, and triggers bounded backtracking for iterative self-correction, while a SELECTOR agent supplies high-fidelity, low-noise candidate pools from the edit store to enable effective revision. Experiments on MQuAKE-2002 and MQuAKE-hard demonstrate that CARE effectively mitigates error propagation, achieving a new state-of-the-art.
AB - Knowledge within large language models (LLMs) inevitably lags behind an evolving world, motivating knowledge editing methods that update facts without expensive retraining. In multi-hop knowledge editing, models must not only recall updated facts but also correctly propagate them through multi-hop reasoning chains. However, most existing approaches rely on unidirectional, feed-forward pipelines, decomposing questions and retrieving edited facts in a rigid hop-wise sequence. This design is brittle: a minor retrieval error or logical mismatch at an early hop can become a silent failure that cascades to the final answer without an explicit recovery mechanism. To address this limitation, we propose Critic-Guided Multi-Agent Reasoning for Knowledge Editing (CARE), a framework for closed-loop post-edit reasoning. A CRITIC agent performs chain-level verification by checking both global coherence and step-wise correctness, and triggers bounded backtracking for iterative self-correction, while a SELECTOR agent supplies high-fidelity, low-noise candidate pools from the edit store to enable effective revision. Experiments on MQuAKE-2002 and MQuAKE-hard demonstrate that CARE effectively mitigates error propagation, achieving a new state-of-the-art.
UR - https://www.scopus.com/pages/publications/105053544240
M3 - Conference contribution
AN - SCOPUS:105053544240
T3 - Proceedings of the Annual Meeting of the Association for Computational Linguistics
SP - 17259
EP - 17276
BT - 64th Annual Meeting of the Association for Computational Linguistics, Findings of the Association for Computational Linguistics
A2 - Liakata, Maria
A2 - Moreira, Viviane P.
A2 - Zhang, Jiajun
A2 - Jurgens, David
PB - Association for Computational Linguistics (ACL)
T2 - 64th Annual Meeting of the Association for Computational Linguistics, ACL 2026
Y2 - 2 July 2026 through 7 July 2026
ER -