DeepEdit: Knowledge Editing as Decoding with Constraints

We propose a new perspective of knowledge editing (KE) for large language models (LLMs) that treats it as a constrained decoding problem. We design decoding constraints to regulate LLMs, ensuring coherence between reasoning steps when incorporating new knowledge. To enforce these constraints, we utilize a depth-first search to adaptively substitute new knowledge for the LLMs' original reasoning steps, greedily seeking the optimal path of multi-hop reasoning with new knowledge. From this vantage, we propose DEEPEDIT: Depth-first Search-based Decoding for Knowledge Editing. DEEPEDIT improves the KE of LLMs by enhancing the conciseness, coherence, pertinence, and receptiveness of reasoning with new knowledge. DEEPEDIT is flexibly applicable to any black-box LLM without requiring access to model parameters or token-wise distributions. In addition to DEEPEDIT, we propose two new KE benchmarks: MQuAKE-2002 and MQuAKE-hard, which are designed to provide more precise and challenging assessments of KE approaches. Qualitatively, DEEPEDIT enables LLMs to produce more succinct reasoning outputs in accordance with new knowledge. Quantitatively, it yields significant improvements on multiple KE benchmarks.