Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Validator-Guided Decoding for Multi-Hop Knowledge Editing

AbstractKnowledge editing aims to update specific factual knowledge in large language models without full retraining. Although existing editors can often modify local factual associations, edited models still struggle to apply updated knowledge in multi-hop reasoning. In this work, we study this problem through the lens of the generate-verification gap: a model may fail to directly generate the correct updated answer, while a verifier conditioned on the edit facts can still judge whether a proposed answer is compatible with the updated knowledge. Motivated by this observation, we propose Validator-Guided Decoding (VGD), a test-time method that uses edit facts as verification memory to improve answer generation after knowledge editing. VGD employs a trained validator as an external correctness signal over candidate answers, allowing the model to retain reliable outputs while selectively revising uncertain or inconsistent predictions. It is editor-agnostic and can be applied to base models or on top of existing editors without modifying the editor or further updating the edited model. Experiments on multi-hop benchmarks show that VGD improves hop-level consistency and final-answer accuracy, generally outperforming ICL and CoT baselines. Our code will be available at https://github.com/JoveReCode/VGD.

More from EMNLP 2026

All EMNLP 2026 papers →