In multi-hop reasoning, multi-round retrieval-augmented generation (RAG) methods typically rely on LLM-generated content as the retrieval query. However, these approaches are inherently vulnerable to knowledge overshadowing—a phenomenon where critical information is overshadowed during generation. As a result, the LLM-generated content may be incomplete or inaccurate, leading to irrelevant retrieval and causing error accumulation during the iteration process. To address this challenge, we propose ActiShade, which detects and activates overshadowed knowledge to guide large language models(LLMs) in multi-hop reasoning. Specifically, ActiShade iteratively detects the overshadowed keyphrase in the given query, retrieves documents relevant to both the query and the overshadowed keyphrase, and generates a new query based on the retrieved documents to guide the next-round iteration. By supplementing the overshadowed knowledge during the formulation of next-round queries while minimizing the introduction of irrelevant noise, ActiShade reduces the error accumulation caused by knowledge overshadowing. Extensive experiments show that ActiShade outperforms existing methods across multiple datasets and LLMs.

Yuhang Tian, Dandan Song, Zhijing Wu, Pan Yang, Changzhi Zhou, Jun Yang, Hao Wang, Huipeng Ma, Chenhao Li, Luan Zhang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

In zero-shot cross-lingual event argument extraction(EAE) task, a model is typically trained on source language datasets and then applied on task language datasets. There is a trend to regard the zero-shot cross-lingual EAE task as a sequence generation task with manual prompts or discrete prompts. However, there are some problems with these prompts, including using suboptimal prompts and difficult to transfer from source language to target language. To overcome these issues, we propose a method called SPTEAE(A Soft Prompt Transfer model for zero-shot cross-lingual Event Argument Extraction). SPTEAE utilizes a sequence of tunable vectors which are tuned in source language as event type prompts. These source language event type prompts can be transferred as target prompts to perform target EAE task by key-value selection mechanism. For each event type, SPTEAE learns a special target prompt by attending to highly relevant source prompts. Experiment results show that the average performance of SPTEAE with soft prompt transfer is 2.6% higher than the current state-of-the-art model on the ACE2005 dataset.