ActiShade: Activating Overshadowed Knowledge to Guide Multi-Hop Reasoning in Large Language Models

TL;DR

Knowledge overshadowing can cause critical information to be overlooked in multi-hop reasoning, leading to cascading errors in retrieval-augmented LLMs. ActiShade tackles this by three integrated components: GaP detects overshadowed keyphrases via Gaussian perturbations, a contrastively trained dense retriever fetches documents tied to both the query and the overshadowed keyphrase, and a query formulation module generates explicit next-step queries from retrieved documents. Across HotpotQA, 2WikiMQA, and MuSiQue, ActiShade delivers significant performance gains over state-of-the-art baselines and scales with larger LLMs, while maintaining generalization across datasets. The work demonstrates that targeted activation of overlooked knowledge reduces error accumulation and improves multi-hop QA effectiveness, offering practical benefits for robust RAG systems.

Abstract

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.

Figures (13)

• Figure 1: Illustration of error accumulation caused by knowledge overshadowing.The keyphrase Gloria in the query is overshadowed, leading the LLM to generate inaccurate content, such as Te Deum. This results in the retrieval of irrelevant documents, which in turn causes LLM to generate more inaccurate content in the next-round iteration.
• Figure 2: Overview of ActiShade. ActiShade first detects the overshadowed keyphrase in the query, then retrieves relevant documents based on it, and finally formulates a new query for the next-round retrieval.
• Figure 3: Sensitivity analysis of the Gaussian noise standard deviation $\sigma$.
• Figure 4: Visualization analysis of the Gaussian noise standard deviation $\sigma$.
• Figure 5: Prompt template for relevant document selection in HotpotQA.