Scholarlink

<p>Multi-hop reasoning on knowledge graphs (KGs) has emerged as a cornerstone for inferring missing facts, offering an explainable reasoning process and accurate reasoning results via reinforcement learning (RL) paradigms. However, their reasoning effectiveness drops sharply in few-shot relations (containing only a few triples) due to insufficient training instances. Although previous works have utilized meta-reinforcement learning (MRL) and reward engineering to improve the adaptability of models in few-shot scenarios, they often overlook two critical challenges: (1) coarse entity representations that fail to capture high-order contextual dependencies essential for nuanced reasoning, and (2) inherently sparse action spaces caused by few-shot relations, which restrict the exploration of valid multi-hop paths. To address these limitations, we introduce a <b>m</b>ulti-hop reasoning method with f<b>i</b>ne-g<b>ra</b>ined entity representations and LLM-au<b>g</b>m<b>e</b>nted actions (MIRAGE), a novel framework that fuses fine-grained entity representations with an action-augmented MRL paradigm. Specifically, our framework consists of two core modules: (1) a context-aware representation module that captures query-related dependencies from neighboring entities to produce fine-grained embeddings, and (2) an action-augmented meta-reinforcement learning module that leverages large language models (LLMs) to generate supplementary actions for the agent to expand the action space. By combining fine-grained semantics with dynamically supplementary actions, MIRAGE can conduct multi-hop reasoning by effectively combining multi-step paths over few-shot knowledge graphs. Finally, comprehensive evaluations on benchmark datasets reveal that MIRAGE outperforms state-of-the-art baselines in few-shot scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-hop reasoning with fine-grained entity representations and LLM-augmented actions over few-shot knowledge graphs

  • 期刊论文

  • 发表日期：

    2026年2月28日

查看全文
  • Shangfei Zheng,
  • Yancheng Zhu,
  • Yuchao Zhang,
  • Xiaotong Nie,
  • Jian Hou

摘要

Multi-hop reasoning on knowledge graphs (KGs) has emerged as a cornerstone for inferring missing facts, offering an explainable reasoning process and accurate reasoning results via reinforcement learning (RL) paradigms. However, their reasoning effectiveness drops sharply in few-shot relations (containing only a few triples) due to insufficient training instances. Although previous works have utilized meta-reinforcement learning (MRL) and reward engineering to improve the adaptability of models in few-shot scenarios, they often overlook two critical challenges: (1) coarse entity representations that fail to capture high-order contextual dependencies essential for nuanced reasoning, and (2) inherently sparse action spaces caused by few-shot relations, which restrict the exploration of valid multi-hop paths. To address these limitations, we introduce a multi-hop reasoning method with fine-grained entity representations and LLM-augmented actions (MIRAGE), a novel framework that fuses fine-grained entity representations with an action-augmented MRL paradigm. Specifically, our framework consists of two core modules: (1) a context-aware representation module that captures query-related dependencies from neighboring entities to produce fine-grained embeddings, and (2) an action-augmented meta-reinforcement learning module that leverages large language models (LLMs) to generate supplementary actions for the agent to expand the action space. By combining fine-grained semantics with dynamically supplementary actions, MIRAGE can conduct multi-hop reasoning by effectively combining multi-step paths over few-shot knowledge graphs. Finally, comprehensive evaluations on benchmark datasets reveal that MIRAGE outperforms state-of-the-art baselines in few-shot scenarios.