ReCore: Reflective Cooperative Reasoning Framework for Knowledge-Intensive Question Answering Tasks

ACL ARR 2026 January Submission7196 Authors

Keywords: Knowledge-intensive Question Answering;Large Language Models;Multi-Document Reasoning;Information Extraction

Abstract: Knowledge-intensive question-answering (QA) aim to collect dispersed knowledge from a large number of documents related to the question and derive answers through reasoning. However, traditional methods often suffer from information loss and difficulty in establishing logical connections across documents, making reasoning challenging. To address these issues, we propose Reflective Cooperative Reasoning (ReCore), a framework that combines Role-Based Multi-Perspective Collaboration with Reflective Knowledge Refinement.The former dynamically generates multiple roles based on the task and integrates their insights through a voting mechanism to identify the optimal structuring strategy. This approach facilitates the organized, concise extraction of relevant information and helps reveal relationships among fragmented content from multiple perspectives. The latter introduces a reflective construction mechanism to perform intensive reading over related documents, providing richer background knowledge for the task and enabling further refinement, segmentation, and simplification of useful content. This process mitigates information loss and enhances reasoning effectiveness. Experiments show that ReCore outperforms baselines, achieving a 12-point gain in average answer score (LLM Score) and a 27-percentage-point improvement in Exact Match (EM). It excels in complex tasks such as spotlight locating, comparison, clustering, and chain-of-thought reasoning.

Paper Type: Long

Research Area: Information Extraction and Retrieval

Research Area Keywords: document-level extraction;passage retrieval; dense retrieval; document representation;

Contribution Types: NLP engineering experiment

Languages Studied: English, Chinese

Submission Number: 7196