Reproduce “Co-Evolving Graph and Text Memory for Training-Free Multi-Hop Question Answering” on OpenResearch

About this paper

Hieu Man, Thien Huu Nguyen

Researchers at the University of Oregon present Co-E, a training-free system that dynamically co-evolves graph and textual memory for multi-hop question answering. This approach achieves competitive performance across various KGQA and text-QA benchmarks, including 74.9 Hits@1 on CWQ and 72.6 EM on 2WikiMultiHopQA, outperforming comparable training-free open-backbone baselines.

Key insights

• Bidirectional, within-step co-evolution of textual and graph memory is crucial for dynamically updating the reasoning state in complex multi-hop question answering.
• Active memory maintenance, including consolidation, deduplication, retraction, and pruning of both text snippets and graph triples, is essential for coherence and to prevent memory bloat, proving more effective than monotonic accumulation.
• Integrating both textual evidence (atomic snippets, intermediate responses) and graph facts (KB-schema and open-vocabulary triples) in a unified, continuously updated memory provides complementary information for robust reasoning across diverse QA settings.

Summary by alphaXiv.