Co-Evolving Graph and Text Memory for Training-Free Multi-Hop Question Answering
Proposes Co-E system with bidirectional graph-text memory for training-free multi-hop QA, outperforming baselines.
Key Findings
Methodology
The Co-E system employs a synchronized bidirectional graph-text working memory, integrating a four-stage cycle: consolidating textual evidence, extracting relational triples, injecting facts into graph memory, and re-consolidating text. Each step involves retrieving evidence from text and knowledge bases, filtering via entity linking and relation extraction, then merging new triples into the graph. These graph facts are converted into natural language and injected back into the textual memory, creating a feedback loop. The system supports multiple reasoning strategies, including Monte Carlo Tree Search (MCTS) and chain-of-thought (CoT), leveraging shared memory for dynamic path selection and evidence refinement without training. This design enables continuous, mutual revision of graph and text, improving multi-hop reasoning accuracy.
Key Results
- On six benchmarks, Co-E (MCTS) achieved 72.6% EM on 2WikiMultiHopQA, 70.0% EM on Bamboogle, and 74.9% Hits@1 on CWQ, surpassing comparable open-backbone baselines and rivaling larger trained models. It especially excels in tasks requiring long relation chains and bridge entity maintenance, outperforming several state-of-the-art methods.
- In KGQA, it scored 85.5% Hits@1 on WebQSP and 74.9% on CWQ, demonstrating robustness in structured knowledge reasoning. Ablation studies confirmed that bidirectional synchronization and active memory maintenance are crucial for performance gains.
- Ablation results indicated that removing either memory modality or synchronization significantly degrades accuracy, emphasizing the importance of the coupled graph-text memory loop.
Significance
This work addresses a fundamental bottleneck in multi-hop reasoning—the dynamic, mutual updating of relational and textual evidence—without requiring training. It introduces a flexible, scalable framework that enhances reasoning coherence and accuracy, with broad implications for knowledge-intensive AI applications. By enabling continuous, real-time revision of reasoning states, it paves the way for more adaptive and interpretable AI systems capable of complex multi-step inference.
Technical Contribution
The core technical innovation is the bidirectional synchronization mechanism, which integrates four phases—consolidation, extraction, injection, and re-consolidation—within a shared memory structure. This allows the model to adaptively expand and refine its reasoning graph while maintaining textual context, all without training. The system’s design supports multiple inference strategies, including Monte Carlo Tree Search and chain reasoning, making it versatile and effective across diverse tasks.
Novelty
This is the first demonstration of a training-free, dynamically synchronized graph-text memory system for multi-hop QA, emphasizing mutual, step-by-step evolution of reasoning states. Unlike prior works that rely on static graphs or unidirectional retrieval, this approach actively couples textual and relational information, enabling continuous refinement and expansion of reasoning chains in real-time.
Limitations
- The system's performance heavily depends on the accuracy of entity linking and relation extraction; errors in these steps can propagate and reduce reasoning quality, especially in noisy or ambiguous contexts.
- Handling extremely long or complex relation chains may lead to memory overload or information loss, limiting scalability.
- Current implementation does not incorporate multi-modal data, which could further enhance reasoning in real-world applications. Future work should address these limitations to improve robustness and generalization.
Future Work
Future directions include integrating multi-modal data such as images and audio, improving extraction robustness, and optimizing the synchronization process for longer reasoning chains. Additionally, combining this framework with few-shot learning or fine-tuning could further enhance adaptability to domain-specific tasks, broadening its practical deployment.
AI Executive Summary
Multi-hop question answering (QA) challenges AI systems to perform complex reasoning over multiple steps, integrating relational and textual evidence dynamically. Traditional methods often rely on static knowledge graphs or retrained models, limiting flexibility and scalability. This paper introduces Co-E, a novel, training-free system that employs a bidirectional graph-text working memory, enabling continuous mutual refinement during inference.
The core innovation lies in a four-stage synchronization cycle—consolidating textual evidence, extracting relational triples, injecting facts into a graph memory, and re-consolidating text—forming a feedback loop that dynamically expands and refines the reasoning state. This mechanism allows the system to adaptively incorporate new information, prune irrelevant data, and maintain a coherent reasoning process. Supported by multiple inference strategies, including Monte Carlo Tree Search and chain-of-thought, Co-E effectively handles tasks requiring long relation chains and bridge entities.
Experimental results across six benchmarks demonstrate that Co-E significantly outperforms comparable zero-shot baselines, achieving 72.6% EM on 2WikiMultiHopQA, 70.0% on Bamboogle, and 74.9% on CWQ, rivaling larger trained models. Ablation studies confirm the importance of bidirectional synchronization and active memory maintenance, with performance drops observed when these components are removed. The approach’s flexibility, efficiency, and robustness suggest broad applicability in knowledge-intensive AI tasks.
This work marks a substantial step toward scalable, adaptable, and interpretable multi-hop reasoning systems, addressing longstanding challenges in dynamic evidence integration. Despite current limitations related to extraction errors and scalability, future enhancements could incorporate multi-modal data and domain adaptation, paving the way for more intelligent and autonomous AI systems in real-world applications.
Deep Analysis
Background
多跳问答作为知识推理的核心任务，近年来随着大型预训练模型的发展逐渐成为研究热点。早期方法多依赖预定义的知识图谱（如Freebase、Wikidata）进行结构化推理，代表算法包括路径搜索和图神经网络（GNN）等。近年来，检索增强生成（RAG）模型通过结合文本检索提升知识覆盖，但受限于单向检索和静态图结构，难以应对复杂推理链。多跳问答的核心难点在于证据的动态整合和推理状态的持续更新，传统方法多采用逐步检索或预先构建的子图，缺乏灵活性和实时性。近年来，Memory-augmented和Graph-augmented方法试图解决这一问题，但多为单向或静态维护，未实现图与文本的双向动态同步。本文在此基础上提出同步双向记忆机制，旨在实现推理状态的持续修正和知识的动态扩展，推动多跳问答向更高效、更灵活的方向发展。
Core Problem
多跳问答的核心难题在于如何在推理过程中动态整合关系图和文本证据，避免信息孤岛和静态限制。传统模型多依赖预定义图或单向检索，导致推理链断裂或信息遗漏，难以应对复杂、多层次的推理需求。尤其在涉及长链关系或桥实体时，单一记忆结构难以保持信息一致性和连续性。现有方法在推理过程中缺乏有效的反馈机制，无法实现图和文本的双向修正，限制了模型的推理深度和准确率。因此，设计一种能够实时同步、动态修正推理状态的机制成为亟待解决的问题。
Innovation
本文的创新点主要在于提出同步双向的图-文本工作记忆机制，结合四阶段同步循环，实现无训练、多跳推理中的动态记忆更新。具体包括：• 通过整合文本证据，提取关系三元组，注入图记忆，形成闭环；• 利用实体链接和关系抽取，将文本信息转化为关系三元组，扩展推理边界；• 将图事实转化为自然语言，注入文本记忆中，增强上下文；• 引入多策略推理（MCTS和链式推理），利用共享记忆引导路径选择，避免预定义图结构限制。这些创新共同实现了推理状态的持续修正和知识的动态扩展，显著提升多跳问答的效果。
Methodology
- �� 输入问题和上下文，生成子查询；• 从文本和知识库中检索证据流（文本片段和关系三元组）；• 通过实体链接和关系抽取，将文本证据转化为关系三元组，合并到图记忆中，扩展推理边界；• 将存留的图三元组自然语言化，注入文本记忆，增强上下文；• 在每个推理步骤中，利用同步操作（整合文本、关系提取、注入图、再整合文本）实现双向记忆修正；• 支持多策略推理（MCTS和链式推理），利用共享记忆引导路径选择和结果优化；• 最终输出答案，并更新记忆状态，形成闭环推理。
Experiments
采用六个公开基准（包括WebQSP、CWQ、2WikiMultiHopQA、HotpotQA、MuSiQue、Bamboogle），评估模型的多跳问答性能。对比训练模型和无训练基线，采用EM和Hits@1指标。通过消融实验验证双向同步、记忆机制和检索策略的重要性。参数设置包括Qwen3-8B作为推理基础模型，关系抽取和检索采用特定算法，确保公平比较。实验还分析不同同步策略和记忆维护对性能的影响，验证系统在复杂推理场景中的优势。
Results
在六个基准上，Co-E（MCTS）在2WikiMultiHopQA达72.6%的EM，在Bamboogle达70.0%，在CWQ达74.9%的Hits@1，均优于无训练基线。特别是在桥实体和长链关系任务中表现优越，超越多种基于图和文本的对比方法。消融实验显示，双向同步和主动记忆维护是性能提升的关键，单向或缺少维护会导致性能显著下降。结果验证了同步机制在复杂推理中的有效性，为无训练问答提供新思路。
Applications
该系统适用于需要高精度、多跳推理的知识问答场景，如智能助手、自动问答系统、知识库维护等。无需大量训练数据，适合资源有限环境，能动态扩展知识边界，提升问答的准确性和灵活性。未来可结合多模态信息，实现跨模态推理，推动行业智能化升级。
Limitations & Outlook
系统依赖高质量的实体链接和关系抽取，抽取错误会影响整体效果。面对极长链或复杂关系时，记忆同步可能导致信息冗余或遗漏。当前未充分融合多模态信息，未来需增强鲁棒性和泛化能力。此外，推理过程计算成本较高，需优化算法以适应大规模应用。
Plain Language Accessible to non-experts
想象你在厨房里做饭。每次你拿到食材（文本信息）时，会根据食谱（知识图）来判断怎么搭配。以前，厨师只会用一份食谱，按照固定步骤做菜，但如果遇到新食材或特殊口味，就会出错。现在，厨师用一种聪明的方法：每次尝试后，都能根据实际情况调整食谱（图）和用料（文本），两者互相学习和修正。这样，不管遇到什么新菜，都能灵活应对，做出美味佳肴。这个系统就像厨师一样，能不断调整食谱和材料，确保每次都做出最合适的菜肴，完全不用提前准备好所有的菜谱。
ELI14 Explained like you're 14
想象你在玩一个超级复杂的拼图游戏。每次你拼完一块，就会根据拼图的样子，决定下一块应该放在哪里。有时候，你发现一块拼图可以帮你找到隐藏的线索，但你也需要记住之前拼的部分，才能继续拼下去。这个系统就像你一样，不仅记住每一块拼图，还会根据拼图的样子，调整下一步的策略。它可以不断修正自己，确保每次拼图都越来越接近完整。这样，无论拼多复杂的图片，它都能一步步搞定，甚至不用提前知道全部拼图的样子。
Glossary
同步双向记忆 (Bidirectional Memory Synchronization)
一种同时更新图和文本记忆的机制，确保两者在推理过程中相互修正。技术上通过四阶段循环实现，包括整合、抽取、注入和再整合。
论文中的核心机制，用于动态修正推理状态。
多跳问答 (Multi-Hop Question Answering)
一种需要多轮推理、逐步整合证据以得出答案的任务。涉及长链关系和桥实体，难度较高。
系统设计和性能评估的主要任务类型。
关系三元组 (Relational Triples)
由实体和关系组成的结构化信息，用于表示知识图中的关系。抽取自文本或知识库，支持推理扩展。
构建图记忆的基础单元。
共享记忆 (Shared Memory)
在推理过程中同时维护文本和图的记忆结构，支持双向信息流动和动态修正。
系统实现的核心技术。
MCTS (蒙特卡洛树搜索)
一种探索多路径推理的搜索策略，通过模拟和评估选择最优路径。适用于复杂推理树的路径选择。
多策略推理中的路径搜索方法。
Open Questions Unanswered questions from this research
- 1 如何进一步提升同步机制在极长链推理中的效率和准确性，尤其在多模态信息融合场景下的应用尚未充分探索。
Applications
Immediate Applications
智能问答系统
可用于企业客服、智能助手等场景，提升多轮复杂问答的准确性和效率，无需大量训练数据，适应多变知识环境。
知识库维护
自动扩展和修正知识图，支持动态知识更新和推理，降低人工维护成本。
Long-term Vision
通用智能推理平台
结合多模态信息，实现跨领域、跨模态的动态推理，推动AI在科学研究、医疗诊断等领域的深度应用。
Abstract
Multi-hop question answering requires coordinating relational and textual evidence across reasoning steps, a combination neither a text corpus nor a knowledge graph can supply alone. Prior work often emphasizes only part of this loop: graph-augmented RAG retrieves from a pre-built or query-updated graph, KGQA systems search within topic-centered subgraphs, and memory-augmented agents maintain evolving memories without continuously reconciling graph memory with textual context. We propose Co-E, a training-free system built around synchronized bidirectional graph-text working memory. A synchronization cycle consolidates textual memory, extracts relational triples into graph memory, and injects graph facts back into the generation context. Because both memories are maintained, they shape subsequent retrieval and generation. Evaluated on six multi-hop QA benchmarks, Co-E improves over comparable training-free open-backbone baselines and is competitive with larger or trained systems.