Paper deep dive
Cognition on Graph: Navigating Massive Knowledge Space via Cognitive Cycles and Bidirectional Graph-Text Synergy
Gengxian Zhou, Jian Xu, Zichen Tang, Shiming Xiang, Haihong E, Cheng-Lin Liu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 9/19/2026, 3:03:18 AM
Summary
The paper introduces CoG (Cognition on Graph), a training-free framework for adaptive knowledge exploration in Retrieval-Augmented Generation (RAG). CoG addresses limitations in existing GraphRAG methods by implementing a cognitive-inspired plan-explore-reflect cycle. It establishes bidirectional synergy between structured Knowledge Graphs (KGs) and unstructured text corpora, using entities extracted from text to dynamically guide graph exploration. This approach allows for proactive planning, dual-source retrieval, and dynamic strategy adjustment via reflection, significantly outperforming state-of-the-art methods on multi-hop QA benchmarks.
Entities (20)
Relation Signals (11)
Gengxian Zhou → affiliatedwith → Beijing University of Posts and Telecommunications
confidence 95% · Gengxian Zhou Affiliation: Beijing University of Posts and Telecommunications
Haihong E → affiliatedwith → Beijing University of Posts and Telecommunications
confidence 95% · Haihong E ... Affiliation: Beijing University of Posts and Telecommunications
Jian Xu → affiliatedwith → Chinese Academy of Sciences
confidence 95% · Jian Xu Affiliation: MAIS, Institute of Automation, Chinese Academy of Sciences
CoG → establishes → bidirectional synergy
confidence 95% · Crucially, it establishes deep bidirectional synergy between structured graph and unstructured text
CoG → uses → cognitive cycle
confidence 95% · CoG performs a continuous plan-explore-reflect cycle
CoG → evaluatedon → multi-hop QA benchmarks
confidence 90% · Extensive experiments on seven multi-hop QA benchmarks demonstrate that CoG significantly outperforms
CoG → outperforms → State-of-the-Art Methods
confidence 90% · CoG significantly outperforms state-of-the-art methods while achieving superior exploration efficiency.
VectorRAG → retrievesvia → vector similarity
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Retrieval-Augmented Generation (RAG) has empowered Large Language Models (LLMs) to tackle knowledge-intensive tasks. However, navigating global, heterogeneous knowledge bases (large-scale knowledge graphs and text corpora) for complex reasoning remains a challenge. Existing methods typically employ reactive, graph-driven exploration strategies, which blindly follow graph topology without adapting to the question context or evolving exploration progress, and lack deep bidirectional synergy between graph and text. To address these limitations, we propose CoG (Cognition on Graph), a cognitive-inspired, training-free framework for adaptive knowledge exploration. Drawing inspiration from human problem-solving, CoG performs a continuous plan-explore-reflect cycle, where it proactively formulates investigation plans, performs dual-source retrieval, and dynamically reflects on progress to adjust strategies. Crucially, it establishes deep bidirectional synergy between structured graph and unstructured text, where entities extracted from text dynamically guide graph exploration to bridge knowledge gaps. Extensive experiments on seven multi-hop QA benchmarks demonstrate that CoG significantly outperforms state-of-the-art methods while achieving superior exploration efficiency. Our code and datasets are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2609.12791v1
- Canonical: https://arxiv.org/abs/2609.12791v1
Trouble viewing inline? Open PDF directly →
Full Text
113,969 characters extracted from source content.
Expand or collapse full text
Cognition on Graph: Navigating Massive Knowledge Space via Cognitive Cycles and Bidirectional Graph-Text Synergy Gengxian Zhou Affiliation: Beijing University of Posts and Telecommunications Affiliation: Zhongguancun Academy, Beijing Email: liucl@nlpr.ia.ac.cn Jian Xu Affiliation: MAIS, Institute of Automation, Chinese Academy of Sciences Affiliation: School of Artificial Intelligence, University of Chinese Academy of SciencesCorrespondence:ehaihong@bupt.edu.cn Zichen Tang Affiliation: Beijing University of Posts and Telecommunications Shiming Xiang Affiliation: MAIS, Institute of Automation, Chinese Academy of Sciences Affiliation: School of Artificial Intelligence, University of Chinese Academy of SciencesCorrespondence:ehaihong@bupt.edu.cn Haihong E †thanks: Corresponding authors. Affiliation: Beijing University of Posts and Telecommunications Cheng-Lin Liu11footnotemark: 1 Affiliation: Zhongguancun Academy, Beijing Affiliation: MAIS, Institute of Automation, Chinese Academy of Sciences Affiliation: School of Artificial Intelligence, University of Chinese Academy of SciencesCorrespondence:ehaihong@bupt.edu.cn Abstract Retrieval-Augmented Generation (RAG) has empowered Large Language Models (LLMs) to tackle knowledge-intensive tasks. However, navigating global, heterogeneous knowledge bases (large-scale knowledge graphs and text corpora) for complex reasoning remains a challenge. Existing methods typically employ reactive, graph-driven exploration strategies, which blindly follow graph topology without adapting to the question context or evolving exploration progress, and lack deep bidirectional synergy between graph and text. To address these limitations, we propose CoG (Cognition on Graph), a cognitive-inspired, training-free framework for adaptive knowledge exploration. Drawing inspiration from human problem-solving, CoG performs a continuous plan-explore-reflect cycle, where it proactively formulates investigation plans, performs dual-source retrieval, and dynamically reflects on progress to adjust strategies. Crucially, it establishes deep bidirectional synergy between structured graph and unstructured text, where entities extracted from text dynamically guide graph exploration to bridge knowledge gaps. Extensive experiments on seven multi-hop QA benchmarks demonstrate that CoG significantly outperforms state-of-the-art methods while achieving superior exploration efficiency. Our code and datasets are available at https://github.com/zhougengxian/CoG. 1 Introduction While Large Language Models (LLMs) excel in various tasks, their reliance on static parametric knowledge causes factual hallucinations and opaque reasoning in knowledge-intensive scenarios Huang et al. (2025b); Li et al. (2025b); Wang et al. (2026). Retrieval-Augmented Generation (RAG) mitigates this by grounding response in external knowledge sources to improve accuracy and traceability Zhang et al. (2025b); Song et al. (2025). Figure 1: Comparison of exploration strategies. CoG integrates all prior capabilities in a cognitive cycle, uniquely establishing a bidirectional synergy where text actively bridges graph gaps, enabling robust navigation in global knowledge bases. Effective retrieval is the core of RAG. VectorRAG Lewis et al. (2020) retrieves text chunks via vector similarity, while GraphRAG Gutiérrez et al. (2025); Luo et al. (2025a); Gutiérrez et al. (2024); Yang et al. (2025); Wu et al. (2026); Dong et al. (2026) constructs Knowledge Graphs (KGs) for structural guidance. However, existing methods predominantly operate in simplified, local settings. They restrict the knowledge scope to pre-selected oracle paragraphs known to contain answers, rather than raw, complete entity documents. This masks the inherent noise of real-world retrieval, failing to assess true robustness Luo et al. (2025b). When transitioning to the global setting, the core challenge escalates from simple information retrieval to complex knowledge exploration. It requires systems to progressively construct evidence chains by synthesizing structural associations from KGs and semantic information from text corpora. Throughout this process, the systems must effectively navigate the noise, ambiguity and distractions inherent in large-scale knowledge bases. Existing works attempting to address this challenge, however, fall short in two critical aspects (illustrated in Figure 1). First, exploration remains heavily KG-dependent with insufficient text utilization. Most methods rely solely on KG traversal for exploration Sun et al. (2024); Chen et al. (2024b); Ma et al. (2025a); Wang et al. (2025). Even hybrid methods Ma et al. (2025b) merely use text as auxiliary context, failing to leverage entities within text to actively guide the exploration, leaving the inherent sparsity and incompleteness of KGs unaddressed. Second, the exploration strategies are fragile and rigid. Relying on pre-annotated entities as rigid starting points Xu et al. (2026) and lacking mechanisms for backtracking Chen et al. (2024b) or proactive adjustment Ma et al. (2025a), these systems are prone to getting lost in irrelevant subgraphs or stranded by linking failures or knowledge gaps. To bridge these gaps, we draw inspiration from the cognitive process of human problem-solving. Facing complex questions in open domains, humans do not blindly traverse information; instead, they engage in a dynamic cycle of planning, exploration, and reflection. They proactively formulate investigation plans, flexibly synthesize information from diverse sources, and continuously reflect on accumulated evidence to adjust their strategy. Guided by this mechanism, we propose CoG (Cognition on Graph), a framework that instantiates this human-like reasoning process into a continuous cognitive loop comprising Planning, Exploration, Synthesis, and Reflection. In the Planning phase, instead of relying on pre-annotated entities, CoG proactively analyzes the question to formulate an investigation strategy, from which it derives exploratory sub-queries and identifies relevant entities to anchor the search. During the Exploration phase, CoG performs dual-source knowledge gathering. It retrieves 1-hop relevant subgraphs from the KG while simultaneously conducting hierarchical reading of entity-related documents in the text corpus, capturing both structural relations and rich contextual details. The Synthesis module then aggregates evidence from both sources into a unified retrieval summary, distilling key insights that advance the reasoning process and achieving deep bidirectional synergy between structured and unstructured knowledge. Finally, the Reflection module evaluates the utility of the retrieved information: for productive explorations, it updates the strategy and generates the next query based on the progress made; for unproductive ones, it diagnoses the failure using the global interaction history and adjusts the course, ensuring robust navigation through vast knowledge spaces. Our main contributions are threefold: • We propose CoG, a cognitive-inspired, closed-loop, training-free GraphRAG framework. It leverages continuous cognitive cycles of plan-explore-reflect for adaptive knowledge exploration in noisy global knowledge bases. • We establish deep bidirectional synergy between structured and unstructured knowledge, where entities dynamically extracted from text actively guide graph exploration to effectively address KG incompleteness. • Extensive experiments on seven knowledge-intensive, multi-hop QA benchmarks demonstrate that CoG significantly outperforms state-of-the-art methods, showcasing superior reasoning capability and robustness in navigating large-scale knowledge bases. 2 Related Work Method Scope Text Corpus Knowledge Graph Key Capabilities Exploration Strategy #Words #Ent. Source #Ent. #Trip. Text Guide Refl. & Bk. Pro. Plan. Adapt. EL. VectorRAG Local 1.5M 21K – – – – – – – Dense Retrieval GraphRAG Series† Local 0.9M 12K Constr.‡ 97K 1.4M – – – – Graph-Guided Retrieval StructGPT, KG-Agent Global – – Freebase 39M 1.9B ✗ ✗ ✗ ✗ Interaction History ToG, ReKnoS Global – – Wikidata 115M 1.7B ✗ ✗ ✗ ✗ Beam Search PoG Global – – Freebase 39M 1.9B ✗ ✓ ✗ ✗ Beam Search w/ Backtrack DoG Global – – Freebase 39M 1.9B ✗ ✗ ✓ ✗ Multi-Agent Debate AnchorRAG Global – – Freebase 39M 1.9B ✗ ✗ ✗ ✓ Parallel Anchor Exploration KERAG Global – – Wikidata 115M 1.7B ✗ ✗ ✗ ✓ Schema-Guided Expansion ToG-2 Global 5B 7.1M Wikidata 115M 1.7B ✗ ✗ ✗ ✗ Beam Search CoG (Ours) Global 5B 7.1M Wikidata 115M 1.7B ✓ ✓ ✓ ✓ Cognitive Cycle Table 1: Comparison of retrieval and exploration methods. We report knowledge base scales as total Words and Entity documents for Text, versus Entities and Triples for KG (Units: K=Thousand, M=Million, B=Billion). CoG uniquely operates on a massive scale while integrating all key exploration capabilities: Text Guide (Text entity guides graph exploration), Refl. & Bk. (Reflection & Backtrack), Pro. Plan. (Proactive planning), and Adapt. EL. (Adaptive entity linking). † : Includes HippoRAG, HippoRAG2, GFM-RAG, et al. ‡ : Graphs constructed from local text corpora. ✓: Present, ✗: Absent, “–”: not applicable. RAG enhances LLMs by retrieving external knowledge. Beyond VectorRAG Lewis et al. (2020), GraphRAG methods construct structured indices like knowledge graphs Gutiérrez et al. (2024); Guo et al. (2024); Gutiérrez et al. (2025); Zhu et al. (2025), hierarchical communities Edge et al. (2024); Huang et al. (2025a), or document graphs Wang et al. (2024); Chen et al. (2025); Li et al. (2024a) to guide retrieval. Recent advancements employ reinforcement learning Luo et al. (2026); Yu et al. (2026) or hybrid dense retrieval Lei et al. (2025) over graphs. However, these works primarily operate in local settings, restricting the knowledge base to highly relevant oracle documents or domain-specific KGs. This simplifies the exploration challenge, failing to assess robustness in noisy global environments. In contrast, CoG addresses the realistic and massive scale of global knowledge bases (Wikidata and Wikipedia), navigating an exponentially larger and noisier search space. For global knowledge exploration, prior works focus on iterative KG reasoning. Path-based methods like ToG Sun et al. (2024), StructGPT Jiang et al. (2023), and KG-Agent Jiang et al. (2025) perform prompt-driven exploration, while ReKnoS Wang et al. (2025) and KERAG Sun et al. (2025) improve coverage via super-relations or schemas. Strategy-based methods introduce specific heuristics, such as backtracking in PoG Chen et al. (2024b) or multi-agent debate and parallelism in DoG Ma et al. (2025a) and AnchorRAG Xu et al. (2026). However, these approaches treat advanced reasoning behaviors as fragmented add-ons to fixed pipelines. Meanwhile, learning-based methods Chang et al. (2026); Lin et al. (2026); Cui et al. (2025) train specialized navigation modules, introducing high training costs and instability while lacking explicit interpretability. Distinguishing itself from these rigid pipelines and costly training, CoG introduces a training-free, closed-loop cognitive cycle. It natively unifies “plan-explore-reflect”, enabling proactive planning and adaptive error-correction to emerge autonomously. Despite advanced reasoning, KG-centric methods remain vulnerable to KG sparsity and incompleteness. To address this, hybrid methods integrate KGs with text corpora Lee et al. (2025); Sarmah et al. (2024); Tan et al. (2025). CoK Li et al. (2024b) verifies reasoning across multi-source bases, while ToG-2 Ma et al. (2025b) uses graph entities to anchor document retrieval. Other approaches utilize text merely for relevance scoring Lei et al. (2025) or rely on pre-constructed hyperedges Luo et al. (2026). However, these methods treat text as passive context or require rigid offline constructions, lacking dynamic structural guidance. In contrast, CoG introduces a deep bidirectional synergy where unstructured text actively guides graph traversal. By extracting entities from text to dynamically bridge KG structural gaps, it achieves robust exploration in noisy global spaces. Table 1 provides a systematic comparison of these methods across key capabilities. 3 Method 3.1 Task Definition Figure 2: Overview of the CoG framework. It employs a continuous cognitive cycle of plan-explore-reflect to adaptively navigate structured and unstructured knowledge for complex QA. We focus on the task of multi-hop Question Answering (QA) over a global, hybrid knowledge base. Given a complex question q, the system must construct an evidence chain by iteratively retrieving and reasoning over heterogeneous knowledge sources to generate the final answer a. The knowledge base =,K=\G,D\ comprises two complementary components: Structured Knowledge Graph =(ℰ,ℛ,)G=(E,R,T), where ℰE and ℛR denote the sets of entities and relations. The KG provides structured information in two forms: 1) relational fact triples =(h,r,t)∣h,t∈ℰ,r∈ℛT=\(h,r,t) h,t ,r \ that connect pairs of entities, and 2) entity attributes (e.g., date of birth, gender) that describe the properties of individual entities. Unstructured Text Corpus D, a large-scale collection of documents. Each document d∈d corresponds to an entity and is hierarchically structured, containing a summary paragraph, a table of contents, and the full text organized into sections. It provides rich, fine-grained semantic descriptions that complement the factual statements in G. 3.2 CoG Framework Figure 2 illustrates the overall architecture of CoG. Given a question q, CoG iteratively executes a cognitive cycle until sufficient evidence is gathered to generate an answer. Each cycle consists of four tightly integrated phases: (1) Planning formulates an investigation strategy and generates a sub-query; (2) Exploration retrieves evidence from both KG and text; (3) Synthesis aggregates and distills the retrieved information; (4) Reflection evaluates progress and decides the next action. Throughout this process, CoG maintains three long-term memory components to enable coherent multi-turn reasoning: a Notebook N to accumulate verified facts, a Candidate Entity Pool P to manage promising but deferred exploration leads, and an Interaction History ℋH to enable global reflection and strategic adjustment. 3.2.1 Planning The Planning Module proactively analyzes the question and formulates an initial exploration strategy. Unlike methods that rely on pre-annotated entities, CoG dynamically derives a strategic plan directly from the question q. This process comprises three key components: Question Analysis. It analyzes the question q to identify information needs and reasoning dependencies, determining what knowledge must be gathered first. This analysis A0A_0 serves as a high-level guide for subsequent actions. Query Decomposition. Based on the analysis, the question is decomposed into a set of focused sub-queries Q0=q1(0),…,qB(0)Q_0=\q_1^(0),…,q_B^(0)\, each targeting a specific information need that can be explored independently as the immediate next step. Entity Identification. It identifies a list of anchor entities E0=e1(0),…,eB(0)E_0=\e_1^(0),…,e_B^(0)\ corresponding one-to-one with Q0Q_0. Each ei(0)e_i^(0) serves as the core anchor point for initiating the exploration of query qi(0)q_i^(0) in both the KG and text corpus. Formally, the initial plan P0P_0 is generated by: P0=(A0,Q0,E0)=Promptinit(q,LLM).P_0=(A_0,Q_0,E_0)=Prompt_init(q,LLM). (1) The prompt is provided in Appendix K, Table 16. 3.2.2 Exploration Given a sub-query qi(t)q_i^(t) and its anchor entity ei(t)e_i^(t) from the Planning phase, the Exploration module performs dual-source retrieval to gather complementary evidence from both the structured KG G and the unstructured text corpus D. KG Exploration. To navigate massive global KGs with inherent ambiguity and noise, we design a three-stage pipeline that progressively narrows the search scope from entity disambiguation to relation filtering and fine-grained fact pruning. (1) Entity Linking. We design a context-aware coarse-to-fine recall strategy. First, we retrieve top-k candidate entities C via vector similarity between the query context Ci(t)=(q,At,qi(t),ei(t))C_i^(t)=(q,A_t,q_i^(t),e_i^(t)) and entity profiles (aliases, description and neighbors). Then, the LLM performs precise disambiguation by evaluating each candidate’s profile against the query context to determine the optimal match: =VectorRecall(Ci(t),ℰ,k), =VectorRecall(C_i^(t),E,k), e∗ e^* =Promptlink(Ci(t),,LLM). =Prompt_link(C_i^(t),C,LLM). (2) (2) Relation Discovery. First, we retrieve all relations ReR_e connected to the linked entity e∗e^*. Then, the LLM selects a subset Rsel⊂ReR_sel⊂ R_e relevant to the query context Ci(t)C_i^(t) by considering relation labels, connectivity and potential to lead toward the answer, thereby pruning the search space at the schema level: Re∗ R_e^* =GetRelations(e∗,), =GetRelations(e^*,G), Rsel R_sel =Promptrel(Ci(t),e∗,Re∗,LLM). =Prompt_rel(C_i^(t),e^*,R_e^*,LLM). (3) (3) Fact Pruning. We retrieve all facts FrawF_raw (triples and attributes) associated with the selected relations RselR_sel. The LLM performs fine-grained pruning to retain only facts that either directly answer the query qi(t)q_i^(t) or serve as promising stepping stones for further exploration, filtering out noise while preserving critical evidence: Fraw F_raw =(e∗,r,t)∈∣r∈Rsel,t∈ℰ =\(e^*,r,t) r∈ R_sel,t \ ∪(h,r,e∗)∈∣r∈Rsel,h∈ℰ, ∪\(h,r,e^*) r∈ R_sel,h \, FiKG F_i^KG =Promptfact(Ci(t),Fraw,LLM). =Prompt_fact(C_i^(t),F_raw,LLM). (4) Text Exploration. We propose a hierarchical reading strategy that mimics human reading behavior, narrowing the focus from document retrieval to global skimming and fine-grained section analysis, balancing retrieval efficiency with depth. (1) Adaptive Page Selection. We employ an adaptive mechanism to retrieve the optimal entity document did_i for ei(t)e_i^(t). When encountering disambiguation or missing pages, the LLM progressively refines the search query by analyzing query context Ci(t)C_i^(t) and Wikipedia API feedback, ensuring robust page localization despite real-world noise. (2) Global Skimming. The LLM skims the high-level structure of the retrieved page did_i ( summary, table of contents and infobox) to assess the page’s relevance to the query context Ci(t)C_i^(t) and extract useful evidence IskimI_skim. If deeper investigation is needed, it selects a focused subset of promising sections SselS_sel likely to contain missing details: Ssel,Iskim=Promptskim(Ci(t),di,LLM).S_sel,I_skim=Prompt_skim(C_i^(t),d_i,LLM). (5) (3) Detailed Reading. For each selected section s∈Ssels∈ S_sel, we retrieve its full content, split it into chunks, and identify the top-k relevant chunks via dense retrieval. Concurrently, tables within the section are extracted and serialized. Finally, the LLM performs joint analysis of text chunks and tables to synthesize fine-grained evidence IdetailsI_detail^s: Idetail(s)=Promptdetail(Ci(t),s,LLM).I_detail^(s)=Prompt_detail(C_i^(t),s,LLM). (6) Combining skimmed and detailed extractions, the final text evidence for entity ei(t)e_i^(t) is Fitext=Iskim∪Idetail(s)∣s∈SselF_i^text=I_skim∪\I_detail^(s) s∈ S_sel\. Dual-Source Evidence. The final retrieved evidence for sub-query qi(t)q_i^(t) is the aggregation of dual-source knowledge: Ki(t)=FiKG∪FitextK_i^(t)=F_i^KG∪ F^text_i. Prompts for KG (Tables 17–20) and text (Tables 21–24) exploration are provided in Appendix K. 3.2.3 Synthesis & Reflection The Synthesis module aggregates dual-source evidence and evaluates progress. We design structured formats to present KG and text retrieval Kt=K1(t),…,KB(t)K_t=\K_1^(t),…,K_B^(t)\. The LLM first distills key facts and promising leads into a unified retrieval summary MtM_t, achieving deep synergy between structured KG and unstructured text. Simultaneously, it makes a judgment JtJ_t based on the notebook N and the current plan Pt=(At,Qt,Et)P_t=(A_t,Q_t,E_t) to determine whether to generate the final answer, continue exploration, or adjust the strategy: Jt,Mt=Promptsyn(q,Pt,Kt,,LLM).J_t,M_t=Prompt_syn(q,P_t,K_t,N,LLM). (7) Depending on JtJ_t, the Reflection module dynamically directs exploration through two pathways: (1) Continue Exploration. When progress is made but evidence remains incomplete, it refines the analysis to reflect new findings, formulates a new plan Pt+1P_t+1 targeting remaining knowledge gaps. Simultaneously, it updates two long-term memory structures: Notebook N, which accumulates verified key facts directly relevant to q, and Candidate Entity Pool P, which stores promising but deferred entities due to exploration width constraints: Pt+1,,=Promptcont(q,Pt,Mt,,,LLM).P_t+1,N,P=Prompt_cont(q,P_t,M_t,N,P,LLM). (8) (2) Strategy Adjustment. When facing dead ends, it performs critical reflection by analyzing the global interaction history ℋH to diagnose strategic errors. Then it pivots the strategy by backtracking to deferred entities in P or reformulating queries from entirely new angles based on the q and current progress, thereby ensuring resilient exploration: Pt+1,=Promptrecov(q,Pt,Mt,ℋ,,,LLM).P_t+1,P=Prompt_recov(q,P_t,M_t,H,N,P,LLM). (9) Interaction History Update. After each turn, the system appends the current retrieval summary MtM_t and plan PtP_t to the interaction history: ℋ←ℋ∪(Pt,Mt)H ∪\(P_t,M_t)\. This cumulative memory enables global retrospection, allowing the system to avoid repeating failed strategies and to make globally informed decisions across multiple turns. Prompts for synthesis (Table 25) and reflection (Tables 28, 29) are provided in Appendix K. 3.2.4 Overall Workflow & Answer Generation CoG executes the aforementioned cognitive cycle iteratively, until the Reflection module deems the evidence sufficient or a maximum turn limit is reached. The complete algorithmic workflow is detailed in Algorithm 1 in Appendix A. Upon sufficiency, the system synthesizes accumulated facts in Notebook N and final retrieval result MtM_t. The LLM constructs a comprehensive reasoning chain connecting these facts to directly address the question, producing the final answer a. 4 Experiments 4.1 Experimental Setup Datasets. We evaluate on seven knowledge-intensive multi-hop QA benchmarks: (1) KG-based QA: KGQAGen Zhang et al. (2025a), CWQ Talmor and Berant (2018), QALD10-en Usbeck et al. (2024), and WebQSP Yih et al. (2016), which require reasoning over structured facts for entity-centric answers. (2) Text-based QA: 2WikiMQA Ho et al. (2020), AdvHotpotQA Ye and Durrett (2022), and MusiQue Trivedi et al. (2022), which demand deep semantic understanding and multi-document synthesis for span extraction and free-form answer generation. Detailed descriptions and statistics are provided in Appendix B. We use Exact Match (EM) as the primary evaluation metric following previous work Ma et al. (2025b), and additionally report token-level F1 and reference-based LLM-as-judge results in Appendix D. For repeated runs, we report the median of three independent runs to ensure robustness. Baselines. We compare CoG against representative methods across four categories: (1) LLM-only: Direct few-shot prompting, Chain-of-Thought Wei et al. (2022), and Self-Consistency Wang et al. (2023), relying solely on internal parametric knowledge; (2) Text-based RAG: VectorRAG (top-33 chunk dense retrieval), ReAct Yao et al. (2023) and IRCoT Trivedi et al. (2023) for iterative text exploration, and Search-o1 Li et al. (2025a) for on-demand retrieval during reasoning; (3) KG-based RAG: ToG Sun et al. (2024) for iterative KG traversal, PoG Chen et al. (2024b) with backtracking, and DoG Ma et al. (2025a) using debate-driven query evolution; (4) Hybrid RAG: ToG-2 Ma et al. (2025b), the current state-of-the-art framework combining KG traversal with text retrieval under the same global knowledge setting as CoG. Implementation Details. Following previous works Ma et al. (2025b); Li et al. (2024b), we adopt the global knowledge setting, utilizing the full Wikidata and Wikipedia as structured and unstructured knowledge bases, respectively (details in Appendix C). This setting presents a realistic and challenging environment compared to simplified local settings. We employ Qwen3-32B (non-thinking mode) as the backbone LLM for CoG and baselines to ensure fair comparison. For retrieval, we utilize bge-m3 Chen et al. (2024a) as the embedding model for text chunk retrieval, and Qwen3-Embedding-4B Zhang et al. (2025c) for entity linking in the KG. We set the maximum exploration depth to 4 and width (sub-queries per turn) to 5. For entity linking, we retrieve the top-20 candidate entities. In text exploration, we extract up to 3 relevant chunks per section. 4.2 Main Results Baseline Type Method Multi-hop KG-based QA Multi-hop text-based QA KGQAGen CWQ QALD10-en WebQSP 2WikiMQA AdvHotpotQA MusiQue LLM-only Direct 36.70 32.42 47.45 72.45 47.20 20.13 8.40 CoT 38.28 33.59 45.35 70.84 47.40 23.70 9.60 CoT-SC 40.50 33.40 46.55 71.40 49.60 24.03 10.60 Text-based RAG Vector RAG 48.66 30.76 39.04 70.49 52.60 32.14 12.40 ReAct 50.79 34.18 41.44 76.15 63.20 41.88 17.00 IRCoT 53.75 33.30 41.14 71.40 71.60 47.08 20.60 Search-o1 46.34 36.72 43.54 72.03 44.60 40.26 19.40 KG-based RAG ToG 46.90 28.42 48.65 73.28 50.40 18.83 9.20 PoG 50.93 27.15 48.95 78.53 73.20 15.26 14.00 DoG 54.77 33.11 55.86 82.38 61.20 20.13 11.40 Hybrid RAG ToG-22 55.61 41.41 53.45 81.89 73.00 35.71 16.20 Proposed CoG 74.70 42.87 55.26 81.75 85.20 51.95 27.80 Table 2: Main results (Exact Match, %) of different methods on seven multi-hop QA datasets. All methods use Qwen3-32B (non-thinking mode) as the backbone LLM. The best results are highlighted in bold. As shown in Table 2, CoG substantially outperforms baselines across most benchmarks, with advantages particularly pronounced on complex tasks. On the challenging text-heavy MusiQue, CoG (27.8%) significantly surpasses the strongest text agent IRCoT (20.6%). Similarly, it achieves a 34.3% relative gain over ToG-2 on the high-quality KG-based KGQAGen. Meanwhile, CoG maintains comparable performance on simpler WebQSP (1-2 hops). These results validate our cognitive framework’s effectiveness in unifying structured and unstructured knowledge exploration. Additional F1 and LLM-as-judge results in Appendix D confirm that these gains persist under both lexical-overlap and semantic evaluation. Analysis of Baselines VectorRAG struggles with KG-based tasks, as its flat retrieval lacks structural guidance. Agentic text baselines (ReAct, IRCoT, Search-o1) and strong KG agents (PoG, DoG) improve over basic RAG but exhibit single-source limitations: text agents falter on KG-heavy tasks, while DoG struggles on text-heavy ones. ToG-2 effectively combines KG and text to outperform single-source baselines, but its reactive exploration lacks long-term planning and bidirectional synergy, faltering on complex tasks. CoG consistently outperforms single-source and hybrid baselines, proving that our Graph-Text Synergy effectively overcomes pure graph limitations and text retrieval noise, yielding benefits beyond a generic agentic loop. 4.3 Ablation Study Variant KGQA CWQ QALD 2Wiki AdvHot MusiQ Avg. CoG (Full) 74.70 42.87 55.26 85.20 51.95 27.80 56.30 Impact of Knowledge Sources w/o Text 75.90 39.06 51.65 76.20 27.27 20.20 48.38 w/o KG 56.44 40.92 48.65 79.80 49.68 20.80 49.38 Impact of Bidirectional Synergy w/o T-guide 74.14 42.58 52.85 76.20 39.61 22.80 51.36 w/o Synthesis 73.12 41.11 50.15 83.60 43.83 24.20 52.67 Impact of Cognitive Strategy w/o Reflect 72.57 42.38 52.85 85.00 48.05 25.80 54.44 w/o Plan 78.78 43.85 50.75 81.00 47.73 25.40 54.59 w/o Adap.EL 70.81 44.14 53.15 86.00 47.73 23.40 54.21 Table 3: Ablation results (EM, %) of CoG across three groups: (1) Knowledge Sources: removing either KG (w/o KG) or Text (w/o Text); (2) Bidirectional Synergy: disabling text-guided KG exploration (w/o T-guide), or removing summary synthesis module (w/o Synthesis); (3) Cognitive Strategy: excluding reflection-based adjustment (w/o Reflect), proactive planning (w/o Plan), or adaptive entity linking (w/o Adap.EL). To dissect CoG’s components, we conduct ablation studies across three dimensions (Table 3). Knowledge Sources. Removing either source causes significant drops (−6.92%-6.92\% for w/o KG, −7.92%-7.92\% for w/o Text). The severe degradation of w/o Text on text-heavy datasets (AdvHotpotQA, MusiQue) confirms that while KG provides structural guidance, rich text semantics are indispensable for bridging knowledge gaps. Bidirectional Synergy. Simply accessing both sources is insufficient. Disabling text-to-KG feedback (w/o T-guide) drops performance by 4.94%, indicating that entities discovered from text serve as effective new anchors for bridging disconnected KG paths. Furthermore, removing the synthesis module (w/o Synthesis) degrades performance, indicating that explicitly distilling heterogeneous evidence is crucial for filtering noise and informing subsequent reasoning. Cognitive Strategy. Cognitive mechanisms ensure robust exploration. Removing reflection, planning, or adaptive entity linking all degrades average performance. These results suggest that CoG benefits not only from access to KG and text, but also from its closed-loop control: reflection helps recover from noisy paths, planning guides goal-oriented search, and adaptive entity linking provides reliable anchors under ambiguous mentions. 4.4 In-depth Analysis Generalizability and Cost-Effectiveness. Table 4 demonstrates CoG’s effectiveness across diverse LLMs (detailed in Appendix E). Notably, CoG unlocks the reasoning potential of smaller models, yielding a massive +113.2% gain on Qwen3-8B. Remarkably, CoG-8B (51.0%) outperforms all strong 32B baselines (e.g., IRCoT 44.6%, ToG-2 45.9%). This confirms that CoG’s superiority stems from its robust cognitive architecture rather than mere parameter scale, effectively mitigating error propagation on smaller LLMs while offering a highly cost-effective solution. Backbone Method KG Avg. Text Avg. Overall Avg. Qwen3-32B Direct 41.48 25.24 33.36 CoG 59.31 54.98 57.15 (+71.3%) Qwen3-8B Direct 31.09 16.73 23.91 CoG 53.39 48.55 50.97 (+113.2%) gpt-oss-120b Direct 28.33 15.88 22.10 CoG 42.87 40.28 41.58 (+88.1%) DeepSeek-V3.2 Direct 47.91 28.06 37.99 CoG 64.79 60.50 62.64 (+64.9%) GPT-5.1 Direct 45.67 38.67 42.17 CoG 56.00 55.33 55.67 (+32.0%) Table 4: Performance comparison (EM, %) of CoG versus Direct Prompting across different backbone LLMs. Proprietary models are evaluated on a sampled subset (100 samples per dataset) due to cost constraints. Impact of Exploration Depth. Figure 3 illustrates the performance trends across varying depths (see Appendix F for full results). On KG tasks, CoG consistently outperforms VectorRAG across all depths, confirming the superiority of structured cognitive exploration over flat retrieval. Similarly, text tasks show steady gains as depth increases, indicating that deeper exploration effectively gathers scattered semantic evidence. Notably, the optimal depth correlates strongly with dataset difficulty. Simpler datasets (WebQSP, 2WikiMQA) achieve optimal performance at shallower depths and then stabilize, whereas challenging ones (KGQAGen, AdvHotpotQA) continue to benefit from deeper exploration. It demonstrates CoG’s adaptability to varying task complexity. Figure 3: Impact of exploration depth on KG-based and Text-based tasks and representative datasets. Exploration Efficiency. We evaluate efficiency across four dimensions: accessed documents, token consumption, latency, and wall-clock time per retrieval round (detailed breakdown in Appendix G). Despite multiple LLM calls per turn, CoG’s goal-oriented planning achieves remarkable precision, accessing 12.4×12.4× fewer documents than ToG-2’s exhaustive beam search (Table 5). Regarding absolute cost, CoG exhibits complexity adaptability: on simpler datasets (e.g., WebQSP), it maintains low token usage (∼ 15k) comparable to text agents like IRCoT (∼ 22k) while achieving superior accuracy. On complex tasks (e.g., MusiQue), costs increase to support deep reasoning, trading computation for capability where efficient methods fail. Crucially, this overhead can be mitigated by using smaller models, as CoG-8B already outperforms 32B baselines (Appendix E), improving the accuracy-cost trade-off for deployment. Method KGQA CWQ QALD WebQ 2Wiki Hotpot MusiQ Avg. ToG-2 37.1 40.4 31.2 22.9 17.7 36.1 42.2 32.5 CoG 3.1 3.1 2.3 1.5 2.4 2.5 3.5 2.6 Gain ×11.9 ×13.2 ×13.3 ×15.5 ×7.3 ×14.8 ×12.1 ×12.4 Table 5: Comparison of exploration efficiency (Average number of unique documents accessed per question). Hyperparameter Robustness. We study the sensitivity of adaptive entity linking by varying the retrieved entity candidates from k=6k=6 to 3030 (Appendix H). CoG remains stable across settings, with average EM varying only slightly from 59.1% to 60.3%. Increasing k does not degrade performance, suggesting that CoG’s context-aware disambiguation can filter noisy candidates while benefiting from broader entity recall. Manual Evaluation. We manually analyze sampled reasoning paths in Appendix I to evaluate reasoning faithfulness and diagnose error causes. For EM-correct cases, CoG yields the highest proportion of True Positives (89.1%) with faithful reasoning chains. In contrast, CoT-SC and ToG-2 suffer more severe false positives (Weak FP + FP: 29.1% and 12.7% vs. CoG’s 3.6%), often reaching correct answers via hallucinated or flawed intermediate steps. For EM-incorrect cases, 71.1% of CoG’s errors are attributable to external factors such as metric false negatives, dataset flaws, or temporal mismatch, while 28.9% reflect internal exploration or reasoning failures. These findings support both the reliability of CoG’s EM gains and the need for stronger exploration focus and disambiguation. Qualitative Analysis. We provide a detailed case study in Appendix J, demonstrating how CoG’s cognitive cycle and bidirectional synergy enable it to recover from dead ends and discover critical text evidence to solve complex multi-hop queries. 5 Conclusion In this paper, we introduced CoG, a cognitive-inspired framework for autonomous knowledge exploration in global, hybrid knowledge bases. Under the continuous “plan-explore-reflect” human cognitive cycle, CoG shifts the paradigm from passive, reactive retrieval to proactive knowledge discovery, enabling adaptive strategy adjustment and robust recovery from exploration failures. Crucially, it establishes a deep bidirectional synergy between structured KGs and unstructured text, leveraging their complementary strengths to bridge knowledge gaps. Extensive experiments demonstrate that CoG significantly outperforms state-of-the-art baselines on multiple multi-hop QA tasks while achieving superior exploration efficiency. Limitations While CoG demonstrates superior performance and exploration efficiency in global knowledge exploration, several limitations remain. Latency and Cost. The iterative cognitive loop inevitably incurs higher inference latency and token cost than passive retrieval methods. Although CoG reduces redundant document access through goal-oriented exploration and can improve the cost-accuracy trade-off when paired with smaller backbones, its multi-call workflow may still limit real-time deployment. Dependency on Backbone LLM. The framework’s effectiveness still depends on the backbone LLM’s instruction following and long-context reasoning abilities. Although CoG substantially improves smaller models, weaker LLMs may still struggle with complex planning, evidence synthesis, or reflection over noisy intermediate results. Acknowledgments This work is supported by the Zhongguancun Academy (Grant No. 02012501), in part by the National Natural Science Foundation of China (Grant Nos. 62473271, 62506357), and the Fundamental Research Funds for the Beijing University of Posts and Telecommunications (Grant No. 2025AI4S03). References Chang et al. (2026) Ge Chang, Jinbo Su, Jiacheng Liu, Pengfei Yang, Yuhao Shang, Huiwen Zheng, Hongli Ma, Yan Liang, Yuanchun Li, and Yunxin Liu. 2026. Enhancing agentic textual graph retrieval with synthetic stepwise supervision. In ACL, pages 25510–25525, San Diego, California, United States. Association for Computational Linguistics. Chen et al. (2024a) Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024a. M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. In Findings of ACL, pages 2318–2335, Bangkok, Thailand. Association for Computational Linguistics. Chen et al. (2024b) Liyi Chen, Panrong Tong, Zhongming Jin, Ying Sun, Jieping Ye, and Hui Xiong. 2024b. Plan-on-graph: Self-correcting adaptive planning of large language model on knowledge graphs. In NeurIPS, volume 37, pages 37665–37691. Curran Associates, Inc. Chen et al. (2025) Weijie Chen, Ting Bai, Jinbo Su, Jian Luan, Wei Liu, and Chuan Shi. 2025. Kg-retriever: Efficient knowledge indexing for retrieval-augmented large language models. In 2025 IEEE International Conference on Knowledge Graph (ICKG), pages 27–34. Cui et al. (2025) Yuanning Cui, Zequn Sun, Wei Hu, and Zhangjie Fu. 2025. Kgfr: A foundation retriever for generalized knowledge graph question answering. Preprint, arXiv:2511.04093. Dong et al. (2026) Junnan Dong, Siyu An, Yifei Yu, QIANWEN ZHANG, Linhao Luo, Xiao Huang, di yin, Yunsheng Wu, and Xing Sun. 2026. Youtu-graphrag: Vertically unified agents for graph retrieval-augmented complex reasoning. In ICLR, volume 2026, pages 143735–143755. Edge et al. (2024) Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization. Preprint, arXiv:2404.16130. Guo et al. (2024) Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2024. Lightrag: Simple and fast retrieval-augmented generation. Preprint, arXiv:2410.05779. Gutiérrez et al. (2025) Bernal Jiménez Gutiérrez, Yiheng Shu, Weijian Qi, Sizhe Zhou, and Yu Su. 2025. From RAG to memory: Non-parametric continual learning for large language models. In ICML, volume 267 of Proceedings of Machine Learning Research, pages 21497–21515. PMLR. Gutiérrez et al. (2024) Bernal Jiménez Gutiérrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. 2024. HippoRAG: Neurobiologically inspired long-term memory for large language models. In NeurIPS, volume 37, pages 59532–59569. Curran Associates, Inc. Ho et al. (2020) Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing a multi-hop QA dataset for comprehensive evaluation of reasoning steps. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6609–6625, Barcelona, Spain (Online). International Committee on Computational Linguistics. Huang et al. (2025a) Haoyu Huang, Yongfeng Huang, Yang Junjie, Zhenyu Pan, Yongqiang Chen, Kaili Ma, Hongzhi Chen, and James Cheng. 2025a. Retrieval-augmented generation with hierarchical knowledge. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 6044–6060, Suzhou, China. Association for Computational Linguistics. Huang et al. (2025b) Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2025b. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Trans. Inf. Syst., 43(2). Jiang et al. (2023) Jinhao Jiang, Kun Zhou, Zican Dong, Keming Ye, Xin Zhao, and Ji-Rong Wen. 2023. StructGPT: A general framework for large language model to reason over structured data. In EMNLP, pages 9237–9251, Singapore. Association for Computational Linguistics. Jiang et al. (2025) Jinhao Jiang, Kun Zhou, Xin Zhao, Yang Song, Chen Zhu, Hengshu Zhu, and Ji-Rong Wen. 2025. KG-agent: An efficient autonomous agent framework for complex reasoning over knowledge graph. In Proceedings of the 63rd annual meeting of the association for computational linguistics (volume 1: Long papers), pages 9505–9523. Association for Computational Linguistics. Lee et al. (2025) Meng-Chieh Lee, Qi Zhu, Costas Mavromatis, Zhen Han, Soji Adeshina, Vassilis N. Ioannidis, Huzefa Rangwala, and Christos Faloutsos. 2025. HybGRAG: Hybrid retrieval-augmented generation on textual and relational knowledge bases. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 879–893. Association for Computational Linguistics. Lei et al. (2025) Yongjia Lei, Haoyu Han, Ryan A. Rossi, Franck Dernoncourt, Nedim Lipka, Mahantesh M Halappanavar, Jiliang Tang, and Yu Wang. 2025. Mixture of structural-and-textual retrieval over text-rich graph knowledge bases. In Findings of the Association for Computational Linguistics: ACL 2025, pages 18306–18321, Vienna, Austria. Association for Computational Linguistics. Lewis et al. (2020) Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. In NeurIPS, volume 33, pages 9459–9474. Curran Associates, Inc. Li et al. (2024a) Shilong Li, Yancheng He, Hangyu Guo, Xingyuan Bu, Ge Bai, Jie Liu, Jiaheng Liu, Xingwei Qu, Yangguang Li, Wanli Ouyang, Wenbo Su, and Bo Zheng. 2024a. GraphReader: Building graph-based agent to enhance long-context abilities of large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 12758–12786. Association for Computational Linguistics. Li et al. (2025a) Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. 2025a. Search-o1: Agentic search-enhanced large reasoning models. In EMNLP, pages 5420–5438, Suzhou, China. Association for Computational Linguistics. Li et al. (2024b) Xingxuan Li, Ruochen Zhao, Yew Ken Chia, Bosheng Ding, Shafiq Joty, Soujanya Poria, and Lidong Bing. 2024b. Chain-of-knowledge: Grounding large language models via dynamic knowledge adapting over heterogeneous sources. In ICLR, volume 2024, pages 9565–9587. Li et al. (2025b) Yangning Li, Weizhi Zhang, Yuyao Yang, Wei-Chieh Huang, Yaozu Wu, Junyu Luo, Yuanchen Bei, Henry Peng Zou, Xiao Luo, Yusheng Zhao, Chunkit Chan, Yankai Chen, Zhongfen Deng, Yinghui Li, Hai-Tao Zheng, Dongyuan Li, Renhe Jiang, Ming Zhang, Yangqiu Song, and Philip S. Yu. 2025b. Towards agentic rag with deep reasoning: A survey of rag-reasoning systems in llms. Preprint, arXiv:2507.09477. Lin et al. (2026) Junhong Lin, Shicheng Liu, Jinyeop Song, Song Wang, Julian Shun, and Yada Zhu. 2026. Efficient and transferable agentic knowledge graph rag via reinforcement learning. Preprint, arXiv:2509.26383. Luo et al. (2026) Haoran Luo, Haihong E, Guanting Chen, Qika Lin, Yikai Guo, Fangzhi Xu, Zemin Kuang, Meina Song, Xiaobao Wu, Yifan Zhu, and Luu Anh Tuan. 2026. Graph-r1: Towards agentic graphrag framework via end-to-end reinforcement learning. Preprint, arXiv:2507.21892. Luo et al. (2025a) Linhao Luo, Zicheng Zhao, Reza Haffari, Dinh Phung, Chen Gong, and Shirui Pan. 2025a. Gfm-rag: Graph foundation model for retrieval augmented generation. In NeurIPS, volume 38, Main Conference, pages 36371–36405. Curran Associates, Inc. Luo et al. (2025b) Qi Luo, Xiaonan Li, Tingshuo Fan, Xinchi Chen, and Xipeng Qiu. 2025b. Towards global retrieval augmented generation: A benchmark for corpus-level reasoning. Preprint, arXiv:2510.26205. Ma et al. (2025a) Jie Ma, Zhitao Gao, Qi Chai, Wangchun Sun, Pinghui Wang, Hongbin Pei, Jing Tao, Lingyun Song, Jun Liu, Chen Zhang, and Lizhen Cui. 2025a. Debate on graph: A flexible and reliable reasoning framework for large language models. In AAAI, volume 39, pages 24768–24776. Ma et al. (2025b) Shengjie Ma, Chengjin Xu, Xuhui Jiang, Muzhi Li, Huaren Qu, Cehao Yang, Jiaxin Mao, and Jian Guo. 2025b. Think-on-graph 2.0: Deep and faithful large language model reasoning with knowledge-guided retrieval augmented generation. In ICLR, volume 2025, pages 52782–52806. Sarmah et al. (2024) Bhaskarjit Sarmah, Dhagash Mehta, Benika Hall, Rohan Rao, Sunil Patel, and Stefano Pasquali. 2024. HybridRAG: Integrating knowledge graphs and vector retrieval augmented generation for efficient information extraction. In Proceedings of the 5th ACM International Conference on AI in Finance, ICAIF ’24, pages 608–616. Association for Computing Machinery. Song et al. (2025) Zirui Song, Bin Yan, Yuhan Liu, Miao Fang, Mingzhe Li, Rui Yan, and Xiuying Chen. 2025. Injecting domain-specific knowledge into large language models: A comprehensive survey. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 25297–25311, Suzhou, China. Association for Computational Linguistics. Sun et al. (2024) Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel Ni, Heung-Yeung Shum, and Jian Guo. 2024. Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph. In ICLR, volume 2024, pages 3868–3898. Sun et al. (2025) Yushi Sun, Kai Sun, Yifan Ethan Xu, Xiao Yang, Xin Luna Dong, Nan Tang, and Lei Chen. 2025. KERAG: Knowledge-enhanced retrieval-augmented generation for advanced question answering. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 6194–6216. Association for Computational Linguistics. Talmor and Berant (2018) Alon Talmor and Jonathan Berant. 2018. The web as a knowledge-base for answering complex questions. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 641–651, New Orleans, Louisiana. Association for Computational Linguistics. Tan et al. (2025) Xingyu Tan, Xiaoyang Wang, Qing Liu, Xiwei Xu, Xin Yuan, Liming Zhu, and Wenjie Zhang. 2025. HydraRAG: Structured cross-source enhanced large language model reasoning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 14431–14459, Suzhou, China. Association for Computational Linguistics. Trivedi et al. (2022) Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. MuSiQue: Multihop questions via single-hop question composition. Transactions of the Association for Computational Linguistics, 10:539–554. Trivedi et al. (2023) Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2023. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10014–10037. Association for Computational Linguistics. Usbeck et al. (2024) Ricardo Usbeck, Xi Yan, Aleksandr Perevalov, Longquan Jiang, Julius Schulz, Angelie Kraft, Cedric Möller, Junbo Huang, Jan Reineke, Axel-Cyrille Ngonga Ngomo, et al. 2024. Qald-10–the 10th challenge on question answering over linked data: Shifting from dbpedia to wikidata as a kg for kgqa. Semantic Web, 15(6):2193–2207. Wang et al. (2025) Song Wang, Junhong Lin, Xiaojie Guo, Julian Shun, Jundong Li, and Yada Zhu. 2025. Reasoning of large language models over knowledge graphs with super-relations. In ICLR, volume 2025, pages 4208–4226. Wang et al. (2026) Xinming Wang, Jian Xu, Sheng Lian, Yi Chen, Haiyang Guo, Fei Zhu, Yuanqi Shao, Minsi Ren, Hongzhu Yi, Aslan H. Feng, Hongming Yang, Tailin Wu, Han Hu, Shiming Xiang, Xu-Yao Zhang, and Cheng-Lin Liu. 2026. The hitchhiker’s guide to autonomous research: A survey of scientific agents. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–20. Wang et al. (2023) Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. In ICLR. Wang et al. (2024) Yu Wang, Nedim Lipka, Ryan A. Rossi, Alexa Siu, Ruiyi Zhang, and Tyler Derr. 2024. Knowledge graph prompting for multi-document question answering. In AAAI, volume 38, pages 19206–19214. Wei et al. (2022) Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, volume 35, pages 24824–24837. Curran Associates, Inc. Wu et al. (2026) Xiaojun Wu, Cehao Yang, Xueyuan Lin, Chengjin Xu, Xuhui Jiang, Yuanliang Sun, Hui Xiong, Jia Li, and Jian Guo. 2026. Think-on-graph 3.0: Efficient and adaptive llm reasoning on heterogeneous graphs via multi-agent dual-evolving context retrieval. Preprint, arXiv:2509.21710. Xu et al. (2026) Jiasheng Xu, Mingda Li, Yongqiang Tang, Peijie Wang, and Wensheng Zhang. 2026. Towards open-world retrieval-augmented generation on knowledge graph: A multi-agent collaboration framework. In Proceedings of the ACM Web Conference 2026, W ’26, page 2171–2182, New York, NY, USA. Association for Computing Machinery. Yang et al. (2025) Cehao Yang, Xiaojun Wu, Xueyuan Lin, Chengjin Xu, Xuhui Jiang, Yuanliang Sun, Jia Li, Hui Xiong, and Jian Guo. 2025. Graphsearch: An agentic deep searching workflow for graph retrieval-augmented generation. Preprint, arXiv:2509.22009. Yang et al. (2018) Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369–2380. Association for Computational Linguistics. Yao et al. (2023) Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. In ICLR. Ye and Durrett (2022) Xi Ye and Greg Durrett. 2022. The unreliability of explanations in few-shot prompting for textual reasoning. In NeurIPS, volume 35, pages 30378–30392. Yih et al. (2016) Wen-tau Yih, Matthew Richardson, Chris Meek, Ming-Wei Chang, and Jina Suh. 2016. The value of semantic parse labeling for knowledge base question answering. In ACL, pages 201–206, Berlin, Germany. Association for Computational Linguistics. Yu et al. (2026) Chuanyue Yu, Kuo Zhao, Yuhan Li, Heng Chang, Mingjian Feng, Xiangzhe Jiang, Yufei Sun, Jia Li, Yuzhi Zhang, Qingyun Sun, Jianxin Li, and Ziwei Zhang. 2026. Graphrag-r1: Graph retrieval-augmented generation with process-constrained reinforcement learning. In Proceedings of the ACM Web Conference 2026, W 2026, Dubai, United Arab Emirates, originally scheduled for April 13-17, 2026, rescheduled for June 29 - July 3, 2026, pages 1398–1409. ACM. Zhang et al. (2025a) Liangliang Zhang, Zhuorui Jiang, Hongliang Chi, Haoyang Chen, Mohammed ElKoumy, Fali Wang, Qiong Wu, Zhengyi Zhou, Shirui Pan, Suhang Wang, and Yao Ma. 2025a. Diagnosing and addressing pitfalls in kg-rag datasets: Toward more reliable benchmarking. In NeurIPS, volume 38, Main Conference. Curran Associates, Inc. Zhang et al. (2025b) Qinggang Zhang, Shengyuan Chen, Yuanchen Bei, Zheng Yuan, Huachi Zhou, Zijin Hong, Hao Chen, Yilin Xiao, Chuang Zhou, Junnan Dong, Yi Chang, and Xiao Huang. 2025b. A survey of graph retrieval-augmented generation for customized large language models. Preprint, arXiv:2501.13958. Zhang et al. (2025c) Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. 2025c. Qwen3 embedding: Advancing text embedding and reranking through foundation models. Preprint, arXiv:2506.05176. Zhu et al. (2025) Xiangrong Zhu, Yuexiang Xie, Yi Liu, Yaliang Li, and Wei Hu. 2025. Knowledge graph-guided retrieval augmented generation. In NAACL, pages 8912–8924, Albuquerque, New Mexico. Association for Computational Linguistics. Appendix A CoG Algorithm We provide the pseudo-code of the CoG framework in Algorithm 1. It outlines the iterative workflow, illustrating how the Planning, Exploration, Synthesis, and Reflection modules coordinate to achieve autonomous knowledge exploration in large-scale knowledge bases. Algorithm 1 The CoG Algorithm 0: Question q, KG G, Text Corpus D, Max Turns M 0: Answer a Initialize: ←∅N← , ←∅P← , ℋ←∅H← // Phase 1: Initial Planning (§3.2.1) P0=(A0,Q0,E0)←InitPlan(q)P_0=(A_0,Q_0,E_0)← InitPlan(q) for t=0t=0 to M−1M-1 do // Phase 2: Exploration Dual-Source (§3.2.2) Kt←∅K_t← for each (qi(t),ei(t))∈(Qt,Et)(q_i^(t),e_i^(t))∈(Q_t,E_t) do Ci(t)←(q,At,qi(t),ei(t))C_i^(t)←(q,A_t,q_i^(t),e_i^(t)) // Query context FiKG←ExploreKG(Ci(t),)F_i^KG← ExploreKG(C_i^(t),G) // Entity Link → Relation Discovery → Fact Prune Fitext←ExploreText(Ci(t),)F_i^text← ExploreText(C_i^(t),D) // Page Selection → Skimming → Detailed Reading Kt←Kt∪FiKG∪FitextK_t← K_t∪\F_i^KG∪ F_i^text\ // Aggregate dual-source evidence end for // Phase 3: Synthesis (§3.2.3) Jt,Mt←Synthesize(q,Pt,Kt,)J_t,M_t← Synthesize(q,P_t,K_t,N) // Phase 4: Reflection & Re-Planning (§3.2.3) if Jt=SUFFICIENTJ_t= SUFFICIENT then break // Proceed to Answer Generation else if Jt=INSUFFICIENT_USEFULJ_t= INSUFFICIENT\_USEFUL then // Pathway 1: Continue Exploration Pt+1,,←ContinuePlan(q,Pt,Mt,,)P_t+1,N,P← ContinuePlan(q,P_t,M_t,N,P) else // Pathway 2: Strategy Adjustment Pt+1,←AdjustPlan(q,Pt,Mt,ℋ,,)P_t+1,P← AdjustPlan(q,P_t,M_t,H,N,P) end if ℋ←ℋ∪(Pt,Mt)H ∪\(P_t,M_t)\ // Update Interaction History end for // Answer Generation (§3.2.4) a←GenerateAnswer(q,,Pt,Mt)a← GenerateAnswer(q,N,P_t,M_t) return a Appendix B Evaluation Dataset In this section, we provide detailed descriptions of the seven multi-hop question answering datasets used in our experiments, covering both KG-based and text-based multi-hop reasoning tasks. KG-based QA Datasets. These benchmarks primarily require reasoning over structured facts in the knowledge graph and usually expect entity-centric answers grounded in KG nodes or literal attributes. WebQSP Yih et al. (2016) is a widely used knowledge base QA benchmark derived from WebQuestions, containing questions paired with SPARQL queries over Freebase (mapped to Wikidata in our setting). It primarily focuses on questions that require 1-2 hop reasoning over structured KG facts to retrieve target entities. CWQ Talmor and Berant (2018) extends the WebQSP by generating questions with higher complexity. It necessitates multi-hop reasoning (up to 4 hops) and challenges models to handle intricate logical constraints, such as composition, conjunctions, and superlatives over the graph structure. QALD10-en Usbeck et al. (2024) is the English subset of the 10th Question Answering over Linked Data challenge, specifically adapted for Wikidata. It is characterized by highly complex natural language questions that map to intricate SPARQL queries. The dataset specifically tests a system’s ability to handle advanced structural operations, such as aggregations, filters, and multi-hop relationships. KGQAGen Zhang et al. (2025a) is a highly reliable, Wikidata-grounded benchmark generated via an LLM-in-the-loop framework with symbolic verification, featuring a wide range of topics and structural complexities. It systematically addresses the inaccurate annotations and ambiguity in existing evaluation datasets, making it a superior standard for evaluating multi-hop reasoning capabilities in KG-based RAG systems. Text-based QA Datasets. These benchmarks demand semantic understanding and multi-document synthesis, covering span extraction and more free-form answer types such as dates, numbers, boolean answers, and short descriptive phrases. 2WikiMQA Ho et al. (2020) is a large-scale dataset synthesized from Wikidata and Wikipedia, specifically designed to test multi-hop reasoning across multiple documents. It focuses on questions that require comparison, aggregation, and bridging reasoning types. The questions typically require reasoning chains of 2 to 4 hops, challenging models to perform cross-document integration and logical inference over unstructured text. MusiQue Trivedi et al. (2022) is a highly challenging multi-hop QA dataset constructed by compositing single-hop questions in a bottom-up manner. It is explicitly designed to minimize reasoning shortcuts and enforce connected reasoning, ensuring that answering the final question strictly requires resolving intermediate sub-questions. The dataset features 2-4 hop questions that are significantly harder to solve via disconnected reasoning compared to HotpotQA or 2WikiMQA. AdvHotpotQA Ye and Durrett (2022) is an adversarial version of the HotpotQA Yang et al. (2018) dataset. It introduces distractor paragraphs that share high lexical overlap with the question but do not contain the answer, rigorously testing the system’s robustness against retrieval noise and its ability to identify supporting evidence accurately. Table 6 summarizes the statistics of the evaluation datasets. For WebQSP, AdvHotpotQA, and QALD10-en, we use the same test splits as in ToG-2 Ma et al. (2025b) for fair comparison. For CWQ, we adopt the first 1,024 questions from the test set used in ToG Sun et al. (2024). For 2WikiMQA and MusiQue, we employ the evaluation sets provided by IRCoT Trivedi et al. (2023). For KGQAGen, we use the full test set from the KGQAGen-10k benchmark. Category Dataset # Questions KG-based QA WebQSP 1,430 CWQ 1,024 QALD10-en 333 KGQAGen-10k 1,079 Text-based QA 2WikiMQA 500 MusiQue 500 AdvHotpotQA 308 Table 6: Statistics of the evaluation datasets. Appendix C Knowledge Bases CoG operates over a global knowledge environment that combines structured symbolic facts from Wikidata with entity-centric textual evidence from Wikipedia. This design allows the model to alternate between precise graph-based reasoning and flexible textual exploration, which is important for open-domain multi-hop questions where either source alone may be incomplete. Structured Knowledge Graph (Wikidata). We use the full Wikidata dump11 1 https://dumps.wikimedia.org/wikidatawiki/entities/ (May 2024) as the structured knowledge graph. After preprocessing, the resulting graph contains 89.3 million entities and 1.38 billion edges. We distinguish two types of edges: relational edges between entities (e.g., (Steve Jobs, founded, Apple)), and attribute edges that connect entities to literal values such as dates, quantities, or strings (e.g., (Steve Jobs, date of birth, 1955-02-24)). In total, the processed graph contains 823.7 million relational triples and 562.0 million attribute triples. This large-scale setting is substantially different from task-specific or locally constructed KGs. It requires entity linking and retrieval to operate over a global entity space rather than a small candidate graph. Knowledge Graph Preprocessing. We process Wikidata into a retrieval-oriented representation to efficiently support CoG’s core operations. For entity grounding, we extract English labels, aliases, and descriptions. These are encoded via Qwen3-Embedding-4B into a FAISS index, enabling dense candidate recall across nearly 90 million entities with only ∼ 10GB of storage. Since many Wikidata entities share similar or identical names, each candidate is further supplemented with a compact entity profile, including textual metadata, popularity signals, and local graph context, for fine-grained LLM-based disambiguation. For relation exploration and fact acquisition, we organize relational and attribute triples into an in-memory key-value database (Redis) for efficient neighbor lookups. Additionally, we pre-compute the global occurrence frequencies of all relations, which serves as a crucial structural heuristic for the agent to prune high fan-out paths during exploration. Unstructured Text Corpus (Wikipedia). We employ the real-time English Wikipedia22 2 https://en.wikipedia.org/w/api.php as our unstructured text corpus. It comprises approximately 7.1 million entity-centric articles, totaling over 5 billion words. Distinct from traditional RAG that chunks text flatly, we process each page to retain its hierarchical structure, including the summary paragraph, table of contents, and sectioned full text. This structure supports CoG’s hierarchical reading strategy (global skimming followed by detailed reading), enabling efficient navigation through long documents. Wikipedia complements Wikidata by providing richer natural-language context, especially for facts that are difficult to normalize into triples. In contrast, Wikidata provides normalized entities and typed relations for precise symbolic traversal. CoG exploits this complementarity by using KG facts as high-precision reasoning anchors and Wikipedia passages as high-recall contextual evidence. Appendix D Additional Evaluation Metrics To complement the primary EM evaluation, Tables 7 and 8 report LLM-as-judge Correct rates and token-level F1 scores, respectively, for the same model outputs as in Table 2. For LLM-as-judge, Qwen3.5-122B-A10B (non-thinking mode, temperature 0) classifies each response as Correct, Partial, Incorrect, or No Answer, given only question, reference answers, and model response. We conservatively report only the Correct rate. For token-level F1, the same model extracts a concise final answer using only the question and response. We then normalize predictions and references by lowercasing, removing punctuation and articles, and normalizing whitespace. For multiple references, we retain the maximum F1. Baseline Type Method Multi-hop KG-based QA Multi-hop text-based QA Avg. KGQAGen CWQ QALD10-en WebQSP 2WikiMQA AdvHotpotQA MusiQue LLM-only Direct 38.74 34.57 40.84 64.06 29.20 23.38 14.20 35.00 CoT 38.28 35.06 44.14 60.28 31.20 25.97 15.40 35.76 CoT-SC 39.48 32.71 43.54 61.05 32.40 26.62 14.80 35.80 Text-based RAG Vector RAG 50.23 33.20 37.84 64.83 48.80 45.45 18.20 42.65 ReAct 48.19 35.74 42.04 69.79 57.60 54.87 22.80 47.29 IRCoT 52.55 34.57 37.84 66.43 68.60 61.69 25.00 49.53 Search-o1 52.27 39.84 48.65 68.60 47.80 47.73 27.40 47.47 KG-based RAG ToG 42.63 29.20 45.95 62.87 34.80 22.08 11.20 35.53 PoG 51.44 29.59 49.85 70.42 79.20 19.80 16.40 45.24 DoG 43.65 25.68 51.65 63.29 62.00 24.68 13.80 40.68 Hybrid RAG ToG-22 53.57 41.50 49.85 70.21 65.20 42.86 22.40 49.37 Proposed CoG 66.82 44.92 59.76 76.57 88.20 62.34 32.80 61.63 Table 7: LLM-as-judge Correct rate (%) on seven datasets. The best result in each column is highlighted in bold. Baseline Type Method Multi-hop KG-based QA Multi-hop text-based QA Avg. KGQAGen CWQ QALD10-en WebQSP 2WikiMQA AdvHotpotQA MusiQue LLM-only Direct 38.64 33.17 28.14 49.59 33.19 21.63 15.77 31.45 CoT 40.28 35.25 34.35 55.96 35.94 25.13 17.68 34.94 CoT-SC 41.91 34.59 35.29 56.30 36.14 24.58 18.11 35.27 Text-based RAG Vector RAG 48.49 28.98 24.43 47.81 44.61 36.84 17.50 35.52 ReAct 46.15 29.90 25.58 49.78 51.55 42.78 21.99 38.25 IRCoT 51.26 30.34 25.19 48.08 59.86 47.93 23.13 40.83 Search-o1 49.18 35.43 31.39 51.48 44.22 39.44 25.65 39.54 KG-based RAG ToG 43.31 27.24 35.58 58.21 34.31 19.51 13.22 33.05 PoG 51.30 26.46 39.84 59.35 73.47 16.47 15.56 40.35 DoG 46.00 27.20 43.20 56.01 61.84 21.15 14.35 38.54 Hybrid RAG ToG-22 54.96 39.46 38.55 60.58 62.80 37.85 23.45 45.38 Proposed CoG 67.69 38.46 37.96 50.36 76.22 49.95 28.48 49.87 Table 8: Token-level F1 (%) on seven datasets. The best result in each column is highlighted in bold. As shown in Tables 7 and 8, the complementary metrics are consistent with the EM-based findings. Under LLM-as-judge, CoG achieves the best result on all seven datasets, with an average Correct rate of 61.63%, outperforming the strongest baseline, IRCoT, by 12.10 points. Under token-level F1, CoG also obtains the highest average score of 49.87%, improving over ToG-2 by 4.49 points. These results show that CoG’s gains are not merely caused by exact string matching, but remain clear under both semantic equivalence judgment and lexical-overlap evaluation. Appendix E Detailed Results on Backbone LLMs This section supplements the backbone LLM generalizability analysis in Section 4.4. Table 9 presents the detailed performance breakdown across different backbone LLMs, corresponding to the summarized results in Table 4. Evaluation Setup. To balance computational cost with evaluation rigor, we adopted the following dataset selection and sampling strategies: Dataset Selection. We excluded the WebQSP dataset from this analysis, as modern LLMs already achieve near-saturated performance on its simple 1-2 hop queries in the Direct Prompting setting, making it less discriminative for measuring improvement. Open-source Models. For large-scale KG-based datasets (KGQAGen and CWQ), we sampled 300 instances each to align with the scale of QALD10-en. For the remaining datasets (QALD10-en, 2WikiMQA, AdvHotpotQA, and MusiQue), we evaluated on the full test sets. Proprietary Models. We evaluated on a sampled subset of 100 examples for each dataset. Key Observations. The detailed results reveal three key insights regarding CoG’s generalizability and cost-effectiveness: (1) Broad Generalizability: CoG consistently improves performance across all backbone models, regardless of their scale (from 8B to 235B) or type (open-source vs. proprietary). (2) Unlocking Small Models: The relative performance gain is most pronounced on the smallest model, Qwen3-8B (+113.2%), showing that CoG’s cognitive architecture effectively compensates for the limited parametric knowledge of smaller LLMs. (3) Cost-Effectiveness: CoG-8B (Avg. 51.0%) outperforms direct prompting on massive models (e.g. Qwen3-235B 43.7%, DeepSeek-V3.2 38.0%) and surpasses strong 32B baselines (e.g., IRCoT 44.6%, ToG-2 45.9%). Since 8B models are cheaper and faster, CoG-8B offers a highly cost-effective deployment solution that neutralizes the token overhead of multi-turn exploration. Backbone Method KG-based QA Text-based QA Average KGQAGen CWQ QALD 2Wiki Hotpot MusiQ KG Text Total Open-Source Models Qwen3-32B Direct 40.00 37.00 47.45 47.20 20.13 8.40 41.48 25.24 33.36 IRCoT 53.80 33.30 41.10 71.60 47.10 20.60 42.73 46.43 44.58 Search-o1 46.34 36.72 43.54 44.60 40.26 19.40 42.20 34.75 38.48 DoG 54.80 33.10 55.90 61.20 20.10 11.40 47.93 30.90 39.42 ToG-2 55.60 41.40 53.50 73.00 35.70 16.20 50.17 41.63 45.90 CoG 76.33 46.33 55.26 85.20 51.95 27.80 59.31 54.98 57.15 Qwen3-8B Direct 29.10 27.83 36.34 31.80 12.99 5.40 31.09 16.73 23.91 CoG 68.03 42.29 49.85 80.00 45.45 20.20 53.39 48.55 50.97 Qwen3-235B Direct 54.67 44.33 54.95 55.20 32.47 20.80 51.32 36.16 43.74 CoG 79.00 55.11 56.46 87.60 54.55 32.80 63.52 58.32 60.92 gpt-oss-120b Direct 18.67 29.67 36.64 20.80 18.83 8.00 28.33 15.88 22.10 CoG 42.33 40.33 45.95 63.40 38.64 18.80 42.87 40.28 41.58 DeepSeek-V3.2 Direct 53.67 45.33 44.74 39.20 31.17 13.80 47.91 28.06 37.99 CoG 85.33 51.67 57.36 88.40 59.09 34.00 64.79 60.50 62.64 Proprietary Models (Sampled Set, N=100) GPT-5.1 Direct 54.00 47.00 36.00 59.00 40.00 17.00 45.67 38.67 42.17 CoG 81.00 49.00 38.00 82.00 55.00 29.00 56.00 55.33 55.67 Gemini-2.5-Flash Direct 62.00 46.00 27.00 65.00 46.00 22.00 45.00 44.33 44.67 CoG 73.00 44.00 31.00 78.00 58.00 23.00 49.33 53.00 51.17 Table 9: Detailed performance comparison (EM, %) of different backbone LLMs with and without the CoG framework across six datasets (WebQSP excluded). For Qwen3-32B, we additionally include strong baselines from various categories (Text-based, KG-based, and Hybrid RAG) to demonstrate that CoG with a lightweight 8B model can outperform these 32B baselines. Appendix F Detailed Results on Exploration Depth Table 10 provides the full results corresponding to the exploration depth analysis discussed in Section 4.4 and visualized in Figure 3. The table lists the Exact Match (EM) scores for CoG across exploration depths from 1 to 6 on all seven benchmark datasets, alongside the average performance on KG-based and Text-based tasks. Method KG-based QA Text-based QA Avg. KGQAGen CWQ QALD10 WebQSP Avg. 2Wiki AdvHotpot MusiQue Avg. VectorRAG 48.66 30.76 39.04 70.49 47.24 52.60 32.14 12.40 32.38 40.87 CoG (Depth=1) 67.47 37.99 45.35 80.63 57.86 48.00 26.62 4.60 26.41 44.38 CoG (Depth=2) 71.83 42.58 51.95 81.05 61.85 85.20 46.43 20.00 50.54 57.01 CoG (Depth=3) 74.24 42.58 53.75 81.75 63.08 86.80 49.35 25.00 53.72 59.07 CoG (Depth=4) 74.70 42.87 55.26 81.75 63.65 85.20 51.95 27.80 54.98 59.93 CoG (Depth=5) 74.51 42.97 51.65 81.47 62.65 84.00 52.92 26.00 54.31 59.07 CoG (Depth=6) 75.90 43.65 54.35 80.56 63.62 83.20 53.25 25.60 54.02 59.50 Table 10: Full results of exploration depth analysis. We report Exact Match (EM) scores for CoG with maximum exploration depths ranging from 1 to 6 on all seven datasets, alongside the VectorRAG baseline. The best result for each dataset is highlighted in bold. Appendix G Detailed Efficiency Profile This section details CoG’s computational overhead across token consumption, end-to-end latency, and wall-clock time per retrieval round. Method KGQAGen CWQ QALD10-en WebQSP 2WikiMQA AdvHotpotQA MusiQue Token Latency Token Latency Token Latency Token Latency Token Latency Token Latency Token Latency ReAct 3,850 14.1 3,996 13.7 3,908 13.8 3,870 12.3 3,984 18.8 3,631 13.2 7,758 20.0 IRCoT 23,832 32.9 26,107 28.9 23,483 33.7 22,798 19.5 22,350 41.5 23,038 38.2 29,798 46.1 Search-o1 12,262 98.4 13,505 119.8 10,611 84.0 9,891 62.6 19,196 140.5 11,932 77.9 20,913 120.0 ToG 6,552 68.7 4,168 41.9 4,154 69.7 2,963 39.1 4,779 70.5 4,946 68.4 4,695 75.2 PoG 8,752 16.7 5,117 12.7 3,660 10.9 2,862 7.0 4,909 16.6 7,220 23.7 11,899 28.4 DoG 15,494 168.7 10,304 133.9 3,088 20.6 3,320 18.1 6,671 105.4 10,364 158.7 8,575 134.0 ToG-2 4,193 109.7 3,424 78.6 3,273 53.3 3,127 51.3 3,067 42.7 3,398 51.0 4,008 95.0 CoG 36,323 285.1 29,465 194.2 15,636 95.7 14,204 80.7 15,508 121.8 18,958 145.8 38,876 263.5 Table 11: Token consumption and latency (seconds) per question across different methods. Values represent the median of the dataset to reflect the typical cost for a standard query. Complexity Adaptability. CoG dynamically scales resources to query difficulty (Table 11). For simpler tasks (WebQSP, 2WikiMQA), it uses fewer tokens (∼ 14-15k) than IRCoT (∼ 22k) yet achieves higher accuracy. For complex tasks (MusiQue, KGQAGen), it invests more tokens into deep reasoning, yielding substantial gains where efficient methods fail (+11.6% on MusiQue over ToG-2). Latency Breakdown. Figure 4 details the average wall-clock time per retrieval round. The latency primarily stems from sequential LLM calls in the cognitive cycle. However, this extended reasoning time supports a fully open-ended retrieval setting: CoG starts from the raw question without pre-annotated entities, directly grounds and retrieves evidence, and adaptively adjusts KG/text exploration according to the evolving reasoning state. Furthermore, this deep filtering ensures precise localization, reducing accessed documents by 12.4×12.4× compared to ToG-2 (Table 5). The current wall-clock time can be further optimized by parallelizing the dual-source exploration modules and concurrent sub-queries. Figure 4: Average wall-clock time (seconds) breakdown per retrieval round across different datasets. Cost-Effectiveness. While CoG consumes more tokens on complex tasks, this is offset by its ability to utilize smaller, cheaper models. As shown in Appendix E, CoG with an 8B model outperforms strong 32B baselines, suggesting a more favorable accuracy-cost trade-off for practical deployment. Appendix H Sensitivity Analysis on Entity Linking To evaluate the robustness of CoG to the candidate retrieval size (top-k) in the entity linking module, we vary the number of candidate entities from 66 to 3030. We conduct this analysis on four representative datasets, covering both KG-based and text-based QA tasks. For KGQAGen, we evaluate on a 300-sample subset to control computational cost. k KGQAGen 2Wiki Hotpot MusiQ Avg. 6 79.3 84.8 47.7 25.6 59.4 12 78.3 84.2 49.7 25.2 59.4 18 74.3 85.8 50.0 26.2 59.1 20 76.3 85.2 52.0 27.8 60.3 30 78.0 85.6 51.3 26.2 60.3 Table 12: Sensitivity analysis of entity-linking candidate size k. We report EM (%) on four representative datasets. As shown in Table 12, CoG is robust to the entity-linking candidate size. Across k=6k=6 to 3030, the average EM remains within a narrow range of 59.1%–60.3%. Even with a narrow retrieval scope (k=6k=6), CoG achieves a strong average EM of 59.4%, while increasing k provides broader entity recall without causing overall performance degradation. We set k=20k=20 as the default because it achieves the best average performance, tied with k=30k=30, while using fewer candidate entities. This indicates that CoG’s context-aware disambiguation effectively filters noisy candidates while benefiting from broader entity recall. Appendix I Manual Evaluation and Error Study As the Exact Match (EM) metric strictly relies on string matching, it often fails to evaluate the validity of the underlying reasoning process. To gain deeper insights, we conduct a manual analysis on a sampled subset. We assess Reasoning Faithfulness for cases where the EM score is correct (EM=1), and perform an Error Diagnosis for cases where the model is judged incorrect (EM=0). I.1 Reasoning Faithfulness For predictions judged correct by EM, we further examine whether the model reaches the answer through a faithful reasoning process. We categorize each EM-correct prediction into four types: • True Positive (TP): The reasoning chain is logically sound, and the extracted evidence fully supports the final answer. • Flawed True Positive (Flawed TP): The final answer and core reasoning direction are correct, but the reasoning contains minor verification gaps, shortcut steps, or non-critical factual imperfections. • Weak False Positive (Weak FP): The final answer is correct by coincidence, but the core intermediate entities or facts are substantially incorrect, i.e., the model is right for the wrong reasons. • False Positive (FP): The model fails to deduce the correct answer, but the EM metric falsely judges it correct due to substring matching in the reasoning text. Table 13 shows that CoG produces the most faithful EM-correct predictions. Its severe false-positive rate (Weak FP + FP) is only 3.6%, much lower than ToG-2 (12.7%) and Self-Consistency (29.1%). Even when Flawed TP is treated as non-ideal reasoning, CoG remains the most reliable method. Self-Consistency is vulnerable to ungrounded parametric reasoning, often producing hallucinated bridge entities or fabricated attributes (Weak FPs), or long CoT traces that accidentally match the gold string (FPs). ToG-2 reduces such hallucinations through retrieval, but its reactive exploration can still preserve wrong intermediate entities or unresolved constraints. CoG’s graph-text exploration and reflection better ground intermediate steps, so most of its imperfect cases involve incomplete verification of secondary constraints (Flawed TPs) rather than broken core reasoning chains. Method TP Flawed TP Weak FP FP Self-Consistency 60.0% 10.9% 21.8% 7.3% ToG-2 81.8% 5.5% 3.6% 9.1% CoG (Ours) 89.1% 7.3% 1.8% 1.8% Table 13: Distribution of reasoning faithfulness for predictions evaluated as correct by EM. Figure 5 further shows that identical EM scores can hide different reasoning quality. While SC and ToG-2 may reach the correct answer through hallucinated or weakly supported intermediate links, CoG derives the answer from a more grounded and faithful reasoning path. This example also involves a temporal update: although the dataset reference path is based on an older coach, CoG identifies the updated coach and reaches the correct nationality through verifiable evidence. Question: What nationality is the sport club that the current coach of Werder Bremen played professionally for? Gold Answer: German Reference Reasoning Path: Werder Bremen → (coach) Alexander Nouri → (played for) VfL Osnabrück → (nationality) German ✗SC (Weak FP): Werder Bremen → Florian Kohfeldt → [Hallucinated] VfL Wolfsburg → German Hallucinates that coach F. Kohfeldt played for VfL Wolfsburg (he only coached them). Hits the answer “German” purely by coincidence. ✗ToG-2 (Weak FP): Werder Bremen → Florian Kohfeldt → [Guessed] Werder Bremen → German Fails to retrieve his actual former clubs. Circularly assumes that since he coaches Werder Bremen (a German club), he must have played for them, guessing “German”. ✓CoG (True Positive): Werder Bremen → Daniel Thioune → VfL Osnabrück → German Identifies current coach D. Thioune from updated knowledge. Retrieves his actual former clubs (e.g., VfL Osnabrück) and confirms they are German. Figure 5: A qualitative example showing CoG’s superior reasoning faithfulness despite identical EM scores across models. I.2 Error Diagnosis To understand the limitations of CoG, we conduct a primary-cause error diagnosis on a sampled subset of instances where CoG fails the EM evaluation. As illustrated in Figure 6, we categorize the errors into External Factors and Internal Errors. Figure 6: Taxonomy of error causes for CoG based on manual analysis. External Factors (71.1%): A large portion of the errors are attributable to benchmark or evaluation limitations. • Metric False Negative (37.8%): The model successfully completes the multi-hop reasoning but fails the strict string matching. Common instances include surface-form mismatch (aliases or pseudonyms), paraphrases, or answer granularity differences (city vs. metro area). • Dataset Flaws (26.7%): These include typos in the questions, grammatically ambiguous prompts leading to misaligned relation extraction, and noisy ground-truth annotations. • KB Temporal Mismatch (6.7%): Temporal mismatch occurs when the model retrieves updated facts from the knowledge source (e.g., recent sports team transfers), which conflict with the outdated static dataset ground truth. Internal Errors (28.9%): The remaining errors represent genuine limitations in the agentic exploration process. • Exploration Distraction (13.3%): In environments with long texts or ambiguous clues, the agent occasionally gets trapped by irrelevant entities or over-thinks historical details, exhausting its maximum exploration turns. • Disambiguation (6.7%): The model sometimes overlooks rigid constraints during entity disambiguation (e.g., ignoring a "state school" constraint and erroneously selecting a private school). • Reasoning Hallucination (4.4%) & Parametric Guess (4.4%): The model occasionally struggles with counting logic over complex long tables, or bypasses retrieval entirely by relying on over-confident parametric guesses. Overall, this diagnosis shows that raw EM can underestimate CoG, with many failures caused by rigid string matching, noisy annotations, or temporal mismatch. The remaining internal errors highlight important directions for improvement in exploration focus, entity disambiguation, and evidence-grounded verification. Appendix J Case Studies Table 14 and Table 15 compare the reasoning trajectories of ToG-2 and CoG on a complex multi-hop question from AdvHotpotQA. ToG-2 Analysis. It fails due to its passive, graph-driven exploration strategy. The system starts by retrieving general information about “Stanford Cardinal” but fails to locate the specific player due to the absence of statistical attributes in the KG. Lacking a reflection mechanism, ToG-2 cannot diagnose this failure. Instead, it blindly continues its beam search, pivoting to high-ranking but irrelevant entities like “quarterback” in Turn 3, ultimately leading to an irreversible dead end. CoG Analysis. In contrast, CoG demonstrates robust navigation through its cognitive cycle and bidirectional synergy. CoG initially encounters the same dead end as ToG-2 in Turn 1. However, Reflection diagnoses the strategy as too narrow. Proactive Planning then pivots to broader queries like “football receiving leaders” instead of the player directly (Turn 2). Crucially, Text Entity Utilization bridges KG gaps by identifying the entity “James Lofton”, which aligns with the statistical constraints in the question, from Wikipedia related sections. This bidirectional synergy enables CoG to dynamically identify the target player and verify the specific statistics in Turn 3, successfully deriving the answer “David”. Question: What is the middle name of the former Stanford Cardinals player who caught 57 passes for 1010 yards and 14 TDs? Pre-annotated Entities: Stanford Cardinal (Q2938060), Touchdown (Q650807) Ground Truth: David Turn ToG-2 Reasoning Process turn 1 Starting Entities: Stanford Cardinal (Q2938060), Touchdown (Q650807) Relations Explored: member of sports team, sport, educated at, facet of, named after (6 relations) Retrieved Entities: 57 entities (49 from “member of sports team”, 4 from “sport”, 1 from “educated at”, 2 from “Touchdown” relations) Notable: 49 Stanford athletes including Greg Camarillo, John Elway, Mark Marquess, etc. Document Retrieval: Retrieved Wikipedia pages for 55 entities (2 pages failed: sports terminology, Margaret Hayden) Issue: No specific statistics matching 57/1010/14 found in documents Reasoning: Information insufficient - cannot identify the specific player Entity Ranking (for next turn): [0.28] Greg Camarillo | [0.12] John Elway | [0.05] Mark Marquess turn 2 Starting Entities: Greg Camarillo (Q1248387), John Elway (Q432931), Mark Marquess (Q6768732) Relations Explored: given name, educated at, member of sports team, position played, occupation (9 relations) Retrieved Entities: 11 entities (Greg, Menlo-Atherton HS, Stanford Univ., 5 NFL teams, John, quarterback, Mark, baseball player, Stanford Cardinal) Document Retrieval: Retrieved Wikipedia pages for all 11 entities Issue: Found Greg Camarillo’s given name but no middle name; no confirmation of statistics Reasoning: Cannot confirm Greg Camarillo is the target player or find his middle name Entity Ranking (for next turn): [0.45] Stanford Cardinal | [0.00] quarterback turn 3 Starting Entities: Stanford Cardinal (Q2938060), quarterback (Q622747) Relations Explored: member of sports team, position played, sport (6 relations) Retrieved Entities: 106 entities (50 from “member of sports team”, 4 from “sport”, 1 from “educated at”, 50 from “position played”, 1 from “part of”) Notable: 50 quarterbacks (Logan Thomas, John Brantley, Charlie Whitehurst, etc.), 50 Stanford athletes Document Retrieval: Retrieved Wikipedia pages for 103 entities (3 pages failed: Cameron Dantley, Dai Minagawa, Margaret Hayden) Critical Failure: Explored wrong position (quarterback vs. wide receiver) No Reflection: System continued beam search without recognizing strategic misalignment Entity Ranking: [0.31] Logan Thomas | [0.12] John Brantley | [0.02] Charlie Whitehurst Final Reasoning: Information insufficient to identify the player or middle name Outcome: FAILURE - Generated answer without evidence Table 14: Reasoning trace of ToG-2. The method employs a rigid beam search where the next exploration steps are determined by ranking score based on entity document relevance. Without reflection, the pruning mechanism blindly discards potentially correct paths and misguides the search towards irrelevant high-ranking entities, leading to an irreversible failure. Red highlights indicate critical failure analysis for the system. Question: What is the middle name of the former Stanford Cardinals player who caught 57 passes for 1010 yards and 14 TDs? Ground Truth: David Turn CoG Reasoning Process Turn 1 Planning Phase: Analysis: “To determine the middle name, I first need to identify the specific player who caught 57 passes for 1010 yards and 14 touchdowns.” Query: “Stanford Cardinals player with 57 catches, 1010 yards, 14 TDs” Entity: “Stanford Cardinals player” Exploration Phase: KG: Entity linking failed - 9 candidates all team-level entities (e.g., Stanford Cardinal football, Stanford Cardinal men’s soccer), no individual player entities Wikipedia: Searched “Stanford Cardinal football” page, extracted team context (Jim Plunkett, Andrew Luck) but no player matching the statistics Synthesis: No player-specific information found Reflection: Judgment = INSUFFICIENT_USELESS “The initial approach was too narrow… The core strategic error was not breaking the problem into smaller, more general steps.” Turn 2 Recovery Planning: Strategy Shift Reflection: “Instead of starting with the player’s name, we should first identify the season or year when these stats might have been achieved, or search for a list of Stanford receivers with notable seasons.” Updated Analysis: “Shift focus to general sources of Stanford football statistics or notable players.” Next Queries: [“Stanford football all-time receiving leaders”, “Stanford football players with 14 TDs in a season”] Exploration (Query: “Stanford football all-time receiving leaders”): KG: Linked to “Stanford Cardinal football statistical leaders” (Q19878473) Relation Discovery: Found 1 relation - “sport” (P641) → American football Fact Pruning: Kept 0 facts (too general, not useful for identifying specific players) Wikipedia: Found “Stanford Cardinal football statistical leaders” page with “Receiving” section “Receiving” section containing 9 statistical tables, extracted the following information from table: “Single season 6” (receiving yards): James Lofton - 1010 yards (1977) “Single season 7” (receiving TDs): James Lofton - 14 TDs (1977); J.J. Arcega-Whiteside - 14 TDs (2018) Key Discovery: Identified “James Lofton” as candidate with 2/3 statistics matched (1010 yards + 14 TDs in 1977) Synthesis: Strong candidate found but 57 receptions unconfirmed Reflection: Judgment = INSUFFICIENT_USEFUL - Continue with James Lofton Turn 3 Planning Phase: Query: “James Lofton 1977 season stats” Entities: “James Lofton” Exploration Phase: KG: Linked to James Lofton (Q1680714) Identity Check: educated at → Stanford; member of team → Stanford Cardinal. Attribute Retrieval: Found alias “James David Lofton” (Directly reveals middle name) Wikipedia: Found “James Lofton” page From “College career” section: Confirmed “57 receptions for 1, 010 yards and 14 touchdowns during the 1977 season” Synthesis: KG confirmed identity & name; Text verified the missing statistical constraint. Middle name derivation, complete. Reflection: Judgment = SUFFICIENT Final Answer: “The middle name of the former Stanford Cardinals player is David.” Outcome: SUCCESS in 3 turns Table 15: Reasoning trace of CoG. Reflection and Proactive Planning diagnose the initial failure and recover from dead ends. Adaptive Entity Linking identifies implicitly useful entities (“1977 Stanford Cardinals football team”) not directly mentioned in the question. Finally, Text Entity Utilization effectively extracts the target entity (“James Lofton”) from text to guide the final answer derivation. Purple highlights show successful reflection and strategic adjustment; blue highlights show key information extraction from dual sources (KG structure + Wikipedia text). Appendix K Prompts This section provides the main prompt used in CoG to guide LLM reasoning at each stage of the cognitive cycle. Table 16 presents the prompt used for the initial planning phase (Section 3.2.1). The prompt includes four examples to guide the LLM in handling diverse multi-hop reasoning scenarios. Tables 17–20 present the prompts for the three-stage KG Exploration (Section 3.2.2): Entity Linking, Relation Discovery, and Fact Pruning, respectively. These prompts demonstrate how CoG leverages in-context learning to achieve precise, schema-aware navigation over large-scale KGs. Tables 21–24 detail the prompts used in the Text Exploration module (Section 3.2.2), covering Adaptive Page Selection (Tables 21 and 22) and Hierarchical Reading (Tables 23 and 24). Table 25 presents the prompt for the Synthesis module. Tables 26 and 27 provide examples of structured evidence formats for KG and text, respectively, that populate the $evidence_blocks placeholder in the Synthesis prompt. Tables 28 and 29 detail the prompts for the Reflection module (Section 3.2.3), enabling the system to either deepen exploration based on new findings or pivot its strategy by diagnosing failures from the interaction history. ROLE You are a highly strategic retrieval planner. Your function is to analyze a complex question and identify the immediate, parallelizable search actions required to proceed. TASK Given a user’s question, do NOT provide the final answer. Instead, determine the next logical search queries. A single question might require multiple parallel queries to gather the necessary initial information. Your plan must include your reasoning, a list of search queries, and a corresponding list of entities. CONSTRAINTS - The number of queries in the ‘Query’ list must not exceed 5. - [CRITICAL] The ‘Query’ and ‘Entities’ lists must have the exact same number of items. For each query, the corresponding item in the ‘Entities’ list must be the single, most specific, and central named entity being investigated in that query. - Only output the query/queries for the immediate next step. Do not plan, describe, or allude to any subsequent steps. OUTPUT FORMAT Your output must follow this exact structure, with no additional commentary: Analysis: Your thought process explaining why these next steps are necessary and what information is being sought. Query: [A Python-style list of concise and effective search query strings.] Entities: [A Python-style list of the core entities in the Query, corresponding one-to-one with the Query list.] EXAMPLES Example 1: Question: What is the nationality of one of the men for whom the Schumann–Runge bands are named? Output: Analysis: The question requires the nationality of a person. Before I can find their nationality, I must first identify the individuals the "Schumann–Runge bands" are named after. Query: ["who are the Schumann-Runge bands named after"] Entities: ["Schumann–Runge bands"] … (Other examples omitted for brevity) … Example 4: Question: Where did the leader of the largest European country after the collapse of the country that denied anything more than an advisory role in the Korean war die? Output: Analysis: The question is complex and requires multiple pieces of information. First, I need to identify which country denied anything more than an advisory role in the Korean War. […] I will begin by identifying the country that denied a combat role in the Korean War. Query: ["country that denied anything more than an advisory role in Korean War"] Entities: ["Korean War"] YOUR TASK Question: $question Output: Table 16: Prompt template for the Initial Planning module. The LLM is instructed to analyze the question, generate parallelizable sub-queries, and identify corresponding anchor entities. This structured output initializes the exploration cycle. ROLE You are an expert in Knowledge Graph entity linking. Your task is to disambiguate an entity mention from a user’s question by matching it to the correct entity in a knowledge graph, using the surrounding context. CONTEXT - Original Question: $question - Overall Plan (Analysis): $analysis - Current Sub-Query: $query - Entity Mention to Link: $entity CANDIDATE ENTITIES Here are the top candidate entities in the Knowledge Graph, sorted by a preliminary relevance score. Each candidate includes its Wikidata QID, label, description, aliases, popularity, some of its neighbors, and relevance score. $formatted_candidates YOUR TASK Your task is to critically evaluate the candidate entities based on the provided CONTEXT. Your goal is to either identify the single correct entity OR determine that no suitable match exists. • Analyze the Context: Carefully review the ’Original Question’ and ’Overall Plan’. What are the key details about the entity $entity (e.g., their time period, relationships, role)? • Evaluate Each Candidate: For each candidate, compare its Description, Aliases, and Neighborhood against the context. A correct match should be consistent with the context. • Make a Decision: – If, and only if, you find one candidate that is a confident and accurate match for $entity based on all available information, your output should be its QID on a single line. – If none of the candidates are a confident match, or if their key details contradict the context, you MUST output the single word: NO_MATCH. OUTPUT FORMAT • If a confident match is found, your entire output MUST be only the QID of that entity (e.g., Q12345). • If no confident match is found, your entire output MUST be the single word: NO_MATCH. • Your output must not contain any other text, explanation, or reasoning. YOUR RESPONSE: Table 17: Prompt template for the Entity Linking module. The LLM is tasked with precise disambiguation by comparing candidate profiles against the question context. The placeholder $formatted_candidates is populated with structured candidate data, as illustrated in Table 18. Example of Formatted Candidates --- [Entity] - QID: Q716680 - Label: Steve McQueen - Description: American actor (1930-1980) - Aliases: Terrence Stephen McQueen, King of Cool - Popularity (Degree): In: 154, Out: 45, Attr: 22 [Neighborhood] - Outgoing Relations (as Head): - spouse: Neile Adams, Ali MacGraw, Barbara Minty - child: Chad McQueen, Terry Leslie McQueen [Scores] - Final: 0.952 (Name similarity=1.000, Popularity=0.850, Description=0.920, Neighbor=0.880) --- ... (Other candidates omitted) ... Table 18: An example of the formatted candidate data populated into the $formatted_candidates slot in Table 17. This structured format provides the LLM with comprehensive entity details (profile, neighborhood, and retrieval scores) for disambiguation. ROLE You are an expert Knowledge Graph analyzer. Your task is to select the most relevant relations (properties) of a given entity that will help answer a specific question. CONTEXT - Original Question: $question - Overall Plan (Analysis): $analysis - Current Sub-Query: $query - Entity in Focus: $entity_label ($entity_qid): $entity_description AVAILABLE RELATIONS The entity $entity_label is connected to $total_relations_count relations (Outgoing: $outgoing_count, Incoming: $incoming_count) in the Knowledge Graph. Below is a list of all relations connected to the entity. Each relation includes its label, PID, and a frequency indicating how many other entities in the graph are linked via this relation. Outgoing Relations: $outgoing_relations Incoming Relations: $incoming_relations YOUR TASK Based on the CONTEXT, select relations that will help find the answer using a two-pronged approach: 1. Direct Relations: Select relations that seem to directly contain the answer. 2. Exploratory Relations: If no direct relations exist, select relations that could lead to intermediate entities, which might then contain the answer. OUTPUT FORMAT Your output must follow this exact structure, with no additional commentary: Reasoning: Your brief thought process explaining why you chose these specific relations based on the question and the entity. Selected PIDs: [A Python-style list of strings, where each string is the PID of a selected relation (e.g., ["P31", "P17"]). If no relations seem relevant, output []] YOUR RESPONSE: Table 19: Prompt template for the Relation Discovery module. The LLM selects a subset of relations relevant to the sub-query to prune the search space at the schema level. ROLE You are an expert assistant specializing in knowledge graph analysis for question answering. Your mission is to intelligently prune a list of facts, guiding a multi-step exploration process. CONTEXT - Original Question: $question - Overall Plan (Analysis): $analysis - Current Sub-Query: $query - Entity in Focus: $entity_label AVAILABLE FACTS Below is a structured list of all facts retrieved from the Knowledge Graph for the entity $entity_label. The facts are categorized by relation type (Outgoing or Incoming). $formatted_facts YOUR TASK Your goal is to decide which facts to keep. Evaluate the facts using a two-pronged approach: 1. Directly Relevant Facts: Keep any facts that directly help answer the Current Sub-Query. 2. Promising Intermediate Entities: Keep facts that represent entities which are promising stepping stones. A fact is "promising" if exploring it further is highly likely to lead to the answer or provide crucial context. Discard facts that are clearly irrelevant, noisy, or are just general information. OUTPUT FORMAT Your output MUST be a single, valid JSON object with the following structure: "reasoning": "Structured thought process...", "pruned_facts": "outgoing": "P17": ["kept fact labels"], ... , "incoming": "P802": ["kept fact labels"], ... YOUR RESPONSE: Table 20: Prompt template for the Fact Pruning module. The LLM performs fine-grained filtering of facts to retain only critical evidence or promising exploration leads. ROLE You are a research agent resolving an ambiguous search query on Wikipedia. TASK Your search for "$entity" led to a disambiguation page with multiple possible meanings. You must carefully analyze the full context provided below to choose the single most relevant link from the options. CONTEXT - Original Question: $question - Overall Plan (Analysis): $analysis - Current Sub-Task (Query): $query - Ambiguous Search Query: "$entity" OPTIONS $options_list YOUR DECISION Analyze the options based on the context to select the best fit. Your output must be a single line containing only the exact title of the chosen page from the list. If none of the options seem relevant for answering the original question, output the string "NO_MATCH". YOUR RESPONSE: Table 21: Prompt template for the Disambiguation step in Adaptive Page Selection. The LLM resolves ambiguous entities by selecting the correct page from disambiguation options based on context. ROLE You are a research agent trying to find the right information on Wikipedia after a search query failed. TASK Your search for "$entity" did not find a matching Wikipedia page. Use the contextual information and the suggestions provided to decide on the best query for your next attempt. CONTEXT - Original Question: $question - Overall Plan (Analysis): $analysis - Current Sub-Task (Query): $query - Failed Search Query: "$entity" SUGGESTIONS Below are suggestions from the Wikipedia search API to help you refine your query. "Did you mean?": This is the API’s top recommendation, often correcting a typo or suggesting a more standard page title. - Suggestion: $suggestion Similar Pages Found: These are pages with titles that are textually similar to your failed query. One of them might be the correct page under a slightly different name. - Similar Pages: $search_results_list YOUR DECISION Based on the context and the suggestions, choose the best query for your next search attempt. This could be the "Did you mean?" suggestion, one of the similar pages, or a completely new query you formulate based on the feedback from this failed search. Your output must be a single line containing only the new search query. If none of the suggestions seem useful for answering the original question, output the string "NO_MATCH". YOUR RESPONSE: Table 22: Prompt template for the Search Refinement step in Adaptive Page Selection. The LLM uses API feedback (corrections and similar pages) to recover from search failures. ROLE You are a research agent tasked with answering a complex question by navigating Wikipedia. TASK You have been provided with the summary, an infobox table, and the section list of a Wikipedia page relevant to your current query. Your goal is to extract all useful information from the summary and infobox, identify promising sections for deeper investigation, or determine that the page is irrelevant. CONTEXT - Original Question: $question - Overall Plan (Analysis): $analysis - Current Search Query: "$query" - Retrieved Wikipedia Page Title: "$page_title" AVAILABLE INFORMATION Page Summary: $page_summary Summary Infobox: $infobox_table Page Sections (Top-Level): $page_sections YOUR DECISION Based on the information above, what is your next best action? Choose one of the following two options. 1. If the page seems relevant: • Extract Direct Information and Promising Clues from the summary/infobox. • Identify sections for deeper investigation. 2. If the page is clearly irrelevant: Explain why and suggest a better search query. YOUR RESPONSE: Table 23: Prompt template for the Global Skimming module. The LLM assesses page relevance, extracts high-level evidence from the summary and infobox, and selects promising sections for detailed reading. ROLE You are a meticulous research agent. Your task is to analyze a section of a Wikipedia page that contains both text and tables to extract relevant information for a given query. TASK You have been given several retrieved text chunks and a preview of all tables found within a specific Wikipedia section. Your goal is to: 1. Extract Direct Information and Promising Clues from BOTH the Retrieved Text Chunks and the Table Previews. 2. Analyze table previews. If a preview is insufficient but promising, select the table for a full read. 3. Provide a single, unified rationale that explains both your information extraction findings and your table selections. CONTEXT - Original Question: $question - Current Search Query: "$query" - Section Being Investigated: "$section_title" - Section Exploration Rationale (during summary reading): $exploration_rationale AVAILABLE INFORMATION 1. Retrieved Text Chunks: (Top $k relevant chunks) $context_chunks 2. Table Previews: (Previews of tables: $table_names_in_section) $tables_preview OUTPUT FORMAT Rationale: A consolidated explanation covering both text and tables. Extracted Info: Key info from text chunks and table previews. State "None" if nothing relevant is found. Selected Tables: List of table names requiring a full read (e.g., ["Awards"]). Empty list [] if none. YOUR RESPONSE: Table 24: Prompt template for the Detailed Reading phase. The LLM performs joint analysis of retrieved text chunks and table previews to synthesize fine-grained evidence. ROLE You are a master AI strategist leading a multi-hop question-answering mission. Your task is to synthesize retrieved information from various sources, evaluate progress against the overall plan, and decide the most logical next step. CONTEXT - Original Question: $question - Overall Plan (Analysis): $analysis - Notebook (Summary of Known Facts): $notebook - Current Sub-Queries: $queries - Core Entities in Sub-Queries: $entities EVIDENCE This section contains the information retrieved from different sources for all sub-queries executed in this turn. $evidence_blocks YOUR TASK Carefully review all evidence, and in conjunction with the Original Question and your Overall Plan, complete the following three steps: 1. Verbatim Information Extraction: Meticulously extract all useful information. Your goal is to faithfully transfer potentially relevant information to your ‘Extracted Content’. Include: • Direct Facts: Core facts that directly contribute to answering the original question. • Promising Leads: New entities or critical factual clues essential for guiding the next step. • [CRITICAL] Do NOT summarize or rephrase. This is for extraction only. 2. Think Step-by-Step: Document your thought process. • Explain how the direct facts help answer the question. • Discuss the potential value of promising leads and how they might be explored. • Identify what key information is still missing. • If evidence is useless, explain why the current sub-queries failed. 3. Make a Judgment: Choose one of the following options: • SUFFICIENT: Information is adequate to generate a final, complete answer. • INSUFFICIENT_USEFUL: Valuable clues found, but more information is needed. Continue investigation based on the current findings. • INSUFFICIENT_USELESS: Information is irrelevant or has led to a dead end. A new strategy is needed. OUTPUT FORMAT Thought Process: Your step-by-step thinking process… Extracted Content: A structured collection of key facts and promising new leads… Judgment: SUFFICIENT, INSUFFICIENT_USEFUL, or INSUFFICIENT_USELESS. YOUR RESPONSE: Table 25: Prompt template for the Synthesis module. The LLM aggregates dual-source evidence, extracts key insights, and judges the current progress to determine the next action (answer, continue, or recover). Examples of the content for $evidence_blocks are shown in Tables 26 and 27. Example of Formatted Candidates KG Exploration for Mention: "John Cage" [Entity Linking] - Linked Entity: John Cage (Q180727) - Description: American composer and music theorist - Candidates (top-4 preview): - John Cage (Q180727): American composer and music theorist - John Cage (Q5347597): Fictional character from Ally McBeal - Linking Rationale: The query context discusses "4’33" and "prepared piano", which are signature works of the composer John Cage. [Retrieved Facts for "John Cage"] [Attributes] - date of birth: [’1912-09-05’] - occupation: [’composer’, ’philosopher’, ’artist’] [Outgoing] - notable work: ["4’33’", ’Music of Changes’, ’Imaginary Landscape No. 4’] - student of: [’Arnold Schoenberg’, ’Henry Cowell’] [Incoming] - influenced by: [’Erik Satie’, ’Marcel Duchamp’, ’D. T. Suzuki’] [Filtering Summary] - Relations explored: 5 of 21 total - Facts kept: 12 of 43 total [Reasoning] - Relation Selection: Selected relations regarding works, influences, and personal life relevant to his artistic development. - Fact Pruning: Kept only major works and key figures like Schoenberg and Cunningham; discarded minor administrative categories. Table 26: An example of structured KG evidence formatted for the $evidence_blocks in the Synthesis module. It includes the linked entity, verified facts (attributes and relations), and a reasoning trace for the pruning process. Example of Formatted Candidates Wikipedia Retrieval Results for entity: "John Cage" Information Extraction: - Page Title: John Cage - All Page Sections: [’Life’, ’Music’, ’Visual art...’, ...] - A. Skimming Summary and Section Selection: - Rationale: The summary highlights Cage as a pioneer of indeterminacy and the prepared piano. To understand his specific methods, we need to explore the "Music" section. - Selected Sections for Deeper Analysis: [’Music’] - B. Extracted Content: - From Summary: John Cage was a pioneer of indeterminacy in music. Best known for "4’33’", a silent composition. He developed the "prepared piano" and used the "I Ching" as a standard composition tool. - From Sections: - Section: "Music" - Rationale for section processing: Contains detailed descriptions of his composition techniques like chance operations and rhythmic structures. - Extracted from Text: - In 1951, Cage started using the I Ching to compose using chance, imitating nature’s manner of operation. - Music of Changes (1951) was the first major work created using this method. - Cheap Imitation (1969) is a chance-controlled reworking of Erik Satie’s Socrate. Table 27: An example of Text evidence formatted for the $evidence_blocks in the Synthesis module. It summarizes the high-level page content and provides fine-grained details extracted from specific sections relevant to the query. ROLE You are a master AI strategist leading a multi-hop question-answering mission. Your task is to plan the next step of the investigation after an information-gathering turn that was useful but insufficient. CONTEXT - Original Question: $question - Current Notebook: $notebook - Previous Overall Plan: $analysis - Previous Sub-Queries: $queries - Candidate Entities Pool: $candidate_entities_pool NEW FINDINGS - Thought Process: $thought_process - Extracted Content: $extracted_content YOUR TASK Plan the next round of investigation to dig deeper and bridge information gaps. 1. Update Notebook: Combine current notebook with new findings into a single, coherent summary, retaining all unique and relevant details. 2. Update Analysis: Revise the overall plan to reflect new understanding. Explain what the immediate next step should focus on and why. 3. Plan Next Queries: Define a new set of sub-queries and corresponding core entities based on the updated analysis. 4. Manage Candidate Pool: Review existing leads and identify new promising entities from the new findings. Update the pool by adding new leads and removing promoted ones. OUTPUT FORMAT Your output must follow this exact structure: Thought Process: Your step-by-step reasoning. Updated Notebook: The comprehensive summary of all known facts. Updated Analysis: Your revised analysis and plan. Next Queries: [List of concise and effective search query.] Next Entities: [List of core entities corresponding to queries] Updated Candidate Pool: [List of dictionaries, e.g., "entity": "Name", "reason": "…"] YOUR RESPONSE: Table 28: Prompt template for Continue Exploration in the Reflection module. The LLM updates the notebook and analysis based on new findings and plans the next steps to deepen the investigation. ROLE You are a master AI strategist leading a multi-hop question-answering mission. Your task is to recover from a failed information-gathering turn where the retrieved evidence was useless. CONTEXT - Original Question: $question - Current Notebook: $notebook - Candidate Entities Pool: $candidate_entities_pool - Interaction History: $interaction_history FAILED TURN DETAILS The last exploration turn was deemed unproductive. Here’s what went wrong: - Previous Overall Plan: $analysis - Failed Sub-Queries: $queries - Failed Core Entities: $entities - Extracted Content: $extracted_content - Reasoning for Failure: $thought_process YOUR TASK The previous approach hit a dead end. Critically reflect on the history, diagnose the error, and pivot the strategy. 1. Critical Reflection: Analyze why the previous approach failed (e.g., flawed plan, wrong entities). Summarize the core strategic error and derive actionable guiding principles for the next attempt. 2. Update Analysis: Propose a fundamentally new plan that leverages insights from reflection. Consider promoting leads from the Candidate Pool, re-examining the original question for missed keywords, or formulating queries from an entirely different angle. 3. Plan Next Queries: Define a new set of queries that represent a clear change in direction from the failed ones. 4. Manage Candidate Pool: Review the pool. Decide if existing leads are now high-priority, add any new promising leads discovered incidentally, and generate the updated pool by carrying over unused leads while removing promoted ones. OUTPUT FORMAT Your output must follow this exact structure: Thought Process: Critical reflection on failure and reasoning for the new plan. Updated Analysis: The revised analysis and new strategic direction. Next Queries: [List of concise and effective search query.] Next Entities: [List of core entities corresponding to queries.] Updated Candidate Pool: [List of updated candidate dictionaries, e.g., "entity": "Name", "reason": "…"] YOUR RESPONSE: Table 29: Prompt template for Strategy Adjustment in the Reflection module. The LLM diagnoses the failure using interaction history and pivots the strategy by generating entirely new queries or backtracking to candidate entities.