Cognition on Graph: Navigating Massive Knowledge Space via Cognitive Cycles and Bidirectional Graph-Text Synergy

1 Introduction

• RAG uses external knowledge to ground LLM responses, addressing factual hallucinations and opaque reasoning in knowledge-intensive scenarios.The cited motivation contrasts retrieval-grounded responses with reliance on static parametric knowledge.Source: §1 · p. 1
• Global exploration requires synthesizing graph structure and text semantics into evidence chains while navigating noise, ambiguity, and distractions.This differs from retrieval in simplified, local settings.Source: §1 · pp. 1–2
• Existing methods often rely on KG traversal, use text only as auxiliary context, or lack proactive adjustment and backtracking, leaving KG gaps unresolved.KG-only methods include Sun et al. (2024), Chen et al. (2024b), Ma et al. (2025a), and Wang et al. (2025); hybrid methods such as Ma et al. (2025b) treat text as auxiliary context.Source: §1 · p. 2
• CoG proactively analyzes the question to plan sub-queries and anchor entities, then retrieves relevant subgraphs and reads entity-related documents hierarchically.Synthesis combines the two sources into a unified retrieval summary; reflection uses exploration utility and interaction history to continue, adjust, or diagnose failure.Source: §1 · p. 2
• CoG is a training-free framework that adapts exploration through a continuous plan-explore-reflect cycle.It draws inspiration from human problem-solving and is designed for noisy global knowledge bases.Source: §1 · p. 2

2 Related Work

• Prior retrieval and GraphRAG methods often use oracle documents or domain-specific knowledge graphs, limiting assessment of robustness in noisy global environments.CoG targets global-scale sources such as Wikidata and Wikipedia.Source: §2 · p. 2
• Table 1 compares knowledge-base scale and four exploration capabilities: text guidance, reflection and backtracking, proactive planning, and adaptive entity linking.Its caption identifies CoG as operating at massive scale while integrating all four capabilities.Source: Table 1 · p. 3
• Global exploration methods use path-based prompting, coverage heuristics, backtracking, debate, or parallelism, while learning-based navigation incurs training costs and may lack interpretability.Examples include ToG (Sun et al., 2024), StructGPT (Jiang et al., 2023), KG-Agent (Jiang et al., 2025), PoG (Chen et al., 2024b), DoG (Ma et al., 2025a), and AnchorRAG (Xu et al., 2026).Source: §2 · pp. 2–3
• Hybrid methods use text as passive context or rely on offline structures, whereas CoG extracts text entities to dynamically guide graph traversal across structural gaps.The cited approaches include CoK (Li et al., 2024b), ToG-2 (Ma et al., 2025b), and methods using relevance scoring or pre-constructed hyperedges.Source: §2 · p. 3

3.1 Task Definition

• The task is multi-hop QA over a global hybrid knowledge base, building evidence chains through iterative retrieval and reasoning across KG facts and entity documents.The KG supplies triples and attributes; entity documents provide fine-grained semantic descriptions.Source: §3.1 · p. 3

3.2 CoG Framework

• CoG iterates planning, dual-source exploration, synthesis, and reflection, supported by a notebook, deferred-entity pool, and interaction history.The cycle adapts exploration as evidence accumulates.Source: §3.2 · pp. 3–4

3.2.1 Planning

• Planning analyzes question dependencies, decomposes information needs into focused sub-queries, and assigns each an anchor entity directly from the question.This plan does not rely on pre-annotated entities.Source: §3.2.1 · p. 4

3.2.2 Exploration

• Exploration pairs KG filtering with hierarchical text reading, narrowing from page skimming to relevant-section chunk and table analysis before aggregating complementary evidence.The KG pipeline filters entities, relations, and facts; the text strategy proceeds from page selection and skimming to detailed reading.Source: §3.2.2 · pp. 4–5

3.2.3 Synthesis & Reflection

• Synthesis distills retrieved facts and leads into a unified summary; reflection either targets remaining gaps or recovers from dead ends using deferred entities and interaction history.The system updates its notebook and candidate pool as it continues exploration.Source: §3.2.3 · p. 5

3.2.4 Overall Workflow & Answer Generation

• The cycle stops when evidence is sufficient or the turn limit is reached, then the system connects notebook facts and final retrieval results into a reasoning chain and answer.Source: §3.2.4 · pp. 5–6

4.1 Experimental Setup

• Evaluation spans seven knowledge-intensive multi-hop QA benchmarks—four KG-based and three text-based—and uses Qwen3-32B in non-thinking mode as the backbone for CoG and baselines.Baselines cover LLM-only, text-based RAG, KG-based RAG, and hybrid RAG; maximum exploration depth is 4 and width (sub-queries per turn) is 5.Source: §4.1 · p. 6

4.2 Main Results

• CoG outperforms baselines across most benchmarks, especially MusiQue (27.8% vs. IRCoT 20.6%) and KGQAGen (34.3% relative gain over ToG-2).It maintains comparable performance on simpler WebQSP; additional F1 and LLM-as-judge results confirm the gains under lexical-overlap and semantic evaluation.Source: §4.2 · p. 6, Table 2 · p. 7

4.3 Ablation Study

• Ablations show both KG and text matter (−6.92%, −7.92%); disabling text-to-KG guidance drops performance by 4.94%, and removing reflection, planning, or adaptive entity linking also degrades it.The synthesis module is also important for distilling heterogeneous evidence, filtering noise, and informing exploration.Source: §4.3 · p. 7, Table 3 · p. 7

4.4 In-depth Analysis

• On Qwen3-8B, CoG yields a +113.2% gain; CoG-8B (51.0%) exceeds the 32B IRCoT (44.6%) and ToG-2 (45.9%) baselines.The comparison reports performance across backbone LLMs; proprietary models use a 100-sample-per-dataset subset.Source: §4.4 · p. 7, Table 4 · p. 8
• Deeper exploration benefits text tasks; WebQSP and 2WikiMQA stabilize at shallower depths, while KGQAGen and AdvHotpotQA continue benefiting from deeper exploration.On KG tasks, CoG outperforms VectorRAG across all tested depths; optimal depth correlates with dataset difficulty.Source: §4.4 · pp. 7–8, Figure 3 · p. 8
• CoG accesses 12.4× fewer documents than ToG-2; costs rise on complex tasks, while smaller CoG-8B improves the accuracy-cost trade-off.On simpler WebQSP, CoG uses about 15k tokens, compared with about 22k for IRCoT.Source: §4.4 · p. 8, Table 5 · p. 8
• Across entity-candidate counts k = 6 to 30, average EM varies only from 59.1% to 60.3%, and increasing k does not degrade performance.The authors attribute this stability to context-aware disambiguation filtering noisy candidates while benefiting from broader entity recall.Source: §4.4 · p. 8
• Manual review finds 89.1% True Positives among CoG's EM-correct cases; 71.1% of its EM-incorrect cases are attributed to external factors.A case study illustrates recovery from dead ends and discovery of critical text evidence for complex multi-hop queries.Source: §4.4 · p. 8

5 Conclusion

• CoG is a cognitive-inspired framework for autonomous knowledge exploration in global hybrid knowledge bases.Its continuous plan-explore-reflect cycle shifts retrieval from passive, reactive behavior toward proactive knowledge discovery.Source: §5 · p. 8
• CoG combines structured knowledge graphs and unstructured text through bidirectional synergy to bridge knowledge gaps.The approach supports adaptive strategy adjustment and recovery from exploration failures.Source: §5 · pp. 8–9
• Experiments show that CoG significantly outperforms state-of-the-art baselines on multiple multi-hop QA tasks while achieving superior exploration efficiency.Source: §5 · p. 9

Limitations

• CoG’s multi-call workflow may limit real-time deployment despite reducing redundant document access and potentially improving the cost-accuracy trade-off with smaller backbones.The iterative loop incurs higher inference latency and token cost than passive retrieval methods.Source: Limitations · p. 9
• CoG’s effectiveness depends on the backbone LLM’s instruction following and long-context reasoning; weaker models may struggle with complex planning, synthesis, or reflection over noisy results.Source: Limitations · p. 9

A CoG Algorithm

• Algorithm 1 outlines how Planning, Exploration, Synthesis, and Reflection coordinate in CoG’s iterative workflow for autonomous knowledge exploration.Source: A CoG Algorithm · p. 12

B Evaluation Dataset

• The seven evaluation datasets span KG-based reasoning over structured facts and text-based tasks requiring semantic understanding and multi-document synthesis.Source: B Evaluation Dataset · pp. 12–13
• WebQSP targets 1–2-hop KG questions, CWQ adds up-to-4-hop logical complexity, and QALD10-en tests advanced SPARQL-like structural operations.Source: B Evaluation Dataset · p. 12
• KGQAGen is a Wikidata-grounded benchmark generated with symbolic verification, designed to address annotation inaccuracy and ambiguity in existing datasets.Source: B Evaluation Dataset · pp. 12–13
• 2WikiMQA tests cross-document multi-hop reasoning, MusiQue enforces connected reasoning through intermediate questions, and AdvHotpotQA adds high-overlap distractor paragraphs.Source: B Evaluation Dataset · p. 13
• The evaluation uses benchmark-specific test sets: ToG-2 splits for WebQSP, AdvHotpotQA, and QALD10-en; 1,024 CWQ questions; IRCoT evaluation sets for 2WikiMQA and MusiQue; and the full KGQAGen-10k test set.Source: B Evaluation Dataset · p. 13

C Knowledge Bases

• CoG’s global environment combines Wikidata’s structured facts with entity-centric textual evidence from Wikipedia, requiring retrieval across a global rather than small candidate entity space.Source: C Knowledge Bases · p. 13
• Wikipedia’s approximately 7.1 million articles retain page hierarchy to support global skimming followed by detailed reading.Source: C Knowledge Bases · p. 14
• CoG uses Wikidata for normalized entities and typed relations as precise reasoning anchors, while Wikipedia supplies richer natural-language context as high-recall evidence.Source: C Knowledge Bases · p. 14

• The LLM-as-judge evaluator labels responses CORRECT, PARTIAL, INCORRECT, or NO ANSWER, while F1 uses normalized extracted answers and keeps the maximum score across references.Source: D Additional Evaluation Metrics · p. 14
• 61.63% average CORRECT rate gives CoG the best LLM-as-judge result on all seven datasets, 12.10 points above IRCoT.Source: D Additional Evaluation Metrics · p. 14
• 49.87% average token-level F1 is CoG’s highest score, 4.49 points above ToG-2.Source: D Additional Evaluation Metrics · p. 14

E Detailed Results on Backbone LLMs

• The backbone analysis excludes WebQSP as near-saturated under direct prompting, samples 300 instances each for KGQAGen and CWQ, and uses 100 examples per dataset for proprietary models.Source: E Detailed Results on Backbone LLMs · p. 14
• CoG consistently improves performance across backbone models ranging from 8B to 235B, including open-source and proprietary models.Source: E Detailed Results on Backbone LLMs · p. 14
• 113.2% is the relative performance gain reported for Qwen3-8B, the largest among the evaluated backbones.Source: E Detailed Results on Backbone LLMs · p. 14
• CoG-8B averages 51.0%, exceeding direct prompting on Qwen3-235B (43.7%) and DeepSeek-V3.2 (38.0%), as well as 32B IRCoT (44.6%) and ToG-2 (45.9%).The authors describe 8B models as cheaper and faster, making CoG-8B a cost-effective deployment option.Source: E Detailed Results on Backbone LLMs · pp. 14–15

F Detailed Results on Exploration Depth

• Table 10 reports CoG Exact Match scores at exploration depths 1–6 across all seven datasets, with separate averages for KG-based and text-based tasks.Source: F Detailed Results on Exploration Depth · p. 15

Complexity Adaptability

• +11.6% on MusiQue over ToG-2 accompanies deeper token investment; on WebQSP and 2WikiMQA, CoG uses ∼14-15k tokens versus IRCoT’s ∼22k while achieving higher accuracy.For complex tasks, CoG invests more tokens in deep reasoning; for simpler tasks, it uses fewer tokens than IRCoT.Source: G Detailed Efficiency Profile · p. 15

Latency Breakdown

• Figure 4 reports average wall-clock time per retrieval round; sequential LLM calls are the primary latency source, and parallelizing dual-source exploration and sub-queries could optimize it.CoG’s open-ended retrieval starts from raw questions and adapts KG/text exploration to the evolving reasoning state; deep filtering reduces accessed documents 12.4× versus ToG-2.Source: G Detailed Efficiency Profile · p. 15, Figure 4 · p. 16

Cost-Effectiveness

• CoG with an 8B model outperforms strong 32B baselines, suggesting a more favorable accuracy-cost trade-off for practical deployment.This can offset the greater token consumption on complex tasks.Source: G Detailed Efficiency Profile · pp. 15–16, Table 9 · p. 16

H Sensitivity Analysis on Entity Linking

• 59.1%–60.3% average EM across k = 6 to 30; k = 20 ties k = 30 for best average performance while using fewer candidate entities.Even at k = 6, CoG achieves 59.4% average EM.Source: H Sensitivity Analysis on Entity Linking · p. 16, Table 12 · p. 17

I.1 Reasoning Faithfulness

• Among EM-correct predictions, the audit distinguishes sound reasoning, flawed-but-correct chains, coincidental answers, and false positives from substring matching.CoG’s imperfect cases mostly involve incomplete verification of secondary constraints rather than broken core reasoning chains.Source: §I.1 · p. 17
• 3.6% severe false-positive rate (Weak FP + FP) for CoG, versus 12.7% for ToG-2 and 29.1% for Self-Consistency, is the lowest among these methods.CoG remains the most reliable method even when Flawed TP is treated as non-ideal reasoning.Source: §I.1 · p. 17, Table 13 · p. 18
• Figure 5 illustrates that identical EM scores can conceal different reasoning quality; CoG’s example reaches the answer through verifiable evidence despite an outdated dataset reference path.The example involves a temporal update to the coach identified by CoG.Source: Figure 5 · p. 18, §I.1 · p. 18

I.2 Error Diagnosis

• 71.1% of sampled errors are external factors, including metric false negatives (37.8%), dataset flaws (26.7%), and knowledge-base temporal mismatch (6.7%).The remaining 28.9% are internal errors in the agentic exploration process.Source: §I.2 · p. 18, Figure 6 · p. 18

K Prompts

• The prompts cover initial planning, KG exploration, text exploration, synthesis, and reflection, including reflection prompts for continuing exploration or pivoting strategy.The KG stages are entity linking, relation discovery, and fact pruning; text exploration includes adaptive page selection and hierarchical reading.Source: K Prompts · p. 19

J Case Studies

• In the AdvHotpotQA case, CoG recovers from an initial dead end by broadening its query and using a text-extracted entity to guide graph exploration to the answer.ToG-2 instead continues its beam search toward irrelevant high-ranking entities without reflection.Source: J Case Studies · p. 19, Table 14 · p. 20, Table 15 · p. 21