Co-Evolving Graph and Text Memory for Training-Free Multi-Hop Question Answering

# Co-Evolving Graph and Text Memory for Training-Free Multi-Hop Question Answering
Source: [https://arxiv.org/html/2607.23278](https://arxiv.org/html/2607.23278)
Thien Huu Nguyen University of Oregon, OR, USA thienn@uoregon\.edu

###### Abstract

Multi\-hop question answering requires coordinating relational and textual evidence across reasoning steps, a combination neither a text corpus nor a knowledge graph can supply alone\. Prior work often emphasizes only part of this loop: graph\-augmented RAG retrieves from a pre\-built or query\-updated graph, KGQA systems search within topic\-centered subgraphs, and memory\-augmented agents maintain evolving memories without continuously reconciling graph memory with textual context\. We propose Co\-E, a training\-free system built around synchronized bidirectional graph\-text working memory\. A synchronization cycle consolidates textual memory, extracts relational triples into graph memory, and injects graph facts back into the generation context\. Because both memories are maintained, they shape subsequent retrieval and generation\. Evaluated on six multi\-hop QA benchmarks, Co\-E improves over comparable training\-free open\-backbone baselines and is competitive with larger or trained systems\.111Codebase available athttps://github\.com/hieum98/wemg

Co\-Evolving Graph and Text Memory for Training\-Free Multi\-Hop Question Answering

Hieu ManUniversity of Oregon, OR, USAhieum@uoregon\.eduThien Huu NguyenUniversity of Oregon, OR, USAthienn@uoregon\.edu

## 1Introduction

Retrieval\-augmented generation \(RAG\) has become the dominant paradigm for knowledge\-intensive question answeringZhaoet al\.\([2024](https://arxiv.org/html/2607.23278#bib.bib1)\); Singhet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib2)\)\. Multi\-hop questions, however, expose a limitation deeper than retrieval recall\. Later evidence often depends on entities, relations, or constraints discovered only after earlier hops are resolved\. A passage may reveal a bridge entity that should expand the graph frontier; a graph edge may reveal a constraint that should reshape the next textual query\. Standard RAG pipelines cannot perform this update because they select context before reasoning unfolds\. Likewise, graph\-augmented methods that rely on a pre\-built corpus graph or a fixed topic\-centered candidate subgraph cannot fully revise their relational state during reasoning, leaving missed bridge entities, noisy edges, and incomplete paths uncorrected\. The central challenge is to maintain a reasoning state in which textual evidence and relational structure revise each other as the answer is constructed\.

Existing systems address parts of this problem, but usually privilege one side of the reasoning state\. Corpus\-graph RAG systems build a graph index before inference and retrieve from itDonget al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib7)\); Maet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib3)\); Zhouet al\.\([2026b](https://arxiv.org/html/2607.23278#bib.bib6)\)\. KGQA systems such as ToGSunet al\.\([2024](https://arxiv.org/html/2607.23278#bib.bib8)\)and ToG 2\.0Maet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib9)\)improve graph\-based reasoning over knowledge bases, but rely on pre\-existing topology and do not maintain paired textual working memory\. Graph search methods such as MCTS\-KBQAXionget al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib41)\)and ReKG\-MCTSSonget al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib45)\)navigate a fixed frontier rather than extending it\. A parallel line equips agents with evolving memories: HGMemZhouet al\.\([2026a](https://arxiv.org/html/2607.23278#bib.bib12)\)builds hypergraph memory from retrieved text, SubQRAGLiet al\.\([2025a](https://arxiv.org/html/2607.23278#bib.bib52)\)accumulates extracted triples as graph memory, while SE\-SearchLiet al\.\([2026a](https://arxiv.org/html/2607.23278#bib.bib13)\), MemSearch\-o1Zhanget al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib14)\), and A\-MEMXuet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib15)\)maintain textual or reflective memories\. These systems show that maintained memory is useful; what remains less explored is how textual and graph memories should revise one another in same reasoning loop\.

We argue that this coupling requiresbidirectional, within\-step co\-evolution of graph and text\. When a passage mentions a new entity, it should become available for graph expansion; when a triple is extracted or retrieved, it should enter the textual context for the next generation step\. Without both directions, the system either accumulates unstructured text that must be reinterpreted from scratch, or traverses a graph whose useful facts are not exposed to the generator\. We introduceCo\-E\(Co\-Evolving\), a training\-free multi\-hop QA system built around synchronized graph\-text working memory\. Co\-E maintains a shared memory with two coupled stores: question\-relevant text snippets and graph memory containing KB\-schema and open\-vocabulary triples\. At each step, Co\-E retrieves from text and knowledge\-base streams, filters evidence, and applies a four\-operation synchronization cycle: it consolidates textual evidence, extracts and merges triples into graph memory, injects surviving graph facts back into textual memory, and re\-consolidates the context\. This cycle lets text expand the graph frontier while graph structure grounds subsequent generation\. Because the same mechanism operates during inference, Co\-E requires no fine\-tuning and serves both multi\-hop text\-QA and KGQA settings\.

Empirically, Co\-E achieves strong results across six benchmarks spanning text\-QA and KGQA\. With a Qwen3\-8B backbone and no training, Co\-E reaches 72\.6 EM on 2WikiMultiHopQA, 70\.0 EM on Bamboogle and 74\.9 Hits@1 on CWQ, outperforming comparable open\-backbone baselines and competing with larger or trained systems\.

## 2Related Work

We organize prior work by how evidence is represented and updated during multi\-hop reasoning\.

Corpus\-based retrieval and agentic search\.Corpus\-based RAG methods decompose complex questions into sub\-queries and interleave retrieval with intermediate reasoning\. IRCoTTrivediet al\.\([2023](https://arxiv.org/html/2607.23278#bib.bib23)\)alternates retrieval with chain\-of\-thought reasoning, while HopRAGLiuet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib24)\), ComposeRAGWuet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib26)\), and RT\-RAGShiet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib25)\)improve query decomposition, structured prompting, or reasoning\-tree construction\. A related line trains or prompts agents to search over multiple turns, such as Search\-R1Jinet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib28)\)and MR\-SearchXiaoet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib29)\)\. These systems make retrieval adaptive, but evidence remains primarily textual\. Relational structure must be inferred from snippets rather than maintained as explicit state\. Co\-E complements adaptive retrieval with graph memory that makes entities and relations available to later retrieval and generation\.

Graph\-augmented RAG and KG reasoning\.Graph\-augmented approaches add structured evidence to RAG, but many rely on graph structure fixed before or outside the current reasoning trajectory\. Corpus\-graph systems such as HippoRAG 2 and C2RAGgutiérrez2025ragmemorynonparametriccontinual; Maet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib3)\)build graph\-like indices over text and retrieve from them at inference time\. EA\-GraphRAGDonget al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib7)\), GraphAnchorLiuet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib34)\), PAGERLiet al\.\([2026b](https://arxiv.org/html/2607.23278#bib.bib5)\), and Graph\-R1 \+ EKAWanget al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib4)\)improve when or how structured evidence is retrieved\. KGQA systems differ in how they choose the graph frontier\. Learned/subgraph systems such as SubgraphRAGLiet al\.\([2025b](https://arxiv.org/html/2607.23278#bib.bib36)\), GNN\-RAGMavromatis and Karypis \([2025](https://arxiv.org/html/2607.23278#bib.bib39)\), and iQUESTWang and Yu \([2025](https://arxiv.org/html/2607.23278#bib.bib44)\)learn or compute query\-specific subgraphs before handing evidence to the LLM\. Live\-KG methods such as ToGSunet al\.\([2024](https://arxiv.org/html/2607.23278#bib.bib8)\), ToG 2\.0Maet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib9)\), PoGTanet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib35)\), Interactive\-KBQAXionget al\.\([2024](https://arxiv.org/html/2607.23278#bib.bib16)\), MCTS\-KBQAXionget al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib41)\), ReKG\-MCTSSonget al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib45)\), and KERAGSunet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib46)\)focus on traversing or quering a knowledge graph during inference\. Trained graph\-reasoning systems such as RoGLuoet al\.\([2024](https://arxiv.org/html/2607.23278#bib.bib38)\), KBQA\-R1Sunet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib43)\), and GraphWalkerXuet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib40)\)improve retrieval or reasoning through supervised or reinforcement learning\. These systems provide strong graph retrieval; Co\-E focuses on synchronizing structured and textual working state during reasoning\.

![Refer to caption](https://arxiv.org/html/2607.23278v1/x1.png)Figure 1:Co\-E framework\. Left: MCTS mode, where synchronized shared memory guides tree expansion\. Right: the four\-stage reasoning loop\. Each step generates a sub\-query from shared graph\-text memoryℳ\(t\)\\mathcal\{M\}^\{\(t\)\}, retrieves text snippetsS\(t\)S^\{\(t\)\}and graph triplesR\(t\)R^\{\(t\)\}, generates an intermediate answer, and applies bidirectionalSync\\mathrm\{Sync\}to produce the next step’s memoryℳ\(t\+1\)\\mathcal\{M\}^\{\(t\+1\)\}\.Memory\-augmented reasoning\.Memory\-augmented reasoning systems maintain intermediate state across reasoning or search steps, whether in text, graph, or both\. HGMemZhouet al\.\([2026a](https://arxiv.org/html/2607.23278#bib.bib12)\)builds hypergraph memory from retrieved text\. SubQRAGLiet al\.\([2025a](https://arxiv.org/html/2607.23278#bib.bib52)\)is closest to Co\-E among graph\-memory systems: it decomposes questions, consults source documents when the graph is insufficient, and accumulates extracted triples into graph memory\. SE\-SearchLiet al\.\([2026a](https://arxiv.org/html/2607.23278#bib.bib13)\)purifies evidence through a Think\-Search\-Memorize loop, MemSearch\-o1Zhanget al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib14)\)grows reasoning\-aligned memory paths, and A\-MEMXuet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib15)\)revises textual notes retroactively\. These works show that memory should be maintained, not merely accumulated\. Co\-E is complementary: instead of treating memory as a textual note store or graph\-only trace, it maintains paired textual and graph memories and reconciles them after each reasoning step\. Graph facts are therefore injected back into the textual context that conditions later retrieval and generation\.

## 3Methodology

### 3\.1System Overview

Given a questionqq, Co\-E answers it through an iterative reasoning process over a shared graph\-text memory\. At steptt, the memory is

ℳ\(t\)=\(𝒯\(t\),𝒢\(t\)\),\\mathcal\{M\}^\{\(t\)\}=\\bigl\(\\mathcal\{T\}^\{\(t\)\},\\mathcal\{G\}^\{\(t\)\}\\bigr\),\(1\)where𝒯\(t\)\\mathcal\{T\}^\{\(t\)\}is textual memory and𝒢\(t\)\\mathcal\{G\}^\{\(t\)\}is graph memory\. Co\-E produces a final answeraawhile updating both memories after each reasoning step\. Unlike pipelines that treat retrieved graph context as fixed, Co\-E actively maintains𝒢\\mathcal\{G\}: it can add text\-derived triples, prune irrelevant edges, and write surviving facts back into textual memory\.

Each reasoning step has four stages\. First,Request Generationproduces a focused sub\-queryq\(t\)q^\{\(t\)\}from the original question and current memory\. Second,Dual\-Stream Retrieval/Searchretrieves textual snippetsS\(t\)S^\{\(t\)\}and graph triplesR\(t\)R^\{\(t\)\}\. Third,Response Generationproduces an intermediate answer or reasoning statey\(t\)y^\{\(t\)\}\. Fourth,Memory Synchronizationupdates both memories through a bidirectional operator:

\(𝒯\(t\+1\),𝒢\(t\+1\)\)←Sync​\(𝒯\(t\),𝒢\(t\),S\(t\),R\(t\),y\(t\)\)\\bigl\(\\mathcal\{T\}^\{\(t\+1\)\},\\mathcal\{G\}^\{\(t\+1\)\}\\bigr\)\\leftarrow\\mathrm\{Sync\}\\bigl\(\\mathcal\{T\}^\{\(t\)\},\\mathcal\{G\}^\{\(t\)\},S^\{\(t\)\},R^\{\(t\)\},y^\{\(t\)\}\\bigr\)\(2\)Co\-E supports both MCTS and CoT reasoning modes; both use the same retrieval and synchronization machinery, differing only in how they choose the next sub\-query\.

### 3\.2Shared Graph\-Text Memory

Co\-E maintains a single shared memory object rather than independent per\-hop contexts\. Every step reads from the current memory and writes verified evidence back the same memory, so early evidence can affect later retrieval and generation\.

Textual memory\.The textual memory𝒯\\mathcal\{T\}stores self\-contained, question\-relevant snippets: extracted evidence, intermediate sub\-answers, verifier feedback, and natural\-language statements produced from graph triples\. Raw documents never enter memory; an extractor filters each retrieved document into atomic snippets\.

Graph memory\.The graph memory𝒢\\mathcal\{G\}is a directed graph over entities and relations\.KB\-schema triplesare canonical and retrieved directly from a knowledge base\.Open\-vocabulary triplesare extracted from textual memory and use natural\-language relation labels\. The former provide reliable schema\-grounded structure, while the latter let Co\-E represent relations expressed in text but absent from the KB schema\. A shared entity dictionary links surface mentions in𝒯\\mathcal\{T\}to KB identifiers in𝒢\\mathcal\{G\}, allowing textual and graph evidence to refer to the same entities\.

Before triples are merged into𝒢\\mathcal\{G\}, a triple pruner removes irrelevant, redundant, or contradicted edges\. Nodes that resolve to the same KB identifier are collapsed, and isolated low\-value nodes are discarded\. Thus graph memory is actively maintained rather than monotonically accumulated\.

### 3\.3Bidirectional Memory Synchronization

The synchronization operator is the core mechanism of Co\-E\. It co\-evolves text and graph memory through four operations\.

Textual consolidation\.Co\-E first consolidates𝒯\(t\)\\mathcal\{T\}^\{\(t\)\}, newly retrieved snippetsS\(t\)S^\{\(t\)\}, and the intermediate responsey\(t\)y^\{\(t\)\}\. The consolidator deduplicates overlapping snippets, merges complementary evidence, and retracts contradicted intermediate predictions\. This prevents the memory from becoming an ever\-growing block of noisy context\.

Text\-to\-graph propagation\.From the consolidated textual memory, an entity linker identifies entities and maps them to KB identifiers when possible\. A relation extractor then proposes open\-vocabulary triples from the text\. These triples are combined with KB triplesR\(t\)R^\{\(t\)\}, pruned for relevance and consistency, and merged into𝒢\(t\)\\mathcal\{G\}^\{\(t\)\}\. This step lets newly discovered textual evidence expand the relational frontier\.

Graph\-to\-text propagation\.Co\-E textualizes surviving graph triples and appends them to textual memory as explicit relational statements\. This gives the generator direct access to graph constraints in natural language, rather than requiring it to infer them from graph structure separately\.

Textual re\-consolidation\.Finally, the consolidator runs a second pass over the textual memory after graph\-derived statements have been injected\. The result is a coherent memory stateℳ\(t\+1\)\\mathcal\{M\}^\{\(t\+1\)\}in which textual evidence and graph evidence are reconciled\. This updated memory conditions the next sub\-query, retrieval step, and response generation\.

### 3\.4Dual\-Stream Retrieval

Co\-E retrieves from both text and graph sources:Corpus retrieval\.A query generator rewritesq\(t\)q^\{\(t\)\}into search\-oriented queries conditioned onℳ\(t\)\\mathcal\{M\}^\{\(t\)\}\. These queries are sent to either web search or dense retrieval over a local corpus\. Retrieved documents are passed to the extractor, which produces the snippet setS\(t\)S^\{\(t\)\}\. Only atomic extracted snippets enter memory\.Graph retrieval\.Graph retrieval uses the current memory to choose which KB regions to explore\. Co\-E links entities fromq\(t\)q^\{\(t\)\}andℳ\(t\)\\mathcal\{M\}^\{\(t\)\}to KB identifiers, augments the seed set with relevant entities already present in𝒢\(t\)\\mathcal\{G\}^\{\(t\)\}, and issueskk\-hop SPARQL queries from this seed set\. To control combinatorial growth, candidate triples are reranked between hops and then filtered by the triple pruner\. The resulting triples formR\(t\)R^\{\(t\)\}\. Because the seed set depends on the current graph memory, graph retrieval is memory\-conditioned rather than a fixed precomputed traversal\.

### 3\.5Reasoning Modes

Co\-E can run the same memory update cycle under two reasoning policies\.

MCTS mode\.MCTS mode searches over candidate reasoning trajectories\. The root is the original questionqq, and child nodes are generated by the LLM\. FollowingQiet al\.\([2024](https://arxiv.org/html/2607.23278#bib.bib51)\), nodes may represent sub\-question answering, self\-correction, synthesis, question rephrasing, or final\-answer generation\. Selection uses a prior\-weighted UCB score, following the PUCT\-style familySilveret al\.\([2017](https://arxiv.org/html/2607.23278#bib.bib53)\)rather than classic UCT:

S​\(s,a\)=Q¯​\(s,a\)\+c⋅P​\(s,a\)⋅Np​\(s\)\+11\+N​\(s,a\)S\(s,a\)=\\bar\{Q\}\(s,a\)\+c\\cdot P\(s,a\)\\cdot\\frac\{\\sqrt\{N\_\{\\mathrm\{p\}\}\(s\)\+1\}\}\{1\+N\(s,a\)\}\(3\)whereQ¯​\(s,a\)\\bar\{Q\}\(s,a\)is the average reward accumulated by backpropagation,P​\(s,a\)P\(s,a\)is a child node type’s heuristic prior,N​\(s,a\)N\(s,a\)is the child visit count,Np​\(s\)N\_\{\\mathrm\{p\}\}\(s\)is the parent visit count, andcccontrols exploration\. Co\-E separates tree statistics from memory state: node visits and values remain branch\-specific, while evidence discovered during MCTS is accumulated in a shared working memory\. After rollout, a verifier scores candidate terminal states under three views: no retrieved context, textual memory, and graph memory\. Its natural\-language assessments are added to textual memory as system\-prediction items\. After backpropagation, Co\-E appliesSync\\mathrm\{Sync\}once to the shared memory, where consolidation can remove unsupported predictions and filters newly introduced triples\. This lets high\-quality search feedback shape later retrieval, while relying on verifier feedback and active maintenance to reduce contamination from weak branches\.

CoT mode\.CoT mode follows a sequential reasoning chain\. At each step, the model generates and answers sub\-questions conditioned onqqandℳ\(t\)\\mathcal\{M\}^\{\(t\)\}, then applies the same synchronization operator before continuing\. CoT mode is cheaper because it avoids tree search and verifier\-guided branch selection, while MCTS mode explores multiple candidate reasoning paths before committing\.

## 4Experiments

Table 1:KGQA Hits@1 on WebQSP and CWQ\.Italicrows use closed\-source or\>\>14B backbones;boldmarks the best open≤\\leq8B score\.Table 2:Multi\-hop text\-QA EM\.Italicrows use closed\-source or\>\>14B backbones;boldmarks the best open≤\\leq8B\.### 4\.1Experimental Setup

Datasets\.We evaluate on six benchmarks spanning KGQA over structured knowledge bases and multi\-hop text\-QA over unstructured corpora\. For KGQA, WebQSPYihet al\.\([2016](https://arxiv.org/html/2607.23278#bib.bib21)\)contains mostly 1\-2 hop Freebase questions, while CWQTalmor and Berant \([2018](https://arxiv.org/html/2607.23278#bib.bib22)\)requires conjunction, composition, and longer relation chains\. The main KGQA comparison follows the standard topic\-subgraph protocol: each question is evaluated against a topic\-centered Freebase candidate graph rather than against the full KB\. For text\-QA, 2WikiMultiHopQAHoet al\.\([2020](https://arxiv.org/html/2607.23278#bib.bib17)\)emphasizes entity\-centric bridge reasoning, HotpotQAYanget al\.\([2018](https://arxiv.org/html/2607.23278#bib.bib18)\)contains Wikipedia questions with stronger lexical cues, MuSiQueTrivediet al\.\([2022](https://arxiv.org/html/2607.23278#bib.bib19)\)composes single\-hop questions into longer chains, and BambooglePresset al\.\([2023](https://arxiv.org/html/2607.23278#bib.bib20)\)contains adversarial multi\-hop questions that stress bridge\-entity reuse\.

Metrics\.For KGQA, we report Hits@1, following prior work on WebQSP and CWQMaet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib9)\); Liet al\.\([2025b](https://arxiv.org/html/2607.23278#bib.bib36)\); a prediction is correct if the top answer matches any gold alias after normalization\. For text\-QA, we report Exact Match \(EM\)Trivediet al\.\([2023](https://arxiv.org/html/2607.23278#bib.bib23)\); Liuet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib24)\)\. Some reported baselines only release F1; these are labeled explicitly and treated as reference points rather than direct EM comparisons\. Because EM penalizes correct paraphrases or formatting differences, ablations and Appendix[A](https://arxiv.org/html/2607.23278#A1)also report LLM\-judged accuracy \(Acc\) using Qwen3\-30B\-A3B\-Thinking\-2507Team \([2025](https://arxiv.org/html/2607.23278#bib.bib47)\), which receives the question, gold answer, and prediction\.

Baselines\.We group baselines by evidence representation\. For text\-QA,Corpus\-Based RAGmethods rely primarily on textual retrieval and reasoning, including IRCoTTrivediet al\.\([2023](https://arxiv.org/html/2607.23278#bib.bib23)\), HopRAGLiuet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib24)\), ComposeRAGWuet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib26)\), RT\-RAGShiet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib25)\), Search\-o1Liet al\.\([2025c](https://arxiv.org/html/2607.23278#bib.bib27)\), Search\-R1Jinet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib28)\), MR\-SearchXiaoet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib29)\), and Search\-P1Xiaet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib30)\)\.Graph\-Augmented Retrievalincludes corpus\-graph and graph\-trained systems such as HippoRAG 2gutiérrez2025ragmemorynonparametriccontinual, C2RAGMaet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib3)\), PAGERLiet al\.\([2026b](https://arxiv.org/html/2607.23278#bib.bib5)\), GraphAnchorLiuet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib34)\), EA\-GraphRAGDonget al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib7)\), Graph\-R1 \+ EKAWanget al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib4)\), and ProGraph\-R1Parket al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib48)\)\.Memory\-Augmentedsystems include SubQRAGLiet al\.\([2025a](https://arxiv.org/html/2607.23278#bib.bib52)\), MemSearch\-o1Zhanget al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib14)\), and SE\-SearchLiet al\.\([2026a](https://arxiv.org/html/2607.23278#bib.bib13)\)\. For KGQA,Learned/Subgraph KG\-Retrieval KGQAsystems learn or compute query\-specific candidate subgraphs before LLM reasoning, including SubgraphRAGLiet al\.\([2025b](https://arxiv.org/html/2607.23278#bib.bib36)\), GNN\-RAGMavromatis and Karypis \([2025](https://arxiv.org/html/2607.23278#bib.bib39)\), and iQUESTWang and Yu \([2025](https://arxiv.org/html/2607.23278#bib.bib44)\)\.Live\-KG KGQAsystems traverse or query a knowledge graph at inference time, including ToG 2\.0Maet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib9)\), PoGTanet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib35)\), Interactive\-KBQAXionget al\.\([2024](https://arxiv.org/html/2607.23278#bib.bib16)\), MCTS\-KBQAXionget al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib41)\), ReKG\-MCTSSonget al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib45)\), KBQA\-o1Luoet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib42)\), RoGLuoet al\.\([2024](https://arxiv.org/html/2607.23278#bib.bib38)\), and KERAGSunet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib46)\)\. Open≤\\leq8B models are our primary comparison group; closed\-source and much larger models are reference\.

Implementation Details\.Co\-E is training\-free\. We use Qwen3\-8BTeam \([2025](https://arxiv.org/html/2607.23278#bib.bib47)\)as the reasoning backbone, Qwen3\-Embedding\-4BZhanget al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib54)\)for dense retrieval over the Wiki23 corpus, and SPARQL for graph retrieval\. The same graph\-text memory mechanism is used across settings, but the graph source differs\. In the main KGQA experiments, graph retrieval is restricted to the benchmark\-provided Freebase candidate subgraph\. In text\-QA, graph retrieval uses full live\-Wikidata\. For detail implementation, see Appendix[B](https://arxiv.org/html/2607.23278#A2)\.

### 4\.2KGQA Results

Table[1](https://arxiv.org/html/2607.23278#S4.T1)reports Hits@1 on WebQSP and CWQ under the standard topic\-subgraph protocol\. Each question is paired with a compact Freebase subgraph centered on the topic entity\. This setting tests reasoning over a noisy but bounded candidate graph: Co\-E does not expand the Freebase search space beyond the provided subgraph, but it can prune candidate relations, preserve intermediate entities, and textualize graph facts for later generation\. On WebQSP, where questions are short 1\-2 hop queries and the answer path is often exposed by the topic graph, the highest scores come from closed\-source or trained systems such as PoG, SubgraphRAG, and iQUEST\. Co\-E \(MCTS\) reaches 85\.5 Hits@1, below those systems but essentially tied with the strongest open 7B trained baselines, RoG and GNN\-RAG \(both 85\.7\)\. Co\-E \(CoT\) reaches a similar 85\.3, suggesting that synchronized memory, rather than tree search alone, accounts for most of Co\-E’s WebQSP performance\. On CWQ, the advantage of maintaining graph\-text state is clearer because questions require conjunctions, composition, and longer relation chains\. Co\-E \(MCTS\) reaches 74\.9 Hits@1, slightly above PoG \(74\.7\) and iQUEST \(73\.9\), and improves over MCTS\-KBQA by 2\.8 points, GNN\-RAG by 8\.1 points, SubgraphRAG by 8\.2 points, RoG by 12\.3 points, and ReKG\-MCTS by 15\.1 points\. The larger MCTS\-CoT gap on CWQ indicates that search helps when several plausible relation paths compete\. Overall, the KGQA results support Co\-E as a training\-free graph\-memory reasoner, with the largest gains when the model must preserve and select among multi\-step relations\.

Table 3:Bamboogle ablations for MCTS and CoT\.
### 4\.3Multi\-hop Text\-QA Results

Table[2](https://arxiv.org/html/2607.23278#S4.T2)reports EM on the four text\-QA benchmarks\. Co\-E \(MCTS\) is strongest among open≤\\leq8B systems, reaching 61\.8 on HotpotQA, 72\.6 on 2WikiMultiHopQA, 52\.5 on MuSiQue, and 70\.0 on Bamboogle\. Co\-E \(CoT\) is lower than MCTS but remains competitive with same scale baselines, suggesting that synchronized memory provides much of the improvement while tree search adds gain when bridge candidates compete\.

The largest gains appear on datasets where a later hop depends on preserving a bridge entity or relation discovered earlier\. On 2WikiMultiHopQA, Co\-E improves over C2RAG by 6\.7 points and SubQRAG by 10\.7 points\. This supports the intended use of text\-to\-graph propagation: once a passage identifies a bridge entity, Co\-E can link it, store it as graph memory, and use it to seed later graph or corpus retrieval\. The advantage is clearer on harder compositional text\-QA benchmarks\. On MuSiQue, Co\-E reaches 52\.5 EM, outperforming HopRAG\-GPT\-4o by 10\.3 points, ProGraph\-R1 by 12\.7 points, and SubQRAG by 22\.8 points\. MuSiQue’s longer 2\-4 hop chains make retrieval noise compound; active consolidation and graph\-text synchronization preserve the resolved chain while discarding stale content\. Bamboogle shows a similar pattern: Co\-E reaches 70\.0 EM, improving over Search\-R1\-7B by 26\.8 points and PAGER\-32B by 7\.6 points\. These questions often cannot be answered until an intermediate entity is resolved, so MCTS explores alternative bridge candidates while synchronization commits verified evidence into both memories\. HotpotQA has the smallest margin: Co\-E reaches 61\.8 EM, ahead of SubQRAG and ProGraph\-R1 but close to HopRAG\-GPT\-4o\. HotpotQA often has stronger lexical cues and shorter chains, making corpus\-only or closed\-source pipelines more competitive\.

Table 4:Co\-E with Freebase topic\-subgraph versus Full\-Wikidata retrieval \(Hits@1\)\.![Refer to caption](https://arxiv.org/html/2607.23278v1/x2.png)Figure 2:Pass@kkon Bamboogle;kkis the number of tree expansions for MCTS and chain depth for CoT\.
### 4\.4Ablation Study

Table[3](https://arxiv.org/html/2607.23278#S4.T3)ablates Co\-E on Bamboogle, where questions often require a resolved bridge entity in a later hop\. The ablations test three design choices: whether synchronization must be bidirectional, whether both memory modalities are needed, and whether retrieval\-time filtering can replace persistent memory\.

Synchronization\.Removing bidirectional synchronization drops MCTS EM from 70\.0 to 61\.1 and CoT EM from 66\.4 to 52\.5\. The drop under both search policies shows that the effect is not merely a byproduct of tree exploration\. One\-way propagation is also insufficient: text→\\rightarrowgraph reaches 64\.0 MCTS EM, while graph→\\rightarrowtext reaches 63\.1\. These variants preserve one half of the loop, but lose the feedback cycle that lets textual evidence expand the graph frontier and graph facts shape the next textual context\. Removing active maintenance causes the largest synchronization\-related degradation, reducing MCTS EM to 49\.5\. This variant can still propagate evidence across modalities, but it no longer reliably deduplicates, merges, or retracts stale content\. The result suggests that co\-evolution must be selective rather than purely accumulative\.

Memory modality\.Removing graph memory lowers MCTS EM to 54\.8, and removing textual memory lowers it to 55\.0\. The degradation suggests that textual memory provides grounding that KB triples may not encode, while graph memory preserves explicit entity\-relation structure that text leaves implicit\. Neither memory alone recovers the full behavior, supporting paired graph\-text state\.

Retrieval and extraction\.Removing textual retrieval causes the largest component drop \(−\-23\.4 MCTS EM\), followed by graph retrieval \(−\-20\.2\) and the extractor \(−\-19\.8\)\. This confirms that Co\-E depends on both evidence streams, but also that retrieval alone is not enough\. The extractor ablation shows that raw documents are too noisy to accumulate directly; filtering them into concise, self\-contained snippets is part of memory construction rather than a minor preprocessing detail\.

![[Uncaptioned image]](https://arxiv.org/html/2607.23278v1/x3.png)

Table 5:Memory walkthrough for"Who is the mother of the father of George Washington?"\. Graph\-injected text grounds step 2 retrieval on Augustine Washington; dead\-end graph branches are gray and⋆\\starmarks the gold answer\.
### 4\.5Full\-Wikidata Evaluation

The KGQA results above follow the standard topic\-subgraph protocol, where each question is paired with a pre\-retrieved Freebase subgraph centered on its topic entity\. This protocol enables controlled comparison with prior KGQA systems, but it assumes question\-specific preprocessing: a topic entity must be identified and a compact candidate graph extracted before reasoning begins\. Such preprocessing is not always available in open\-domain settings, where systems may need to start from a large KG and from entities discovered during retrieval\. To test this harder setting, we rerun Co\-E \(MCTS\) on WebQSP and CWQ using the same Full\-Wikidata retrieval stack as the text\-QA experiments\. This experiment is a stress test rather than a matched replacement for Freebase evaluation, since entity identifiers, relation schemas, and answer aliases differ between Freebase and Wikidata\. Table[4](https://arxiv.org/html/2607.23278#S4.T4)shows that moving from Freebase topic subgraphs to full Wikidata has only a small effect on CWQ, reducing Hits@1 from 74\.9 to 73\.4\. For compositional questions, memory\-conditioned entity discovery and SPARQL expansion can recover much of the useful relational frontier even without a benchmark\-provided candidate graph\. The drop is larger on WebQSP, from 85\.5 to 78\.6, because many WebQSP questions are short and the preprocessed subgraph often exposes the answer path directly \(due to its significant smaller size and better schema match\)\. Without this pruning, retrieval noise has a larger effect on short questions with less compositional structure\. Overall, the Full\-Wikidata setting shows that Co\-E benefits from topic\-subgraph pruning when available, while retaining strong compositional KGQA performance when forced to retrieve from a much larger and schema\-mismatched graph\.

### 4\.6Analysis

Pass@kk\.Figure[2](https://arxiv.org/html/2607.23278#S4.F2)separates search from memory\. MCTS improves rapidly in the first few expansions because it can evaluate multiple bridge candidates before committing, whereas CoT must extend a single chain\. The CoT curve also improves as memory accumulates, indicating that synchronization strengthens later reasoning even without branch exploration\. Both curves eventually plateau, suggesting that most gains come from early memory updates and that the remaining MCTS advantage reflects exploration over alternative bridge entities or relation chains\.

Memory evolution\.Table[5](https://arxiv.org/html/2607.23278#S4.T5)illustrates how bidirectional synchronization changes the next retrieval action, not just the final context\. For the question"Who is the mother of the father of George Washington?", the first step retrieves Augustine Washington as George Washington’s father\. Text→\\rightarrowgraph propagation turns this evidence into parentage edges, and graph→\\rightarrowtext propagation writes the resolved father relation back into textual memory\. The second step can therefore query around Augustine Washington as a grounded entity rather than around an underspecified phrase\. This steers SPARQL expansion toward Augustine Washington’s parentage triples and injects the mother edge to Mildred Warner back into memory\. The example also clarifies the role of pruning and search: memory may contain plausible off\-path facts, while verification and branch selection decide which path answers the question\.

Failure analysis\.Manual inspection of Bamboogle failures, detailed in Appendix[D](https://arxiv.org/html/2607.23278#A4), shows that remaining errors usually come from evidence quality or path selection rather than from the absence of graph\-text memory\. Retrieval gaps account for 45% of failures: the system often identifies the right entity chain, but the corpus repeatedly supports an incorrect value, which synchronization cannot fix when the evidence is internally consistent\. Incomplete reasoning chains account for 21%, usually when relevant facts are present in memory but the synthesizer selects a chain that is too short, too long, or off by one hop\. Wrong entity linking and synthesis errors each account for 14%, and surface\-form mismatch accounts for 7%\. These categories suggest remedies: stronger source cross\-checking for retrieval gaps, better entity disambiguation for linking errors, and explicit chain\-completeness verification for incomplete reasoning\.

## 5Conclusion

We introduced Co\-E, a training\-free multi\-hop QA system built around synchronized graph\-text working memory\. Co\-E treats reasoning as an iterative memory\-maintenance problem: textual evidence can expand the graph frontier when open graph retrieval is available, and graph facts can be written back into the textual context that conditions later retrieval and generation\. This bidirectional synchronization lets the system revise its relational state during inference rather than relying on a static retrieved context or a fixed topic\-centered candidate graph\. Across six KGQA and text\-QA benchmarks, Co\-E consistently improves over comparable training\-free open\-backbone baselines, achieving 72\.6 EM on 2WikiMultiHopQA, 70\.0 EM on Bamboogle, 52\.5 EM on MuSiQue, and 74\.9 Hits@1 on CWQ\. Ablations show that both propagation directions and active memory maintenance are necessary: simply accumulating evidence is not enough\. The main KGQA results show strong performance under the standard Freebase topic\-subgraph protocol, while the Full\-Wikidata stress test shows that Co\-E can retain strong compositional KGQA performance without benchmark\-specific subgraph preprocessing\. Together, these results support graph\-text co\-evolution as a practical mechanism for training\-free multi\-hop reasoning\.

## Limitations

Co\-E is limited by the quality of the evidence it retrieves\. Bidirectional synchronization can reconcile redundant or contradicted memory entries, but it cannot reliably correct a wrong fact that is repeatedly supported by the retrieved corpus\. This appears in the Bamboogle failure analysis, where retrieval gaps account for 45% of errors\. Future systems could cross\-check numerical attributes, superlatives, and entity facts against independent sources, though doing so would add latency\. Co\-E also depends on accurate entity linking and path selection\. Wrong entity linking accounts for 14% of Bamboogle failures, and incomplete reasoning chains account for 21%\. These errors suggest that graph\-text memory is not sufficient by itself: the system still needs better candidate disambiguation, stronger hop\-count constraints, and more explicit verification of whether the selected chain fully answers the question\. MCTS mode uses a shared global working memory rather than a fully branch\-local memory\. Verifier assessments, consolidation, and graph pruning mitigate contamination from weak branches, but they do not provide rollback semantics; future work could compare shared\-memory MCTS with branch\-local or reward\-gated memory commits\. Finally, Co\-E trades additional inference cost for stronger reasoning\. MCTS mode requires multiple LLM calls per expansion; CoT mode is cheaper but less accurate\. Stronger early stopping, cached verification, or adaptive switching between CoT and MCTS could reduce this cost\. Our experiments primarily use a Qwen3\-8B backbone, with a Bamboogle\-only Qwen3\.5\-4B check in Appendix[C](https://arxiv.org/html/2607.23278#A3); evaluating larger or more specialized models, and extending the smaller\-backbone study across all benchmarks, remains future work\.

## References

- S\. Dong, Q\. Zhang, Y\. Xiao, S\. Chen, C\. Zhou, and X\. Huang \(2026\)Use graph when it needs: efficiently and adaptively integrating retrieval\-augmented generation with graphs\.External Links:2602\.03578,[Link](https://arxiv.org/abs/2602.03578)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.19.19.1),[§1](https://arxiv.org/html/2607.23278#S1.p2.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1)\.
- X\. Ho, A\. Duong Nguyen, S\. Sugawara, and A\. Aizawa \(2020\)Constructing a multi\-hop QA dataset for comprehensive evaluation of reasoning steps\.InProceedings of the 28th International Conference on Computational Linguistics,D\. Scott, N\. Bel, and C\. Zong \(Eds\.\),Barcelona, Spain \(Online\),pp\. 6609–6625\.External Links:[Link](https://aclanthology.org/2020.coling-main.580/),[Document](https://dx.doi.org/10.18653/v1/2020.coling-main.580)Cited by:[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p1.1)\.
- M\. Jeon, S\. Cho, and J\. Shu \(2025\)PPoGA: predictive plan\-on\-graph with action for knowledge graph question answering\.External Links:2602\.00007,[Link](https://arxiv.org/abs/2602.00007)Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.14.10.1)\.
- B\. Jin, H\. Zeng, Z\. Yue, J\. Yoon, S\. Arik, D\. Wang, H\. Zamani, and J\. Han \(2025\)Search\-r1: training llms to reason and leverage search engines with reinforcement learning\.External Links:2503\.09516,[Link](https://arxiv.org/abs/2503.09516)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.10.10.1),[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.11.11.1),[§2](https://arxiv.org/html/2607.23278#S2.p2.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.6.6.1)\.
- J\. Li, Y\. Jin, D\. Liu, H\. Ding, J\. Wu, D\. Chen, Y\. Shen, Y\. Qin, Y\. Tai, C\. Wang, X\. Yuan, and Y\. Wang \(2026a\)SE\-search: self\-evolving search agent via memory and dense reward\.External Links:2603\.03293,[Link](https://arxiv.org/abs/2603.03293)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.27.27.1),[§1](https://arxiv.org/html/2607.23278#S1.p2.1),[§2](https://arxiv.org/html/2607.23278#S2.p4.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.16.16.1)\.
- J\. Li, J\. Ruan, S\. Tang, S\. Chen, K\. Chang, Y\. Ge, T\. Xiao, and J\. Zhu \(2025a\)SUBQRAG: sub\-question driven dynamic graph rag\.External Links:2510\.07718,[Link](https://arxiv.org/abs/2510.07718)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.24.24.1),[§1](https://arxiv.org/html/2607.23278#S1.p2.1),[§2](https://arxiv.org/html/2607.23278#S2.p4.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.15.15.1.1)\.
- M\. Li, S\. Miao, and P\. Li \(2025b\)Simple is effective: the roles of graphs and large language models in knowledge\-graph\-based retrieval\-augmented generation\.External Links:2410\.20724,[Link](https://arxiv.org/abs/2410.20724)Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.7.3.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p2.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 1](https://arxiv.org/html/2607.23278#S4.T1.5.1.3.3.1.1)\.
- X\. Li, G\. Dong, J\. Jin, Y\. Zhang, Y\. Zhou, Y\. Zhu, P\. Zhang, and Z\. Dou \(2025c\)Search\-o1: agentic search\-enhanced large reasoning models\.External Links:2501\.05366,[Link](https://arxiv.org/abs/2501.05366)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.8.8.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.4.4.1.1)\.
- X\. Li, Z\. Liu, H\. Xin, Y\. Yan, S\. Wang, Z\. Zeng, S\. Mei, G\. Yu, and M\. Sun \(2026b\)Structured knowledge representation through contextual pages for retrieval\-augmented generation\.External Links:2601\.09402,[Link](https://arxiv.org/abs/2601.09402)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.17.17.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.10.10.1.1)\.
- H\. Liu, Z\. Wang, X\. Chen, Z\. Li, F\. Xiong, Q\. Yu, and W\. Zhang \(2025\)HopRAG: multi\-hop reasoning for logic\-aware retrieval\-augmented generation\.External Links:2502\.12442,[Link](https://arxiv.org/abs/2502.12442)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.3.3.1),[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.4.4.1),[§2](https://arxiv.org/html/2607.23278#S2.p2.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p2.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.3.3.1.1)\.
- Z\. Liu, M\. Wu, X\. Li, Y\. Yan, S\. Wang, C\. Yang, M\. Yu, Z\. Zeng, and M\. Sun \(2026\)Graph\-anchored knowledge indexing for retrieval\-augmented generation\.External Links:2601\.16462,[Link](https://arxiv.org/abs/2601.16462)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.18.18.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.11.11.1)\.
- H\. Luo, H\. E, Y\. Guo, Q\. Lin, X\. Wu, X\. Mu, W\. Liu, M\. Song, Y\. Zhu, and L\. A\. Tuan \(2025\)KBQA\-o1: agentic knowledge base question answering with monte carlo tree search\.External Links:2501\.18922,[Link](https://arxiv.org/abs/2501.18922)Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.3.3.3.4),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 1](https://arxiv.org/html/2607.23278#S4.T1.5.1.11.11.1)\.
- L\. Luo, Y\. Li, G\. Haffari, and S\. Pan \(2024\)Reasoning on graphs: faithful and interpretable large language model reasoning\.External Links:2310\.01061,[Link](https://arxiv.org/abs/2310.01061)Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.18.14.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 1](https://arxiv.org/html/2607.23278#S4.T1.5.1.13.13.1)\.
- S\. Ma, C\. Xu, X\. Jiang, M\. Li, H\. Qu, C\. Yang, J\. Mao, and J\. Guo \(2025\)Think\-on\-graph 2\.0: deep and faithful large language model reasoning with knowledge\-guided retrieval augmented generation\.External Links:2407\.10805,[Link](https://arxiv.org/abs/2407.10805)Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.12.8.1),[§1](https://arxiv.org/html/2607.23278#S1.p2.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p2.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 1](https://arxiv.org/html/2607.23278#S4.T1.5.1.7.7.1.1)\.
- Y\. Ma, S\. Liang, R\. Wang, Jiakai, Q\. Chen, M\. Li, and K\. Qin \(2026\)Mitigating kg quality issues: a robust multi\-hop graphrag retrieval framework\.External Links:2603\.14828,[Link](https://arxiv.org/abs/2603.14828)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.16.16.1),[§1](https://arxiv.org/html/2607.23278#S1.p2.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.9.9.1.1)\.
- C\. Mavromatis and G\. Karypis \(2025\)GNN\-RAG: graph neural retrieval for efficient large language model reasoning on knowledge graphs\.InFindings of the Association for Computational Linguistics: ACL 2025,W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 16682–16699\.External Links:[Link](https://aclanthology.org/2025.findings-acl.856/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-acl.856),ISBN 979\-8\-89176\-256\-5Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.8.4.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 1](https://arxiv.org/html/2607.23278#S4.T1.5.1.5.5.1)\.
- J\. Park, S\. Lee, O\. Z\. Khan, H\. J\. Kim, and J\. Kim \(2026\)HyperGraphPro: progress\-aware reinforcement learning for structure\-guided hypergraph rag\.External Links:2601\.17755,[Link](https://arxiv.org/abs/2601.17755)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.21.21.1),[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.22.22.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.13.13.1)\.
- O\. Press, M\. Zhang, S\. Min, L\. Schmidt, N\. A\. Smith, and M\. Lewis \(2023\)Measuring and narrowing the compositionality gap in language models\.External Links:2210\.03350,[Link](https://arxiv.org/abs/2210.03350)Cited by:[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p1.1)\.
- Z\. Qi, M\. Ma, J\. Xu, L\. L\. Zhang, F\. Yang, and M\. Yang \(2024\)Mutual reasoning makes smaller llms stronger problem\-solvers\.External Links:2408\.06195,[Link](https://arxiv.org/abs/2408.06195)Cited by:[§3\.5](https://arxiv.org/html/2607.23278#S3.SS5.p2.1)\.
- Qwen Team \(2026\)Qwen3\.5: towards native multimodal agents\.External Links:[Link](https://qwen.ai/blog?id=qwen3.5)Cited by:[Appendix C](https://arxiv.org/html/2607.23278#A3.p2.1)\.
- Y\. Shi, M\. Sun, Z\. Liu, M\. Yang, Y\. Fang, T\. Sun, and X\. Gu \(2026\)Reasoning in trees: improving retrieval\-augmented generation for multi\-hop question answering\.External Links:2601\.11255,[Link](https://arxiv.org/abs/2601.11255)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.6.6.1),[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.7.7.1),[§2](https://arxiv.org/html/2607.23278#S2.p2.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1)\.
- D\. Silver, J\. Schrittwieser, K\. Simonyan, I\. Antonoglou, A\. Huang, A\. Guez, T\. Hubert, L\. Baker, M\. Lai, A\. Bolton, Y\. Chen, T\. P\. Lillicrap, F\. Hui, L\. Sifre, G\. van den Driessche, T\. Graepel, and D\. Hassabis \(2017\)Mastering the game of go without human knowledge\.Nature550,pp\. 354–359\.External Links:[Link](https://api.semanticscholar.org/CorpusID:205261034)Cited by:[§3\.5](https://arxiv.org/html/2607.23278#S3.SS5.p2.1)\.
- A\. Singh, A\. Ehtesham, S\. Kumar, T\. T\. Khoei, and A\. V\. Vasilakos \(2026\)Agentic retrieval\-augmented generation: a survey on agentic rag\.External Links:2501\.09136,[Link](https://arxiv.org/abs/2501.09136)Cited by:[§1](https://arxiv.org/html/2607.23278#S1.p1.1)\.
- X\. Song, S\. Zhang, and T\. Yu \(2025\)ReKG\-MCTS: reinforcing LLM reasoning on knowledge graphs via training\-free Monte Carlo tree search\.InFindings of the Association for Computational Linguistics: ACL 2025,W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 9288–9306\.External Links:[Link](https://aclanthology.org/2025.findings-acl.484/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-acl.484),ISBN 979\-8\-89176\-256\-5Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.16.12.1),[§1](https://arxiv.org/html/2607.23278#S1.p2.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 1](https://arxiv.org/html/2607.23278#S4.T1.5.1.10.10.1)\.
- J\. Sun, C\. Xu, L\. Tang, S\. Wang, C\. Lin, Y\. Gong, L\. M\. Ni, H\. Shum, and J\. Guo \(2024\)Think\-on\-graph: deep and responsible reasoning of large language model on knowledge graph\.External Links:2307\.07697,[Link](https://arxiv.org/abs/2307.07697)Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.11.7.1),[§1](https://arxiv.org/html/2607.23278#S1.p2.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1)\.
- X\. Sun, Z\. Chen, X\. Zheng, Q\. Liu, S\. Wu, B\. Song, Z\. Wang, W\. Wang, and L\. Wang \(2026\)KBQA\-r1: reinforcing large language models for knowledge base question answering\.External Links:2512\.10999,[Link](https://arxiv.org/abs/2512.10999)Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.19.15.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1)\.
- Y\. Sun, K\. Sun, Y\. E\. Xu, X\. Yang, X\. L\. Dong, N\. Tang, and L\. Chen \(2025\)KERAG: knowledge\-enhanced retrieval\-augmented generation for advanced question answering\.InFindings of the Association for Computational Linguistics: EMNLP 2025,pp\. 6194–6216\.Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.4.2),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1)\.
- A\. Talmor and J\. Berant \(2018\)The web as a knowledge\-base for answering complex questions\.External Links:1803\.06643,[Link](https://arxiv.org/abs/1803.06643)Cited by:[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p1.1)\.
- X\. Tan, X\. Wang, Q\. Liu, X\. Xu, X\. Yuan, and W\. Zhang \(2025\)Paths\-over\-graph: knowledge graph empowered large language model reasoning\.External Links:2410\.14211,[Link](https://arxiv.org/abs/2410.14211)Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.13.9.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 1](https://arxiv.org/html/2607.23278#S4.T1.5.1.8.8.1.1)\.
- Q\. Team \(2025\)Qwen3 technical report\.External Links:2505\.09388,[Link](https://arxiv.org/abs/2505.09388)Cited by:[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p2.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p4.1)\.
- H\. Trivedi, N\. Balasubramanian, T\. Khot, and A\. Sabharwal \(2022\)MuSiQue: multihop questions via single\-hop question composition\.Transactions of the Association for Computational Linguistics10,pp\. 539–554\.External Links:[Link](https://aclanthology.org/2022.tacl-1.31/),[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00475)Cited by:[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p1.1)\.
- H\. Trivedi, N\. Balasubramanian, T\. Khot, and A\. Sabharwal \(2023\)Interleaving retrieval with chain\-of\-thought reasoning for knowledge\-intensive multi\-step questions\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 10014–10037\.External Links:[Link](https://aclanthology.org/2023.acl-long.557/),[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.557)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.9.9.1),[§2](https://arxiv.org/html/2607.23278#S2.p2.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p2.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.5.5.1)\.
- S\. Wang and Y\. Yu \(2025\)IQUEST: an iterative question\-guided framework for knowledge base question answering\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 15616–15628\.External Links:[Link](https://aclanthology.org/2025.acl-long.760/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.760),ISBN 979\-8\-89176\-251\-0Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.9.5.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 1](https://arxiv.org/html/2607.23278#S4.T1.5.1.4.4.1.1)\.
- Y\. Wang, S\. Fang, B\. Wang, Q\. Luo, X\. Huang, Y\. Zheng, and X\. Qiu \(2026\)Multi\-hop reasoning via early knowledge alignment\.External Links:2512\.20144,[Link](https://arxiv.org/abs/2512.20144)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.20.20.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.12.12.1)\.
- R\. Wu, Y\. Lee, F\. Shu, D\. Xu, S\. Hwang, Z\. Yao, Y\. He, and F\. Yan \(2025\)ComposeRAG: a modular and composable rag for corpus\-grounded multi\-hop question answering\.External Links:2506\.00232,[Link](https://arxiv.org/abs/2506.00232)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.5.5.1),[§2](https://arxiv.org/html/2607.23278#S2.p2.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1)\.
- T\. Xia, M\. Xu, L\. Hu, Y\. Sun, W\. Li, L\. Shang, L\. Liu, P\. Shu, H\. Yu, and J\. Jiang \(2026\)Search\-p1: path\-centric reward shaping for stable and efficient agentic rag training\.External Links:2602\.22576,[Link](https://arxiv.org/abs/2602.22576)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.12.12.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1)\.
- T\. Xiao, Y\. Yuan, H\. Ivison, H\. Zhu, F\. Brahman, N\. Lambert, P\. Dasigi, N\. A\. Smith, and H\. Hajishirzi \(2026\)Meta\-reinforcement learning with self\-reflection for agentic search\.External Links:2603\.11327,[Link](https://arxiv.org/abs/2603.11327)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.13.13.1),[§2](https://arxiv.org/html/2607.23278#S2.p2.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 2](https://arxiv.org/html/2607.23278#S4.T2.5.1.7.7.1)\.
- G\. Xiong, J\. Bao, and W\. Zhao \(2024\)Interactive\-KBQA: multi\-turn interactions for knowledge base question answering with large language models\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 10561–10582\.External Links:[Link](https://aclanthology.org/2024.acl-long.569/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.569)Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.15.11.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 1](https://arxiv.org/html/2607.23278#S4.T1.5.1.9.9.1.1)\.
- G\. Xiong, H\. Li, and W\. Zhao \(2025\)MCTS\-kbqa: monte carlo tree search for knowledge base question answering\.External Links:2502\.13428,[Link](https://arxiv.org/abs/2502.13428)Cited by:[Table 6](https://arxiv.org/html/2607.23278#A1.T6.4.4.17.13.1),[§1](https://arxiv.org/html/2607.23278#S1.p2.1),[§2](https://arxiv.org/html/2607.23278#S2.p3.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1),[Table 1](https://arxiv.org/html/2607.23278#S4.T1.5.1.12.12.1)\.
- S\. Xu, Y\. Xu, J\. Liu, C\. Yuan, W\. Peng, J\. Zhao, and K\. Liu \(2026\)GraphWalker: agentic knowledge graph question answering via synthetic trajectory curriculum\.External Links:2603\.28533,[Link](https://arxiv.org/abs/2603.28533)Cited by:[§2](https://arxiv.org/html/2607.23278#S2.p3.1)\.
- W\. Xu, Z\. Liang, K\. Mei, H\. Gao, J\. Tan, and Y\. Zhang \(2025\)A\-mem: agentic memory for llm agents\.External Links:2502\.12110,[Link](https://arxiv.org/abs/2502.12110)Cited by:[§1](https://arxiv.org/html/2607.23278#S1.p2.1),[§2](https://arxiv.org/html/2607.23278#S2.p4.1)\.
- Z\. Yang, P\. Qi, S\. Zhang, Y\. Bengio, W\. Cohen, R\. Salakhutdinov, and C\. D\. Manning \(2018\)HotpotQA: a dataset for diverse, explainable multi\-hop question answering\.InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing,E\. Riloff, D\. Chiang, J\. Hockenmaier, and J\. Tsujii \(Eds\.\),Brussels, Belgium,pp\. 2369–2380\.External Links:[Link](https://aclanthology.org/D18-1259/),[Document](https://dx.doi.org/10.18653/v1/D18-1259)Cited by:[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p1.1)\.
- W\. Yih, M\. Richardson, C\. Meek, M\. Chang, and J\. Suh \(2016\)The value of semantic parse labeling for knowledge base question answering\.InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\),pp\. 201–206\.Cited by:[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p1.1)\.
- S\. Zhang, J\. Li, Y\. Zhang, P\. Jia, Y\. Wang, X\. Qian, W\. Zhang, M\. Wang, Y\. Liu, and X\. Zhao \(2026\)MemSearch\-o1: empowering large language models with reasoning\-aligned memory growth in agentic search\.External Links:2604\.17265,[Link](https://arxiv.org/abs/2604.17265)Cited by:[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.25.25.1),[Table 7](https://arxiv.org/html/2607.23278#A1.T7.1.1.26.26.1),[§1](https://arxiv.org/html/2607.23278#S1.p2.1),[§2](https://arxiv.org/html/2607.23278#S2.p4.1),[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p3.1)\.
- Y\. Zhang, M\. Li, D\. Long, X\. Zhang, H\. Lin, B\. Yang, P\. Xie, A\. Yang, D\. Liu, J\. Lin, F\. Huang, and J\. Zhou \(2025\)Qwen3 embedding: advancing text embedding and reranking through foundation models\.arXiv preprint arXiv:2506\.05176\.Cited by:[§4\.1](https://arxiv.org/html/2607.23278#S4.SS1.p4.1)\.
- P\. Zhao, H\. Zhang, Q\. Yu, Z\. Wang, Y\. Geng, F\. Fu, L\. Yang, W\. Zhang, J\. Jiang, and B\. Cui \(2024\)Retrieval\-augmented generation for ai\-generated content: a survey\.External Links:2402\.19473,[Link](https://arxiv.org/abs/2402.19473)Cited by:[§1](https://arxiv.org/html/2607.23278#S1.p1.1)\.
- C\. Zhou, C\. Zhang, G\. Yu, F\. Meng, J\. Zhou, W\. Lam, and M\. Yu \(2026a\)Improving multi\-step rag with hypergraph\-based memory for long\-context complex relational modeling\.External Links:2512\.23959,[Link](https://arxiv.org/abs/2512.23959)Cited by:[§1](https://arxiv.org/html/2607.23278#S1.p2.1),[§2](https://arxiv.org/html/2607.23278#S2.p4.1)\.
- P\. Zhou, H\. Li, Z\. Nie, J\. Chen, Q\. Gong, W\. Zhang, and C\. Yu \(2026b\)Understand then memory: a cognitive gist\-driven rag framework with global semantic diffusion\.External Links:2602\.15895,[Link](https://arxiv.org/abs/2602.15895)Cited by:[§1](https://arxiv.org/html/2607.23278#S1.p2.1)\.

## Appendix AFull Baseline Results

Tables[6](https://arxiv.org/html/2607.23278#A1.T6)and[7](https://arxiv.org/html/2607.23278#A1.T7)provide the full baseline comparison used to construct the main result tables\. We include each method’s backbone, training setting, reported metric, and any supplementary semantic\-accuracy score when available\. The tables include both training\-free and trained systems, with closed\-source or substantially larger backbones retained as reference points rather than primary comparisons\.

MethodBackboneTrainingWebQSPCWQMetric noteLearned/Subgraph KG\-Retrieval KGQASubgraphRAGLiet al\.\([2025b](https://arxiv.org/html/2607.23278#bib.bib36)\)GPT\-4o\-mini/ GPT\-4oYes \(MLP\)90\.1 / 77\.566\.7 / 59\.1Hits@1 / F1GNN\-RAGMavromatis and Karypis \([2025](https://arxiv.org/html/2607.23278#bib.bib39)\)LLaMA\-2\-7BYes \(GNN\)85\.7 / 71\.366\.8 / 60\.4Hits@1 / F1iQUESTWang and Yu \([2025](https://arxiv.org/html/2607.23278#bib.bib44)\)GPT\-4oYes \(GNN\)88\.9373\.85Hits@1Live\-KG KGQAToGSunet al\.\([2024](https://arxiv.org/html/2607.23278#bib.bib8)\)GPT\-3\.5\-turboNo76\.257\.1Hits@1ToG 2\.0Maet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib9)\)GPT\-3\.5\-turboNo81\.1—Hits@1PoGTanet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib35)\)GPT\-3\.5\-turboNo93\.974\.7Hits@1PPoGAJeonet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib37)\)GPT\-3\.5\-turboNo83\.164\.5Hits@1Interactive\-KBQAXionget al\.\([2024](https://arxiv.org/html/2607.23278#bib.bib16)\)GPT\-4No72\.47 / 71\.2059\.17 / 49\.07Hits@1 / F1ReKG\-MCTSSonget al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib45)\)LLaMA\-3\-8BNo72\.2059\.80Hits@1KBQA\-o1Luoet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib42)\)LLaMA\-3\.1\-8BNo68\.3∗/ 59\.857\.8∗/ 43\.1∗Hits@1 / F1MCTS\-KBQAXionget al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib41)\)LLaMA\-3\.1\-8BNo73\.5 / 72\.572\.1 / 64\.9Hits@1 / F1KERAGSunet al\.\([2025](https://arxiv.org/html/2607.23278#bib.bib46)\)LLaMA\-3\.1\-70BYes \(SFT\)\+84\.370\.20Hits@1RoGLuoet al\.\([2024](https://arxiv.org/html/2607.23278#bib.bib38)\)LLaMA\-2\-7BYes \(SFT\)85\.7 / 70\.862\.6 / 56\.2Hits@1 / F1KBQA\-R1Sunet al\.\([2026](https://arxiv.org/html/2607.23278#bib.bib43)\)LLaMA\-3\.1\-8BYes \(GRPO\)83\.4—F1Co\-E \(MCTS\)Qwen3\-8BNo85\.5 / 94\.5574\.9 / 89\.85Hits@1 / AccCo\-E \(CoT\)Qwen3\-8BNo85\.3 / 92\.770\.1 / 87\.78Hits@1 / Acc

Table 6:Full KGQA baseline results\.∗denotes reproduced results\.\+KERAG trains an 8B model to summarize retrieved evidence\.Table 7:Full multi\-hop text\-QA baseline results\.
## Appendix BImplementation Details

LLM inference runs on 4 NVIDIA A100 80GB GPUs\. The Qwen3\-8B reasoning backbone is served on 2 GPUs with SGLang model parallelism and thinking enabled, then exposed through a LiteLLM OpenAI\-compatible proxy\. Embedding and reranking models run on a separate GPU, while evaluation runs on a CPU server with 128 GB RAM and 32 cores\. Dense retrieval uses an in\-memory FAISS index over the Wiki23 corpus\. SPARQL queries target either a local Virtuoso endpoint for KGQA or the public Wikidata endpoint for text\-QA\. Redis caches repeated calls during MCTS simulations\.

GroupParameterValueLLM GenerationBackboneQwen3\-8BTemperature1\.0top\-p0\.95top\-k20Presence penalty1\.5ThinkingenabledRetrievalEmbedding modelQwen3\-Embedding\-4BReranker modelQwen3\-Reranker\-4BReranker top\-k10Web search top\-k5Graph hops \(kk\)2Entity candidates per mention3Triple pruner top\-k128MCTSMax iterations20Max tree depth10Exploration weight \(cc\)2\.5Max simulation depth5Min iterations before early stop5High\-confidence threshold0\.9Convergence patience5Semantic sufficiency count5MCTS node priorsSub\-question/Answer0\.60Self\-Correction0\.50Synthesis0\.45Question Rephrasing0\.40Final Answer0\.30CoTMax chain depth10Max sub\-questions per step3Working MemoryMax textual memory tokens8192

Table 8:Configuration hyperparameters applied across the Co\-E framework\.### B\.1Knowledge Base Setup

KGQA benchmarks \(WebQSP, CWQ\)\.We follow the standard topic\-subgraph protocol: each question is paired with a pre\-retrieved Freebase candidate subgraph centered on its topic entity, and graph retrieval is restricted to that subgraph\. This reduces the full\-KG search space but still leaves noisy and off\-path relations for the system to filter\. SPARQL queries target a local Virtuoso endpoint loaded with the per\-question subgraphs\. Entity linking maps surface mentions to Freebase MIDs using a dictionary derived from the benchmark training set; no new entities are added during evaluation\. Corpus retrieval uses FAISS over the Wiki23 corpus with Qwen3\-Embedding\-4B\.

Text\-QA benchmarks \(2WikiMultiHopQA, HotpotQA, MuSiQue, Bamboogle\)\.We use full Wikidata without pre\-selection\. Graph retrieval issueskk\-hop SPARQL queries against the public Wikidata endpoint and prunes candidates to the top 128 triples\. Entity linking maps surface mentions to Wikidata QIDs; unlike the KGQA setting, the shared entity dictionary grows as new entities are discovered\. Corpus retrieval again uses FAISS over the Wiki23 corpus\.

Full\-Wikidata KGQA\.The Full\-Wikidata evaluation in §[4\.5](https://arxiv.org/html/2607.23278#S4.SS5)applies the text\-QA protocol to WebQSP and CWQ: the same Wikidata endpoint, no pre\-selected subgraph, and the same dual\-stream retrieval stack\. This setting tests robustness when the knowledge\-base source is not benchmark\-adapted\.

## Appendix CInference Cost Analysis

Table[10](https://arxiv.org/html/2607.23278#A3.T10)reports role\-level inference cost on Bamboogle: average LLM calls per question, output \(non\-thinking\) tokens per call, and total output tokens per question\. These counts include every LLM invocation, both cache hits and misses; our implementation caches repeated LLM and retrieval calls, which reduces wall\-clock latency but does not lower the totals reported here, since they reflect the underlying compute demand\. Co\-E \(MCTS\) uses 204\.1 calls and 27\.0K output tokens per question, compared with 58\.0 calls and 7\.8K tokens for Co\-E \(CoT\), and 9\.3 calls and 1\.2K tokens for IR\-CoT\. In output tokens, MCTS is therefore about 3\.5×\\timesCoT and 22×\\timesIR\-CoT\. The cost is concentrated in retrieval\-side construction, not reasoning\. Retrieval roles account for 57\.1% of MCTS output tokens and 68\.7% of CoT output tokens\. Within retrieval, the extractor is the largest contributor at 6\.2K tokens per MCTS question and 2\.7K per CoT question; within reasoning, answer generation contributes 8\.0K and 2\.1K tokens\. MCTS further repeats this loop over alternative branches and adds verifier calls\. The added cost is therefore tied to making intermediate evidence explicit, pruning noisy state, and exploring multiple reasoning paths\.

To probe how much of Co\-E’s accuracy depends on the 8B backbone, we re\-run both modes on Bamboogle using Qwen3\.5\-4BQwen Team \([2026](https://arxiv.org/html/2607.23278#bib.bib55)\)as the backbone for every role\. Results are shown in Table[9](https://arxiv.org/html/2607.23278#A3.T9)\. The smaller backbone loses 1\.8 EM in MCTS mode and 3\.3 EM in CoT mode\. Co\-E therefore degrades gracefully when the backbone is roughly halved, consistent with the design hypothesis that most of the gain comes from synchronized graph\-text memory rather than from raw backbone capacity\. This also provides direct evidence for the deployment argument below: in settings where 8B inference is too expensive, a 4B\-class backbone preserves most of the accuracy at a fraction of the per\-call compute\.

Table 9:Bamboogle backbone\-size sensitivity for Co\-E\.This cost can be reduced without changing the training\-free setup\. We use Qwen3\-8B for all roles because it is a relatively small, locally deployable backbone and keeps evaluation controlled\. The Qwen3\.5\-4B result above shows that a roughly halved backbone preserves most of Co\-E’s Bamboogle accuracy, so the cheaper\-backbone direction is empirically supported and not just hypothetical\. In deployment, Co\-E’s role\-based prompting makes model routing straightforward: retrieval\-side roles such as NER, extraction, and pruning can use cheaper 2B\-4B models or deterministic tools, while stronger models are reserved for harder reasoning roles such as question generation, verification, and consolidation\. Thus, inference cost can be significantly reduced because most computation is on the retrieval side\. Caching can also reuse repeated entity\-linking, SPARQL, extraction, verification, and LLM\-call outputs\. Finally, CoT provides the same graph\-text memory mechanism at much lower cost, while MCTS can be reserved for ambiguous or long\-chain questions\.

Table 10:Bamboogle inference cost by LLM role\. Call counts are averaged per question; the role\-level token column reports tokens per call, and the final row reports total output tokens per question\.
## Appendix DFailure Mode Analysis

Table[11](https://arxiv.org/html/2607.23278#A4.T11)reports the full failure breakdown for Co\-E \(MCTS\) on Bamboogle: 29 incorrect predictions out of 125 questions\. Categories are assigned by manual inspection of the prediction, working memory, and retrieved evidence\.Retrieval gapdenotes cases where the system identifies the right entity but the corpus supports an incorrect fact\.Incomplete chaincovers failures where the needed evidence is present but the final chain is too short, too long, or off by one hop\.Wrong entity linkingcovers scope\-sensitive mentions resolved to the wrong entity\.Synthesis errormeans the correct answer appears in memory but the final generator outputs a different answer\.Surface\-form mismatchcaptures strict\-EM artifacts where the prediction is semantically correct but formatted differently from the gold string\.

Table 11:Failure mode breakdown for Co\-E \(MCTS\) on Bamboogle\.
## Appendix EPrompt Templates

This appendix reproduces the prompt templates used at each stage of Co\-E\. All LLM calls use JSON\-mode structured output; each prompt’sOutput Formatblock gives the required schema\. Prompts are shared across KGQA and text\-QA benchmarks, with only runtime context variables changed \(question, memory state, and retrieved evidence\)\.

### E\.1Request Generation

Used at the start of each reasoning step \(§[3\.1](https://arxiv.org/html/2607.23278#S3.SS1), stage 1\) to generate a focused sub\-queryq\(t\)q^\{\(t\)\}from the current memoryℳ\(t\)\\mathcal\{M\}^\{\(t\)\}and questionqq\. In CoT mode, this prompt also decides whether the question is already answerable from memory\. In MCTS mode, it serves as theSub\-question/Answernode type\.

Request Generation Prompt[⬇](data:text/plain;base64,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)Youareanexpertassistantformulti\-hopquestionansweringandreasoningdecomposition\.Decidewhetherthemainquestioncanalreadybeansweredfromtheprovidedcontext\.Ifnot,generatestrategicsubquestionstoadvancethereasoning\.\#\#IntermediateAnswerIf‘intermediate\_answer‘isprovided,itistheresolvedresultofthePREVIOUShop\.Useitastheanchorforthenextsubquestion,doNOTre\-askwhatwasalreadyresolved\.\#\#CorePrinciplesEachsubquestionmust:\-targetarealknowledgegapnotanswerablefromtheprovidedcontext\-beatomic,self\-contained,andunderstandablewithoutthemainquestion\-benon\-redundantwithothergeneratedsubquestions\*\*Sequentialchainsareallowed\*\*whenthemainquestionexplicitlylinkshops\.Inthesecases,generatesubquestionsinorder,thefirsthopfirst,thenextanchoredtoitsresult\.Sequentialsubquestionsmaydependoneachother\.\#\#Instructions1\.Analyzethemainquestion:identifycoreintent,keyentities,constraints,andrequiredreasoningsteps\.2\.Checkthecontext:ifsufficienttoanswer,set‘is\_answerable‘totrueandstop\.3\.Identifymissingknowledge:onlygapsthatmeaningfullyadvancereasoningtowardtheanswer\.4\.Generatesubquestions:\-Forparallelgaps:eachmustbeindependentlyanswerable\.\-Forchainedhops:generateinsequentialorder;use‘intermediate\_answer‘toanchorstep2\+\.\-Ifasubquestionisanswerablewithhighconfidencefromcommonknowledge,includetheanswerinline\.5\.Validate:removesubquestionsanswerablefromcontext,redundant,orlow\-value\.Keepatmost3\.\#\#ScopeConsistencyPreservethegeographicorcategoricalscopeofthemainquestion\.DoNOTsilentlynarrowaglobalscopetoaspecificregionwithoutjustification\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-is\_answerable:boolean\-subquestions:arrayofstringsornull

### E\.2Corpus Search

#### E\.2\.1Query Generator

Rewrites the sub\-queryq\(t\)q^\{\(t\)\}into search\-optimized queries for web or dense\-retrieval search \(§[3\.4](https://arxiv.org/html/2607.23278#S3.SS4), Corpus Search\)\.

Query Generator Prompt[⬇](data:text/plain;base64,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)YouareaReasoningEnginethatdeconstructsuserinputintoprecise,self\-containedsearchqueries\.\#\#Principles1\.Self\-Contained:Eachqueryunderstandablewithoutoriginalinput\.2\.Atomic:Onesinglefactperquery\.3\.Essential&Non\-Redundant:Everyquerynecessaryandunique\.\#\#Instructions1\.ParsetheInput:identifytype\(factual,comparative,causal,temporal\),keyentities,andrequiredreasoningsteps\.2\.GenerateStrategicQueries:formulatequeriesthatcollectivelycoverallnecessaryinformationtoanswertheinput\.3\.EnsureSelf\-Containment:eachquerymustbeindependentlyanswerable\.4\.ReviewforCompletenessandNon\-Redundancy\.5\.TemporalGrounding:iftheinputcontains"current","now",orapresent\-tensesuperlative,add"asof\[currentyear\]"toatleastonequery\.6\.FallbackQueries:add1\-2fallbackqueriesusingcommonaliasesoralternativephrasingssothatretrievalsucceedsiftheprimaryphrasingreturnsnoresults\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-queries:arrayofstrings

#### E\.2\.2Extractor

Filters each retrieved document into self\-contained, question\-relevant snippets that formS\(t\)S^\{\(t\)\}\(§[3\.4](https://arxiv.org/html/2607.23278#S3.SS4), Corpus Search\)\. Raw documents never enter the working memory directly\.

Extractor Prompt[⬇](data:text/plain;base64,WW91IGFyZSBhIG1ldGljdWxvdXMgcmVzZWFyY2ggYW5hbHlzdC4gQnVpbGQgYSBjb21wcmVoZW5zaXZlIGRvc3NpZXIgb2YKaW5mb3JtYXRpb24gZnJvbSB0aGUgcHJvdmlkZWQgdGV4dCB0aGF0IGNvdWxkIGhlbHAgYW5zd2VyIHRoZSBxdWVzdGlvbi4KClJ1bGVzOgotIENvbnNpZGVyIGJvdGggZGlyZWN0IGFuZCBpbmRpcmVjdCByZWxldmFudCBpbmZvcm1hdGlvbi4gSW5mb3JtYXRpb24gaXMKICByZWxldmFudCBpZiBpdCBjb250YWlucyBjbHVlcyB0aGF0IGNvdWxkIGhlbHAgYW5zd2VyIHRoZSBxdWVzdGlvbiAobm90CiAgbmVjZXNzYXJpbHkgZGlyZWN0bHkgYW5zd2VyaW5nIGl0LCBidXQgcHJvdmlkaW5nIGluZm9ybWF0aW9uIHRoYXQgY291bGQKICBoZWxwIGFuc3dlciB0aGUgcXVlc3Rpb24pLgotIEV4dHJhY3RlZCBpbmZvcm1hdGlvbiBtdXN0IGJlIHNlbGYtY29udGFpbmVkIGFuZCBjbGVhciwgaS5lLiwgdW5kZXJzdGFuZGFibGUKICB3aXRob3V0IGFueSBleHRlcm5hbCBjb250ZXh0LgoKSW5zdHJ1Y3Rpb25zOgoxLiBRdWVzdGlvbiBEZWNvbnN0cnVjdGlvbjogaWRlbnRpZnkgcHJpbWFyeSBzdWJqZWN0LCBrZXkgZW50aXRpZXMsIGFuZAogICBzcGVjaWZpYyBpbmZvcm1hdGlvbiBzb3VnaHQuCjIuIENhbmRpZGF0ZSBJZGVudGlmaWNhdGlvbjogaWRlbnRpZnkgYW5kIHF1b3RlIEFMTCBwYXNzYWdlcyBwb3RlbnRpYWxseQogICByZWxhdGVkIHRvIGNvbmNlcHRzIGluIHRoZSBxdWVzdGlvbi4gQmUgbGliZXJhbCBhbmQgaW5jbHVzaXZlIGluIHRoaXMKICAgaW5pdGlhbCBwYXNzLgozLiBSZWxldmFuY2UgRXZhbHVhdGlvbjogYXNzZXNzIGVhY2ggcXVvdGUgYWdhaW5zdCBjcml0ZXJpYSAoZGlyZWN0bHkKICAgYW5zd2VyaW5nLCBjb250ZXh0dWFsLCBzdXBwb3J0aW5nIGV2aWRlbmNlLCBldGMuKS4KNC4gRXh0cmFjdGlvbjogZXh0cmFjdCBBTEwgcmVsZXZhbnQgaW5mb3JtYXRpb24uIEFkZCBjb250ZXh0IGZvciBjbGFyaXR5IGJ1dAogICBwcmVzZXJ2ZSBvcmlnaW5hbCBtZWFuaW5nLiBFYWNoIGV4dHJhY3RlZCBpdGVtIG11c3QgYmUgRlVMTFkgVU5ERVJTVEFOREFCTEUKICAgb24gaXRzIG93bi4KNS4gRmluYWwgZXZhbHVhdGlvbjogZXhhbWluZSBlYWNoIGl0ZW0gZm9yIHNlbGYtY29udGFpbm1lbnQgYW5kIHJlbGV2YW5jZTsKICAgcmV3cml0ZSBvciByZW1vdmUgYXMgbmVlZGVkLgoKIyMgT3V0cHV0IEZvcm1hdApSZXNwb25kIHdpdGggYSBKU09OIG9iamVjdCB3aXRoIGV4YWN0bHkgdGhlc2Uga2V5czoKLSByZWxldmFudF9pbmZvcm1hdGlvbjogYXJyYXkgb2Ygc3RyaW5ncw==)Youareameticulousresearchanalyst\.Buildacomprehensivedossierofinformationfromtheprovidedtextthatcouldhelpanswerthequestion\.Rules:\-Considerbothdirectandindirectrelevantinformation\.Informationisrelevantifitcontainscluesthatcouldhelpanswerthequestion\(notnecessarilydirectlyansweringit,butprovidinginformationthatcouldhelpanswerthequestion\)\.\-Extractedinformationmustbeself\-containedandclear,i\.e\.,understandablewithoutanyexternalcontext\.Instructions:1\.QuestionDeconstruction:identifyprimarysubject,keyentities,andspecificinformationsought\.2\.CandidateIdentification:identifyandquoteALLpassagespotentiallyrelatedtoconceptsinthequestion\.Beliberalandinclusiveinthisinitialpass\.3\.RelevanceEvaluation:assesseachquoteagainstcriteria\(directlyanswering,contextual,supportingevidence,etc\.\)\.4\.Extraction:extractALLrelevantinformation\.Addcontextforclaritybutpreserveoriginalmeaning\.EachextracteditemmustbeFULLYUNDERSTANDABLEonitsown\.5\.Finalevaluation:examineeachitemforself\-containmentandrelevance;rewriteorremoveasneeded\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-relevant\_information:arrayofstrings

### E\.3Response Generation

Used in stage 3 of each reasoning iteration \(§[3\.1](https://arxiv.org/html/2607.23278#S3.SS1)\) to produce an intermediate answer from the current memory augmented with newly retrieved evidence\.

Response Generation Prompt[⬇](data:text/plain;base64,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)Youareanexpertassistantspecializinginprecise,well\-reasonedquestionanswering\.Deliveradirect,accurateanswerwithtransparent,step\-by\-stepreasoning\.\#\#Instructions1\.Analyzethequestion:identifycoreintent,keyentities,andspecificinformationsought\.2\.Contextpriority:whencontextisprovided,groundyouranswerexclusivelyinthecontext,donotintroduceexternalfacts\.Onlyuseyourownknowledgewhencontextisabsentorclearlyincomplete,andexplicitlystatewhendoingso\.3\.Synthesizeaclear,well\-reasonedanswer\.Stateanyassumptionsclearly\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-answer:string\-concise\_answer:string\-reasoning:string\-confidence\_level:string

### E\.4Memory Updating

#### E\.4\.1Memory Consolidation

Runs twice per synchronization cycle \(§[3\.3](https://arxiv.org/html/2607.23278#S3.SS3)\): once after evidence retrieval \(primary consolidation\) and once after graph\-to\-text injection \(re\-consolidation\)\.

Memory Consolidation Prompt[⬇](data:text/plain;base64,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)YouareanexpertMemoryConsolidationAgent\.Processaninputmemory\(alistofinformationitems\)andconsolidateitintoarefinedmemorycontainingonlyinformationrelevantandusefulforansweringthegivenquestion\.\#\#Instructions1\.QuestionAnalysis:identifyprimarysubject,keyentities,andinformationsought\.2\.MemoryAtomization:atomizememoryintoatomic,self\-containeditems\.Nopronounsorreferencestoexternalcontext\.3\.Deduplication:iftwoitemshavethesamecontent,keepone\.Ifoneitemiscompletelycontainedinanother,removeit\.4\.RelevanceEvaluation:keepitemsthatcontainANYcluethatcouldhelpanswerthequestion\.4b\.ProvenanceAudit:foreach\[SystemPrediction\]item,checkifany\[Retrieval\]itemcoversthesameclaim\.Ifcontradicted\-\>removethe\[SystemPrediction\]item\.Ifsupported\-\>upgradeprovenanceto"Retrieval"\.4c\.HopDepthFiltering:itemstagged\[hop=N\]wereretrievedatreasoningstepN\.Ifalower\-hopitemfullyanswersthequestion,discardhigher\-hopitemsnotneeded\.Strip\[hop=N\]prefixesfromoutput‘content‘;recorddepthin‘hop\_depth‘\.5\.IrrelevantInformationRemoval\.6\.ConflictResolution:whena\[Retrieval\]itemanda\[SystemPrediction\]itemstateconflictingspecificfacts,ALWAYSkeep\[Retrieval\]anddiscard\[SystemPrediction\]\.Iftwo\[Retrieval\]itemsconflict,keepbothandnotetheconflict\.7\.Refinement:ensureeachitemisself\-containedandclear\.8\.Finalcheck:verifyeverykeptitemisself\-contained,non\-redundant,andhascorrectprovenance\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-consolidated\_memory:arrayofobjects;eachobjecthas:\-content:string\-provenance:string,exactly"SystemPrediction"or"Retrieval"\-hop\_depth:integerornull

#### E\.4\.2Named Entity Recognition and Entity Linking

Identifies entities in the consolidated memory and links them to Wikidata QIDs for inclusion in the shared entity dictionary \(§[3\.3](https://arxiv.org/html/2607.23278#S3.SS3), text\-to\-graph propagation\)\.

NER and Entity Linking Prompt[⬇](data:text/plain;base64,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)YouareanexpertNamedEntityRecognitionspecialist\.Extractallnamedentitiesfromthetext\.YoumaybegivenanoptionallistofKNOWNENTITIES,eachwith:\-id:WikidataQID\-name:officialWikidatalabel\-description:briefdescriptionfordisambiguationWikidatalinkingrules\(strict\):\-IfanextractedentityclearlymatchesaKNOWNENTITY,setitsidtothatQIDandusetheKNOWNENTITY’sofficialname\.\-Ifthereisanyambiguityoryouarenotfullycertain,setidtonull\.\-NEVERguessorinventaQIDifitisnotprovided\.Instructions:1\.Iftextisaquestion,focusONLYonentitieswhicharemaincluestoanswerthequestion\.2\.Definepreciseboundaries\(includemodifiers\)\.3\.Handleambiguityusingcontext\.4\.Extractuniqueentitiesonlyonce\(deduplicatebyreal\-worldidentity\)\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-entities:arrayofobjects;eachobjecthas:\-id:stringornull\-name:string\-description:stringornull

#### E\.4\.3Relation Extraction

Extracts open\-vocabulary \(subject, relation, object\) triples from the consolidated memory to extend the graph memory𝒢\\mathcal\{G\}\(§[3\.3](https://arxiv.org/html/2607.23278#S3.SS3), text\-to\-graph propagation\)\.

Relation Extraction Prompt[⬇](data:text/plain;base64,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)YouareanexpertRelationExtractionspecialist\.Extractallmeaningfulrelationshipsbetweenentities\.Eachrelationshipmustbeself\-contained\.YoumaybegivenanoptionallistofKNOWNENTITIES\(withWikidataQIDs\);applythesamestrictlinkingrulesasNER\.Instructions:1\.Identifyentitypairswithdirectrelationships\.2\.Breakdowncomplexrelationshipsintosimplerones\.3\.Onlyextractexplicitlystatedorstronglyimpliedrelationships\.4\.Useclear,concise,active\-voicerelationtypes\.5\.Ensurerelationsareself\-containedandnon\-duplicated\.\#\#CanonicalRelationDirectionAlwaysusetheACTIVEform\.NeverusepassiveorinverseformsRules,converttoactiveformwhenapredicate:1\.endsin"\_of"\-\>invert:has\_child,contains,has\_capital2\.startswith"is\_"\-\>inverttoactiveform3\.endsin"\_by"\-\>invert:precedes,directed,owns,succeeds\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-relations:arrayofobjects;eachobjecthas:\-subject:string\-subject\_id:stringornull\-relation:string\-object:string\-object\_id:stringornull\-context:stringornull

#### E\.4\.4Triple Pruner

Filters extracted and retrieved triples for question\-relevance before they are merged into𝒢\\mathcal\{G\}\(§[3\.3](https://arxiv.org/html/2607.23278#S3.SS3)and §[3\.4](https://arxiv.org/html/2607.23278#S3.SS4)\)\. Applied as the LLM stage of the two\-stage pruner \(cross\-encoder reranker followed by this LLM filter\)\.

Triple Pruner Prompt[⬇](data:text/plain;base64,WW91IGFyZSBhIEtub3dsZWRnZSBHcmFwaCBFeHBlcnQuIEdpdmVuIGEgcXVlc3Rpb24gYW5kIGEgbGlzdCBvZiB0cmlwbGVzLAprZWVwIG9ubHkgdGhlIHRyaXBsZXMgdGhhdCBhcmUgZ2VudWluZWx5IHVzZWZ1bCBmb3IgYW5zd2VyaW5nIHRoZSBxdWVzdGlvbi4KCiMjIFJlbGV2YW5jZSBDcml0ZXJpYQpBIHRyaXBsZSAoU3ViamVjdCwgUmVsYXRpb24sIE9iamVjdCkgaXMgcmVsZXZhbnQgT05MWSBpZiBpdCBtZWV0cyBvbmUgb2Y6CgoxLiBEaXJlY3QgcmVsZXZhbmNlOiBib3RoIHN1YmplY3QgQU5EIG9iamVjdCBhcmUgZGlyZWN0bHkgcmVsYXRlZCB0byB0aGUKICAgcXVlc3Rpb24sIGFuZCB0aGUgcmVsYXRpb24gY29ubmVjdHMgdGhlbSBpbiBhIHdheSB0aGF0IGhlbHBzIGFuc3dlciBpdC4KICAgLSBFeGFtcGxlIFE6ICJXaGF0IGlzIHRoZSBjYXBpdGFsIG9mIEZyYW5jZT8iCiAgICAgS2VlcDogKEZyYW5jZSwgY2FwaXRhbCwgUGFyaXMpCiAgICAgRHJvcDogKEZyYW5jZSwgYm9yZGVycywgR2VybWFueSkKCjIuIENoYWluIHJlbGV2YW5jZTogdGhlIHRyaXBsZSBmb3JtcyBwYXJ0IG9mIGEgcmVhc29uaW5nIGNoYWluIHdpdGggYW5vdGhlcgogICBrZXB0IHRyaXBsZS4gT25lIGVudGl0eSBvZiB0aGlzIHRyaXBsZSBtdXN0IG1hdGNoIGFuIGVudGl0eSBpbiBhbm90aGVyCiAgIHJlbGV2YW50IHRyaXBsZSwgYW5kIHRvZ2V0aGVyIHRoZXkgaGVscCBhbnN3ZXIgdGhlIHF1ZXN0aW9uLgogICAtIEV4YW1wbGUgUTogIldobyBpcyB0aGUgc3BvdXNlIG9mIHRoZSBwcmVzaWRlbnQgb2YgRnJhbmNlPyIKICAgICBLZWVwOiAoRnJhbmNlLCBwcmVzaWRlbnQsIE1hY3JvbikgKyAoTWFjcm9uLCBzcG91c2UsIEJyaWdpdHRlKQoKIyMgS2V5IFJ1bGUKRG8gTk9UIGtlZXAgYSB0cmlwbGUganVzdCBiZWNhdXNlIG9uZSBlbnRpdHkgc3VwZXJmaWNpYWxseSBtYXRjaGVzIGEgd29yZCBpbgp0aGUgcXVlc3Rpb24uIERPIGtlZXAgYSB0cmlwbGUgaWYgb25lIGVudGl0eSBpcyBjbGVhcmx5IHJlbGV2YW50IEFORCB0aGUgb3RoZXIKY291bGQgcGxhdXNpYmx5IGJlIGFuIGludGVybWVkaWF0ZSBzdGVwIG9yIGFuc3dlciBpbiB0aGUgcmVhc29uaW5nIGNoYWluLgoKIyMgT3V0cHV0IEZvcm1hdApSZXNwb25kIHdpdGggYSBKU09OIG9iamVjdCB3aXRoIGV4YWN0bHkgdGhlc2Uga2V5czoKLSBrZWVwX2luZGljZXM6IGFycmF5IG9mIGludGVnZXJzLCAwLWJhc2VkIGluZGljZXMgb2YgdHJpcGxlcyB0byByZXRhaW4=)YouareaKnowledgeGraphExpert\.Givenaquestionandalistoftriples,keeponlythetriplesthataregenuinelyusefulforansweringthequestion\.\#\#RelevanceCriteriaAtriple\(Subject,Relation,Object\)isrelevantONLYifitmeetsoneof:1\.Directrelevance:bothsubjectANDobjectaredirectlyrelatedtothequestion,andtherelationconnectstheminawaythathelpsanswerit\.\-ExampleQ:"WhatisthecapitalofFrance?"Keep:\(France,capital,Paris\)Drop:\(France,borders,Germany\)2\.Chainrelevance:thetripleformspartofareasoningchainwithanotherkepttriple\.Oneentityofthistriplemustmatchanentityinanotherrelevanttriple,andtogethertheyhelpanswerthequestion\.\-ExampleQ:"WhoisthespouseofthepresidentofFrance?"Keep:\(France,president,Macron\)\+\(Macron,spouse,Brigitte\)\#\#KeyRuleDoNOTkeepatriplejustbecauseoneentitysuperficiallymatchesawordinthequestion\.DOkeepatripleifoneentityisclearlyrelevantANDtheothercouldplausiblybeanintermediatesteporanswerinthereasoningchain\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-keep\_indices:arrayofintegers,0\-basedindicesoftriplestoretain

### E\.5MCTS\-Specific Prompts

These prompts are used exclusively in MCTS mode \(§[3\.5](https://arxiv.org/html/2607.23278#S3.SS5)\)\. Each corresponds to one of the five MCTS node types\.

#### E\.5\.1Self\-Correction Node

Self\-Correction Prompt[⬇](data:text/plain;base64,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)Youareanexpertinanswerverificationandrefinement\.Givenaquestion,proposedanswer,andcontext,verifycorrectnessandprovidearefinedresponse\.\#\#Instructions1\.Parsequestionrequirements\.2\.Extractrelevantfactsfromcontext\.3\.Evaluateproposedansweras:CORRECT/PARTIAL/INCORRECT/UNSUPPORTED\.4\.Generaterefinedanswer\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-status:string,oneof:correct,partial,incorrect,unsupported\-refined\_answer:string\-confidence\_level:string

#### E\.5\.2Question Rephrasing Node

Question Rephrasing Prompt[⬇](data:text/plain;base64,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)YouareaQuestionRefinerthattransformsunclearquestionsintoprecise,clearquestions\.Therephrasedquestionmustbefullyunderstandableonitsownwithoutneedingtoreferbacktotheoriginal\.\#\#Principles1\.ClarityFirst:eliminateambiguityandjargon\.2\.PreserveIntent:donotalterthecoreinquiry\.3\.EnhanceAnswerability:makespecificandself\-contained\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-rephrased\_question:string

#### E\.5\.3Synthesis Node

Synthesis Prompt[⬇](data:text/plain;base64,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)YouareaspecializedAIformulti\-stepreasoning\.Performasingle,focusedreasoningstepbyanalyzingcontextandproducingaconsolidatedsynthesis\.\#\#Instructions1\.Analyzethemainquestionobjective\.2\.Reviewallinformationincontext\.3\.Ifcontextissufficienttodirectlyanswer,statethisandformulatethedefinitiveanswer\.4\.Otherwise,synthesizenewthoughtsthatadvancereasoning:\-Causalortemporallinks\-Corerelationshipidentification\-Progresssummary\-Contradictionidentification\-Hypothesisformulation\#\#CriticalConstraints1\.NoExternalInformation:doNOTintroducefactsnotinthecontext\.2\.NoNewQuestions:synthesize,donotquery\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-is\_answerable:boolean\-step\_conclusion:string\-confidence\_level:string

#### E\.5\.4Final Answer Synthesis

Synthesizes the final answer from all candidate terminal nodes, weighted by their MCTS reward scores\.

Final Answer Synthesis Prompt[⬇](data:text/plain;base64,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)Youareanexpertinargumentativesynthesis\.Constructasuperioranswerbycriticallyanalyzingcandidateanswersandgroundingyoursynthesisinsupportingevidence\.\#\#UnderstandingyourinputsCandidateanswersareMCTSoutputsprefixedwith\[quality\_score=X\]whereXin\[\-1,1\]\.Score\>=0\.5:primaryevidence\.Score0\-0\.5:supportingevidence\.Score<0:treatskeptically\.supporting\_evidencehastwocomponents:\-Textualfactstagged\[Retrieval\]\(highreliability\)or\[SystemPrediction\]\(mediumreliability\)\.\-Knowledgegraphtriples\(under\*\*InformationN\*\*sections\):structuredentity\-relationshipfacts\.Useforverifyingentityattributes;discardtriplesunrelatedtothequestion\.\#\#SynthesisProcedurePhaseI,CandidateAnalysis:extractcoreclaim,premises,reasoningchainforeachcandidate\.Weightbyquality\_score\.PhaseII,EvidenceCross\-Check&ConflictResolution\.Adjudicationhierarchy:1\.High\-score\(\>=0\.5\)\+\[Retrieval\]corroboration2\.Convergentpositive\-scorecandidates3\.\[Retrieval\]factsregardlessofcandidatescore4\.Logicallysoundreasoningconsistentwithgraphtriples5\.MajorityagreementaslastresortPhaseIII,Synthesis&Self\-Critique:verifyeveryfactualclaimistraceabletoatleastonereliablesource\.Reviseifnot\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-final\_answer:string\-concise\_answer:string\-reasoning:string\-confidence\_level:string,oneof:high,medium,low,uncertain

#### E\.5\.5Verifier

The verifier runs three parallel calls conditioned on \(i\) no retrieved context, \(ii\) the textual memory𝒯\\mathcal\{T\}, and \(iii\) the graph memory𝒢\\mathcal\{G\}serialized as triples; the normalized mean score becomes the MCTS rewardr∈\[−1,1\]r\\in\[\-1,1\]\.

Verifier Prompt[⬇](data:text/plain;base64,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)Youareanexpertverifier\.Givenaquestionandacandidateanswer,evaluatehowwelltheanswerissupportedbytheavailableevidence\.\-Ifcontextisprovided:useitastheprimarysourcetoverifytheanswer\.\-Ifcontextisnotprovided:useyourownknowledgetoverifyindependently\.Ratefrom0\.0to10\.0howwelltheansweriscorrect/supported\.Scoring:\-9\.0\-10\.0:Fullyverified/stronglysupported\-7\.0\-8\.9:Mostlysupported,minorgaps\-5\.0\-6\.9:Partiallysupportedoruncertain\-3\.0\-4\.9:Weaklysupported,significantdoubts\-0\.0\-2\.9:Contradictedorcompletelyunsupported\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-rating:number,floatfrom0\.0to10\.0\-reasoning:string

### E\.6LLM\-Judged Accuracy \(Acc\)

Used offline to compute the supplementary Acc metric\. This judge is not part of the Co\-E inference pipeline\.

LLM\-Judged Accuracy Prompt[⬇](data:text/plain;base64,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)Youareanexpertevaluator\.Ratethesystem\_answeronascalefrom0\.0to10\.0basedonhoweffectivelyitaddressestheuser\_question\.EvaluationCriteria:1\.Correctness\(60%\):Istheinformationfactuallyaccurate?\-Ifcorrect\_answerisprovidedandsystem\_answermatches,award10\.0\.\-Otherwise,verifyaccuracyusingyourknowledge\.2\.Helpfulness&Relevance\(40%\):Doesitaddresstheuser’scoreneed?ScoringGuidelines:\-9\.0\-10\.0:Correctandcomprehensive\-7\.0\-8\.9:Mostlycorrectwithminorissues\-5\.0\-6\.9:Partiallyaddressesorhasaccuracyconcerns\-3\.0\-4\.9:Significantcorrectnessorrelevanceissues\-0\.0\-2\.9:Incorrectorcompletelyoff\-topicUncertaintyrule:ifyouarenotconfidentaboutthecorrectanswer\(obscurefacts,precisenumbers,rareentities\),assign5\.0ratherthanaconfidenthighorlowscore\.\#\#OutputFormatRespondwithaJSONobjectwithexactlythesekeys:\-rating:number,floatfrom0\.0to10\.0\-reasoning:string