Towards a Learnable Retriever for Textual Graph Question Answering

# Towards a Learnable Retriever for Textual Graph Question Answering
Source: [https://arxiv.org/html/2610.33511](https://arxiv.org/html/2610.33511)
Yuchen YanZhichen ZengXiao LinLihui LiuYinglong XiaHanghang Tong

###### Abstract

Textual graph question answering \(GraphQA\) and graph\-based retrieval\-augmented generation \(GraphRAG\) have gained increasing attention as a way to ground large language models \(LLMs\) on structured knowledge\. Despite extensive efforts, existing methods still face two key gaps\. The first is anarchitectural gap: existing methods typically fine\-tune an LLM\-based generator, but do not also make the retriever itself a learnable LLM for graph evidence selection within one end\-to\-end retrieve\-then\-generate pipeline\. The second is aretrieval\-supervision gap: existing retrievers rarely use gold answer signals to supervise evidence selection\. As a result, they may fail to recall evidence nodes required for reasoning, resulting in insufficientcoverageof the gold answer set; meanwhile, naive attempts to boost recall often introduce excessive irrelevant information that compromisescompactnessand overwhelms the LLM\. To address these gaps, we proposeTaLenT, an LLM\-based end\-to\-end retriever\-generator framework for textual GraphQA\. To bridge the architectural gap,TaLenTpairs a learnable LLM\-based retriever for evidence selection with a fine\-tuned LLM\-based generator within one unified retrieve\-then\-generate pipeline\. To bridge the retrieval\-supervision gap, we leverage gold answer signals as retriever supervision\. Specifically, we stratify graph nodes into gold answers, answer\-related nodes, and irrelevant noise\. We then adopt a weighted pseudo\-label loss to prioritize gold answers and preserve answer\-related nodes for coverage, while discarding irrelevant noise for compactness\. Experimental results across three GraphQA datasets demonstrate thatTaLenTconsistently improves downstream performance\.

1University of Illinois Urbana\-Champaign

2Wayne State University

3Meta

## 1Introduction

![Refer to caption](https://arxiv.org/html/2610.33511v1/figure_1.png)Figure 1:An illustrative example of two limitations of query\-based graph retrieval\. For the query “What was Isaac Newton’s profession?”, a query\-similarity retriever returns a Newton\-centric subgraph \(blue circle\)\. We partition the nodes into three groups:*Gold answers*\(only a small subset of the gold answer set is retrieved by the subgraph in the blue circle, showing limitedcoverage\),*Answer\-related nodes*\(topically relevant but not exact gold answers\), and*Irrelevant noise*\(nodes that match the query tokens such as “Newton” yet are not answer\-related\)\. The presence of substantial noise alongside partial gold coverage highlights thecompactnessissue: improving recall by expanding query\-similar candidates can inflate the subgraph with distractors\.Large language models \(LLMs\) have achieved impressive progress in general\-purpose generation and reasoning\([Thoppilan et al\. 2022](https://arxiv.org/html/2610.33511#bib.bib14);[Bang et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib15)\), but their outputs are bounded by what is encoded in model parameters\. Trained on large\-scale open\-domain corpora collected before a fixed cutoff date, LLMs may hallucinate or omit critical evidence when a query requires precise, up\-to\-date, or domain\-specific knowledge\([Zhao et al\. 2026](https://arxiv.org/html/2610.33511#bib.bib6);[Wang et al\. 2023b](https://arxiv.org/html/2610.33511#bib.bib16)\)\.

To mitigate these limitations, Retrieval\-Augmented Generation \(RAG\) has emerged as a dominant paradigm that grounds LLM outputs in external evidence by dynamically retrieving relevant knowledge snippets for context\-aware generation\([Lewis et al\. 2020](https://arxiv.org/html/2610.33511#bib.bib17);[Shi et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib18);[Gao et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib7)\)\. While conventional RAG methods\([Lewis et al\. 2020](https://arxiv.org/html/2610.33511#bib.bib17);[Shi et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib18)\)mainly retrieve evidence from unstructured text corpora, recent work has shifted toward textual graphs as a more expressive and structured knowledge source\. This trend has spurred the development of Graph\-based RAG \(GraphRAG\) and Graph Question Answering \(GraphQA\)\. By exploiting the explicit semantics and relational structure of textual graphs, graph\-based RAG methods support more grounded reasoning and complex information synthesis\([IV et al\. 2019](https://arxiv.org/html/2610.33511#bib.bib9);[Luo et al\. 2025](https://arxiv.org/html/2610.33511#bib.bib11);[He et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib10)\)\.

A graph\-based RAG pipeline generally consists of two stages: graph evidence retrieval and answer generation\. On the generation side, existing methods often use an LLM\-based generator and fine\-tune it over retrieved graph contexts\. This makes graph evidence retrieval a critical step, since the generator can only reason over the evidence selected for it\.

Conventional retrieval methods in GraphRAG can be categorized into non\-parametric, graph neural network \(GNN\)\-based, and LM\-based approaches\. Non\-parametric retrievers construct subgraphs using graph search algorithms or predefined rules\([Taunk et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib20);[Yasunaga et al\. 2021](https://arxiv.org/html/2610.33511#bib.bib21)\)\. GNN\-based methods encode graph structure with GNNs and identify task\-relevant components to form the retrieved subgraph\([Mavromatis and Karypis 2024](https://arxiv.org/html/2610.33511#bib.bib19)\)\. LM\-based methods often use LMs to assist graph retrieval, for example by reading structured data, generating retrieval cues, or using LM reasoning to guide evidence selection\([Baek et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib3);[Jiang et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib22)\)\.

However, with these retrieval designs, existing graph\-based RAG pipelines still leave an architectural gap\. While the generator is often fine\-tuned as an LLM\-based module, the retriever itself is usually not made a learnable LLM for graph evidence selection\. Instead, evidence selection is still handled by non\-parametric, GNN\-based, or LM\-assisted procedures\. As a result, the retriever and generator are not both trained as LLM modules within one end\-to\-end retrieve\-then\-generate pipeline\.

Beyond this architectural gap, these conventional retrievers also face a retrieval\-supervision gap\. They rely primarily on query similarity while rarely using gold answers to supervise evidence selection\. Query similarity can favor nodes matching query terms over answer\-critical evidence, leading to two limitations:coverage, where a query\-based retriever may fail to cover the gold answer set in the retrieved subgraph, andcompactness, where it may introduce excessive irrelevant noise\. An illustrative example is shown in Figure[1](https://arxiv.org/html/2610.33511#S1.F1)\. A key observation is that the*gold answer set*provides precisely the missing supervision for retrieval\. Gold answers specify what should be covered and also help distinguish answer\-related evidence from irrelevant noise, and can therefore help alleviate these two limitations\. However, such gold signals are rarely exploited to guide or train graph retrievers\.

In this work, we address both gaps by proposingTaLenT, an LLM\-based end\-to\-end retriever\-generator framework for textual GraphQA\. To bridge the architectural gap,TaLenTmakes the retriever itself a learnable LLM for graph evidence selection, and pairs it with a fine\-tuned LLM\-based generator within one unified retrieve\-then\-generate pipeline\. Within the full pipeline, we first apply a rule\-based preprocessing step to construct a candidate subgraph, providing a high\-recall pool of nodes under the LLM context budget\. The learnable LLM\-based retriever is then trained to select relevant evidence nodes, and the generator is fine\-tuned on contexts constructed from the trained retriever’s outputs\. To bridge the retrieval\-supervision gap, we use gold answer signals as retriever supervision to improve both retrieval coverage and compactness\. To improve retrieval coverage, we train a retriever based on a weighted pseudo\-label objective, incorporating gold answer signals to guide the retrieval process\. To promote answer\-critical evidence, we prioritize gold answer and down\-weight other rest signals to improve retrieval compactness\. Intuitively, the loss assigns the highest weight to gold answers to ensure coverage, assigns a much smaller weight to answer\-related nodes to retain useful context, and ignores irrelevant noise to enforce compactness\. In this way,TaLenTimproves evidence selection before generation, leading to better downstream QA performance\. Our contributions are summarized as follows:

- •Idea\.We introduce an LLM\-based end\-to\-end retriever\-generator framework for textual GraphQA, where a learnable LLM\-based retriever and a fine\-tuned LLM\-based generator are placed within one unified retrieve\-then\-generate pipeline\. This design fills the architectural gap by making the retriever itself a learnable LLM for graph evidence selection, rather than only fine\-tuning the downstream generator\.
- •Method\.We train the learnable LLM\-based retriever with a weighted pseudo\-label objective that leverages gold answer signals\. By prioritizing gold answer nodes and down\-weighting weakly related nodes, the objective fills the retrieval\-supervision gap and leads to better downstream QA performance\.
- •Evaluation\.We conduct extensive experiments on three GraphQA benchmarks to validate both retrieval effectiveness and end\-task QA performance\. Our method outperforms the best competitor by at least 3\.18%\.

## 2Preliminaries

Notation\.We use calligraphic letters to denote sets, e\.g\.,𝒱\\mathcal\{V\}andℰ\\mathcal\{E\}for node and edge sets\. A textual graph is represented as𝒢=\(𝒱,ℰ\)\\mathcal\{G\}=\(\\mathcal\{V\},\\mathcal\{E\}\), where each nodev∈𝒱v\\in\\mathcal\{V\}has a textual attributetext⁡\(v\)\\mathrm\{text\}\(v\), e\.g\., an entity name or concept, and each edgee=\(u,r,v\)∈ℰe=\(u,r,v\)\\in\\mathcal\{E\}denotes a relationrrconnecting the source nodeuuto the destination nodevv\. The textual attribute of an edge is denoted astext⁡\(e\)\\mathrm\{text\}\(e\)\. We use𝒮\\mathcal\{S\}to denote the subgraph of𝒢\\mathcal\{G\}, where𝒮⊆𝒢\\mathcal\{S\}\\subseteq\\mathcal\{G\}indicates that its node set𝒱𝒮⊆𝒱\\mathcal\{V\}\_\{\\mathcal\{S\}\}\\subseteq\\mathcal\{V\}and its edge setℰ𝒮⊆ℰ\\mathcal\{E\}\_\{\\mathcal\{S\}\}\\subseteq\\mathcal\{E\}\. We useqqfor the query andaafor the final answer from the generator\. We usea⋆a^\{\\star\}to denote the token sequence of the gold answer in the dataset and use𝒱⋆\\mathcal\{V\}^\{\\star\}as the set of nodes appearing in the gold answer, i\.e\., the gold answer set\. We denote the retriever and generator byRθR\_\{\\theta\}andGϕG\_\{\\phi\}, respectively\.

### 2\.1Problem Setup

We consider a general retrieve\-then\-generate setup\([He et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib10)\)for textual graph question answering \(GraphQA\)\. Each instance in the dataset consists of a natural\-language queryqqand a textual graph𝒢=\(𝒱,ℰ\)\\mathcal\{G\}=\(\\mathcal\{V\},\\mathcal\{E\}\)as the knowledge base\. Due to the large scale of𝒢\\mathcal\{G\}, it is impractical to directly serialize𝒢\\mathcal\{G\}into an LLM prompt, hence the system first retrieves a relevant*subgraph*, conditioned on which an answer is generated\.

Formally, a retrieverRθR\_\{\\theta\}extracts a subgraph𝒮\\mathcal\{S\}from the textual graph𝒢\\mathcal\{G\}based on the queryqq:

𝒮=Rθ​\(q,𝒢\)\.\\mathcal\{S\}=R\_\{\\theta\}\(q,\\mathcal\{G\}\)\.\(1\)
A generatorGϕG\_\{\\phi\}then predicts the answeraabased on the queryqqand the retrieved subgraph𝒮\\mathcal\{S\}:

a=Gϕ​\(q,𝒮\)\.a=G\_\{\\phi\}\(q,\\mathcal\{S\}\)\.\(2\)

### 2\.2Candidate Connected Subgraph

For large textual graphs, directly serializing the full graph often exceeds the LLM context budget\. As in prior graph\-based QA pipelines, we include a preprocessing stage\. Specifically, following G\-Retriever\([He et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib10)\), we first construct a*candidate connected subgraph*𝒮\{\\mathcal\{S\}\}using the Prize\-Collecting Steiner Tree \(PCST\) algorithm\([Bienstock et al\. 1993](https://arxiv.org/html/2610.33511#bib.bib12)\), and use𝒮\\mathcal\{S\}as the input context for downstream modules\.

Indexing and similarity retrieval\.Given a queryqqand a textual graph𝒢=\(𝒱,ℰ\)\\mathcal\{G\}=\(\\mathcal\{V\},\\mathcal\{E\}\), we embed the query and textual attributes of node𝒱\\mathcal\{V\}and edgesℰ\\mathcal\{E\}using a sentence encoder such as SentenceBERT\([Reimers and Gurevych 2019](https://arxiv.org/html/2610.33511#bib.bib13)\):

𝐳q\\displaystyle\\mathbf\{z\}\_\{q\}=LM⁡\(q\),\\displaystyle=\\mathrm\{LM\}\(q\),\(3\)𝐳v\\displaystyle\\mathbf\{z\}\_\{v\}=LM⁡\(text⁡\(v\)\),v∈𝒱,\\displaystyle=\\mathrm\{LM\}\(\\mathrm\{text\}\(v\)\),\\quad v\\in\\mathcal\{V\},\(4\)𝐳e\\displaystyle\\mathbf\{z\}\_\{e\}=LM⁡\(text⁡\(e\)\),e∈ℰ\.\\displaystyle=\\mathrm\{LM\}\(\\mathrm\{text\}\(e\)\),\\quad e\\in\\mathcal\{E\}\.\(5\)
We compute cosine similarity scores and retrieve the top\-kvk\_\{v\}nodes and top\-kek\_\{e\}edges:

𝒱sim\\displaystyle\\mathcal\{V\}\_\{\\text\{sim\}\}=arg​topkv∈𝒱,k=kvcos\(𝐳q,𝐳v\),\\displaystyle=\\argtopk\_\{v\\in\\mathcal\{V\},\\,k=k\_\{v\}\}\\ \\cos\(\\mathbf\{z\}\_\{q\},\\mathbf\{z\}\_\{v\}\),\(6\)ℰsim\\displaystyle\\mathcal\{E\}\_\{\\text\{sim\}\}=arg​topke∈ℰ,k=kecos\(𝐳q,𝐳e\),\\displaystyle=\\argtopk\_\{e\\in\\mathcal\{E\},\\,k=k\_\{e\}\}\\ \\cos\(\\mathbf\{z\}\_\{q\},\\mathbf\{z\}\_\{e\}\),\(7\)
wherekvk\_\{v\}andkek\_\{e\}control the number of retrieved nodes and edges, respectively\.

Subgraph construction\.We assign*prizes*to retrieved candidates based on their ranks, so higher\-ranked items receive larger prizes\. Concretely, forv∈𝒱simv\\in\\mathcal\{V\}\_\{\\text\{sim\}\}with rankrvr\_\{v\}we setp⁡\(v\)=kv−rvp\(v\)=k\_\{v\}\-r\_\{v\}to prioritize those with higher similarity, and analogously definep⁡\(e\)=ke−rep\(e\)=k\_\{e\}\-r\_\{e\}fore∈ℰsime\\in\\mathcal\{E\}\_\{\\text\{sim\}\}; items outside these retrieved lists receive zero prize\. PCST then selects a connected subgraph𝒮=\(𝒱𝒮,ℰ𝒮\)\\mathcal\{S\}=\(\\mathcal\{V\}\_\{\\mathcal\{S\}\},\\mathcal\{E\}\_\{\\mathcal\{S\}\}\)that trades off prize coverage and compactness:

𝒮=arg​max𝒮⊆𝒢connected⁡\(∑v∈𝒱𝒮p⁡\(v\)\+∑e∈ℰ𝒮p⁡\(e\)−c​\|ℰ𝒮\|\)\.\\displaystyle\\mathcal\{S\}=\\argmax\_\{\\begin\{subarray\}\{c\}\\mathcal\{S\}\\subseteq\\mathcal\{G\}\\\\ \\text\{connected\}\\end\{subarray\}\}\\Big\(\\sum\_\{v\\in\\mathcal\{V\}\_\{\\mathcal\{S\}\}\}p\(v\)\+\\sum\_\{e\\in\\mathcal\{E\}\_\{\\mathcal\{S\}\}\}p\(e\)\-c\\,\|\\mathcal\{E\}\_\{\\mathcal\{S\}\}\|\\Big\)\.\(8\)
wherecccontrols the size\-coverage trade\-off\.

## 3Method

![Refer to caption](https://arxiv.org/html/2610.33511v1/figure_2.png)Figure 2:Overview ofTaLenT, an LLM\-based end\-to\-end retriever\-generator framework for textual GraphQA\. The fullTaLenTpipeline contains three stages: \(1\) Preprocessing, which applies PCST to construct an initial candidate connected subgraph𝒮\\mathcal\{S\}from the knowledge base𝒢\\mathcal\{G\}based on queryqq\. \(2\) Retriever training, which trains a learnable LLM\-based retrieverRθR\_\{\\theta\}with weighted pseudo\-label supervision, given the candidate subgraph𝒮\\mathcal\{S\}and queryqq, to produce𝒱^\\hat\{\\mathcal\{V\}\}\. \(3\) Generator training, which fine\-tunes an LLM\-based generatorGϕG\_\{\\phi\}with standard SFT given retrieved node set𝒱^\\hat\{\\mathcal\{V\}\}andqq, to produce the final answeraa\.Overview\.TaLenTis an LLM\-based end\-to\-end retriever\-generator framework for textual GraphQA\. Its two core trainable LLM modules are: \(i\) a learnable nodes retrieverRθR\_\{\\theta\}trained with weighted pseudo\-label supervision \(Section[3\.1](https://arxiv.org/html/2610.33511#S3.SS1)\), and \(ii\) an answer generatorGϕG\_\{\\phi\}fine\-tuned with standard Supervised Fine\-Tuning \(SFT\) \(Section[3\.2](https://arxiv.org/html/2610.33511#S3.SS2)\)\. The fullTaLenTpipeline is illustrated in Figure[2](https://arxiv.org/html/2610.33511#S3.F2)\. Given a queryqqand a full textual graph𝒢\\mathcal\{G\}, a rule\-based PCST preprocessing step first constructs an initial candidate connected subgraph𝒮\\mathcal\{S\}that fits the LLM context budget\. We then train a learnable retrieverRθR\_\{\\theta\}with weighted pseudo\-label supervision to produce an answer\-relevant node set𝒱^\\hat\{\\mathcal\{V\}\}\(Section[3\.1](https://arxiv.org/html/2610.33511#S3.SS1)\)\. Using answer\-relevant nodes𝒱^\\hat\{\\mathcal\{V\}\}, we construct textual contexts and fine\-tune a generatorGϕG\_\{\\phi\}with standard supervised learning \(Section[3\.2](https://arxiv.org/html/2610.33511#S3.SS2)\)\.

### 3\.1Retriever Training

As discussed earlier, PCST and other query\-similarity based retrievers often fail to cover the answer\-related nodes and include irrelevant noise, resulting in limited*coverage*and poor*compactness*\. We therefore use gold answer signals to guide retriever training, encouraging prioritization of gold answers while down\-weighting other signals\. We first describeTaLenT’s retriever prompting and then introduce the weighted pseudo\-label objective\.

Retriever prompting\.We implement our learnable retrieverRθR\_\{\\theta\}as an instruction\-following LLM \(with LoRA\) that performs node selection based on a provided candidate subgraph\. Given a queryqqand the PCST subgraph𝒮\\mathcal\{S\}, we prompt the LLM to produce the nodes that are most helpful for answeringqq\. The LLM outputs should be exactly one line in the format of “name1\|name2\|\.\.\.”\. Eachnamesuch asname1andname2represents the names of the nodesRθR\_\{\\theta\}finally selects\. This constrained format makes the output directly interpretable as a retrieved nodes set, which we denote by𝒱^=Rθ​\(q,𝒮\)\\hat\{\\mathcal\{V\}\}=R\_\{\\theta\}\(q,\{\\mathcal\{S\}\}\)\. A concrete retriever prompt example is provided in Appendix[E\.1](https://arxiv.org/html/2610.33511#A5.SS1)\. It is worth noting that since our retrieverRθR\_\{\\theta\}is an LLM, it leverages its inherent parametric knowledge during the selection process\. Consequently, the output node set𝒱^\\hat\{\\mathcal\{V\}\}may include nodes that were not present in the initial candidate subgraph𝒮\\mathcal\{S\}\(i\.e\.,𝒱^⊈𝒱𝒮\\hat\{\\mathcal\{V\}\}\\not\\subseteq\\mathcal\{V\}\_\{\\mathcal\{S\}\}\)\. In our method,𝒱^\\hat\{\\mathcal\{V\}\}encompasses the complete set of nodes generated byRθR\_\{\\theta\}, reflecting both the information from𝒮\\mathcal\{S\}and the LLM’s internal knowledge\.

Fine\-tuning with weighted pseudo\-label supervision\.Recall that existing works in GraphRAG use retrievers that are largely query\-similarity driven, which leads to two limitations: coverage and compactness\. Our goal is to train the retriever to output the node set𝒱^\\hat\{\\mathcal\{V\}\}that achieves both high*coverage*and*compactness*: it should reliably include nodes that appear in the gold answer set𝒱⋆\\mathcal\{V\}^\{\\star\}\(coverage\) while avoiding redundant or weakly related candidates \(compactness\)\. The gold answer provides precisely the missing supervision, i\.e\., the gold signal, for retrieval\. To inject gold signals into retrieval, we leverage the dataset\-provided gold answers and stratify candidate nodes into three levels: \(i\)*gold answers*, which are candidate nodes that overlap with the gold answer set𝒱⋆\\mathcal\{V\}^\{\\star\}, and provides evidence that the retriever should prioritize to support correct reasoning, \(ii\)*answer\-related nodes*, which are candidate nodes within𝒮\\mathcal\{S\}but are not gold answers, and may still be helpful for providing context, though they are less directly tied to the final answer, and \(iii\)*irrelevant noise*, which are nodes not contained in𝒮\\mathcal\{S\}, and are completely irrelevant and should be suppressed to maintain compactness\. Given the candidate subgraph𝒮\\mathcal\{S\}with node set𝒱𝒮\\mathcal\{V\}\_\{\\mathcal\{S\}\}, and the gold answer set𝒱⋆\\mathcal\{V\}^\{\\star\}, we utilize these nodes as weighted pseudo label supervision by defining

𝒫=𝒱⋆∩𝒱𝒮\(strong positives; level\-1\),\\mathcal\{P\}=\\mathcal\{V\}^\{\\star\}\\cap\\mathcal\{V\}\_\{\\mathcal\{S\}\}\\quad\\text\{\(strong positives; level\-1\)\},\(9\)𝒲=𝒱𝒮∖𝒱⋆\(weak positives; level\-2\)\.\\mathcal\{W\}=\\mathcal\{V\}\_\{\\mathcal\{S\}\}\\setminus\\mathcal\{V\}^\{\\star\}\\quad\\text\{\(weak positives; level\-2\)\}\.\(10\)All nodes outside𝒱𝒮\\mathcal\{V\}\_\{\\mathcal\{S\}\}are treated as irrelevant noise \(level\-3\) during retriever training\.

We trainRθR\_\{\\theta\}as conditional generation of a target stringyythat concatenates nodes in𝒫\\mathcal\{P\}and𝒲\\mathcal\{W\}\(separated by the pipe symbol “\|”; strong positives first, then weak positives\)

y=p1​∣⋯∣​p\|𝒫\|\|w1​∣⋯∣​w\|𝒲\|,y=p\_\{1\}\\mid\\cdots\\mid p\_\{\|\\mathcal\{P\}\|\}\\mid w\_\{1\}\\mid\\cdots\\mid w\_\{\|\\mathcal\{W\}\|\},\(11\)wherepi,∀i=1,…​\|𝒫\|p\_\{i\},\\forall i=1,\\dots\|\\mathcal\{P\}\|are nodes in𝒫\\mathcal\{P\}andwj,∀j=1​…​\|𝒲\|w\_\{j\},\\forall j=1\\dots\|\\mathcal\{W\}\|are nodes in𝒲\\mathcal\{W\}\. Letxxdenote the retriever input prompt, which includes the queryqqand the textualized candidate subgraph𝒮\\mathcal\{S\}, and lety1:Ty\_\{1:T\}be the tokenized target sequence, whereTTis its token length\. To prioritize level\-1 \(gold answer\) nodes while still preserving level\-2 \(answer\-related\) candidates as a weak learning signal, we minimize the token\-weighted negative log\-likelihood:

ℒretr\(θ\)=−∑t=1Tωtlogpθ\(yt∣x,y<t\)\.\\mathcal\{L\}\_\{\\text\{retr\}\}\(\\theta\)=\-\\sum\_\{t=1\}^\{T\}\\omega\_\{t\}\\log p\_\{\\theta\}\(y\_\{t\}\\mid x,y\_\{<t\}\)\.\(12\)
Since we want to encourageRθR\_\{\\theta\}to prioritize gold answer nodes while still allowing potentially useful answer\-related nodes to contribute as a weak learning signal, we adopt the following token weights to achieve this

ωt=\{1,if​yt​belongs to a node in​𝒫,λ,if​yt​belongs to a node in​𝒲,0,otherwise,\\omega\_\{t\}=\\begin\{cases\}1,&\\text\{if \}y\_\{t\}\\text\{ belongs to a node in \}\\mathcal\{P\},\\\\ \\lambda,&\\text\{if \}y\_\{t\}\\text\{ belongs to a node in \}\\mathcal\{W\},\\\\ 0,&\\text\{otherwise,\}\\end\{cases\}\(13\)whereλ\\lambdais a hyperparameter that controls the contribution of weak positives relative to strong positives\.

Table 1:Benchmark results \(%\)\.
### 3\.2Generator Training

Generator Prompting\.Given the predicted node set𝒱^=Rθ​\(q,𝒮\)\\hat\{\\mathcal\{V\}\}=R\_\{\\theta\}\(q,\\mathcal\{S\}\)output by the trainedRθR\_\{\\theta\}, we form the generator input as a*prompt*that concatenates the queryqqwith subgraphs derived from𝒱^\\hat\{\\mathcal\{V\}\}\. These subgraphs contain \(i\) the retrieved nodes in𝒱^\\hat\{\\mathcal\{V\}\}and \(ii\) their associated edgesℰ^\\hat\{\\mathcal\{E\}\}that appear in𝒮\\mathcal\{S\}\. Concretely, we include edges for which at least one endpoint is in𝒱^\\hat\{\\mathcal\{V\}\}:

ℰ^=\{\(u,r,v\)∈ℰ:u∈𝒱^​or​v∈𝒱^\}\.\\hat\{\\mathcal\{E\}\}=\\\{\(u,r,v\)\\in\\mathcal\{E\}:u\\in\\hat\{\\mathcal\{V\}\}\\ \\text\{or\}\\ v\\in\\hat\{\\mathcal\{V\}\}\\\}\.\(14\)We then serialize the retrieved nodes𝒱^\\hat\{\\mathcal\{V\}\}together with the selected edgesℰ^\\hat\{\\mathcal\{E\}\}into a*serialized graph text*and prepend the queryqqto form the generator prompt\. Importantly, this*serialized graph text*is not required to form one whole connected subgraph\. This design avoids forcing unnecessary “bridge” nodes or edges solely to maintain connectivity, which often increases irrelevant context and hurts compactness\. Instead, we allow the retrieved evidence to appear as multiple disconnected components, as long as it covers the nodes in𝒱^\\hat\{\\mathcal\{V\}\}\. Moreover, as we mentioned in Section[3\.1](https://arxiv.org/html/2610.33511#S3.SS1),𝒱^\\hat\{\\mathcal\{V\}\}may include nodes that come fromRθR\_\{\\theta\}’s parametric knowledge and do not appear in the initial candidate subgraph𝒮\\mathcal\{S\}, we retain these nodes as additional isolated nodes in the*serialized graph text*\. A concrete generator prompt example together with the graph serialization format is provided in Appendix[E\.2](https://arxiv.org/html/2610.33511#A5.SS2)\.

Fine\-tuning with standard SFT\.We fine\-tune an LLM\-based generatorGϕG\_\{\\phi\}using standard supervised fine\-tuning \(SFT\)\. For each training instance, we first run the trained retrieverRθR\_\{\\theta\}to obtain the generator prompt constructed in the previous paragraph \(i\.e\., the*serialized graph text*formed from𝒱^\\hat\{\\mathcal\{V\}\}and the associatedℰ^\\hat\{\\mathcal\{E\}\}, together with the queryqq\)\. We then trainGϕG\_\{\\phi\}to generate the gold answera⋆a^\{\\star\}conditioned on this prompt:

ℒgen\(ϕ\)=−∑t=1Talogpϕ\(at⋆∣q,𝒱^,ℰ^,a<t⋆\),\\mathcal\{L\}\_\{\\text\{gen\}\}\(\\phi\)=\-\\sum\_\{t=1\}^\{T\_\{a\}\}\\log p\_\{\\phi\}\(a^\{\\star\}\_\{t\}\\mid q,\\hat\{\\mathcal\{V\}\},\\hat\{\\mathcal\{E\}\},a^\{\\star\}\_\{<t\}\),\(15\)wherea⋆=\(a1⋆,…,aTa⋆\)a^\{\\star\}=\(a^\{\\star\}\_\{1\},\\ldots,a^\{\\star\}\_\{T\_\{a\}\}\)denotes the*token*sequence of the gold answer, andTaT\_\{a\}is its length\. This SFT step transfers retriever’s improved evidence selection into downstream QA\.

### 3\.3Inference and Evaluation

At inference time, we apply the full retrieve\-then\-generate pipeline using the retriever and generator fine\-tuned on the training \(and validation\) splits\. For each test instance\(q,𝒢\)\(q,\\mathcal\{G\}\), we first apply the fixed PCST\-based preprocessing step to construct the candidate subgraph𝒮\\mathcal\{S\}, then run the learned retriever to obtain𝒱^=Rθ​\(q,𝒮\)\\hat\{\\mathcal\{V\}\}=R\_\{\\theta\}\(q,\\mathcal\{S\}\)\. We materialize\(𝒱^,ℰ^\)\(\\hat\{\\mathcal\{V\}\},\\hat\{\\mathcal\{E\}\}\)into a textual graph context and feed it to the generator to produce the final answera=Gϕ​\(q,𝒱^,ℰ^\)a=G\_\{\\phi\}\(q,\\hat\{\\mathcal\{V\}\},\\hat\{\\mathcal\{E\}\}\)\. We report task metrics computed from the generator outputs on the test split\.

## 4Experiments

### 4\.1Experiment Setup

Datasets and Metrics\.To evaluate our method on textual QA tasks, we conduct experiments on three textual graph question answering datasets: ExplaGraphs and WebQSP from the GraphQA benchmark\([He et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib10)\), and CWQ\([Talmor and Berant 2018](https://arxiv.org/html/2610.33511#bib.bib46)\)\. Specifically, \(1\) ExplaGraphs’s task is to determine whether two arguments are supportive or contradictory\. \(2\) WebQSP is a knowledge graph question answering dataset whose primary task is to answer questions that require multi\-hop reasoning\. \(3\) CWQ is a more compositional KGQA benchmark than WebQSP, with more complex questions and longer reasoning chains\. For evaluation, we adopt Accuracy for ExplaGraphs and Hit@1 for WebQSP and CWQ\. Dataset statistics and additional dataset details are provided in Appendix[A](https://arxiv.org/html/2610.33511#A1)\.

Baselines\.We compare against two groups of baselines\. \(1\) Inference\-Only methods that include Zero\-shot, Zero\-CoT\([Kojima et al\. 2022](https://arxiv.org/html/2610.33511#bib.bib1)\), CoT\-BAG\([Wang et al\. 2023a](https://arxiv.org/html/2610.33511#bib.bib2)\), KAPING\([Baek et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib3)\), Graph\-based Inference, and StructGPT\([Jiang et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib22)\)\. \(2\) Prompt\-Tuning methods that include Frozen LLM \+ Prompt Tuning \(PT\), GraphToken\([Perozzi et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib4)\), G\-Retriever\([He et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib10)\), and TaoNa\-A\([Yan et al\. 2025](https://arxiv.org/html/2610.33511#bib.bib5)\)\. For G\-Retriever, we include a variant that fine\-tunes the LLM\-based generator with LoRA \(G\-Retriever w/ LoRA\)\. Additional discussion of our differences from other answer\-aware or answer\-related methods and recent graph retrieval baselines not included in Table[1](https://arxiv.org/html/2610.33511#S3.T1)is provided in Appendix[D](https://arxiv.org/html/2610.33511#A4)\.

Experiment pipeline\.For the main experiments, we use the open\-source LLaMA2\-Chat\-7B model\([Touvron et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib23)\)as the backbone for both the retriever and the generator, and conduct all experiments on a single NVIDIA A100 GPU\. We follow the same train/validation/test splits as prior work for each dataset\. For WebQSP and CWQ, since the full KG exceeds the context budget, we first construct a PCST candidate subgraph withkv=ke=10k\_\{v\}\{=\}k\_\{e\}\{=\}10and then apply our learnable retriever\. During retriever training, we use the weighted pseudo\-label objective with strong\-positive weight11\. The weak\-positive weight is set toλ=0\.001\\lambda=0\.001for WebQSP andλ=0\.01\\lambda=0\.01for ExplaGraphs and CWQ\. Additional implementation details, including hyperparameter configurations and computing infrastructure, are provided in Appendix[B](https://arxiv.org/html/2610.33511#A2)\. Appendix[C](https://arxiv.org/html/2610.33511#A3)reports sensitivity analyses forλ\\lambdaand PCST size parameterskvk\_\{v\}andkek\_\{e\}\. A backbone ablation with Llama\-3\-8B\([Grattafiori et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib50)\)is provided in Appendix[G](https://arxiv.org/html/2610.33511#A7)\.

### 4\.2Benchmark results

Figure 3:WebQSP test\-set retrieval results \(macro\-averaged\)\. We measure how well PCST subgraphs andTaLenT’s learned retriever output𝒱^\\hat\{\\mathcal\{V\}\}cover the gold answer set𝒱⋆\\mathcal\{V\}^\{\\star\}\. We compare: \(1\) the PCST candidate subgraph withkv=ke=10k\_\{v\}\{=\}k\_\{e\}\{=\}10\(this is the retriever input\), \(2\) a much smaller PCST subgraph withkv=ke=3k\_\{v\}\{=\}k\_\{e\}\{=\}3\(a size\-matched reference\), and \(3\) the nodes set𝒱^\\hat\{\\mathcal\{V\}\}selected byTaLenT’s learned retriever\. Our retriever keeps coverage close to PCST\(kv=ke=10k\_\{v\}\{=\}k\_\{e\}\{=\}10\) while using far fewer nodes, and it also covers more gold nodes than the size\-matched PCST\(kv=ke=3k\_\{v\}\{=\}k\_\{e\}\{=\}3\)\.Table[1](https://arxiv.org/html/2610.33511#S3.T1)reports the benchmark results on ExplaGraphs \(Accuracy\), WebQSP \(Hit@1\), and CWQ \(Hit@1\)\. Overall,TaLenToutperforms the strongest competitors across all three datasets\. We draw the following three observations\.

\(1\)TaLenToutperforms the strongest competitors\.TaLenTachieves 91\.79% accuracy on ExplaGraphs, 76\.97% Hit@1 on WebQSP, and 55\.71% Hit@1 on CWQ, improving over the strongest baseline by 4\.74% on ExplaGraphs, 3\.18% on WebQSP, and 6\.52% on CWQ\. These results show thatTaLenTgeneralizes well under different graph QA objectives and metrics, with the largest gain obtained on the more challenging CWQ benchmark\.

\(2\) Comparison to inference\-only methods\.Inference\-only baselines remain behindTaLenTon all three datasets\. Even compared with the strongest inference\-only baseline, StructGPT,TaLenTimproves by 13\.45% on ExplaGraphs, 4\.37% on WebQSP, and 23\.14% on CWQ\. This performance gap stems from their reliance on test\-time prompting and non\-trainable evidence acquisition, without task\-specific fine\-tuning of either the retriever or generator\. In contrast,TaLenTtrains both components, enabling more reliable evidence selection and more accurate answers\.

\(3\) Comparison to prompt\-tuning methods\.Prompt\-tuning methods form substantially stronger baselines than inference\-only methods overall, with G\-Retriever w/ LoRA being the strongest prior prompt\-tuning baseline on all three datasets\. Nevertheless,TaLenTconsistently improves over it, from 87\.05% to 91\.79% on ExplaGraphs, from 73\.79% to 76\.97% on WebQSP, and from 49\.19% to 55\.71% on CWQ\. These gains support the two central design choices ofTaLenT\. First,TaLenTmoves beyond rule\-based, query\-driven evidence selection by introducing a learnable LLM\-based retriever for task\-adaptive graph evidence selection within an end\-to\-end retriever\-generator framework\. Second, it trains this retriever with gold\-answer\-guided weighted pseudo\-label supervision to improve evidence coverage and compactness, thereby benefiting downstream generation\.

### 4\.3Coverage\-Compactness Tradeoff

In this section, we study howTaLenT’s retriever achieves a better coverage\-compactness tradeoff for its output node set𝒱^\\hat\{\\mathcal\{V\}\}\. We conduct experiments on WebQSP, with results shown in Figure[3](https://arxiv.org/html/2610.33511#S4.F3), and make the following observations\.

Compactness: similar coverage with substantially fewer nodes\.We first compare the compactness of nodes retrieved by the initial PCST candidate subgraph \(kv=ke=10k\_\{v\}\{=\}k\_\{e\}\{=\}10\) andTaLenT, as shown in Figure[3](https://arxiv.org/html/2610.33511#S4.F3)\. Under similar gold\-node coverage,TaLenT’s retriever output𝒱^\\hat\{\\mathcal\{V\}\}contains significantly fewer nodes\. Both methods achieve comparable Hit@1 and Recall on WebQSP, yet𝒱^\\hat\{\\mathcal\{V\}\}contains less than half as many nodes: on average 12\.79 nodes forTaLenTversus 31\.75 nodes for the PCST subgraph\. Meanwhile,𝒱^\\hat\{\\mathcal\{V\}\}attains notably higher precision and F1, improving precision from 9\.26% to 24\.55% and F1 from 12\.22% to 26\.63%\. These results suggest that the weighted pseudo\-label supervision effectively incorporates gold signals, enabling the retriever to filter redundant or weakly related nodes and produce a more compact context without sacrificing coverage\.

Coverage: better gold answer coverage under comparable node number\.To further compare coverage, we report results for a much smaller PCST subgraph withkv=ke=3k\_\{v\}\{=\}k\_\{e\}\{=\}3, whose size is comparable to that ofTaLenT\. As shown in Figure[3](https://arxiv.org/html/2610.33511#S4.F3), under a similar node budget,TaLenT’s retriever achieves substantially higher coverage\. Specifically,TaLenTimproves Hit@1 from 61\.92% to 80\.28% and Recall from 51\.76% to 67\.84% compared to PCST\. This demonstrates that, given a similar number of nodes, our gold\-signal\-involved weighted pseudo\-label supervision enables the retriever to achieve significantly higher coverage than the rule\-based, query\-similarity\-driven PCST algorithm\.

Table 2:Breakdown of gold answer coverage contributed by different parts of the learned retriever output on WebQSP test\. All metrics are macro\-averaged over examples and reported as percentages\.Table 3:Retriever ablation on WebQSP test set \(macro\-averaged over examples\)\. “Ablate retriever training” replacesTaLenT’s learnable retriever by a frozen retriever without LoRA fine\-tuning\.Table 4:QA ablations on WebQSP \(Hit@1, %\)\. All systems follow the same pipeline asTaLenTexcept for the ablated component stated in each row\.
### 4\.4Retrieval Beyond PCST

Fine\-tuning on an LLM\-based retriever may leverage parametric knowledge to retrieve nodes beyond the PCST candidate subgraph\. We therefore examine whetherTaLenT’s retriever can retrieve such beyond\-PCST nodes and how they contribute to performance\. Specifically, we split the retriever output𝒱^\\hat\{\\mathcal\{V\}\}into nodes inside and outside the PCST subgraph\. Shown in Table[2](https://arxiv.org/html/2610.33511#S4.T2), nodes outside PCST alone recover a non\-trivial portion of gold answer nodes, with a Hit@1 of 10\.26% and a Recall of 5\.46%\. Moreover, incorporating beyond\-PCST nodes improves Hit@1 from75\.3175\.31to80\.28%80\.28\\%and Recall from64\.19%64\.19\\%to67\.84%67\.84\\%\. Results indicate that the trained LLM retriever effectively complements the PCST candidate subgraph with parametric\-knowledge nodes that enhance gold\-node coverage\. An additional groundedness and hallucination analysis of out\-of\-PCST retrieval is provided in Appendix[F](https://arxiv.org/html/2610.33511#A6)\.

## 5Ablation Study

We conduct two ablation studies on WebQSP to illustrate \(i\) the effect of pseudo\-label retriever training \(ii\) the effect of learnable retriever\.

### 5\.1On the Effect of Pseudo\-label Training

Retrieval quality\.Table[3](https://arxiv.org/html/2610.33511#S4.T3)comparesTaLenT’s retriever against an ablated variant, which uses the same backbone LLaMA2\-Chat\-7B model but keeps it*frozen*\(no training\) and directly prompts it to select helpful nodes from the PCST candidate subgraph\. Results show that the frozen retriever fails to recover gold nodes required for answering, with Hit@1, Precision, and Recall of 36\.61%, 14\.97%, and 26\.26%, respectively\. In contrast,TaLenT’s retriever explicitly prioritizes gold answer nodes during retriever fine\-tuning using pseudo label training, substantially improving coverage with a Hit@1 of 80\.28%, a precision of 24\.55% and a recall of 67\.84%, indicating that injecting gold signals into retriever training is crucial for selecting answer\-relevant nodes\.

Downstream QA performance\.We next examine whether the retrieval improvement translates into better end\-task QA\. Table[4](https://arxiv.org/html/2610.33511#S4.T4)summarizes QA Hit@1 under three systems: \(i\)TaLenTwith a*frozen*retriever \(ablate training on retriever\), \(ii\)TaLenT*without*the retriever module \(directly feeding the PCST candidate subgraph to the generator\), and \(iii\) the fullTaLenTsystem\. Switching from a frozen retriever to our pseudo\-label trained retriever improves QA Hit@1 from 68\.37% to 76\.97%\. This gain is consistent with Table[3](https://arxiv.org/html/2610.33511#S4.T3): pseudo\-label training recovers substantially more gold nodes, providing the generator with more reliable evidence\.

### 5\.2On the Effect of Learnable Retriever

We further isolate the contribution of the retriever by removing it from the pipeline and directly feeding the PCST candidate subgraph into the generator, i\.e\., withoutRθR\_\{\\theta\}or𝒱^\\hat\{\\mathcal\{V\}\}selection\. As shown in Table[4](https://arxiv.org/html/2610.33511#S4.T4), removing the learnable retriever reduces QA performance from 76\.97% to 72\.50%, indicating that an explicit retriever provides clear benefits for downstream QA\. Notably, using a frozen retriever performs even worse \(Hit@1 of 68\.37%\) than removing the retriever entirely, suggesting that retrieval is not inherently beneficial when gold answer node coverage is low, as noisy context can mislead the generator\. This result further underscores the importance of our weighted pseudo\-label training, which substantially improves retrieval quality for end\-task QA\.

## 6Related Work

Retrieval Augmented Generation\.RAG grounds LLMs with external evidence to reduce hallucinations and improve faithfulness\([Sun et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib8);[Gao et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib7)\)\. Classic RAG systems follow a straightforward index\-retrieve\-generate workflow\([Ma et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib25)\), while many recent variants strengthen retrieval with additional processing before and after retrieval\. For example, pre\-retrieval techniques such as query transformation, expansion, and rewriting are often adopted to better match the search space\([Zheng et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib26);[Gao et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib7)\), and post\-retrieval reranking is commonly used to refine the evidence set\([Blagojevi 2023](https://arxiv.org/html/2610.33511#bib.bib27)\)\. More broadly, modular RAG systems use LLMs to iteratively refine retrieval queries or evidence selection\([Yu et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib28)\)\.

RAG over Graphs\.Extensive works study how to combine graph\-structured information with LLMs\([Li et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib29);[Wang et al\. 2023a](https://arxiv.org/html/2610.33511#bib.bib2)\)\. Prior efforts span general graph modeling\([Liu et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib31);[Lei et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib30)\), multimodal architectures that incorporate graph\-like structures\([Yoon et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib32);[Li et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib33)\)\. Representative applications include fundamental graph reasoning\([Zhao et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib35)\), node prediction\([Sun et al\. 2025](https://arxiv.org/html/2610.33511#bib.bib36);[Chen et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib37)\), graph\-level prediction\([Qian et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib38)\)\. A related line of work discusses GraphRAG / GraphQA, where the system selects graph evidence and conditions an LLM on graph context to answer questions\([IV et al\. 2019](https://arxiv.org/html/2610.33511#bib.bib9);[Luo et al\. 2025](https://arxiv.org/html/2610.33511#bib.bib11)\)\.

## 7Conclusion

We address two key gaps in existing GraphRAG systems: anarchitectural gapand aretrieval\-supervision gap\. To address both gaps, we proposeTaLenT, an LLM\-based end\-to\-end retriever\-generator framework for textual GraphQA\. To bridge the architectural gap,TaLenTmakes the retriever itself a learnable LLM for graph evidence selection and pairs it with a fine\-tuned LLM\-based generator within one unified retrieve\-then\-generate pipeline\. To bridge the retrieval\-supervision gap, we trainTaLenT’s retriever with a gold\-answer\-guided weighted pseudo\-label loss that prioritizes answer\-related signals to improve coverage while suppressing noise for compactness\. Extensive experiments on three GraphQA benchmarks demonstrate thatTaLenTconsistently improves downstream QA performance\.

## References

- Baeket al\.\(2023\)J\. Baek, A\. F\. Aji, and A\. SaffariKnowledge\-augmented language model prompting for zero\-shot knowledge graph question answering\.InProceedings of the 1st Workshop on Natural Language Reasoning and Structured Explanations \(NLRSE\),pp\. 78–106\.Cited by:[§1](https://arxiv.org/html/2610.33511#S1.p4.1),[Table 1](https://arxiv.org/html/2610.33511#S3.T1.1.5.1),[§4\.1](https://arxiv.org/html/2610.33511#S4.SS1.p2.1)\.
- Banget al\.\(2023\)Y\. Bang, S\. Cahyawijaya, N\. Lee, W\. Dai, D\. Su, B\. Wilie, H\. Lovenia, Z\. Ji, T\. Yu, W\. Chung,et al\.A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity\.InProceedings of the 13th international joint conference on natural language processing and the 3rd conference of the asia\-pacific chapter of the association for computational linguistics \(volume 1: Long papers\),pp\. 675–718\.Cited by:[§1](https://arxiv.org/html/2610.33511#S1.p1.1)\.
- Bienstocket al\.\(1993\)D\. Bienstock, M\. X\. Goemans, D\. Simchi\-Levi, and D\. WilliamsonA note on the prize collecting traveling salesman problem\.Mathematical programming59\(1\),pp\. 413–420\.Cited by:[§2\.2](https://arxiv.org/html/2610.33511#S2.SS2.p1.1)\.
- Blagojevi \(2023\)V\. BlagojeviEnhancing rag pipelines in haystack: introducing diversityranker and lostinthemiddleranker\.August\.Cited by:[§6](https://arxiv.org/html/2610.33511#S6.p1.1)\.
- Bollackeret al\.\(2008\)K\. Bollacker, C\. Evans, P\. Paritosh, T\. Sturge, and J\. TaylorFreebase: a collaboratively created graph database for structuring human knowledge\.InProceedings of the 2008 ACM SIGMOD international conference on Management of data,pp\. 1247–1250\.Cited by:[§A\.1](https://arxiv.org/html/2610.33511#A1.SS1.p1.1)\.
- Chaiet al\.\(2025\)Z\. Chai, T\. Zhang, L\. Wu, K\. Han, X\. Hu, X\. Huang, and Y\. YangGraphllm: boosting graph reasoning ability of large language model\.IEEE Transactions on Big Data\.Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px2.p1.1)\.
- Chenet al\.\(2024\)Z\. Chen, H\. Mao, H\. Li, W\. Jin, H\. Wen, X\. Wei, S\. Wang, D\. Yin, W\. Fan, H\. Liu,et al\.Exploring the potential of large language models \(llms\) in learning on graphs\.ACM SIGKDD Explorations Newsletter25\(2\),pp\. 42–61\.Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Gaoet al\.\(2023\)L\. Gao, X\. Ma, J\. Lin, and J\. CallanPrecise zero\-shot dense retrieval without relevance labels\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 1762–1777\.Cited by:[§1](https://arxiv.org/html/2610.33511#S1.p2.1),[§6](https://arxiv.org/html/2610.33511#S6.p1.1)\.
- Grattafioriet al\.\(2024\)A\. Grattafiori, A\. Dubey, A\. Jauhri, A\. Pandey, A\. Kadian, A\. Al\-Dahle, A\. Letman, A\. Mathur, A\. Schelten, A\. Vaughan,et al\.The llama 3 herd of models\.arXiv preprint arXiv:2407\.21783\.Cited by:[§4\.1](https://arxiv.org/html/2610.33511#S4.SS1.p3.1)\.
- Heet al\.\(2024\)X\. He, Y\. Tian, Y\. Sun, N\. Chawla, T\. Laurent, Y\. LeCun, X\. Bresson, and B\. HooiG\-retriever: retrieval\-augmented generation for textual graph understanding and question answering\.Advances in Neural Information Processing Systems37,pp\. 132876–132907\.Cited by:[§A\.1](https://arxiv.org/html/2610.33511#A1.SS1.p1.1),[§1](https://arxiv.org/html/2610.33511#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.33511#S2.SS1.p1.1),[§2\.2](https://arxiv.org/html/2610.33511#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2610.33511#S4.SS1.p1.1),[§4\.1](https://arxiv.org/html/2610.33511#S4.SS1.p2.1)\.
- Huet al\.\(2022\)E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, W\. Chen,et al\.Lora: low\-rank adaptation of large language models\.\.ICLR1\(2\),pp\. 3\.Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px2.p1.1)\.
- Huet al\.\(2025\)Y\. Hu, Z\. Lei, Z\. Dai, A\. Zhang, A\. Angirekula, Z\. Zhang, and L\. ZhaoCg\-rag: research question answering by citation graph retrieval\-augmented llms\.InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval,pp\. 678–687\.Cited by:[Appendix D](https://arxiv.org/html/2610.33511#A4.SS0.SSS0.Px6)\.
- IVet al\.\(2019\)R\. L\. L\. IV, N\. F\. Liu, M\. E\. Peters, M\. Gardner, and S\. SinghBarack’s wife hillary: using knowledge\-graphs for fact\-aware language modeling\.External Links:1906\.07241,[Link](https://arxiv.org/abs/1906.07241)Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2610.33511#S1.p2.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Jianget al\.\(2023\)J\. Jiang, K\. Zhou, Z\. Dong, K\. Ye, X\. Zhao, and J\. WenStructgpt: a general framework for large language model to reason over structured data\.InProceedings of the 2023 conference on empirical methods in natural language processing,pp\. 9237–9251\.Cited by:[§1](https://arxiv.org/html/2610.33511#S1.p4.1),[Table 1](https://arxiv.org/html/2610.33511#S3.T1.1.12.1),[§4\.1](https://arxiv.org/html/2610.33511#S4.SS1.p2.1)\.
- Kojimaet al\.\(2022\)T\. Kojima, S\. S\. Gu, M\. Reid, Y\. Matsuo, and Y\. IwasawaLarge language models are zero\-shot reasoners\.Advances in neural information processing systems35,pp\. 22199–22213\.Cited by:[Table 1](https://arxiv.org/html/2610.33511#S3.T1.1.3.1),[§4\.1](https://arxiv.org/html/2610.33511#S4.SS1.p2.1)\.
- Leiet al\.\(2023\)B\. Lei, pei\-Hung Lin, C\. Liao, and C\. DingBoosting logical reasoning in large language models through a new framework: the graph of thought\.External Links:2308\.08614,[Link](https://arxiv.org/abs/2308.08614)Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Lesteret al\.\(2021\)B\. Lester, R\. Al\-Rfou, and N\. ConstantThe power of scale for parameter\-efficient prompt tuning\.InProceedings of the 2021 conference on empirical methods in natural language processing,pp\. 3045–3059\.Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px2.p1.1)\.
- Lewiset al\.\(2020\)P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel,et al\.Retrieval\-augmented generation for knowledge\-intensive nlp tasks\.Advances in neural information processing systems33,pp\. 9459–9474\.Cited by:[§1](https://arxiv.org/html/2610.33511#S1.p2.1)\.
- Li and Liang \(2021\)X\. L\. Li and P\. LiangPrefix\-tuning: optimizing continuous prompts for generation\.InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\),pp\. 4582–4597\.Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px2.p1.1)\.
- Liet al\.\(2023\)X\. Li, D\. Lian, Z\. Lu, J\. Bai, Z\. Chen, and X\. WangGraphadapter: tuning vision\-language models with dual knowledge graph\.Advances in Neural Information Processing Systems36,pp\. 13448–13466\.Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Liet al\.\(2024\)Y\. Li, Z\. Li, P\. Wang, J\. Li, X\. Sun, H\. Cheng, and J\. X\. YuA survey of graph meets large language model: progress and future directions\.External Links:2311\.12399,[Link](https://arxiv.org/abs/2311.12399)Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Lianget al\.\(2026\)J\. Liang, P\. Zhou, W\. Zhou, W\. Qing, Q\. Zhao, Z\. Wang, Q\. Song, and X\. LiSentGraph: hierarchical sentence graph for multi\-hop retrieval\-augmented question answering\.External Links:2601\.03014,[Link](https://arxiv.org/abs/2601.03014)Cited by:[Appendix D](https://arxiv.org/html/2610.33511#A4.SS0.SSS0.Px7)\.
- Liuet al\.\(2024\)H\. Liu, J\. Feng, L\. Kong, N\. Liang, D\. Tao, Y\. Chen, and M\. ZhangOne for all: towards training one graph model for all classification tasks\.External Links:2310\.00149,[Link](https://arxiv.org/abs/2310.00149)Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Luoet al\.\(2025\)L\. Luo, J\. Ju, B\. Xiong, Y\. Li, G\. Haffari, and S\. PanChatrule: mining logical rules with large language models for knowledge graph reasoning\.InPacific\-Asia Conference on Knowledge Discovery and Data Mining,pp\. 314–325\.Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2610.33511#S1.p2.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Luoet al\.\(2024\)L\. Luo, Y\. Li, G\. Haffari, and S\. PanReasoning on graphs: faithful and interpretable large language model reasoning\.External Links:2310\.01061,[Link](https://arxiv.org/abs/2310.01061)Cited by:[Appendix D](https://arxiv.org/html/2610.33511#A4.SS0.SSS0.Px5)\.
- Maet al\.\(2023\)X\. Ma, Y\. Gong, P\. He, H\. Zhao, and N\. DuanQuery rewriting in retrieval\-augmented large language models\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 5303–5315\.Cited by:[§6](https://arxiv.org/html/2610.33511#S6.p1.1)\.
- Mavromatis and Karypis \(2024\)C\. Mavromatis and G\. KarypisGNN\-rag: graph neural retrieval for large language model reasoning\.External Links:2405\.20139,[Link](https://arxiv.org/abs/2405.20139)Cited by:[Appendix D](https://arxiv.org/html/2610.33511#A4.SS0.SSS0.Px3),[§1](https://arxiv.org/html/2610.33511#S1.p4.1)\.
- Perozziet al\.\(2024\)B\. Perozzi, B\. Fatemi, D\. Zelle, A\. Tsitsulin, M\. Kazemi, R\. Al\-Rfou, and J\. HalcrowLet your graph do the talking: encoding structured data for llms\.External Links:2402\.05862,[Link](https://arxiv.org/abs/2402.05862)Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2610.33511#S3.T1.1.8.1),[§4\.1](https://arxiv.org/html/2610.33511#S4.SS1.p2.1)\.
- Qianet al\.\(2023\)C\. Qian, H\. Tang, Z\. Yang, H\. Liang, and Y\. LiuCan large language models empower molecular property prediction?\.External Links:2307\.07443,[Link](https://arxiv.org/abs/2307.07443)Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Reimers and Gurevych \(2019\)N\. Reimers and I\. GurevychSentence\-bert: sentence embeddings using siamese bert\-networks\.InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing \(EMNLP\-IJCNLP\),pp\. 3982–3992\.Cited by:[§2\.2](https://arxiv.org/html/2610.33511#S2.SS2.p2.1)\.
- Shiet al\.\(2024\)W\. Shi, S\. Min, M\. Yasunaga, M\. Seo, R\. James, M\. Lewis, L\. Zettlemoyer, and W\. YihReplug: retrieval\-augmented black\-box language models\.InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),pp\. 8371–8384\.Cited by:[§1](https://arxiv.org/html/2610.33511#S1.p2.1)\.
- Sunet al\.\(2024\)F\. Sun, T\. Zheng, A\. Kolekar, R\. Patki, H\. Khazaei, X\. Guo, Z\. Cai, D\. C\. Liu, R\. Li, Y\. Huang, D\. Everaert, H\. Lu, and G\. P\. M\. ChengA product\-aware query auto\-completion framework for e\-commerce search via retrieval\-augmented generation method\.InIR\-RAG@SIGIR,External Links:[Link](https://api.semanticscholar.org/CorpusID:271122382)Cited by:[§6](https://arxiv.org/html/2610.33511#S6.p1.1)\.
- Sunet al\.\(2025\)S\. Sun, Y\. Ren, J\. Chen, and C\. MaLarge language models as topological structure enhancers for text\-attributed graphs\.External Links:2311\.14324,[Link](https://arxiv.org/abs/2311.14324)Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Talmor and Berant \(2018\)A\. Talmor and J\. BerantThe web as a knowledge\-base for answering complex questions\.InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 \(Long Papers\),pp\. 641–651\.Cited by:[§A\.1](https://arxiv.org/html/2610.33511#A1.SS1.p1.1),[§4\.1](https://arxiv.org/html/2610.33511#S4.SS1.p1.1)\.
- Taunket al\.\(2023\)D\. Taunk, L\. Khanna, S\. V\. P\. K\. Kandru, V\. Varma, C\. Sharma, and M\. TapaswiGrapeqa: graph augmentation and pruning to enhance question\-answering\.InCompanion Proceedings of the ACM Web Conference 2023,pp\. 1138–1144\.Cited by:[§1](https://arxiv.org/html/2610.33511#S1.p4.1)\.
- Thoppilanet al\.\(2022\)R\. Thoppilan, D\. D\. Freitas, J\. Hall, N\. Shazeer, A\. Kulshreshtha, H\. Cheng, A\. Jin, T\. Bos, L\. Baker, Y\. Du, Y\. Li, H\. Lee, H\. S\. Zheng, A\. Ghafouri, M\. Menegali, Y\. Huang, M\. Krikun, D\. Lepikhin, J\. Qin, D\. Chen, Y\. Xu, Z\. Chen, A\. Roberts, M\. Bosma, V\. Zhao, Y\. Zhou, C\. Chang, I\. Krivokon, W\. Rusch, M\. Pickett, P\. Srinivasan, L\. Man, K\. Meier\-Hellstern, M\. R\. Morris, T\. Doshi, R\. D\. Santos, T\. Duke, J\. Soraker, B\. Zevenbergen, V\. Prabhakaran, M\. Diaz, B\. Hutchinson, K\. Olson, A\. Molina, E\. Hoffman\-John, J\. Lee, L\. Aroyo, R\. Rajakumar, A\. Butryna, M\. Lamm, V\. Kuzmina, J\. Fenton, A\. Cohen, R\. Bernstein, R\. Kurzweil, B\. Aguera\-Arcas, C\. Cui, M\. Croak, E\. Chi, and Q\. LeLaMDA: language models for dialog applications\.External Links:2201\.08239,[Link](https://arxiv.org/abs/2201.08239)Cited by:[§1](https://arxiv.org/html/2610.33511#S1.p1.1)\.
- Tianet al\.\(2024\)Y\. Tian, H\. Song, Z\. Wang, H\. Wang, Z\. Hu, F\. Wang, N\. V\. Chawla, and P\. XuGraph neural prompting with large language models\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.38,pp\. 19080–19088\.Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px2.p1.1)\.
- Touvronet al\.\(2023\)H\. Touvron, L\. Martin, K\. Stone, P\. Albert, A\. Almahairi, Y\. Babaei, N\. Bashlykov, S\. Batra, P\. Bhargava, S\. Bhosale, D\. Bikel, L\. Blecher, C\. C\. Ferrer, M\. Chen, G\. Cucurull, D\. Esiobu, J\. Fernandes, J\. Fu, W\. Fu, B\. Fuller, C\. Gao, V\. Goswami, N\. Goyal, A\. Hartshorn, S\. Hosseini, R\. Hou, H\. Inan, M\. Kardas, V\. Kerkez, M\. Khabsa, I\. Kloumann, A\. Korenev, P\. S\. Koura, M\. Lachaux, T\. Lavril, J\. Lee, D\. Liskovich, Y\. Lu, Y\. Mao, X\. Martinet, T\. Mihaylov, P\. Mishra, I\. Molybog, Y\. Nie, A\. Poulton, J\. Reizenstein, R\. Rungta, K\. Saladi, A\. Schelten, R\. Silva, E\. M\. Smith, R\. Subramanian, X\. E\. Tan, B\. Tang, R\. Taylor, A\. Williams, J\. X\. Kuan, P\. Xu, Z\. Yan, I\. Zarov, Y\. Zhang, A\. Fan, M\. Kambadur, S\. Narang, A\. Rodriguez, R\. Stojnic, S\. Edunov, and T\. ScialomLlama 2: open foundation and fine\-tuned chat models\.External Links:2307\.09288,[Link](https://arxiv.org/abs/2307.09288)Cited by:[§A\.1](https://arxiv.org/html/2610.33511#A1.SS1.p1.1),[§4\.1](https://arxiv.org/html/2610.33511#S4.SS1.p3.1)\.
- Wanget al\.\(2023a\)H\. Wang, S\. Feng, T\. He, Z\. Tan, X\. Han, and Y\. TsvetkovCan language models solve graph problems in natural language?\.Advances in Neural Information Processing Systems36,pp\. 30840–30861\.Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[Table 1](https://arxiv.org/html/2610.33511#S3.T1.1.4.1),[§4\.1](https://arxiv.org/html/2610.33511#S4.SS1.p2.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Wanget al\.\(2023b\)K\. Wang, F\. Duan, S\. Wang, P\. Li, Y\. Xian, C\. Yin, W\. Rong, and Z\. XiongKnowledge\-driven cot: exploring faithful reasoning in llms for knowledge\-intensive question answering\.External Links:2308\.13259,[Link](https://arxiv.org/abs/2308.13259)Cited by:[§1](https://arxiv.org/html/2610.33511#S1.p1.1)\.
- Wuet al\.\(2024\)S\. Wu, H\. Fei, L\. Qu, W\. Ji, and T\. ChuaNext\-gpt: any\-to\-any multimodal llm\.InForty\-first International Conference on Machine Learning,Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px2.p1.1)\.
- Yanet al\.\(2025\)Y\. Yan, A\. Kolekar, S\. Genc, W\. Xu, E\. W\. Huang, A\. Srinivasan, M\. Jain, Q\. He, and H\. TongTo answer or not to answer \(taona\): a robust textual graph understanding and question answering approach\.InFindings of the Association for Computational Linguistics: EMNLP 2025,pp\. 6360–6376\.Cited by:[Table 1](https://arxiv.org/html/2610.33511#S3.T1.1.11.1),[§4\.1](https://arxiv.org/html/2610.33511#S4.SS1.p2.1)\.
- Yasunagaet al\.\(2021\)M\. Yasunaga, H\. Ren, A\. Bosselut, P\. Liang, and J\. LeskovecQA\-gnn: reasoning with language models and knowledge graphs for question answering\.InProceedings of the 2021 conference of the North American chapter of the association for computational linguistics: human language technologies,pp\. 535–546\.Cited by:[Appendix D](https://arxiv.org/html/2610.33511#A4.SS0.SSS0.Px2),[§1](https://arxiv.org/html/2610.33511#S1.p4.1)\.
- Yoonet al\.\(2023\)M\. Yoon, J\. Y\. Koh, B\. Hooi, and R\. SalakhutdinovMultimodal graph learning for generative tasks\.External Links:2310\.07478,[Link](https://arxiv.org/abs/2310.07478)Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Yuet al\.\(2023\)W\. Yu, D\. Iter, S\. Wang, Y\. Xu, M\. Ju, S\. Sanyal, C\. Zhu, M\. Zeng, and M\. JiangGenerate rather than retrieve: large language models are strong context generators\.External Links:2209\.10063,[Link](https://arxiv.org/abs/2209.10063)Cited by:[§6](https://arxiv.org/html/2610.33511#S6.p1.1)\.
- Zhanget al\.\(2024\)R\. Zhang, J\. Han, C\. Liu, P\. Gao, A\. Zhou, X\. Hu, S\. Yan, P\. Lu, H\. Li, and Y\. QiaoLLaMA\-adapter: efficient fine\-tuning of language models with zero\-init attention\.External Links:2303\.16199,[Link](https://arxiv.org/abs/2303.16199)Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px2.p1.1)\.
- Zhaoet al\.\(2023\)J\. Zhao, L\. Zhuo, Y\. Shen, M\. Qu, K\. Liu, M\. Bronstein, Z\. Zhu, and J\. TangGraphText: graph reasoning in text space\.External Links:2310\.01089,[Link](https://arxiv.org/abs/2310.01089)Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px1.p1.1),[§6](https://arxiv.org/html/2610.33511#S6.p2.1)\.
- Zhaoet al\.\(2026\)W\. X\. Zhao, K\. Zhou, J\. Li, T\. Tang, X\. Wang, Y\. Hou, Y\. Min, B\. Zhang, J\. Zhang, Z\. Dong, Y\. Du, C\. Yang, Y\. Chen, Z\. Chen, J\. Jiang, R\. Ren, Y\. Li, X\. Tang, Z\. Liu, P\. Liu, J\. Nie, and J\. WenA survey of large language models\.External Links:2303\.18223,[Link](https://arxiv.org/abs/2303.18223)Cited by:[§1](https://arxiv.org/html/2610.33511#S1.p1.1)\.
- Zhenget al\.\(2024\)H\. S\. Zheng, S\. Mishra, X\. Chen, H\. Cheng, E\. H\. Chi, Q\. V\. Le, and D\. ZhouTake a step back: evoking reasoning via abstraction in large language models\.External Links:2310\.06117,[Link](https://arxiv.org/abs/2310.06117)Cited by:[§6](https://arxiv.org/html/2610.33511#S6.p1.1)\.
- Zhuet al\.\(2023\)D\. Zhu, J\. Chen, X\. Shen, X\. Li, and M\. ElhoseinyMiniGPT\-4: enhancing vision\-language understanding with advanced large language models\.External Links:2304\.10592,[Link](https://arxiv.org/abs/2304.10592)Cited by:[Appendix H](https://arxiv.org/html/2610.33511#A8.SS0.SSS0.Px2.p1.1)\.

## Appendix AAdditional Dataset Details

### A\.1Datasets

Datasets and Metrics\.To evaluate the effectiveness of our method on textual QA tasks, we conduct experiments on three textual graph question answering datasets, ExplaGraphs and WebQSP from the GraphQA benchmark\([He et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib10)\), and CWQ\([Talmor and Berant 2018](https://arxiv.org/html/2610.33511#bib.bib46)\)\. Specifically, \(1\) ExplaGraphs is developed for generative commonsense reasoning, with a specific focus on the construction of explanation graphs for stance prediction in debates\. By providing detailed and unambiguous commonsense\-augmented graphs, it enables an evaluation of whether arguments either support or refute a specific belief\. ExplaGraphs’s primary task is to determine whether the arguments are supportive or contradictory\. Its node denotes commonsense concepts and the edge denotes Commonsense relations\. In this dataset, the gold answer is a binary label \(i\.e\., ‘support’ or ‘counter’\), which does not provide explicit supervision for the retriever\. Therefore, we leverage the LLaMA2\-Chat\-7B model\([Touvron et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib23)\)to generate pseudo gold reasoning node sets, and treat these annotated nodes as pseudo\-label supervision𝒱⋆\\mathcal\{V\}^\{\\star\}for training the learnable retriever\. Due to the small graph size of the graphs in ExplaGraphs, we directly use the entire graph as the candidate context for retrieval\. Following GraphQA\([He et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib10)\), we evaluate the end task withAccuracy\. \(2\) WebQSP is a large\-scale multi\-hop knowledge graph QA dataset that consists of 4,737 questions\. The primary task is to answer questions that necessitate multi\-hop reasoning\. It employs a subset of Freebase\([Bollacker et al\. 2008](https://arxiv.org/html/2610.33511#bib.bib24)\), with a specific focus on facts within 2 hops of the entities identified in the questions\. The nodes denote Entities in Freebase and the edges denote Relations in Freebase\. The gold answers are lists of Freebase entities, which naturally correspond to node lists and can directly serve as𝒱⋆\\mathcal\{V\}^\{\\star\}without requiring additional annotation\. For metrics, we report Hit@1, counting a prediction as correct if the generated answer matches any of the answers in the gold answer list\. \(3\) CWQ \(ComplexWebQuestions\) is a compositional KGQA dataset constructed from WebQSP by generating more complex queries and rewriting them into natural language\([Talmor and Berant 2018](https://arxiv.org/html/2610.33511#bib.bib46)\)\. It contains 34,689 questions\. Compared with WebQSP, CWQ contains more complex questions and longer reasoning chains\. Like WebQSP, CWQ is grounded in Freebase, where nodes denote entities and edges denote relations\. The gold answers are lists of Freebase entities and can therefore directly serve as𝒱⋆\\mathcal\{V\}^\{\\star\}\. We evaluate CWQ withHit@1\. Dataset statistics are summarized in Table[5](https://arxiv.org/html/2610.33511#A1.T5)\. See Appendix[A\.2](https://arxiv.org/html/2610.33511#A1.SS2)for Pseudo\-Gold Nodes Annotation for ExplaGraphs\.

Table 5:Statistics of datasets\.
### A\.2Pseudo\-Gold Nodes Annotation for ExplaGraphs

ExplaGraphs provides gold supervision only as a stance label \(support/counter\), rather than node\-level evidence annotations\. To make our learnable retriever pseudo label supervision applicable to this dataset, we introduce a lightweight preprocessing step on the training and validation splits: we useLlama\-2\-7b\-chatto annotate a set of pseudo\-gold reasoning nodes for each example\. Concretely, we provide the model with \(i\) the full textual explanation graph, \(ii\) the two arguments, and \(iii\) the gold stance label, and prompt it to select nodes from the graph that form the critical reasoning path for deciding whether the arguments support or counter each other\. The selected nodes are then used as𝒱⋆\\mathcal\{V\}^\{\\star\}in our weighted pseudo\-label objective for training the retriever\. This annotation is used only to train our retriever: the generator is still fine\-tuned with the original task supervision \(predicting support vs\. counter\), and at inference time the pipeline does not require access to any node annotations\.

We further conduct two sanity checks on these pseudo\-gold node annotations\. First, we manually inspected LLM annotations and found these LLM\-selected nodes to be broadly consistent with human judgment\. For example, for the argument pair“Women should be allowed in military combat\.”and“Females as soldiers should be acceptable\.”, whose gold label issupport, the annotator selects critical nodes includingstrength,military combat, andfemalesas “pseudo\-gold” nodes, which we found to be reasonable and aligned with the human interpretation of key evidence needed to justify the support relation\.

Second, we perform a cross\-annotator consistency check usingGPT\-4o\-mini\. Specifically, we re\-annotate the ExplaGraphs dataset withGPT\-4o\-miniusing the same annotation prompt as used forLlama\-2\-7b\-chat, and compare the resulting node sets\. We report two simple agreement statistics\. \(1\)Any\-overlap rate: in80\.44%80\.44\\%of graph instances, the two annotators share at least one selected node\. \(2\)Reference coverage: when treating the originalLlama\-2\-7b\-chatannotations as the reference, theGPT\-4o\-miniannotations cover83\.75%83\.75\\%of those nodes on average\. These results suggest that the pseudo\-gold node annotations are not arbitrary: a second annotator recovers most of their core evidence, while manual inspection further indicates that the selected nodes are reasonable for stance prediction\. These sanity checks suggest that theLlama\-2\-7b\-chatpseudo\-gold annotations are reasonably consistent and provide useful supervision for training the ExplaGraphs retriever\.

## Appendix BImplementation Details

Preprocessing: Candidate\-subgraph construction\.For WebQSP and CWQ, we construct the PCST candidate subgraph withkv=ke=10k\_\{v\}\{=\}k\_\{e\}\{=\}10\. Because the graphs in ExplaGraphs are small, we directly use the full graph as the candidate context\.

Retriever fine\-tuning\.We fine\-tune the retriever with LoRA under our token\-weighted cross\-entropy objective\. Given a question and a candidate subgraph, we construct weighted pseudo labels over candidate nodes: nodes overlapping with gold answer nodes are treated as strong positives with weight1\.01\.0, while the remaining candidate nodes are treated as weak positives\. The weak\-positive weight is set toλ=0\.001\\lambda=0\.001for WebQSP andλ=0\.01\\lambda=0\.01for ExplaGraphs and CWQ\. The training target is a single\-line, pipe\-separated node list, formed by concatenating strong positives and weak positives using the delimiter “\|” \(strong positives first\)\. We deduplicate node names in both retriever outputs and the contexts later used for generator training\.

We use LoRA with rankr=8r\{=\}8, scaling factorα=16\\alpha\{=\}16, and dropout0\.050\.05\. Retriever training uses 3 epochs, learning rate5×10−55\\times 10^\{\-5\}, warmup ratio0\.030\.03, per\-device batch size 1 with gradient accumulation 8\.

Generator fine\-tuning\.We fine\-tune the generator with LoRA using standard supervised fine\-tuning on retrieved textual graph contexts and gold targets\. We use the same LoRA configuration \(r=8r\{=\}8,α=16\\alpha\{=\}16, dropout0\.050\.05\) for generator fine\-tuning\.

Computing infrastructure\.The experiments were run on a server equipped with eight NVIDIA A100\-SXM4 GPUs, each with 80 GB of memory\. Each experimental run used a single GPU\. The server contained two Intel Xeon Gold 6448Y processors \(64 physical cores and 128 hardware threads in total\) and 1 TiB of system memory, and ran Ubuntu 22\.04\.5 LTS\. We used Python 3\.10\.19, NVIDIA driver 535\.183\.06, PyTorch 2\.5\.1 with CUDA runtime 12\.1 and cuDNN 9\.1, Transformers 4\.57\.1, PEFT 0\.17\.1, Accelerate 1\.11\.0, and NumPy 2\.1\.2\.

## Appendix CSensitivity Analysis and Ablation of the Weighted Objective

We provide two complementary analyses on WebQSP\. First, we study the sensitivity of our method to the weak\-positive weightλ\\lambdaand the PCST candidate subgraph size parameterskvk\_\{v\}andkek\_\{e\}\. In theλ\\lambdasweep, the settingλ=1\\lambda=1also serves as an equal\-weight ablation, since strong and weak positives receive the same training weight\.

#### Sensitivity to the weak\-positive weightλ\\lambda\.

We varyλ\\lambdawhile fixing the PCST input tokv=ke=10k\_\{v\}=k\_\{e\}=10\. Table[6](https://arxiv.org/html/2610.33511#A3.T6)reports both downstream QA performance \(Hit@1\) and the average number of retrieved nodes on the test set, where a smaller number indicates better compactness\. Across all tested values ofλ\\lambda, downstream Hit@1 ranges from75\.55%75\.55\\%to76\.97%76\.97\\%, consistently exceeding the strongest baseline reported in the main paper, G\-Retriever w/ LoRA \(73\.79%73\.79\\%\)\. Our default setting,λ=0\.001\\lambda=0\.001, achieves the highest Hit@1 of76\.97%76\.97\\%while retaining a compact output of12\.7912\.79nodes on average\. At the same time, the average number of retrieved nodes increases consistently asλ\\lambdabecomes larger, which is expected since larger weak\-positive weights encourage the retriever to retain more weakly related candidate nodes\. Asλ\\lambdaincreases beyond 0\.001, the retrieved node sets become progressively larger without improving downstream performance\. In particular, the equal\-weight settingλ=1\\lambda=1reduces Hit@1 to75\.55%75\.55\\%and increases the average number of retrieved nodes from12\.7912\.79to25\.2525\.25\. These results support assigning weak positives a substantially smaller weight than strong positives, rather than weighting the two groups equally\.

Table 6:Sensitivity analysis on the weak\-positive weightλ\\lambdaon WebQSP, with PCST input fixed tokv=ke=10k\_\{v\}=k\_\{e\}=10\. We report Hit@1 and the average number of retrieved nodes on the test set; smaller retrieved node sets indicate better compactness\. Largerλ\\lambdavalues lead to less compact retrieval\.
#### Sensitivity to PCST candidate subgraph size\.

We next vary the PCST input sizekv=kek\_\{v\}=k\_\{e\}while fixingλ=0\.001\\lambda=0\.001\. As shown in Table[7](https://arxiv.org/html/2610.33511#A3.T7), downstream QA Hit@1 increases from74\.51%74\.51\\%to76\.97%76\.97\\%as the candidate\-subgraph size increases\. This behavior is expected because a larger PCST subgraph provides a higher\-recall candidate pool from which the learned retriever can select evidence\.

Table 7:Sensitivity analysis on the PCST candidate subgraph size on WebQSP, withλ\\lambdafixed to0\.0010\.001\.

## Appendix DDiscussion of Answer\-Aware or Answer\-Related Methods and Additional Recent Baselines

Table[8](https://arxiv.org/html/2610.33511#A4.T8)summarizes several key distinctions betweenTaLenTand recent related baselines\. Here, in this comparison,*LLM\-based end\-to\-end*refers to a unified retrieve\-then\-generate pipeline that pairs a learnable*LLM\-based retriever*for graph evidence selection with a fine\-tuned*LLM\-based generator*\(i\.e\., the retriever itself is made a trainable LLM, rather than only fine\-tuning the downstream generator\)\.*Coverage\-compactness*indicates whether the method explicitly aims to improve answer\-relevant evidence coverage while suppressing irrelevant noise\.*KGQA*indicates whether the method is directly applicable, as\-is, to open\-ended KGQA benchmarks such as WebQSP and CWQ\.*Label\-prediction GraphQA*indicates whether the method is directly applicable, as\-is, to graph QA tasks whose outputs are discrete labels rather than entity answers, such as ExplaGraphs\-style support/counter prediction\.

Table 8:Comparison ofTaLenTwith answer\-related and other recent graph retrieval methods discussed in Appendix[D](https://arxiv.org/html/2610.33511#A4)\.#### Answer\-aware or answer\-related methods\.

Some related methods do exploit answer\-aware or answer\-related signals, but they do so in task formulations that differ substantially from ours\.

#### QA\-GNN\([Yasunaga et al\. 2021](https://arxiv.org/html/2610.33511#bib.bib21)\)\.

QA\-GNN is designed for multiple\-choice QA rather than our open\-ended GraphQA setting\. Given a question and an answer choicea∈Ca\\in C, it forms a QA context\[q;a\]\[q;a\], links entities mentioned in the question and answer choice to a KG, and performs joint LM\+GNN reasoning over a constructed subgraph for that question\-choice pair\. Therefore, QA\-GNN is conditioned on*candidate answer choices*, rather than using dataset gold answers to supervise a retriever\. Its main focus is answer\-choice scoring in multiple\-choice QA, not training a retriever for open\-ended graph QA generation\. Accordingly, it does not satisfy our*LLM\-based end\-to\-end*criterion, does not explicitly optimize the*coverage\-compactness*trade\-off, and is not directly applicable as\-is to either open\-ended KGQA benchmarks such as WebQSP/CWQ or label\-prediction graph QA tasks such as ExplaGraphs\.

#### GNN\-RAG\([Mavromatis and Karypis 2024](https://arxiv.org/html/2610.33511#bib.bib19)\)\.

GNN\-RAG targets KGQA\. It uses training question\-answer pairs to supervise a GNN that scores answer candidates over a dense KG subgraph, and then retrieves the shortest paths connecting question entities to those candidates as context for a tuned LLM\. Thus, GNN\-RAG does use answer supervision during training, but in the form of answer\-candidate scoring and path retrieval for KGQA, rather than weighted pseudo\-label supervision for graph evidence selection\. In particular, it does not contain a trainable*LLM\-based retriever*, nor does it explicitly optimize the*coverage\-compactness*trade\-off\. It is directly applicable to KGQA, but not to label\-prediction graph QA tasks such as ExplaGraphs, whose outputs are discrete labels rather than KG entities\.

#### Other recent graph retrieval baselines\.

We also discuss several other recent graph retrieval methods that are useful points of comparison but differ substantially in task setting and retrieval target\.

#### RoG\([Luo et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib47)\)\.

RoG is a planning\-retrieval\-reasoning framework tailored to KGQA\. It first generates relation paths as plans, then retrieves valid reasoning paths from the KG and reasons over them with the same LLM\. Although RoG jointly trains its planning and reasoning modules, its retrieval step itself is plan\-guided KG path retrieval rather than a trainable*LLM\-based retriever*under our definition above\. Moreover, RoG is designed for KGQA where answers are expected to be KG entities reachable via KG paths\. Therefore, it does not satisfy our*LLM\-based end\-to\-end*criterion, does not explicitly optimize the*coverage\-compactness*trade\-off, and is not applicable as\-is to label\-prediction graph QA tasks such as ExplaGraphs\.

#### CG\-RAG\([Hu et al\. 2025](https://arxiv.org/html/2610.33511#bib.bib48)\)\.

CG\-RAG targets research question answering over citation graphs\. Its retrieval module ranks relevant chunks by combining sparse and dense retrieval signals with graph encoding over contextual citation graphs, and the retrieved chunk subgraphs are then used for generation\. This differs from our setting in both graph type and retrieval unit: CG\-RAG performs chunk\-level retrieval over citation graphs, rather than node\-level evidence selection over a given QA graph\. Accordingly, it neither provides an*LLM\-based end\-to\-end*retriever\-generator pipeline in our sense nor explicitly optimizes*coverage\-compactness*, and it is not directly applicable as\-is to either KGQA benchmarks such as WebQSP/CWQ or label\-prediction graph QA tasks such as ExplaGraphs\.

#### SentGraph\([Liang et al\. 2026](https://arxiv.org/html/2610.33511#bib.bib49)\)\.

SentGraph is a sentence\-level graph\-based RAG framework for multi\-hop QA over document collections\. It constructs hierarchical sentence graphs with explicit logical dependencies and performs graph\-guided sentence retrieval and path expansion\. In contrast, our setting assumes a given QA graph whose nodes and edges already define the reasoning space, such as Freebase entities/relations in WebQSP/CWQ or commonsense concept graphs in ExplaGraphs\. Thus, SentGraph differs from our setting in graph granularity, retrieval target, and benchmark formulation\. It also does not provide an*LLM\-based end\-to\-end*retriever\-generator pipeline in our sense, does not explicitly optimize*coverage\-compactness*, and is not directly applicable as\-is to either KGQA benchmarks such as WebQSP/CWQ or label\-prediction graph QA tasks such as ExplaGraphs\.

Overall, the distinctive feature ofTaLenTis not merely that it leverages answer\-related signals, but that it combines a unified*LLM\-based end\-to\-end*retriever\-generator pipeline with weighted pseudo\-label supervision, thereby addressing both the*architectural gap*and the*retrieval\-supervision gap*while explicitly improving the*coverage\-compactness*trade\-off\. Moreover, the framework applies to both open\-ended KGQA benchmarks \(e\.g\., WebQSP and CWQ\) and label\-prediction GraphQA tasks \(e\.g\., ExplaGraphs\)\.

## Appendix EConcrete Prompt Templates and Graph Serialization

In this section, we provide one concrete example of the prompts used inTaLenT\. We show the exact retriever prompt and the generator prompt, and clarify how retrieved nodes and associated edges are serialized for answer generation\.

### E\.1Retriever Prompt

The retriever is prompted with the question together with the initial PCST candidate subgraph𝒮\\mathcal\{S\}\. Its task is to output nodes that are most useful for answering the question, and to output them as a single pipe\-separated line\. Below we show one concrete retriever prompt example\.

1<s\>\[INST\]<<SYS\>\>

2YouareanentityretrieverforKGQA\.

3YourjobistoSELECTCANDIDATEENTITIESFROMTHESUBGRAPHthatwouldhelpananswerersolvethequestion\.

4CRITICALCONSTRAINTS:

5\-Returnlowercasenamesonly\.

6\-Selectthemostrelevantnodesfromthesubgraphforansweringthequestion\.

7

8RETURNFORMAT\(MANDATORY\):

9YourENTIREreplymustbeEXACTLYONESINGLELINElike:

10name1\|name2\|name3

11Ifyououtputanythingotherthanthesinglepipe\-separatedline,theanswerisinvalid\.

12

13Hardrules:

14\-OutputEXACTLYoneline;nogreetings,noexplanations,noextrawords\.

15\-ThelinemustcontainONLYlowercaseentitynamesseparatedby’\|’\(nolabelslike’Output:’\)\.

16\-Keepspacesinsidenames;DONOTaddspacesaround’\|’\.

17\-DONOToutputplaceholderslikeA/B/C/unknown\.

18

19<</SYS\>\>

20

21Question:

22wheredidkennedy’sinauguraladdresstakeplace

23

24Subgraph\(CSVbetween<DESC\>tags\):

25<DESC\>

26node\_id,node\_attr

2724,johnf\.kennedy

28282,coldwar

29286,speech

30287,m\.05rrkys

31295,anti\-apartheid

32352,m\.04tnzwc

33397,m\.05r7b2w

34428,johnf\.kennedy1961presidentialinauguration

35429,inauguraladdressofjohnf\.kennedy

36475,m\.0b9zgjq

37661,berlinwall

38737,johnfitzgeraldkennedymemorial

39771,johnfitzgeraldkennedyinauguralpoemreadbyrobertleefrost

40967,unitedstatespresidentialinauguration

411008,johnf\.kennedyspeech

421119,apolloprogram

431282,thekennedypresidentialpressconferences

441335,1957nationalbookawards

451376,m\.0ypz1mv

461458,1957pulitzerprize

471701,jfkinauguration

48

49src,edge\_attr,dst

5024,event\.public\_speaker\.speeches\_or\_presentations,397

51352,award\.award\_honor\.award\_winner,24

5224,award\.award\_nominee\.award\_nominations,475

53428,common\.topic\.image,1701

5424,book\.author\.works\_written,1282

5524,music\.artist\.track,1008

56967,time\.recurring\_event\.instances,428

5724,base\.kwebbase\.kwtopic\.connections\_from,771

58737,symbols\.namesake\.named\_after,24

591376,event\.speech\_or\_presentation\.type\_or\_format\_of\_presentation,286

60287,event\.speech\_or\_presentation\.type\_or\_format\_of\_presentation,286

61397,event\.speech\_or\_presentation\.type\_or\_format\_of\_presentation,286

621119,event\.speech\_topic\.speeches\_or\_presentations\_on\_this\_topic,1376

63295,event\.speech\_topic\.speeches\_or\_presentations\_on\_this\_topic,287

64661,event\.speech\_topic\.speeches\_or\_presentations\_on\_this\_topic,397

65282,event\.speech\_topic\.speeches\_or\_presentations\_on\_this\_topic,397

66428,base\.inaugurations\.inauguration\.inaugural\_address,429

67397,event\.speech\_or\_presentation\.speech\_topic,282

68397,event\.speech\_or\_presentation\.speech\_topic,661

6924,base\.inaugurations\.inauguration\_speaker\.inauguration,428

70475,award\.award\_nomination\.ceremony,1335

71352,award\.award\_honor\.ceremony,1458

72428,base\.inaugurations\.inauguration\.president,24

73287,event\.speech\_or\_presentation\.speech\_topic,295

741376,event\.speech\_or\_presentation\.speech\_topic,1119

75</DESC\>

76

77NowoutputEXACTLYONESINGLELOWERCASELINEwith’\|’separators\.Nothingelse\.\[/INST\]

### E\.2Generator Prompt and Graph Serialization

For answer generation, we serialize the retrieved evidence in the format: nodes are first listed as a table with headernode\_id,node\_attr, followed by edges as a table with headersrc,edge\_attr,dst\.

The serialized node set contains two parts\. First, nodes with regular numeric identifiers are retrieved nodes that appear in the initial PCST candidate subgraph𝒮\\mathcal\{S\}\. Second, nodes whose identifiers start with the prefixna\(e\.g\.,na1,na2\) are retrieved nodes that do not appear in the initial𝒮\\mathcal\{S\}\. These additional nodes are retained in the generator context as isolated nodes\. After listing the nodes, we include the associated edgesℰ^\\hat\{\\mathcal\{E\}\}that appear in𝒮\\mathcal\{S\}\. Therefore, the generator context consists of the retrieved nodes𝒱^\\hat\{\\mathcal\{V\}\}together with their associated edgesℰ^\\hat\{\\mathcal\{E\}\}, and it is not required to form one connected subgraph\. Below we show one concrete generator prompt example\.

1

2<s\>\[INST\]<<SYS\>\>

3YouareaKGQAanswergenerator\.

4PRIMARYGOAL:outputtheFINALANSWERentityname\(s\)tothequestion\.

5UsetheprovidedsubgraphasSTRONGHINTS\.Ifsomecorrectanswersarenotinsubgraphbutanswersthequestion,youmayalsoincludethem\.

6Returnallnamesinlowercase\.

7

8RETURNFORMAT\(MANDATORY\):

9YourENTIREreplymustbeEXACTLYONESINGLELINElike:

10name1\|name2\|name3

11Ifyououtputanythingotherthanthesinglepipe\-separatedline,theanswerisinvalid\.

12

13Guidelines:

14\-OutputONLYanswers;excludedistractorsandthesubjectitselfunlessthesubjectIStheanswer\.

15\-Ifmultipleentitiescorrectlyanswerthequestion,includethemallseparatedby’\|’\.Otherwise,outputasinglename\.

16\-Nogreetings,noexplanations,nolabels\(e\.g\.,’Answer:’\)\.

17\-Keepspacesinsidenames;DONOTaddspacesaround’\|’\.

18\-DONOToutputplaceholders\(A/B/C/unknown\)\.

19

20<</SYS\>\>

21

22Question:

23wheredidkennedy’sinauguraladdresstakeplace

24

25Subgraph\(CSVbetween<DESC\>tags\):

26<DESC\>

27node\_id,node\_attr

2824,johnf\.kennedy

29429,inauguraladdressofjohnf\.kennedy

301008,johnf\.kennedyspeech

311282,thekennedypresidentialpressconferences

321701,jfkinauguration

33na1,washingtond\.c\.

34na2,capitolbuilding

35na3,unitedstates

36na4,1961presidentialinauguration

37na5,unitedstatescapitolbuilding

38

39src,edge\_attr,dst

40428,common\.topic\.image,1701

4124,book\.author\.works\_written,1282

4224,music\.artist\.track,1008

43428,base\.inaugurations\.inauguration\.inaugural\_address,429

4424,base\.inaugurations\.inauguration\_speaker\.inauguration,428

45428,base\.inaugurations\.inauguration\.president,24

46</DESC\>

47

48NowoutputONLYthefinalanswerentityname\(s\)asEXACTLYONESINGLELOWERCASELINEwith’\|’separators\.Nothingelse\.\[/INST\]

Figure 4:WebQSP test\-set retrieval results with Llama\-3\-8B \(macro\-averaged\)\. We measure how well PCST subgraphs andTaLenT’s learned retriever output𝒱^\\hat\{\\mathcal\{V\}\}cover the gold answer set𝒱⋆\\mathcal\{V\}^\{\\star\}\. We compare: \(1\) the PCST candidate subgraph withkv=ke=10k\_\{v\}\{=\}k\_\{e\}\{=\}10\(the retriever input\), \(2\) a smaller PCST subgraph withkv=ke=5k\_\{v\}\{=\}k\_\{e\}\{=\}5\(a size\-matched reference\), and \(3\) the node set𝒱^\\hat\{\\mathcal\{V\}\}selected byTaLenT’s learned retriever\. Our retriever maintains comparable gold\-answer coverage with substantially fewer nodes than the original PCST candidate subgraph \(kv=ke=10k\_\{v\}\{=\}k\_\{e\}\{=\}10\), while achieving higher coverage than the size\-matched PCST subgraph \(kv=ke=5k\_\{v\}\{=\}k\_\{e\}\{=\}5\)\.

## Appendix FGroundedness and Hallucination Analysis of Out\-of\-PCST Retrieval

Retrieving nodes beyond the initial PCST candidate subgraph can improve coverage, but it also raises a natural concern about groundedness: such nodes may introduce ungrounded evidence and potentially contribute to hallucinated answers\. To assess this effect, we conduct an additional analysis on the WebQSP test set\.

We use a practical*hit\-aware hallucination proxy*, since groundedness is ultimately relevant insofar as it affects the final answer\. For each test instance, let𝒮\\mathcal\{S\}denote the initial PCST candidate subgraph with node set𝒱𝒮\\mathcal\{V\}\_\{\\mathcal\{S\}\}, let𝒱^\\hat\{\\mathcal\{V\}\}denote the retrieved node set output by the retriever, and let𝒱⋆\\mathcal\{V\}^\{\\star\}denote the gold answer set\. We define the hallucinated retrieved nodes as

ℋ=\{∅,if Hit@1=1,\{v∈𝒱^∖𝒱𝒮∣v∉𝒱⋆\},if Hit@1=0\.\\mathcal\{H\}=\\begin\{cases\}\\emptyset,&\\text\{if Hit@1\}=1,\\\\ \\\{v\\in\\hat\{\\mathcal\{V\}\}\\setminus\\mathcal\{V\}\_\{\\mathcal\{S\}\}\\mid v\\notin\\mathcal\{V\}^\{\\star\}\\\},&\\text\{if Hit@1\}=0\.\\end\{cases\}\(16\)That is, if the final prediction is correct, we do not count retrieved nodes as hallucinated under this proxy\. If the final prediction is incorrect, we count as hallucinated those retrieved nodes that are outside the PCST candidate subgraph and do not match any gold answer\.

We then compute the overall hallucination rate over the test set as

Hallucination Rate=∑\|ℋ\|∑\|𝒱^\|,\\text\{Hallucination Rate\}=\\frac\{\\sum\|\\mathcal\{H\}\|\}\{\\sum\|\\hat\{\\mathcal\{V\}\}\|\},\(17\)where the sums aggregate over all test instances\.

On the WebQSP test set, the resulting hallucination rate is4\.97%4\.97\\%\. This suggests that although the trained retriever can retrieve nodes beyond PCST, such retrieval introduces only a small fraction of ungrounded nodes overall\. Combined with the results in Section 4\.4 of the main paper, this indicates that out\-of\-PCST retrieval provides useful complementary coverage while maintaining a low level of hallucinated evidence\.

## Appendix GBackbone Ablation with Llama\-3\-8B

We further evaluateTaLenTwith Llama\-3\-8B on WebQSP, comparing against G\-Retriever w/ LoRA, the strongest baseline in Table[1](https://arxiv.org/html/2610.33511#S3.T1)\. As shown in Table[9](https://arxiv.org/html/2610.33511#A7.T9),TaLenTachieves 83\.05% Hit@1, compared with 79\.36% for G\-Retriever w/ LoRA, suggesting that our framework remains effective with a stronger LLM backbone\. We useλ=0\.001\\lambda=0\.001for the learned retriever\.

Table 9:WebQSP downstream QA results with Llama\-3\-8B, comparingTaLenTagainst the strongest baseline G\-Retriever w/ LoRA in Table[1](https://arxiv.org/html/2610.33511#S3.T1)\(Hit@1, %\)\.We further examine the retrieval coverage\-compactness tradeoff under Llama\-3\-8B, as shown in Figure[4](https://arxiv.org/html/2610.33511#A5.F4)\.

Compactness: comparable coverage with fewer nodes\.Compared with the initial PCST candidate subgraph \(kv=ke=10k\_\{v\}\{=\}k\_\{e\}\{=\}10\),TaLenT’s retriever produces a more compact node set while maintaining comparable gold\-answer coverage\. Specifically, the average number of nodes decreases from 31\.75 to 19\.60, while Hit@1 increases from 78\.93% to 81\.45% and Recall from 69\.76% to 70\.40%\. Precision also improves from 9\.26% to 22\.16%, and F1 from 12\.22% to 23\.54%\.

Coverage: better gold\-answer coverage under a comparable node budget\.We further compare against a smaller PCST subgraph withkv=ke=5k\_\{v\}\{=\}k\_\{e\}\{=\}5, which contains 19\.35 nodes on average, close to the 19\.60 nodes retrieved byTaLenT\. Under this comparable node budget,TaLenTimproves Hit@1 from 69\.23% to 81\.45% and Recall from 58\.72% to 70\.40%\. These results exhibit a similar coverage–compactness trend to the main experiments\.

Table 10:Breakdown of gold\-answer coverage contributed by different parts of the learned retriever output on the WebQSP test set with Llama\-3\-8B \(λ=0\.001\\lambda=0\.001\)\. All metrics are macro\-averaged over examples and reported as percentages\.Retrieval beyond PCST\.Table[10](https://arxiv.org/html/2610.33511#A7.T10)further breaks down the gold\-answer coverage of the learned retriever output\. Nodes outside PCST alone achieve 8\.29% Hit@1 and 4\.28% Recall\. Incorporating these nodes increases Hit@1 from 76\.90% to 81\.45% and Recall from 67\.35% to 70\.40%\. This is consistent with the main\-paper observations that the learned retriever can complement the PCST candidate subgraph with additional gold\-answer coverage\.

Overall, these results suggest thatTaLenTmaintains a favorable retrieval coverage–compactness tradeoff and strong downstream QA performance with Llama\-3\-8B\.

## Appendix HMore Related Works

#### Graphs and Large Language Models

A growing body of work studies how to combine graph\-structured information with LLMs\([Li et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib29);[Wang et al\. 2023a](https://arxiv.org/html/2610.33511#bib.bib2)\)\. Prior efforts span general graph modeling\([Liu et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib31);[Lei et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib30)\), multimodal architectures that incorporate graph\-like structures\([Yoon et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib32);[Li et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib33)\)\. Representative applications include fundamental graph reasoning\([Zhao et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib35)\), node prediction\([Sun et al\. 2025](https://arxiv.org/html/2610.33511#bib.bib36);[Chen et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib37)\), graph\-level prediction\([Qian et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib38)\)\. A related line of work discusses GraphRAG / GraphQA, where the system selects graph evidence and conditions an LLM on graph context to answer questions\([IV et al\. 2019](https://arxiv.org/html/2610.33511#bib.bib9);[Luo et al\. 2025](https://arxiv.org/html/2610.33511#bib.bib11)\)\.

#### Parameter\-Efficient Fine\-Tuning

Parameter\-efficient fine\-tuning \(PEFT\) methods enable adapting large pretrained models with limited trainable parameters\. Representative techniques include prompt tuning\([Lester et al\. 2021](https://arxiv.org/html/2610.33511#bib.bib40)\), prefix tuning\([Li and Liang 2021](https://arxiv.org/html/2610.33511#bib.bib41)\), and low\-rank adaptation such as LoRA\([Hu et al\. 2022](https://arxiv.org/html/2610.33511#bib.bib42)\)\. Related adapter\-style methods \(e\.g\., LLaMA\-adapter\([Zhang et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib43)\)\) have also supported rapid adaptation and inspired multimodal instruction\-following systems such as MiniGPT\-4\([Zhu et al\. 2023](https://arxiv.org/html/2610.33511#bib.bib44)\)and NExT\-Chat\([Wu et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib45)\)\. PEFT has recently been explored in graph\+LLM settings as well, including GraphLLM\([Chai et al\. 2025](https://arxiv.org/html/2610.33511#bib.bib34)\)and GraphToken\([Perozzi et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib4)\)for graph reasoning, and GNP\([Tian et al\. 2024](https://arxiv.org/html/2610.33511#bib.bib39)\)for knowledge\-graph QA\.