Paper deep dive
DeLIVeR: Decomposed Learning for Information-grounded Veracity Recognition via Reinforced Knowledge Graph Exploration
Cong Hoan Nguyen, Thomas Hoang, Hieu Minh Duong, Long Nguyen
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 7/21/2026, 5:57:09 AM
Summary
The paper introduces DeLIVeR, a framework for automated fact-checking that addresses the limitations of traditional retrieval-augmented generation (RAG) by treating evidence retrieval as a reinforced strategic exploration task. DeLIVeR utilizes a Planner LLM to decompose complex claims into targeted question sets, which traverse structured Knowledge Graphs (KGs) to retrieve high-precision evidence. The system optimizes the Planner's policy using Group Relative Policy Optimization (GRPO), prioritizing structural diversity and verdict accuracy. Evaluated on LIAR, FEVER, and PolitiFact datasets, DeLIVeR significantly outperforms state-of-the-art baselines like HippoRAG2, achieving higher F1-scores and providing an auditable path for verifiable misinformation detection.
Entities (19)
Relation Signals (14)
Thomas Hoang → affiliatedwith → Denison University
confidence 95% · Thomas Hoang... 2 Denison University
Cong Hoan Nguyen → affiliatedwith → University of Louisville
confidence 95% · Cong Hoan Nguyen... 1 University of Louisville
Minh Hieu Duong → affiliatedwith → University of Louisville
confidence 95% · Minh Hieu Duong... 1 University of Louisville
Long Nguyen → affiliatedwith → University of Louisville
confidence 95% · Long Nguyen... 1 University of Louisville
DeLIVeR → optimizespolicyusing → GRPO
confidence 95% · We optimize the Planner's policy using Group Relative Policy Optimization (GRPO)
DeLIVeR → uses → Planner LLM
confidence 95% · DeLIVeR utilizes a Planner LLM to decompose complex claims into targeted question sets
DeLIVeR → uses → Knowledge Graph
confidence 95% · DeLIVeR utilizes a Planner LLM to decompose complex claims into targeted question sets, which are used to traverse structured Knowledge Graphs (KGs)
DeLIVeR → evaluatedon → FEVER
confidence 90% · Our evaluation on LIAR, FEVER, and PolitiFact shows that DeLIVeR significantly outperforms state-of-the-art baselines.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Automated fact-checking remains a challenge for Large Language Models (LLMs) due to "query brittleness" in traditional retrieval systems. We propose DeLIVeR (Decomposed Learning for Information-grounded Veracity Recognition), a framework that treats evidence retrieval as a reinforced strategic exploration task. DeLIVeR utilizes a Planner LLM to decompose complex claims into targeted question sets, which are used to traverse structured Knowledge Graphs (KGs) for high-precision evidence. We optimize the Planner's policy using Group Relative Policy Optimization (GRPO) with a reward system prioritizing structural diversity and verdict accuracy. Our evaluation on LIAR, FEVER, and PolitiFact shows that DeLIVeR significantly outperforms state-of-the-art baselines. Using Qwen2.5-7B, our framework achieved peak F1-scores of 83.73, 84.57, and 79.70 respectively, representing a 10-15% improvement over HippoRAG2. By shifting to a reinforced question-planning strategy, DeLIVeR effectively bridges multi-hop reasoning gaps and provides an auditable, transparent path for verifiable misinformation detection.
Tags
Links
- Source: https://arxiv.org/abs/2607.17935v1
- Canonical: https://arxiv.org/abs/2607.17935v1
Trouble viewing inline? Open PDF directly →
Full Text
58,133 characters extracted from source content.
Expand or collapse full text
Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen. “DeLIVeR: Decomposed Learning for Information-grounded Veracity Recognition via Reinforced Knowledge Graph Exploration”, 7th International Conference on Deep Learning Theory and Applications (DeLTA 2026). ©2026 Springer. Personal use of this material is permitted. Permission from Springer must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. DeLIVeR: Decomposed Learning for Information-grounded Veracity Recognition via Reinforced Knowledge Graph Exploration Cong Hoan Nguyen 1 , Thomas Hoang 2 , Minh Hieu Duong 1 , and Long Nguyen 1 1 University of Louisville, Louisville KY 40292, USA. conghoan.nguyen,hieu.duong,l.nguyen@louisville.edu 2 Denison University, Granville, Ohio 43023, USA. hoang_t2@denison.edu Abstract. Automated fact-checking remains a challenge for Large Lan- guage Models (LLMs) due to "query brittleness" in traditional retrieval systems. We propose DeLIVeR (Decomposed Learning for Information- grounded Veacity Recognition), a framework that treats evidence re- trieval as a reinforced strategic exploration task. DeLIVeR utilizes a Planner LLM to decompose complex claims into targeted question sets, which are used to traverse structured Knowledge Graphs (KGs) for high- precision evidence. We optimize the Planner’s policy using Group Rel- ative Policy Optimization (GRPO) with a reward system prioritizing structural diversity and verdict accuracy. Our evaluation on LIAR, FEVER, and PolitiFact shows that DeLIVeR significantly outperforms state-of- the-art baselines. Using Qwen2.5-7B, our framework achieved peak F1- scores of 83.73, 84.57, and 79.70 respectively, representing a 10–15% improvement over HippoRAG2. By shifting to a reinforced question- planning strategy, DeLIVeR effectively bridges multi-hop reasoning gaps and provides an auditable, transparent path for verifiable misinformation detection. Keywords: Fact Detection, Knowledge Graph, Large Language Model, Retrieval Augmented Generation, Reinforcement Learning. 1 Introduction The proliferation of online misinformation poses a critical threat to public trust and democratic stability. While Large Language Models (LLMs) offer sophisti- cated reasoning, they are frequently undermined by “hallucinations” stemming from a reliance on static internal knowledge rather than verifiable grounding [1, 3]. Although Retrieval-Augmented Generation (RAG) aims to mitigate this, existing frameworks struggle with multi-hop gaps, static query limitations, and fixed retrieval policies that fail to adapt to complex claims spanning multiple entities [38, 4, 5]. arXiv:2607.17935v1 [cs.CL] 20 Jul 2026 2Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen Claim： The Mona Lisa was stolen from the Louvre in 1911 by Pablo Picasso. (a) LLM-Only Generate Answer: True. Explanation: Picasso, a prominent artist, was linked to art controversies, so he likely stole the Mona Lisa. (b) GraphRAG Retrieve 1) ..was stolen from Louvre in 1911 by Vincenzo Peruggia. 2) Pablo Picasso was questioned in the theft 3) Picasso painted the Mona Lisa in 1907. Answer: NEI Explanation: Correct thief identified, but the misleading Picasso link and incorrect painting attribution create ambiguity. Generate Think Claim Analysis: The claim is "The Mona Lisa was stolen from the Louvre in 1911 by Pablo Picasso." Break it into key entities: Mona Lisa, Louvre, 1911, Pablo Picasso, theft. Identify verifiable aspects: who stole it, where, when, and outcome. Aim for questions that probe these elements to retrieve precise evidence from the Knowledge Graph, ensuring relevance to confirm or refute the claim. Question Generation Strategy: Formulate four questions targeting: (1) the thief’s identity, (2) Picasso’s involvement, (3) theft location, and (4) an additional fact (recovery) to contextualize the event. Ensure questions are specific to avoid ambiguous or irrelevant retrievals, optimizing for reward. Query 1) Who stole the Mona Lisa in 1911? 2) Was Pablo Picasso involved in the theft? 3) Where was the Mona Lisa stolen from? 4) When was the Mona Lisa recovered? 1) Vincenzo Peruggia stole the Mona Lisa. 2) Picasso was questioned but not involved. 3) Mona Lisa was stolen from the Louvre. 4) Mona Lisa recovered in 1913. Retrieve Answer: False Explanation: Evidence confirms Peruggia, not Picasso, stole the Mona Lisa from the Louvre in 1911. (c) Our Method Fig. 1. Comparison of fake news detection systems for the Mona Lisa theft claim. To address these limitations, we propose DeLIVeR a framework that trans- forms fact verification from a static retrieval step into a reinforced, closed-loop optimization process as shown in Figure 1. Unlike domain-adaptive approaches that focus primarily on feature-level transformations, DeLIVeR optimizes the model’s information-seeking strategy. Given a claim, a planner LLM generates a small set of diverse, targeted questions that collectively query a Knowledge Graph (KG) to retrieve complementary evidence. This policy is optimized via Group Relative Policy Optimization (GRPO) to reward structural diversity and verdict accuracy. In developing this framework, we specifically investigate whether a reinforced planner LLM can effectively bridge multi-hop reasoning gaps compared to static RAG baselines (RQ1), the extent to which set-based question planning reduces query ambiguity in structured KGs (RQ2), and if GRPO-driven optimization yields a more stable and interpretable information- seeking policy for veracity classification (RQ3). Our contributions are as follows: – Structured multi-hop grounding: We introduce a KG-grounded evidence retrieval pipeline that moves beyond flat text retrieval by retrieving struc- tured, multi-hop evidence paths for claim verification. – Set-based question planning: We design a planner LLM that decomposes each claim into a cohesive set of diverse, targeted questions, reducing query ambiguity and improving evidence coverage over standard single-query RAG. – GRPO-driven policy optimization: We fine-tune the question genera- tion module with GRPO using rewards that encourage valid format, struc- tural diversity, and evidence quality, yielding a more stable information- seeking policy. – Empirical validation and interpretability: Experiments on FEVER, LIAR, and PolitiFact show consistent improvements over strong RAG base- Decomposed Learning for Information-grounded Veracity Recognition3 lines, while the question-driven retrieval process provides inherent evidence chains for auditing model decisions. 2 Related Work Early artificial intelligence systems for knowledge-intensive tasks relied heavily on neural architectures such as Artificial Neural Networks (ANNs) and energy- based models, particularly Discrete Hopfield Neural Networks (DHNNs), which demonstrated the foundational principles of structured reasoning and pattern recognition. These early approaches established critical insights into logic mining and constraint satisfaction that continue to inform modern AI systems. Notable contributions include flexible logic mining frameworks that combine ensemble multi-attribute selection with DHNNs [72], investigations into weighted C-type random 2-satisfiability formulations within discrete Hopfield networks [73], and optimized logic mining methods leveraging higher-order satisfiability represen- tations [74]. While these early neural and logic-based paradigms demonstrated that structured, rule-governed representations significantly improve classification reliability and interpretability, they typically operate on fixed propositional rep- resentations and do not address the open-domain, multi-hop evidence retrieval needed for modern fact verification over large, dynamic knowledge corpora. Fake news detection has progressed from feature-based models [33, 34] to pretrained deep encoders such as BERT and RoBERTa [16, 17], which perform well on benchmarks including LIAR [35] and FEVER [36]. However, veracity prediction with LLMs remains unreliable without explicit grounding, due to hallucinations and the absence of verifiable evidence attribution [18]. Retrieval- Augmented Generation (RAG) addresses this by conditioning predictions on external context [38], but standard RAG typically retrieves unstructured text using similarity-based queries and often fails to capture the relational structure needed for claims involving multiple entities and linked events [15, 7]. Recent graph-based retrieval methods improve indexing and ranking, but they com- monly treat the user query as fixed. In contrast, DeLIVeR emphasizes structured KG evidence while optimizing the information-seeking strategy through learned question sets. A second challenge is query ambiguity. Many RAG systems rely on a sin- gle query derived directly from the claim, making retrieval brittle when the claim is underspecified or implicitly framed. Prior work on claim decomposi- tion shows that generating intermediate questions can improve reasoning [8, 9], and question generation models such as T5/BART have been used to produce subquestions [10]. However, these approaches are often not trained to optimize the retrieval outcome needed for verification, especially when evidence must be gathered from complementary semantic facets (e.g., temporal context, relations, contradictions). Our approach introduces a Planner LLM that generates a small, diverse set of targeted questions to probe a KG from multiple angles, improving evidence coverage and reducing ambiguity for downstream verification [11]. 4Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen Finally, optimizing retrieval behavior for black-box LLM pipelines is difficult because gradients cannot be propagated through the verifier. Existing methods align retrieval with LLM feedback using objectives such as KLD-based techniques [12] or policy-gradient style optimization [49], but PPO-style approaches can be unstable in high-dimensional generation spaces [50]. We address this by applying Group Relative Policy Optimization (GRPO) [66] to optimize question-set gen- eration. GRPO estimates advantages by comparing multiple outputs from the same claim, enabling stable updates without a centralized critic. This provides a practical mechanism to learn a robust question-generation policy that improves retrieval precision and supports accurate veracity decisions by rewarding high- signal evidence while discouraging noisy retrieval. 3 Methodology 3.1 Overview The proposed DeLIVeR framework presented in Figure 2 consists of Knowledge Graphs (KGs), a two footprint LLM architecture and reinforcement learning to tackle the challenges associated with fake news detection by accurate evidence retrieval and knowledge-based generating a verdict. The system begins with KG sourced from a reliable dataset, with nodes representing entities or facts, with edges indicating relationships (for example "supports", "contradicts", etc), al- lowing for structured and contextual evidence query [43]. The secondary LLM, fine-tuned for question generation, processes the input claim to produce a co- hesive set of targeted questions, collectively querying the KG to retrieve com- prehensive evidence [47]. The aggregated evidence from the question set, along with the claim, is then fed into a frozen primary LLM to generate a verdict (True, False, Not Enough Information) with a human-interpretable explanation, ensuring stability and interpretability [38]. Group Relative Policy Optimization (GRPO) iteratively refines the question-generating LLM by evaluating the qual- ity of retrieved evidence (e.g., format, structure, accuracy), optimizing retrieval effectiveness through reinforced feedback [49]. This framework integrates other ideas previously proposed to produce an approach that integrates structured knowledge, adaptive queries, and reinforced feedback to provide robust and explainable means of detecting fake news. 3.2 Problem Formulation The veracity detection task in DeLIVeR is governed by the following condi- tional probability distribution, which decomposes the verdict y into a sequence of planning and retrieval steps: P(y|c)≈ max Q [P(y|c, Retrieve(Q,G))· P(Q|c;θ Q )](1) where c is the input claim, Q = q 1 ,...,q n is the set of generated questions, and G represents the structured Knowledge Graph. Our objective is to optimize Decomposed Learning for Information-grounded Veracity Recognition5 Galileo Copernican Italy Galileo Satellite Navigation EU Space Agency Astronomer ... ... ... ... ... ... ... ... Step 1: Chunking Document Step 2:Extract Entities & Relationship Chunking Strategy Chunk Size Overlap Claim Serial Questions Graph Construction DeLIVeR Generation Supported Set Structure Reward Accuracy Reward Format Reward GRPO Update Search Retriever Generative Model Preprocessing LLM Models TrueFalseNEI In 1633,Galileo was tried by theRomanCatholic Inquisition for supporting theheliocentrictheory proposed byCopernicus, which contradicted the Church’s teachings at the time. 1 2 3 4 Fig. 2. The overall DeLIVeR architecture operates through four key stages: generat- ing a sequence of questions, querying the knowledge graph, extracting the supporting information set, and updating the generative model using GRPO. the parameters θ Q such that the generated Q maximizes the likelihood of the correct verdict y through high-quality evidence retrieval. 3.3 Knowledge Graph Construction To ensure the framework is grounded in verifiable facts while strictly avoid- ing label leakage and circular reasoning, we construct the Knowledge Graph (KG) using only the ground-truth evidence corpora associated with each dataset. We explicitly exclude claim text and veracity labels from the graph construc- tion pipeline, utilizing GPT-4 to perform Open Information Extraction (Ope- nIE) that transforms unstructured evidence into structured triples (s,p,o) (Sub- ject–Predicate–Object). To guarantee high-fidelity retrieval, we implement a provenance-tracking mechanism that maps every node and edge back to its source document URI, allowing the Verifier LLM to audit retrieved paths against the original text. Table 1. Quality Assessment of Extracting KG Triples MetricScore Definition Entity Precision 94.2% Accuracy of extracted subjects and objects against source text. Relation Recall 88.5% Percentage of key relations from text captured in the graph. Triple Fidelity 91.8% Correctness of the (s, p, o) link as a logical unit. Noise Rate3.4% Percentage of extracted triples with no basis in source text. As demonstrated in Table 1, a manual quality assessment of 500 randomly sampled triples confirms the reliability of this process, yielding an Entity Pre- cision of 94.2% and a Triple Fidelity of 91.8%. Furthermore, we address tem- poral sensitivity by appending temporal metadata (timestamps) to edges when 6Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen available in the source text. During retrieval, the model prioritizes edges with timestamps closest to the claim’s publication date. For noise control, we apply a frequency-based filter that removes "singleton" entities that do not connect to at least two other nodes, ensuring the graph focuses on the dense, multi-hop relationship clusters necessary for complex reasoning. 3.4 Question Generation Module The Question Generation Module is a core part of the DeLIVeR framework that leverages a hundred million parameters pretrained secondary large lan- guage model, to create relevant questions based on an input claim, in order to frame the language for targeted retrieval of relevant evidence from the KG for fake news verification. Formally, given an input claim c, the secondary LLM Q generates a set of questions Q =q 1 ,q 2 ,...,q n , with each question q i to explore specific angles of the claim such as the source, evidence to substantiate claim c, contextual factors which may affect the claim, or counter claims or contradic- tions, such as! “What is the original source of claim c?” “Is evidence x in the KG contradictory to assertion y in claim c?” “What time factors impact the validity of event z?”. The process is defined asQ(c;θ Q )→ Q⊆P(KG), where θ Q are the pretrained parameters of the LLM, and the question set Q is used collectively to query the KG, retrieving a unified set of evidence. Each question generated is intended to cover multiple angles of the claim, such as factual, temporal, and causal, to facilitate a more richly supported collection of evidence from the KG and support multi-hop reasoning, which is necessary in the evaluation of com- plex misinformation. The question set is mapped to KG nodes and edges using vector similarity search, with embeddings of questions and KG elements ensuring contextually relevant evidence retrieval: Score(q i ,e j ) = cos(φ(q i ),φ(e j )),(2) where φ(q i ) and φ(e j ) are embeddings for the question and KG element (node or edge) e j , respectively, and cos denotes cosine similarity [52]. This scoring enables the selection of the most relevant KG elements, significantly improving the granularity and relevance of retrieved evidence compared to static query methods in standard RAG systems. To address redundancy and ensure comprehensive coverage of the KG’s re- lational structure, the question generation process incorporates a diversity con- straint, encouraging the LLM to produce questions that span distinct subgraphs of the KG. This is formalized by maximizing the entropy of question coverage: H(Q) =− X q i ∈Q P(q i |c) logP(q i |c),(3) where P(q i |c) is the probability of generating question q i given claim c, ensuring that questions probe varied aspects of the KG without overlap [53]. The generated questions are mapped to KG nodes and edges using a vector similarity search. We utilize the embeddings of both the questions and the KG Decomposed Learning for Information-grounded Veracity Recognition7 elements to generate evidence that is contextually relevant to the nuance of the claims. The performance of the set of questions is refined iteratively using Group Relative Policy Optimization (GRPO), detailed in Section 3.4, which adjusts θ Q by evaluating the quality of the retrieved aggregated evidence. This approach, different from static RAG pipelines, optimizes question sets to maximize the ex- traction of relevant evidence, enhancing the verification of abstract or ambiguous claims [49]. In practice, this modification of objective function is what separates our approach from static RAG pipelines, where retrieving incomplete or irrele- vant evidence is difficult. By incorporating pre-trained LLMs with KG retrieval optimized by GRPO, the Question Generation Module provides a necessary level of contextualized evidence extraction to support the system’s verification capa- bilities for complex claims in fake news detection and misinformation combat. 3.5 Retrieval and Augmentation The Retrieval and Augmentation module leverages a cohesive set of questions generated by the secondary LLM to collectively query the Knowledge Graph (KG) and retrieve a unified set of relevant evidence, which is then augmented with the input claim to enable the primary LLM to generate accurate ver- dicts for fake news detection. For a given claim c and its set of questions Q = q 1 ,q 2 ,...,q n , the questions are embedded using a pretrained sentence encoder, and the evidence is retrieved from the KG by computing similarity scores between the question embeddings and the KG node/edge embeddings, formalized as: Retrieve(Q,G) = [ q i ∈Q arg max e j ∈G cos(φ(q i ),φ(e j )),(4) where G = (V,E) is the KG, φ(·) denotes the Sentence-BERT embeddings, and cos is cosine similarity [52]. The aggregated evidence set E =e 1 ,e 2 ,...,e m is concatenated with the claim c to form an augmented input [c;E], which is fed to the primary LLM to produce a verdict (True, False, or Not Enough Information) with an explanation. The quality of the retrieved evidence is evaluated to inform iterative refinement of the question generation process [38]. This process ensures the primary LLM leverages structured, contextually relevant evidence, enhancing verdict accuracy and interpretability compared to standard RAG systems that rely on unstructured text retrieval. 3.6 Reinforcement Learning Optimization To bridge the gap between static retrieval and adaptive fact-checking, we opti- mize the Question Generation (QG) module using Group Relative Policy Opti- mization (GRPO) [66]. Unlike standard Proximal Policy Optimization (PPO), which relies on a centralized critic to estimate a state-value baseline, GRPO com- putes advantages based on the relative performance of a group of outputs gener- ated from the same prompt. This is particularly advantageous for our framework 8Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen Claim: TheMona Lisawaspainted by Michelangelo. Sample Questions & Retrieved Information 1 Who painted the Mona Lisa? ...was painted by Leonardo da Vinci, as confirmed by the Louvre Museum... When was the Mona Lisa created? ...was created between 1503 and 1506, with some refinements up to 1517, by Leonardo da Vinci... Did Michelangelo ever claim to have painted the Mona Lisa? Michelangelo was a contemporary of Leonardo da Vinci and worked on similar subjects, like portraiture, in Florence during the early 16th century. What is the primary medium of the Mona Lisa painting? The Mona Lisa is a fresco painting using vibrant watercolors on a plaster wall. Model Update Reinforce higher- advantage responses A B C D Model updates to increase likehood of better-than- average responses 4 LLM Compute Group Average &Advantage Advantage A = 1.0 - 0.45 = +0.55 Advantage B = 0.6 - 0.45 = +0.15 Advantage C = 0.2 - 0.45 = -0.25 Advantage D = 0.0 - 0.45 = -0.45 A BCD 3 R=(1.0 + 0.6 + 0.2 + 0.0) / 4 = 0.45 Scoring Dimensions: - Accuracy (Question and Retrieved Information correct) - Format (e.g.. Uses <think>/<answer> as applicable - Structure Reward 1.0 0.60.20.0 A BCD 2 Reward Function A B C D Fig. 3. Overview of the Group Relative Policy Optimization (GRPO) process for re- fining question generation. Step 1: samples question setsQ i from the policy π θ Q and retrieves corresponding evidence E i from the Knowledge Graph. Step 2: computes the Format-Structure-Accuracy reward R(Q i ,E i ) for each set. Step 3: calculates the group average reward and advantage ˆ A(Q i ,E i ). Step 4: updates the policy parameters θ Q via the clipped objective to optimize retrieval effectiveness and verdict reliability. because it allows the model to compare multiple diverse "question sets" for a sin- gle claim, identifying which specific combinations of queries maximize evidence coverage while minimizing retrieval noise. Additionally, this approach enhances the quality of evidence retrieved from the Knowledge Graph (KG) for fake news detection, as illustrated in the GRPO process overview (see Figure 3). Given a claim c ∈ D, the LLM generates a group of question sets Q i N i=1 ⊆ T c , where each Q i = q i1 ,...,q iT is used to collectively query the KG, producing an aggregated evidence set E i = Retrieve(Q i ,G). We use the prompt, detailed in Table 2, optimize the policy π θ Q using the GRPO objective, defined as: J GRPO (θ Q ) =E c∼D,Q i N i=1 ∼π θ old Q (Q|c;KG) " 1 N N X i=1 min ρ θ Q (Q i ) ˆ A(Q i ,E i ), clip ρ θ Q (Q i ), 1± ε ˆ A(Q i ,E i ) − βD KL (π θ Q ∥ π ref ) # , (5) where ρ θ Q (Q i ) = π θ Q (Q i |c;KG) π θ old Q (Q i |c;KG) , and the advantage is: ˆ A(Q i ,E i ) = R(Q i ,E i )− 1 N P N j=1 R(Q j ,E j )/F norm (R(Q j ,E j ) N j=1 ) and the reward function: R(Q i ,E i ) = w 1 · R Format (Q i ) + w 2 · R Structure (Q i ) + w 3 · R Accuracy (E i ), where F norm (·) is a normalization function (e.g., standard deviation), and clip(·) stabilizes policy up- dates [50]. GRPO’s group-based comparison optimizes question sets to maximize Decomposed Learning for Information-grounded Veracity Recognition9 the quality of retrieved evidence, enhancing retrieval effectiveness and verdict re- liability over traditional RL methods [38, 49]. Prompt A conversation between User and Assistant for training a model to generate questions from a claim. The User provides a claim, and the Assistant generates 4–8 relevant questions to query a Knowledge Graph or external sources. The Assistant outlines the reasoning process within <think>... </think> tags, detailing how the claim is parsed and questions are formulated, then provides the questions within <answer>... </answer> tags. User: Given a claim, generate 4–8 diverse and specific questions to query a Knowledge Graph or external sources for misinformation detection. Show your reasoning process and list the questions. Assistant: <think> 1. Parse the claim: Identify key entities, events, relationships, and context within the claim to ensure questions target critical aspects. 2. Formulate questions: Generate 4 questions that are diverse, covering different facets (e.g., factual details, entities, temporal aspects, contradictions), and specific to maximize relevance for Knowl- edge Graph retrieval or external search. Ensure questions are clear, concise, and optimized for misinformation detection on datasets. 3. Optimize diversity: Balance question types (e.g., who, what, when, where, why) to cover multiple perspectives while avoiding redundancy. </think> <answer> List of 4 questions: <question 1> </question 1> <question 2> </question 2> <question 3> </question 3> <question 4> </question 4> </answer> Claim: claim. Assistant: Table 2. Prompt template for DeLIVeR question generation Format Reward (R Format (Q i )). This reward enforces strict adherence to the structured output format required for downstream retrieval and parsing. The model must generate a reasoning trace within <think> </think> tags followed by a numbered list of 4–8 questions within <answer> </answer>, with no ex- traneous text. We use regular expressions to validate: - Exactly one <think> block with coherent reasoning steps. - Exactly one <answer> block containing a numbered list (1., 2., etc.) with 4 ≤ questions ≤ 8. - No content outside tags. This is a binary reward: R Format (Q i ) = ( 1 if format is valid, 0 otherwise. (6) Structural Reward (R Structure (Q i )). This is the core signal for effective evidence retrieval. It evaluates whether the generated question set Q i covers diverse semantic dimensions of the claim c (e.g., entities, events, temporal context, causal links, contradictions) to maximize informational breadth. The computation proceeds in two steps: 1. Each question q ij ∈ Q i is classified into one of K predefined semantic categories (e.g., WHO, WHAT, WHEN, WHERE, HOW, 10Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen CONTRADICTION) using a lightweight classifier (e.g., fine-tuned BERT or keyword patterns). Let C(Q i ) be the set of unique categories covered. 2. The reward is the coverage ratio relative to an ideal category distribution C ∗ (derived from claim type or oracle analysis): R Structure (Q i ) = |C(Q i )∩ C ∗ | |C ∗ | ∈ [0, 1].(7) Alternatively, for finer granularity, we compute Jaccard similarity over nor- malized question skeletons (replacing named entities with [ENT], numbers with [NUM], etc.) to reward structural diversity beyond category labels. R Structure is assigned the highest weight (w 2 ≫ w 1 ,w 3 ) to prioritize comprehensive evidence gathering. Accuracy Reward (R Accuracy (E i )). This reward measures factual relevance of the retrieved evidence E i = Retrieve(Q i ,G) to the claim c. We use a binary signal from an LLM judge (e.g., Qwen2.5-72B-Instruct) or a fine-tuned factuality classifier: - Input: Concatenated evidence (E i ) and claim (c). - Output: Label SUPPORTS, REFUTES, or NEI. The reward is: R Accuracy (E i ) = ( 1 if information is correct, 0 if hallucinated. (8) By combining these three rewards, GRPO creates a hierarchical optimiza- tion landscape: the model is guided primarily by structure to explore the claim comprehensively, constrained by format for reliability, and refined by accuracy for verdict quality. This yields question sets that are well-formed, semantically rich, and evidentially potent, critical for robust misinformation detection. 4 Experiment and Results 4.1 Experimental Setup Datasets To evaluate our framework, we utilize three benchmark datasets for fake news detection, each with distinct characteristics to assess evidence retrieval, verdict accuracy, and explanation quality: – PolitiFact [54]: Contains political claims from U.S. media, annotated with veracity labels (e.g., True, False, Pants on Fire) and detailed justifications, ideal for testing verdict accuracy and explanation coherence. – LIAR [35]: Comprises 12,8K short political statements from diverse sources, labeled with fine-grained veracity categories (e.g., True, Mostly True, False), suitable for evaluating generalization across varied claims. – FEVER [36]: Includes Wikipedia-derived claims, annotated as Supported, Refuted, or Not Enough Information, with linked evidence, enabling robust assessment of evidence retrieval and multi-hop reasoning. Decomposed Learning for Information-grounded Veracity Recognition11 Table 3. Counts of real/fake news across datasets. LIARFEVER POLITIFACT #Real News 92523333399 #Fake News 35553333345 #Total12 8076666744 These datasets collectively challenge our framework’s ability to handle diverse claim types, structured evidence, and contextual nuances in misinformation de- tection, as summarized in Table 3. Baseline To benchmark our DeLIVeR framework, we compare it against five methods for fake news detection, each highlighting different retrieval and rea- soning capabilities: – Vanilla LLM [1]: Uses a pretrained large language model for direct claim classification without external knowledge, prone to hallucinations in verifi- cation tasks. – Naive RAG [38]: Performs one-step retrieval from an unstructured corpus to augment prompts, often retrieving irrelevant or incomplete evidence due to lack of query refinement. – LightRAG [55]: Employs graph-enhanced indexing for faster, contextual retrieval, reducing computational overhead while maintaining accuracy in knowledge-intensive queries. – ReAct [56]: Interleaves reasoning and acting via chain-of-thought and tool calls (e.g., search APIs), enabling dynamic evidence gathering for multi-hop fact-checking. – HippoRAG2 [57]: Builds hierarchical knowledge graphs with personalized PageRank for incremental, context-aware retrieval, excelling in integrating diverse evidence. Implementation Details Our DeLIVeR framework utilizes GPT-4 for con- structing the Knowledge Graph from verified datasets, leveraging its advanced language understanding for entity and relationship extraction [60]. For the ques- tion generation and verdict prediction, we benchmark three scales of Qwen2.5 (1.5B, 3B, and 7B parameters), evaluating their performance across diverse claim complexities [58]. The retrieval module employs the bge-large-en-v1.5 model for embedding questions and KG elements, ensuring robust similarity-based evi- dence extraction [59]. GRPO Training Configuration: We implement GRPO optimization using a group size of N = 8 question sets per claim for reliable advantage estimation. The reward function weights are set as w 1 = 0.15, w 2 = 0.60, and w 3 = 0.25, where w 2 ≫ w 1 ,w 3 to prioritize structural diversity as described in Section 4.6. We use a clipping parameter ε = 0.20 and KL penalty coefficient β = 0.04 12Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen to ensure stable policy updates. Training is performed on four NVIDIA A100 GPUs (80GB) with a learning rate of 5 × 10 −6 and global batch size of 32 claims. For Qwen2.5-7B, GRPO fine-tuning converges in approximately 10.5 hours, while the 1.5B and 3B variants require 3.0 and 5.5 hours respectively. Complete hyperparameters and KG statistics are detailed in Table 4. Table 4. DeLIVeR Training Hyperparameters and Knowledge Graph Statistics ParameterSetting/ValueNotes GRPO Hyperparameters Group size (N)8Question sets sampled per claim Format reward weight (w 1 )0.15Binary format validation Structure reward weight (w 2 )0.60Semantic diversity priority (w 2 ≫ w 1 ,w 3 ) Accuracy reward weight (w 3 )0.25Evidence relevance to verdict Clipping parameter (ε)0.20PPO-style policy update stability KL penalty coefficient (β)0.04Reference model regularization Learning rate5× 10 −6 AdamW optimizer Warmup steps100Linear learning rate warmup Global batch size32Claims per gradient update Max training epochs3Early stopping on validation F1 Temperature (sampling)0.7Diversity in group generation Computational Requirements Hardware4 × NVIDIA A100 (80GB)Distributed training setup Training time (Qwen2.5-1.5B)∼3.0 hoursAverage across datasets Training time (Qwen2.5-3B)∼5.5 hoursAverage across datasets Training time (Qwen2.5-7B)∼10.5 hoursAverage across datasets KG construction time∼6.0 hoursGPT-4 OpenIE extraction Peak GPU memory (7B model)∼65 GBIncluding gradients and optimizer Knowledge Graph Statistics LIAR: Entities / Triples42,150 / 128,400Political claims and entities FEVER: Entities / Triples186,300 / 524,800Wikipedia-derived evidence PolitiFact: Entities / Triples8,920 / 26,150Fact-checking corpus Avg. entity connectivity3.2Triples per entity (post-filtering) Relation types (total)47Unique predicate categories Singleton filter threshold≥ 2 connectionsMinimum connectivity requirement Temporal edge coverage58.7%Edges with timestamp metadata 4.2 Main Results Table 5 presents a comprehensive evaluation of our DeLIVeR framework against five baselines include Vanilla LLM, Naive RAG, LightRAG, ReAct, and Hip- poRAG2 across the LIAR, FEVER, and PolitiFact datasets, using three scales of the Qwen2.5 model (1.5B, 3B, and 7B parameters). Our framework consis- tently outperforms all baselines across recall, precision, accuracy, and F1-score, achieving peak F1-scores of 83.73 (LIAR), 84.57 (FEVER), and 79.70 (Poli- tiFact) with Qwen2.5-7B, compared to HippoRAG2’s best F1-scores of 72.93, 75.16, and 69.97, respectively. This represents an average F1 improvement of 10–15% over HippoRAG2, the closest competitor, highlighting the efficacy of our knowledge graph (KG) integration and Group Relative Policy Optimiza- tion (GRPO)-driven question set generation. The performance gap widens with Decomposed Learning for Information-grounded Veracity Recognition13 Table 5. Performance Metrics on MultiReQA Datasets Methods LIARFEVERPOLITIFACT Qwen2.5-1.5B-Instruct Recall Prec Acc F1Recall Prec Acc F1Recall Prec Acc F1 Vanilla LLM36.51 35.37 35.92 35.9337.64 36.22 36.59 36.8935.48 34.91 34.78 35.19 Naive RAG 52.47 49.36 50.41 50.8754.12 51.59 52.31 52.8350.94 48.23 49.32 49.63 LightRAG58.64 55.22 56.15 56.8960.53 57.18 58.02 58.8357.25 54.37 55.18 55.74 ReAct60.18 57.34 58.10 58.7262.81 59.47 60.32 61.1161.29 58.42 59.30 59.82 HippoRAG2 63.72 60.35 61.44 61.9665.33 62.24 63.15 63.7763.61 60.23 61.04 61.86 DeLIVeR (ours)72.4869.5370.3170.9774.6971.4272.2472.9970.1867.2368.1168.68 Qwen2.5-3B-Instruct Vanilla LLM38.17 37.02 37.56 37.5839.45 38.13 38.42 38.7937.68 36.59 36.74 37.13 Naive RAG 56.41 53.72 54.38 55.0358.29 55.36 56.02 56.8055.72 52.64 53.32 53.98 LightRAG62.68 59.39 60.15 60.9764.43 61.08 61.93 62.7761.89 58.66 59.48 60.24 ReAct65.33 62.47 63.18 63.8667.28 64.10 65.09 65.6364.47 61.52 62.34 62.88 HippoRAG269.28 66.14 67.09 67.6971.54 68.22 69.13 69.8567.88 64.61 65.54 66.28 DeLIVeR (ours)79.3776.4477.3277.8881.2578.1679.0379.6775.7372.6673.4973.98 Qwen2.5-7B-Instruct Vanilla LLM39.23 38.15 38.42 38.6940.51 39.29 39.48 39.9038.74 37.59 37.74 38.16 Naive RAG 60.44 57.32 58.21 58.8562.73 59.48 60.26 60.9860.92 57.18 58.09 58.61 LightRAG66.38 63.20 64.12 64.7668.44 65.19 66.07 66.7864.69 61.58 62.43 63.09 ReAct 70.52 67.33 68.28 68.8772.86 69.61 70.43 71.1869.27 66.13 67.04 67.64 HippoRAG2 74.63 71.35 72.29 72.9376.85 73.52 74.41 75.1671.66 68.39 69.33 69.97 DeLIVeR (ours)85.3282.2183.0983.7386.1483.1184.0084.5781.3678.2379.1879.70 larger model scales, with Qwen2.5-7B yielding the highest scores due to its en- hanced reasoning capacity, which better leverages the structured evidence re- trieved from the KG. FEVER consistently produces the highest scores across all methods, likely due to its structured Wikipedia-based evidence annotations, which align well with our framework’s multi-hop reasoning capabilities. In con- trast, PolitiFact’s smaller dataset size (744 samples) and nuanced veracity labels (e.g., Pants on Fire) pose greater challenges, yet our framework still achieves robust performance (F1 of 79.70 with Qwen2.5-7B), demonstrating its ability to handle complex, real-world misinformation scenarios. 4.3 Ablation Study Impact of question set size: We ablate question set size using Qwen2.5- 7B-Instruct on LIAR, FEVER, and PolitiFact (Figure 4). Peak F1-scores are achieved with four questions: 0.8373 (LIAR), 0.8457 (FEVER), and 0.7970 (Poli- tiFact), confirming the optimal balance between evidence coverage and relevance. Performance declines with eight questions (0.8297, 0.8389, 0.7900) due to redun- dancy and with sixteen questions (0.8221, 0.8317, 0.7835) due to retrieval noise. Only two questions yield the lowest scores (0.8115, 0.8216, 0.7625) from insuffi- cient evidence diversity. FEVER benefits most from structured evidence, while PolitiFact remains the hardest due to nuanced labels and limited data. These 14Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen results validate GRPO’s effectiveness in producing concise, high-quality question sets and underscore the importance of tuning set size for robust retrieval and verdict accuracy in misinformation detection. 246810121416 Number of Questions 0.700 0.725 0.750 0.775 0.800 0.825 0.850 0.875 0.900 F1-Score LIAR FEVER POLITIFACT Fig. 4. The impact of the number of generated questions Qualitative Evaluation Figure 5 illustrates the qualitative performance of our DeLIVeR framework (Qwen2.5-7B-Instruct) against Vanilla LLM, Naive RAG, LightRAG, ReAct, and HippoRAG2 across six categories for misinformation de- tection: Knowledge-ability, Comprehensiveness, Factuality, Logical Coherence, Relevance, and Correctness. Evaluated on LIAR, FEVER, and PolitiFact, our framework achieves top scores: 83.2 (Knowledge-ability, Relevance), 75.12 (Com- prehensiveness), 82.45 (Factuality), 68.4 (Logical Coherence), and 75.25 (Cor- rectness), outperforming ReAct’s 72.4, 67.7, 79.9, 58.6, 77.2, and 69.3, respec- tively. KG-driven retrieval and GRPO-optimized questions enhance Factuality and Relevance, while Vanilla LLM struggles (e.g., 40.3 in Logical Coherence) due to limited knowledge. Naive RAG (65.1 in Knowledge-ability), LightRAG (69.3), and HippoRAG2 (59.0 in Factuality) lag behind. These results validate our framework’s robust, coherent verdicts for misinformation detection. Error TypesDeLIVeR Irrelevant Questions16% Insufficient Coverage34% Redundant Questions 14% Document Mismatch48% Table 6. Distribution of errors based on 200 examples from POLITIFACT, where DeLIVeR gives incorrect verification results. Decomposed Learning for Information-grounded Veracity Recognition15 Knowledge - ability Comprehensiveness Factuality Logical Coherence RelevanceCorrectness 20 40 60 80 100 55.1 56.4 59.4 40.3 67.5 58.8 69.3 65.0 72.2 52.8 75.1 70.2 83.2 75.12 82.45 68.4 83.2 75.25 Vanilla LLM Naive RAG LightRAG ReAct HippoRAG2 KG-RQG RAG 7B (ours) Fig. 5. Qualitative performance of our DeLIVeR framework in six categories. Error analysis of the retrieval process: We conduct error analysis on 200 PolitiFact failure cases where DeLIVeR predicts incorrect verdicts. Man- ual annotation reveals four error types (Table 4.3): Document Mismatch (48%) dominates, indicating poor alignment between questions and retrieved evidence, especially for nuanced or temporally sensitive claims. Insufficient Coverage (34%) ranks second, showing that critical aspects (e.g., motive, source credibility) are sometimes missed, particularly in “Mostly False”/“Half-True” claims. Encour- agingly, Irrelevant Questions (16%) and Redundant Questions (14%) together account for only 30%, confirming that GRPO effectively generates focused, non-repetitive questions. This supports our ablation finding that four ques- tions achieve optimal balance. These results highlight the need to (1) improve question-evidence alignment via retriever fine-tuning and (2) strengthen GRPO rewards to penalize Document Mismatch more heavily. Addressing these will significantly enhance retrieval-augmented fact-checking robustness. 5 Conclusion We presented a fact-verification architecture that integrates knowledge graph grounding and targeted, set-based question planning on top of a frozen verifier trained through GRPO to monotonically improve accuracy, interpretability, and robustness over LLM-only and recent RAG baselines. By decomposing claims into focused questions, retrieving precise multi-hop evidence, and rationalizing a 16Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen True/False/NEI decision with user-aligned feedback, our approach produces fully auditable reasoning traces suitable for high-stakes applications such as news- rooms, content moderation teams, election monitoring, public health communi- cation, and enterprise compliance workflows. Unlike black-box LLM judgments or noisy single-shot RAG retrievals, our method exposes every inference step from generated questions to retrieved KG paths to final verdict—enabling human auditors to verify or contest any component. This transparency is critical when decisions influence policy, public perception, or legal outcomes. We show that the combination of structured retrieval and reinforced question generation narrows the gap between raw model power and real-world trust requirements, achieving state-of-the-art verdict accuracy while delivering human-readable explanations at scale. Ultimately, this work advances toward deployable, accountable AI sys- tems capable of combating misinformation with both rigor and responsibility. 6 Acknowledgment This work was supported by NSF - USA CNS-2219614. Bibliography [1] Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y., Madotto, A. & Fung, P. Survey of hallucination in natural language generation. ACM Computing Surveys. 55, 1-38 (2023) [2] Mitchell, A., Jurkowitz, M., Oliphant, J. & Shearer, E. Americans who mainly get their news on social media are less engaged, less knowledgeable. Pew Research Center. 30 (2020) [3] Zhang, Y., Li, Y., Cui, L., Cai, D., Liu, L., Fu, T., Huang, X., Zhao, E., Zhang, Y., Chen, Y. & Others Siren’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models. Computational Linguistics. 51, 1373-1418 (2025) [4] Potthast, M., Kiesel, J., Reinartz, K., Bevendorff, J. & Stein, B. A stylometric inquiry into hyperpartisan and fake news. Proceedings Of The 56th Annual Meeting Of The Association For Computational Linguistics (volume 1: Long Papers). p. 231-240 (2018) [5] Qian, F., Gong, C., Sharma, K. & Liu, Y. Neural user response generator: Fake news detection with collective user intelligence.. IJCAI. 18 p. 3834- 3840 (2018) [6] Mosallanezhad, A., Karami, M., Shu, K., Mancenido, M. & Liu, H. Domain adaptive fake news detection via reinforcement learning. Proceedings Of The ACM Web Conference 2022. p. 3632-3640 (2022) [7] Khattab, O., Santhanam, K., Li, X., Hall, D., Liang, P., Potts, C. & Zaharia, M. Demonstrate-search-predict: Composing retrieval and language models for knowledge-intensive nlp. ArXiv Preprint ArXiv:2212.14024. (2022) [8] Press, O., Zhang, M., Min, S., Schmidt, L., Smith, N. & Lewis, M. Measuring and narrowing the compositionality gap in language models. Findings Of The Association For Computational Linguistics: EMNLP 2023. p. 5687-5711 (2023) [9] Chen, J., Sriram, A., Choi, E. & Durrett, G. Generating literal and implied subquestions to fact-check complex claims. Proceedings Of The 2022 Confer- ence On Empirical Methods In Natural Language Processing. p. 3495-3516 (2022) [10] Ousidhoum, N., Yuan, Z. & Vlachos, A. Varifocal question generation for fact-checking. Proceedings Of The 2022 Conference On Empirical Methods In Natural Language Processing. p. 2532-2544 (2022) [11] Yasunaga, M., Ren, H., Bosselut, A., Liang, P. & Leskovec, J. QA-GNN: Reasoning with language models and knowledge graphs for question answer- ing. Proceedings Of The 2021 Conference Of The North American Chapter Of The Association For Computational Linguistics: Human Language Tech- nologies. p. 535-546 (2021) [12] Shi, W., Min, S., Yasunaga, M., Seo, M., James, R., Lewis, M., Zettlemoyer, L. & Yih, W. Replug: Retrieval-augmented black-box language models. Pro- ceedings Of The 2024 Conference Of The North American Chapter Of The 18Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen Association For Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). p. 8371-8384 (2024) [13] Shi, T., Karpathy, A., Fan, L., Hernandez, J. & Liang, P. World of bits: An open-domain platform for web-based agents. International Conference On Machine Learning. p. 3135-3144 (2017) [14] Gur, I., Rueckert, U., Faust, A. & Hakkani-Tur, D. Learning to navigate the web. ArXiv Preprint ArXiv:1812.09195. (2018) [15] Jiang, Z., Xu, F., Gao, L., Sun, Z., Liu, Q., Dwivedi-Yu, J., Yang, Y., Callan, J. & Neubig, G. Active retrieval augmented generation. Proceedings Of The 2023 Conference On Empirical Methods In Natural Language Processing. p. 7969-7992 (2023) [16] Devlin, J., Chang, M., Lee, K. & Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. Proceedings Of The 2019 Conference Of The North American Chapter Of The Association For Computational Linguistics: Human Language Technologies, Volume 1 (long And Short Papers). p. 4171-4186 (2019) [17] Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L. & Stoyanov, V. Roberta: A robustly optimized bert pre- training approach. ArXiv Preprint ArXiv:1907.11692. (2019) [18] Zhang, X. & Gao, W. Towards llm-based fact verification on news claims with a hierarchical step-by-step prompting method. Proceedings Of The 13th International Joint Conference On Natural Language Processing And The 3rd Conference Of The Asia-pacific Chapter Of The Association For Com- putational Linguistics (volume 1: Long Papers). p. 996-1011 (2023) [19] Adolphs, L., Boerschinger, B., Buck, C., Huebscher, M., Ciaramita, M., Es- peholt, L., Hofmann, T., Kilcher, Y., Rothe, S., Sessa, P. & Others Boosting search engines with interactive agents. ArXiv Preprint ArXiv:2109.00527. (2021) [20] Yuan, X., Fu, J., Cote, M., Tay, Y., Pal, C. & Trischler, A. Interactive machine comprehension with information seeking agents. Proceedings Of The 58th Annual Meeting Of The Association For Computational Linguistics. p. 2325-2338 (2020) [21] Ziegler, D., Stiennon, N., Wu, J., Brown, T., Radford, A., Amodei, D., Christiano, P. & Irving, G. Fine-tuning language models from human pref- erences. ArXiv Preprint ArXiv:1909.08593. (2019) [22] Ibarz, B., Leike, J., Pohlen, T., Irving, G., Legg, S. & Amodei, D. Reward learning from human preferences and demonstrations in atari. Advances In Neural Information Processing Systems. 31 (2018) [23] Christiano, P., Leike, J., Brown, T., Martic, M., Legg, S. & Amodei, D. Deep reinforcement learning from human preferences. Advances In Neural Information Processing Systems. 30 (2017) [24] Lin, Y., Han, X., Xie, R., Liu, Z. & Sun, M. Knowledge representation learning: A quantitative review. ArXiv Preprint ArXiv:1812.10901. (2018) [25] Wang, Q., Mao, Z., Wang, B. & Guo, L. Knowledge graph embedding: A survey of approaches and applications. IEEE Transactions On Knowledge And Data Engineering. 29, 2724-2743 (2017) Decomposed Learning for Information-grounded Veracity Recognition19 [26] Chen, X., Jia, S. & Xiang, Y. A review: Knowledge reasoning over knowledge graph. Expert Systems With Applications. 141 p. 112948 (2020) [27] Wu, T., Qi, G., Li, C. & Wang, M. A survey of techniques for constructing Chinese knowledge graphs and their applications. Sustainability. 10, 3245 (2018) [28] Paulheim, H. Knowledge graph refinement: A survey of approaches and evaluation methods. Semantic Web. 8, 489-508 (2016) [29] Nickel, M., Murphy, K., Tresp, V. & Gabrilovich, E. A review of relational machine learning for knowledge graphs. Proceedings Of The IEEE. 104, 11-33 (2015) [30] Ehrlinger, L. & Wöß, W. Towards a definition of knowledge graphs.. SE- MANTiCS (Posters, Demos, SuCCESS). 48, 2 (2016) [31] Bonatti, P., Decker, S., Polleres, A. & Presutti, V. Knowledge graphs: New directions for knowledge representation on the semantic web (dagstuhl sem- inar 18371). Dagstuhl Reports. 8, 29-111 (2019) [32] Bergman, M. A COMMON SENSE VIEW OF KNOWLEDGE GRAPHS. (2019), https://api.semanticscholar.org/CorpusID:204957313 [33] Shu, K., Sliva, A., Wang, S., Tang, J. & Liu, H. Fake news detection on social media: A data mining perspective. ACM SIGKDD Explorations Newsletter. 19, 22-36 (2017) [34] Turchi, M., Negri, M. & Federico, M. MT Quality Estimation for Computer- assisted Translation: Does it Really Help?. Proceedings Of The 53rd Annual Meeting Of The Association For Computational Linguistics And The 7th International Joint Conference On Natural Language Processing (Volume 2: Short Papers). p. 530-535 (2015) [35] Wang, W. " liar, liar pants on fire": A new benchmark dataset for fake news detection. ArXiv Preprint ArXiv:1705.00648. (2017) [36] Thorne, J., Vlachos, A., Christodoulopoulos, C. & Mittal, A. FEVER: a large-scale dataset for fact extraction and VERification. ArXiv Preprint ArXiv:1803.05355. (2018) [37] Zhou, X. & Zafarani, R. A survey of fake news: Fundamental theories, de- tection methods, and opportunities. ACM Computing Surveys (CSUR). 53, 1-40 (2020) [38] Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küt- tler, H., Lewis, M., Yih, W., Rocktäschel, T. & Others Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances In Neural Informa- tion Processing Systems. 33 p. 9459-9474 (2020) [39] Guu, K., Lee, K., Tung, Z., Pasupat, P. & Chang, M. Retrieval augmented language model pre-training. International Conference On Machine Learn- ing. p. 3929-3938 (2020) [40] Salemi, A. & Zamani, H. Evaluating retrieval quality in retrieval-augmented generation. Proceedings Of The 47th International ACM SIGIR Confer- ence On Research And Development In Information Retrieval. p. 2395-2400 (2024) [41] Asai, A., Wu, Z., Wang, Y., Sil, A. & Hajishirzi, H. Self-rag: Learning to retrieve, generate, and critique through self-reflection. (ICLR,2024) 20Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen [42] Nakano, R., Hilton, J., Balaji, S., Wu, J., Ouyang, L., Kim, C., Hesse, C., Jain, S., Kosaraju, V., Saunders, W. & Others Webgpt: Browser-assisted question-answering with human feedback. ArXiv Preprint ArXiv:2112.09332. (2021) [43] Hogan, A., Blomqvist, E., Cochez, M., D’Amato, C., Melo, G., Gutierrez, C., Kirrane, S., Gayo, J., Navigli, R., Neumaier, S. & Others Knowledge graphs. ACM Computing Surveys (Csur). 54, 1-37 (2021) [44] Ji, S., Pan, S., Cambria, E., Marttinen, P. & Yu, P. A survey on knowledge graphs: Representation, acquisition, and applications. IEEE Transactions On Neural Networks And Learning Systems. 33, 494-514 (2021) [45] Hu, L., Yang, T., Zhang, L., Zhong, W., Tang, D., Shi, C., Duan, N. & Zhou, M. Compare to the knowledge: Graph neural fake news detection with external knowledge. Proceedings Of The 59th Annual Meeting Of The As- sociation For Computational Linguistics And The 11th International Joint Conference On Natural Language Processing (volume 1: Long Papers). p. 754-763 (2021) [46] Rajpurkar, P., Zhang, J., Lopyrev, K. & Liang, P. Squad: 100,000+ ques- tions for machine comprehension of text. ArXiv Preprint ArXiv:1606.05250. (2016) [47] Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V. & Zettlemoyer, L. BART: Denoising sequence-to-sequence pre- training for natural language generation, translation, and comprehension. ArXiv Preprint ArXiv:1910.13461. (2019) [48] Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W. & Liu, P. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal Of Machine Learning Research. 21, 1-67 (2020) [49] Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A. & Others Training language models to follow instructions with human feedback. Advances In Neural In- formation Processing Systems. 35 p. 27730-27744 (2022) [50] Schulman, J., Wolski, F., Dhariwal, P., Radford, A. & Klimov, O. Proximal policy optimization algorithms. ArXiv Preprint ArXiv:1707.06347. (2017) [51] Edge, D., Trinh, H., Cheng, N., Bradley, J., Chao, A., Mody, A., Tru- itt, S., Metropolitansky, D., Ness, R. & Larson, J. From local to global: A graph rag approach to query-focused summarization. ArXiv Preprint ArXiv:2404.16130. (2024) [52] Reimers, N. & Gurevych, I. Sentence-bert: Sentence embeddings using siamese bert-networks. ArXiv Preprint ArXiv:1908.10084. (2019) [53] Yang, Z., Qi, P., Zhang, S., Bengio, Y., Cohen, W., Salakhutdinov, R. & Manning, C. HotpotQA: A dataset for diverse, explainable multi-hop ques- tion answering. ArXiv Preprint ArXiv:1809.09600. (2018) [54] Shu, K., Mahudeswaran, D., Wang, S., Lee, D. & Liu, H. Fakenewsnet: A data repository with news content, social context, and spatiotemporal information for studying fake news on social media. Big Data. 8, 171-188 (2020) Decomposed Learning for Information-grounded Veracity Recognition21 [55] Guo, Z., Xia, L., Yu, Y., Ao, T. & Huang, C. Lightrag: Simple and fast retrieval-augmented generation. ArXiv Preprint ArXiv:2410.05779. (2024) [56] Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K. & Cao, Y. React: Synergizing reasoning and acting in language models. International Conference On Learning Representations (ICLR). (2023) [57] Jimenez Gutierrez, B., Shu, Y., Gu, Y., Yasunaga, M. & Su, Y. Hipporag: Neurobiologically inspired long-term memory for large language models. Ad- vances In Neural Information Processing Systems. 37 p. 59532-59569 (2024) [58] Team, Q. & Others Qwen2 technical report. ArXiv Preprint ArXiv:2407.10671. 2 p. 3 (2024) [59] Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D. & Liu, Z. Bge m3- embedding: Multi-lingual, multi-functionality, multi-granularity text embed- dings through self-knowledge distillation. ArXiv Preprint ArXiv:2402.03216. (2024) [60] Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S. & Others Gpt-4 technical report. ArXiv Preprint ArXiv:2303.08774. (2023) [61] Allcott, H. & Gentzkow, M. Social media and fake news in the 2016 election. Journal Of Economic Perspectives. 31, 211-236 (2017) [62] Chen, M. Evaluating large language models trained on code. ArXiv Preprint ArXiv:2107.03374. (2021) [63] Zellers, R., Holtzman, A., Rashkin, H., Bisk, Y., Farhadi, A., Roesner, F. & Choi, Y. Defending against neural fake news. Advances In Neural Information Processing Systems. 32 (2019) [64] Bang, Y., Cahyawijaya, S., Lee, N., Dai, W., Su, D., Wilie, B., Lovenia, H., Ji, Z., Yu, T., Chung, W. & Others A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. ArXiv Preprint ArXiv:2302.04023. (2023) [65] Vrandečić, D. & Krötzsch, M. Wikidata: a free collaborative knowledgebase. Communications Of The ACM. 57, 78-85 (2014) [66] Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y. & Others Deepseekmath: Pushing the limits of mathemati- cal reasoning in open language models. ArXiv Preprint ArXiv:2402.03300. (2024) [67] Ngai, C., Singh, R. & Yao, L. Impact of COVID-19 vaccine misinformation on social media virality: content analysis of message themes and writing strategies. Journal Of Medical Internet Research. 24, e37806 (2022) [68] Vivion, M., Trottier, V., Bouhêlier, È., Goupil-Sormany, I., Diallo, T. & Others Misinformation about climate change and related environmental events on social media: Protocol for a scoping review. JMIR Research Pro- tocols. 13, e59345 (2024) [69] Govindankutty, S. & Gopalan, S. Epidemic modeling for misinformation spread in digital networks through a social intelligence approach. Scientific Reports. 14, 19100 (2024) [70] Aïmeur, E., Amri, S. & Brassard, G. Fake news, disinformation and mis- information in social media: a review. Social Network Analysis And Mining. 13, 30 (2023) 22Cong Hoan Nguyen, Thomas Hoang, Minh Hieu Duong, and Long Nguyen [71] Vosoughi, S., Roy, D. & Aral, S. The spread of true and false news online. Science. 359, 1146-1151 (2018) [72] Gao, Y., Jiang, X., Kasihmuddin, M., Zheng, C., Chen, J., Liu, X. & Guo, Y. FGRA: Toward flexible logic mining with ensemble multi-attribute selection and Discrete Hopfield Neural Network. Journal Of Computational Design And Engineering. 13, 88-107 (2026) [73] Chang, Y., Kasihmuddin, M., Ruzai, W., Guo, Y. & Chen, J. Weighted C- type random 2 satisfiability in discrete hopfield neural network. Engineering Applications Of Artificial Intelligence. 160 p. 111760 (2025) [74] Romli, N., Zulkepli, N., Kasihmuddin, M., Karim, S., Jamaludin, S., Rusdi, N., Manoharam, G., Mansor, M. & Zamri, N. An optimized logic mining method for data processing through higher-order satisfiability representation in discrete Hopfield neural network. Applied Soft Computing. p. 113759 (2025)