PaperlayerView exact revision on arXiv

Source-linked AI summary

TruthfulRAG: Resolving Factual-level Conflicts in Retrieval-Augmented Generation with Knowledge Graphs

Shuyi Liu, Yuming Shang, Xi Zhang

arXiv:2511.10375v1cs.CL

TL;DR

RAG systems can face factual conflicts between retrieved information and LLMs’ parametric knowledge, while existing token- or semantic-level methods may inadequately capture fine-grained discrepancies. TruthfulRAG uses knowledge graphs, query-based retrieval, and entropy filtering to address these conflicts. The paper reports that it consistently outperforms existing state-of-the-art methods, with accuracy improvements of 6.8% to 17.7% over standard RAG.

  • Problem

    Existing token- and semantic-level conflict resolution can rely on fragmented representations that inadequately capture fine-grained factual discrepancies, particularly in knowledge-intensive scenarios.

    Source: Introduction · p. 2

  • Method

    TruthfulRAG extracts triples into knowledge graphs, retrieves query-relevant reasoning paths, and uses entropy-based filtering to identify and mitigate factual conflicts.

    Source: Introduction · p. 2, Graph Retrieval · p. 2, Conflict Resolution · p. 4

  • Results

    TruthfulRAG consistently outperforms existing state-of-the-art methods; its accuracy improvements range from 6.8% to 17.7% compared with standard RAG.

    Source: Ablation Study · p. 6, Conclusion · p. 7

  • Takeaways & Limitations

    TruthfulRAG aims to guide LLMs toward updated, comprehensive external knowledge while correcting temporal and factual errors in internal memory.

    Source: Impact Analysis of Knowledge Conflicts · p. 7

Abstract

from arXiv · show

Retrieval-Augmented Generation (RAG) has emerged as a powerful framework for enhancing the capabilities of Large Language Models (LLMs) by integrating retrieval-based methods with generative models. As external knowledge repositories continue to expand and the parametric knowledge within models becomes outdated, a critical challenge for RAG systems is resolving conflicts between retrieved external information and LLMs' internal knowledge, which can significantly compromise the accuracy and reliability of generated content. However, existing approaches to conflict resolution typically operate at the token or semantic level, often leading to fragmented and partial understanding of factual discrepancies between LLMs' knowledge and context, particularly in knowledge-intensive tasks. To address this limitation, we propose TruthfulRAG, the first framework that leverages Knowledge Graphs (KGs) to resolve factual-level knowledge conflicts in RAG systems. Specifically, TruthfulRAG constructs KGs by systematically extracting triples from retrieved content, utilizes query-based graph retrieval to identify relevant knowledge, and employs entropy-based filtering mechanisms to precisely locate conflicting elements and mitigate factual inconsistencies, thereby enabling LLMs to generate faithful and accurate responses. Extensive experiments reveal that TruthfulRAG outperforms existing methods, effectively alleviating knowledge conflicts and improving the robustness and trustworthiness of RAG systems.

Introduction

RAG can encounter conflicts between changing external sources and LLMs’ static parametric knowledge, undermining response accuracy and reliability. TruthfulRAG addresses the limits of coarse-grained conflict resolution with a knowledge-graph framework for factual-level conflicts.

Introduction

  • Conflicts between dynamic external repositories and static LLM knowledge can undermine the accuracy and reliability of generated content.RAG combines external retrieval with LLM generation, but the two knowledge sources can disagree.Source: Introduction · p. 1
  • Token-level and semantic-level resolution methods use coarse, fragmented representations that can miss factual inconsistencies and complex interdependencies.The stated limitation is especially relevant to knowledge-intensive conflict scenarios.Source: Introduction · pp. 1–2
  • TruthfulRAG uses knowledge graphs to resolve factual-level RAG conflicts through triple-based graph construction, query-based retrieval, and entropy-based filtering.The framework is presented as the first to use knowledge graphs for this purpose.Source: Introduction · p. 2
  • The paper reports that TruthfulRAG outperforms existing methods in mitigating conflicts while improving RAG robustness and trustworthiness.The contribution summary describes the findings as based on extensive experiments.Source: Introduction · p. 2

Methodology

TruthfulRAG converts retrieved content into a knowledge graph, retrieves query-relevant reasoning paths, and applies entropy-based conflict resolution to address factual inconsistencies.

Graph Construction

  • For graph construction, the system segments retrieved content and extracts entity-relation-tail triples, including explicit facts and implicit semantic relationships.The aggregated triples form a structured graph that filters low-information noise and captures factual associations.Source: Graph Construction · p. 2

Graph Retrieval

  • Query-aware retrieval matches query entities, relations, and intent to top-k graph elements, explores two-hop paths, and scores paths by entity and relation coverage.The top-scored paths form the core knowledge paths.Source: Graph Retrieval · pp. 2–3

Methodology

  • TruthfulRAG links graph construction, query-aware graph retrieval, and entropy-based conflict resolution to detect and mitigate factual inconsistencies.The pipeline uses the resulting structured reasoning paths as context for generation.Source: Methodology · p. 2, Figure 2 · p. 3

Graph Retrieval

  • Selected paths are represented with their entity-relation sequence and attribute descriptions for important entities and relations.This representation is intended to preserve structural coherence and enrich context for conflict resolution.Source: Graph Retrieval · p. 3

Conflict Resolution

  • Conflict resolution compares parametric and path-conditioned entropy, treating entropy changes above a threshold as corrective paths for the refined context.The paper describes positive entropy changes as potentially indicating factual inconsistency and negative changes as suggesting alignment.Source: Conflict Resolution · p. 4

Experiments

The experiments are designed to evaluate TruthfulRAG’s factual accuracy, behavior in non-conflicting contexts, effects of structured paths on model confidence, and the contributions of its modules.

Experiments

  • The experiments assess TruthfulRAG’s performance against other methods on factual accuracy and its performance in non-conflicting contexts.These are the first two research questions listed for the evaluation.Source: Experiments · p. 4
  • The evaluation asks how structured reasoning paths affect LLM confidence compared with raw natural-language context.The passage states this as an experimental question, not a reported outcome.Source: Experiments · p. 4
  • The experiments also examine the individual contributions of TruthfulRAG’s modules.The passage identifies this as a research question without reporting its answer.Source: Experiments · p. 4

Experimental Setup

The experiments cover four knowledge-intensive datasets, three LLM backbones, five comparison baselines, and measures of answer accuracy and context precision. Implementation settings specify model-dependent entropy thresholds and fixed generation and retrieval parameters.

Datasets

  • FaithEval tests complex logical conflicts; MuSiQue and SQuAD assess fact-level conflicts requiring multi-hop reasoning; RealtimeQA focuses on temporal conflicts.The four datasets cover unanswerable, inconsistent, or counterfactual contexts, compositional reasoning, and outdated answers against dynamic sources.Source: Datasets · p. 4

Evaluated Models

  • The evaluation uses GPT-4o-mini, Qwen2.5-7B-Instruct, and Mistral-7B-Instruct, spanning open- and closed-source models.Source: Evaluated Models · p. 4

Baselines

Evaluation Metrics

  • Accuracy (ACC) measures correct-answer proportion, while CPR measures answer-related content in processed context; Table 1 reports ACC across datasets and backbones.CPR relates answer-relevant content to total processed-context length; Table 1 also reports average accuracy and improvement over each model’s w/o RAG baseline.Source: Evaluation Metrics · pp. 4–5, Table 1 · p. 5

Implementation Details

  • Dense retrieval uses all-MiniLM-L6-v2 embeddings; entropy thresholds are 1 for GPT-4o-mini and Mistral-7B-Instruct and 3 for Qwen2.5-7B-Instruct.Experiments use NVIDIA V100 GPUs with 32GB memory, generation temperature 0, and Top-K 10.Source: Implementation Details · p. 5

Results and Analysis

TruthfulRAG achieves the highest overall average accuracy and improvement across backbone LLMs, with reported gains over standard RAG. The results also report performance in non-conflicting contexts, confidence with structured reasoning paths, and component ablations.

Overall Performance

  • TruthfulRAG records 81.9% on FaithEval, 79.4% on MuSiQue, and 85.0% on RealtimeQA, the highest accuracies on those datasets.It ranks first or second on SQuAD across models.Source: Overall Performance · p. 5
  • TruthfulRAG achieves accuracy improvements of 3.6% to 29.2% over standard RAG and the best Avg. and Imp. across all backbone LLMs.The authors describe its performance as consistently superior or competitive against baselines.Source: Overall Performance · p. 5

Performance on Non-Conflicting Contexts

  • TruthfulRAG outperforms all baselines on MuSiQue-golden and SQuAD-golden, where retrieved passages are guaranteed non-contradictory.The authors attribute the gains to graph representations that identify fine-grained entities and relations, aiding query-relevant extraction.Source: Performance on Non-Conflicting Contexts · pp. 5–6, Table 2 · p. 6

Impact of Structured Reasoning Paths

  • Across four datasets, structured reasoning paths yield higher model confidence for correct answers than natural-language contexts.Figure 3 compares confidence using negative log-probability; lower values indicate higher actual log-probability and confidence.Source: Impact of Structured Reasoning Paths · p. 6, Figure 3 · p. 6

Ablation Study

  • The full framework improves accuracy by 6.8% to 17.7% over standard RAG; without the graph, CPR falls on most datasets, while graph-only context can add redundant information.The ablations combine graph construction and retrieval as one component; graph-only increases CPR but yields limited accuracy improvements across most datasets.Source: Ablation Study · pp. 6–7, Table 3 · p. 7

Related Work

The review covers the effects of knowledge conflicts and existing resolution strategies. It positions TruthfulRAG as a graph-based approach to resolving factual-level conflicts.

Related Work

  • The review organizes prior research around knowledge-conflict impacts and resolution strategies.These are the two areas covered in the section.Source: Related Work · p. 7

Impact Analysis of Knowledge Conflicts

  • Impact studies examine differences in LLM preferences for parametric knowledge and retrieved external information.TruthfulRAG aims to favor updated, comprehensive external knowledge while correcting temporal and factual errors in internal memory.Source: Impact Analysis of Knowledge Conflicts · p. 7

Solutions to Knowledge Conflicts

  • Token- and semantic-level methods address surface conflicts but do not capture the underlying factual relationships driving inconsistencies.TruthfulRAG instead uses structured triple-based representations to identify and resolve factual-level conflicts in complex expressions.Source: Solutions to Knowledge Conflicts · p. 7

Conclusion

TruthfulRAG applies knowledge graphs to factual-level conflicts in RAG through triple extraction, graph retrieval, and entropy-based filtering. The authors report consistent gains over existing state-of-the-art methods and conclude it is a robust, generalizable solution for improving trustworthiness and accuracy.

Conclusion

  • TruthfulRAG is presented as the first knowledge-graph framework for addressing factual-level conflicts in RAG.It combines systematic triple extraction, query-aware graph retrieval, and entropy-based filtering to form structured reasoning paths.Source: Conclusion · p. 7
  • The authors report that TruthfulRAG consistently outperforms existing state-of-the-art methods.Source: Conclusion · p. 7
  • The authors conclude TruthfulRAG is a robust, generalizable solution for improving RAG trustworthiness and accuracy in knowledge-intensive applications requiring high reliability and precision.Source: Conclusion · p. 7

Problem Statement

The paper defines a RAG knowledge conflict as a query-relevant factual inconsistency between an LLM’s parametric knowledge and retrieved external knowledge. Its objective is to resolve such conflicts while maintaining generation quality and transparent reasoning.

Problem Statement

  • A standard RAG system retrieves documents from an external knowledge base for a query and generates a response using both the query and retrieved context.The LLM also has parametric knowledge acquired during pre-training.Source: Problem Statement · p. 9
  • A conflict occurs when query-relevant factual statements in parametric and retrieved external knowledge are inconsistent.The paper formalizes this as fp ∈ Kp and fe ∈ Ke with fp not equivalent to fe.Source: Problem Statement · p. 9
  • The framework aims to identify and resolve conflicts while maintaining generation quality and transparent reasoning.The stated challenges are representing facts, retrieving and prioritizing relevant information, and making reliable decisions amid conflicting evidence.Source: Problem Statement · p. 9

Case Study

A MuSiQue case study with GPT-4o-mini traces TruthfulRAG through graph construction, retrieval, and conflict resolution. Its example shows the stages isolating corrective knowledge and producing the answer “Sinaloa.”

Case Study

  • In the MuSiQue case study with GPT-4o-mini, Table 9 details intermediate outputs across TruthfulRAG’s processing stages.The case illustrates systematic conflict identification and resolution for consistent, faithful reasoning.Source: Case Study · p. 9
  • Graph construction extracts entity-relation-attribute triples from retrieved context to create a semantically enriched knowledge base.The example graph includes Ciudad Deportiva, the Municipality of Nuevo Laredo, Nuevo Laredo, and Sinaloa.Source: Case Study · p. 9
  • Query-aware graph retrieval selects semantically aligned reasoning paths, including a path from the Municipality of Nuevo Laredo through Nuevo Laredo to Sinaloa.The paths include entity attributes and relational descriptions for semantic coherence and factual completeness.Source: Case Study · p. 9
  • Entropy filtering identifies the Municipality of Nuevo Laredo–Sinaloa path as potentially conflicting, isolates corrective knowledge, and yields “Sinaloa.”The example attributes the potential conflict to a significant entropy increase along that path.Source: Case Study · p. 9

Algorithm Overview

Algorithm 1 transforms raw retrieval context into structured reasoning paths and uses entropy-based confidence filtering to improve the factual consistency of generated responses.

  • Algorithm 1 transforms raw retrieval context into structured reasoning paths.The framework uses entropy-based confidence filtering to improve the factual consistency of model generation.Source: Algorithm Overview · p. 9
  • Entropy-based confidence filtering is the mechanism used to improve factual consistency in model generation.Source: Algorithm Overview · p. 9
  • The algorithm presents the complete TruthfulRAG framework for converting retrieval context into structured reasoning paths.Source: Algorithm Overview · p. 9

Additional Experiments

The additional experiments test threshold robustness, statistical significance, performance on stronger LLMs, and computational cost. They report comparable performance under a unified threshold, significant gains over FaithfulRAG, accuracy improvements on larger models, and moderate overhead alongside practical efficiency and compact contexts.

Additional Experiments

  • The additional experiments cover hyperparameter robustness, significance testing, evaluation on SOTA models, and computational cost.Source: Additional Experiments · p. 9

Hyperparameter Robustness

Significance Testing

Evaluation on SOTA LLMs

Computational Cost Analysis

  • TruthfulRAG adds moderate computational overhead versus FaithfulRAG while maintaining practical efficiency and compact contextual representations.The reported overhead is primarily due to graph-based reasoning and entropy filtering.Source: Computational Cost Analysis · p. 10, Table 7 · p. 11, Table 8 · p. 11
  • The framework builds a graph from extracted triples, retrieves query-relevant paths, and scores paths using fact-aware relevance.It then filters paths by entropy change before generating the response.Source: Computational Cost Analysis · p. 10

Related papers

Loading 2511.10375v1…