pith. sign in

arxiv: 2606.06748 · v2 · pith:QV3BPOCUnew · submitted 2026-06-04 · 💻 cs.CL · cs.AI· cs.LG

Evidence Graph Consistency in Retrieval-Augmented Generation: A Model-Dependent Analysis of Hallucination Detection

Jianru Shen This is my paper

Pith reviewed 2026-06-30 10:49 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.LG
keywords hallucination detectionretrieval-augmented generationevidence graphstructural consistencymodel dependenceRAGTruthlarge language models
0
0 comments X

The pith

Evidence graph consistency detects hallucinations in Llama-2 but reverses direction in GPT and Mistral models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes Evidence Graph Consistency as a way to detect hallucinations in retrieval-augmented generation by building a local graph that links retrieved evidence passages to claims in the generated answer and then extracting five structural consistency measures from that graph. Tested across thousands of responses from six different large language models on the RAGTruth dataset, the measures behave as expected for hallucination detection only in the Llama-2 family while showing the opposite pattern in GPT-4, GPT-3.5, and Mistral-7B. This split implies that hallucination behavior is not uniform across models and that graph-based consistency signals cannot be treated as reliable without reference to the specific model family. A reader would care because current hallucination detectors often assume they can work the same way for any LLM, yet the results suggest that assumption fails in practice.

Core claim

The authors construct a local evidence graph for each response and compute five structural consistency measures as potential hallucination indicators. On the full question-answering split of RAGTruth, these measures align with the expected direction for hallucinations in Llama-2 models but exhibit systematic reversal in GPT-4, GPT-3.5, and Mistral-7B. The reversal indicates qualitatively different hallucination patterns across model families and shows that embedding-based graph consistency cannot function as a model-independent detection signal.

What carries the argument

The Evidence Graph Consistency (EGC) framework, which builds a local evidence graph per response and derives five structural consistency measures from the connections between evidence pieces and answer claims.

If this is right

  • Hallucination detection methods based on graph consistency must be validated separately for each model family rather than assumed to transfer.
  • Qualitatively different hallucination patterns exist between the Llama-2 family and the GPT/Mistral families.
  • Embedding-based structural signals from evidence graphs lose diagnostic value when applied across model families.
  • RAG systems using multiple model families require family-specific hallucination checks rather than a single shared detector.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Model providers could publish family-specific calibration data for graph-based detectors to improve reliability.
  • The reversal might stem from differences in how models integrate retrieved evidence during generation, which could be tested by comparing attention patterns over evidence.
  • Alternative graph definitions that weight claims by model-generated probability might reduce the observed model dependence.

Load-bearing premise

The way the local evidence graph is built and the five consistency measures are calculated does not itself create different connection patterns depending on which model family generated the answer.

What would settle it

Re-running the same graph construction and five measures on responses from a new collection of models that includes both Llama-style and GPT-style families and finding no reversal or model-family split would falsify the claim that the behavior is model-dependent.

Figures

Figures reproduced from arXiv: 2606.06748 by Jianru Shen.

Figure 1
Figure 1. Figure 1: Evidence graph structure for a grounded answer (left) and a hallucinated answer (right) from Llama-2-13B. In the grounded case all claim nodes [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Left: mean EGC score by model and label. Right: per-model diagnostic gap [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: Left: mean EGC score by model and label. Right: per-model diagnostic gap [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: EGC feature distributions for grounded and hallucinated answers across all models (top five panels), and per-model AUROC (bottom right). GPT-3.5 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
read the original abstract

Retrieval-Augmented Generation (RAG) reduces but does not eliminate hallucination in large language models. Existing detection methods rely on flat similarity between generated answers and retrieved passages, ignoring structural relationships among evidence pieces and answer claims. We propose Evidence Graph Consistency (EGC), a framework that constructs a local evidence graph per response and computes five structural consistency measures as hallucination indicators. Evaluated on the full question answering split of RAGTruth across six LLMs (5,767 responses), EGC reveals a consistent model-family split: graph consistency features show the expected diagnostic direction for hallucinations in Llama-2 models but exhibit systematic reversal in GPT-4, GPT-3.5, and Mistral-7B. This reversal suggests qualitatively different hallucination patterns across model families and indicates that embedding-based graph consistency cannot serve as a model-independent hallucination detection signal.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes Evidence Graph Consistency (EGC), a framework that builds a local evidence graph per RAG response and derives five structural consistency measures to detect hallucinations. Evaluated on the full QA split of RAGTruth (5,767 responses across six LLMs), EGC exhibits the expected diagnostic direction for Llama-2 models but a systematic reversal for GPT-4, GPT-3.5, and Mistral-7B, leading to the conclusion that embedding-based graph consistency cannot serve as a model-independent hallucination signal and that hallucination patterns differ qualitatively across model families.

Significance. If the reported reversal is robust to the graph-construction pipeline, the result would be significant for hallucination detection research: it supplies concrete empirical evidence against model-agnostic assumptions in current RAG verification methods and motivates family-specific detectors. The evaluation on a public benchmark with a large response count is a strength.

major comments (2)
  1. [Abstract] Abstract: the central claim of a model-family reversal rests on the five structural consistency measures being computed identically across LLMs. No verification is supplied that claim extraction, relation detection, or edge formation steps are invariant to known model-family differences in output length, fluency, and sentence structure; if these steps embed such differences, the reversal could be an artifact of the measurement pipeline rather than evidence of distinct hallucination mechanisms.
  2. [Evaluation] Evaluation section (implied by the 5,767-response count): the manuscript reports the split but does not describe the exact statistical procedure used to establish that the reversal is systematic across the three non-Llama families (e.g., per-measure sign tests, family-level interaction terms, or correction for multiple comparisons). Without these details the load-bearing claim that the pattern is qualitative rather than noise remains under-supported.
minor comments (2)
  1. [Abstract] The abstract refers to 'embedding-based graph consistency' without clarifying whether the graph edges themselves are embedding-driven or purely syntactic; this notation should be made explicit in the methods.
  2. The paper would benefit from a short table listing the five structural consistency measures with their precise definitions and formulas.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. The two major comments identify areas where additional verification and statistical detail would strengthen the manuscript. We address each point below and commit to revisions that directly respond to the concerns raised.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of a model-family reversal rests on the five structural consistency measures being computed identically across LLMs. No verification is supplied that claim extraction, relation detection, or edge formation steps are invariant to known model-family differences in output length, fluency, and sentence structure; if these steps embed such differences, the reversal could be an artifact of the measurement pipeline rather than evidence of distinct hallucination mechanisms.

    Authors: We agree that explicit verification of pipeline invariance is necessary to support the claim that the observed reversal reflects model-family differences in hallucination mechanisms rather than measurement artifacts. The current pipeline applies a uniform embedding-based similarity threshold for edge formation and a fixed claim-extraction procedure to all responses. In the revised version we will add a dedicated subsection that reports (i) average graph statistics (node count, edge density) broken down by model family, (ii) a sensitivity analysis varying the similarity threshold, and (iii) a qualitative comparison of extracted claims from Llama-2 versus GPT-family outputs. These additions will either confirm invariance or quantify any residual model-specific effects. revision: yes

  2. Referee: [Evaluation] Evaluation section (implied by the 5,767-response count): the manuscript reports the split but does not describe the exact statistical procedure used to establish that the reversal is systematic across the three non-Llama families (e.g., per-measure sign tests, family-level interaction terms, or correction for multiple comparisons). Without these details the load-bearing claim that the pattern is qualitative rather than noise remains under-supported.

    Authors: We acknowledge that the manuscript currently presents the directional reversal descriptively without formal statistical tests. In the revision we will add an explicit statistical analysis subsection that reports (a) per-measure sign tests comparing correlation signs between Llama-2 and the other three families, (b) a family-level interaction term in a mixed-effects model treating model family as a factor, and (c) Bonferroni correction for the five measures. The results of these tests will be included in a new table and discussed in the text. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical reporting on external benchmark without self-referential reduction

full rationale

The paper constructs an evidence graph and five consistency measures from an external RAGTruth dataset and reports observed empirical patterns across model families. No equations, parameters, or claims are defined in terms of the target result (model-family reversal), nor are any 'predictions' fitted to subsets and then re-reported as outputs. No self-citations or uniqueness theorems are invoked. The analysis is self-contained against the benchmark and does not reduce by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract; no explicit free parameters, axioms, or invented entities are detailed beyond the introduction of the EGC framework itself.

pith-pipeline@v0.9.1-grok · 5677 in / 1093 out tokens · 39061 ms · 2026-06-30T10:49:13.072678+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 7 canonical work pages · 5 internal anchors

  1. [1]

    Retrieval-augmented generation for knowledge-intensive NLP tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal,et al., “Retrieval-augmented generation for knowledge-intensive NLP tasks,” in Advances in Neural Information Processing Systems, 2020, pp. 9459– 9474

    2020

  2. [2]

    RAGTruth: A hallucination corpus for developing trustworthy retrieval-augmented language models,

    C. Niu, Y . Wu, J. Zhu, S. Xu, K. Shum, R. Zhong, J. Song, and T. Zhang, “RAGTruth: A hallucination corpus for developing trustworthy retrieval-augmented language models,” inProc. 62nd Annual Meeting of the Association for Computational Linguistics, Bangkok, Thailand, 2024, pp. 10862–10878

    2024

  3. [3]

    RAGAs: Automated evaluation of retrieval augmented generation,

    S. Es, J. James, L. Espinosa Anke, and S. Schockaert, “RAGAs: Automated evaluation of retrieval augmented generation,” inProc. 18th Conference of the European Chapter of the Association for Computa- tional Linguistics, 2024, pp. 150–158

    2024

  4. [4]

    FActScore: Fine-grained atomic evaluation of factual precision in long form text generation,

    S. Min, K. Krishna, X. Lyu, M. Lewis, W. Tau Yih, P. Koh, M. Iyyer, L. Zettlemoyer, and H. Hajishirzi, “FActScore: Fine-grained atomic evaluation of factual precision in long form text generation,” inProc. 2023 Conference on Empirical Methods in Natural Language Processing, Singapore, 2023, pp. 12076–12100

    2023

  5. [5]

    SelfCheckGPT: Zero-resource black-box hallucination detection for generative large language models,

    P. Manakul, A. Liusie, and M. Gales, “SelfCheckGPT: Zero-resource black-box hallucination detection for generative large language models,” inProc. 2023 Conference on Empirical Methods in Natural Language Processing, Singapore, 2023, pp. 9004–9017

    2023

  6. [6]

    HotpotQA: A dataset for diverse, explainable multi-hop question answering,

    Z. Yang, P. Qi, S. Zhang, Y . Bengio, W. W. Cohen, R. Salakhutdinov, and C. D. Manning, “HotpotQA: A dataset for diverse, explainable multi-hop question answering,” inProc. 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, 2018, pp. 2369–2380

    2018

  7. [7]

    Sentence-BERT: Sentence embeddings using Siamese BERT-networks,

    N. Reimers and I. Gurevych, “Sentence-BERT: Sentence embeddings using Siamese BERT-networks,” inProc. 2019 Conference on Empirical Methods in Natural Language Processing, Hong Kong, China, 2019, pp. 3982–3992

    2019

  8. [8]

    spaCy: Industrial-strength natural language processing in Python,

    M. Honnibal, I. Montani, S. Van Landeghem, and A. Boyd, “spaCy: Industrial-strength natural language processing in Python,” Explosion AI, Tech. Rep., 2020

    2020

  9. [9]

    Exploring network structure, dynamics, and function using NetworkX,

    A. A. Hagberg, D. A. Schult, and P. J. Swart, “Exploring network structure, dynamics, and function using NetworkX,” inProc. 7th Python in Science Conference, 2008

    2008

  10. [10]

    MS MARCO: A human generated machine reading comprehension dataset,

    T. Nguyen, M. Rosenberg, X. Song, J. Gao, S. Tiwary, R. Majumder, and L. Deng, “MS MARCO: A human generated machine reading comprehension dataset,” inProc. Workshop on Cognitive Computation: Integrating Neural and Symbolic Approaches, vol. 1773, 2016

    2016

  11. [11]

    Survey of hallucination in natural language generation,

    Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y . Xu, E. Ishii, Y . Bang, A. Madotto, and P. Fung, “Survey of hallucination in natural language generation,”ACM Computing Surveys, vol. 55, no. 12, pp. 1–38, 2023

    2023

  12. [12]

    Wizard of Wikipedia: Knowledge-powered conversational agents,

    E. Dinan, S. Roller, K. Shuster, A. Fan, M. Auli, and J. Weston, “Wizard of Wikipedia: Knowledge-powered conversational agents,” in Proc. International Conference on Learning Representations, 2019

    2019

  13. [13]

    On faithfulness and factuality in abstractive summarisation,

    J. Maynez, S. Narayan, B. Bohnet, and R. McDonald, “On faithfulness and factuality in abstractive summarisation,” inProc. 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 1906–1919

    2020

  14. [14]

    Scikit-learn: Machine learning in Python,

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay, “Scikit-learn: Machine learning in Python,”Journal of Machine Learn- ing Research, vol. 12, 2011

    2011

  15. [15]

    A large annotated corpus for learning natural language inference,

    S. R. Bowman, G. Angeli, C. Potts, and C. D. Manning, “A large annotated corpus for learning natural language inference,” inProc. 2015 Conference on Empirical Methods in Natural Language Processing, Lisbon, Portugal, 2015, pp. 632–642

    2015

  16. [16]

    Retrieval augmentation reduces hallucination in conversation,

    K. Shuster, S. Poff, M. Chen, D. Kiela, and J. Weston, “Retrieval augmentation reduces hallucination in conversation,” inFindings of the Association for Computational Linguistics: EMNLP 2021, 2021, pp. 3784–3803

    2021

  17. [17]

    C. J. Van Rijsbergen,Information Retrieval, 2nd ed. Butterworth- Heinemann, 1979

    1979

  18. [18]

    Retrieval-Augmented Generation for Large Language Models: A Survey

    Y . Gao, Y . Xiong, X. Gao, K. Jia, J. Pan, Y . Bi, Y . Dai, J. Sun, M. Wang, and H. Wang, “Retrieval-augmented generation for large language models: A survey,”arXiv preprint arXiv:2312.10997, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  19. [19]

    BERT: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” inProc. 2019 Conference of the North American Chapter of the Association for Computational Linguistics, Minneapolis, MN, 2019, pp. 4171–4186

    2019

  20. [20]

    QA- GNN: Reasoning with language models and knowledge graphs for question answering,

    M. Yasunaga, H. Ren, A. Bosselut, P. Liang, and J. Leskovec, “QA- GNN: Reasoning with language models and knowledge graphs for question answering,” inProc. 2021 Conference of the North American Chapter of the Association for Computational Linguistics, 2021, pp. 535–546

    2021

  21. [21]

    LRP4RAG: Detecting hal- lucinations in retrieval-augmented generation via layer-wise relevance propagation,

    H. Hu, C. He, X. Xie, and Q. Zhang, “LRP4RAG: Detecting hal- lucinations in retrieval-augmented generation via layer-wise relevance propagation,” unpublished, arXiv:2408.15533, 2024

    work page arXiv 2024

  22. [22]

    A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions

    L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qin, and T. Liu, “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,” arXiv preprint arXiv:2311.05232, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  23. [23]

    D. W. Hosmer, S. Lemeshow, and R. X. Sturdivant,Applied Logistic Regression, 3rd ed. Wiley, 2013

    2013

  24. [24]

    Distributed representations of words and phrases and their compositionality,

    T. Mikolov, I. Sutskever, K. Chen, G. Corrado, and J. Dean, “Distributed representations of words and phrases and their compositionality,” in Advances in Neural Information Processing Systems, vol. 2, 2013, pp. 3111–3119

    2013

  25. [25]

    ARES: An automated evaluation framework for retrieval-augmented generation systems,

    J. Saad-Falcon, O. Khattab, C. Potts, and M. Zaharia, “ARES: An automated evaluation framework for retrieval-augmented generation systems,”arXiv preprint arXiv:2311.09476, 2023

    work page arXiv 2023

  26. [26]

    Ranking generated summaries by correctness: An interesting but chal- lenging application for natural language inference,

    T. Falke, L. F. R. Ribeiro, P. A. Utama, I. Dagan, and I. Gurevych, “Ranking generated summaries by correctness: An interesting but chal- lenging application for natural language inference,” inProc. 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, 2019, pp. 2214–2220

    2019

  27. [27]

    RAG-HAT: A hallucination-aware tuning pipeline for LLM in retrieval- augmented generation,

    J. Song, X. Wang, J. Zhu, Y . Wu, X. Cheng, R. Zhong, and C. Niu, “RAG-HAT: A hallucination-aware tuning pipeline for LLM in retrieval- augmented generation,” inProc. 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, Miami, FL, 2024, pp. 1548–1558

    2024

  28. [28]

    Llama 2: Open Foundation and Fine-Tuned Chat Models

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, et al., “Llama 2: Open foundation and fine-tuned chat models,”arXiv preprintarXiv:2307.09288, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  29. [29]

    Mistral 7B

    A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. de las Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier,et al., “Mistral 7B,”arXiv preprintarXiv:2310.06825, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  30. [30]

    GPT-4 Technical Report

    OpenAI, “GPT-4 technical report,”arXiv preprintarXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  31. [31]

    A survey on RAG meeting LLMs: Towards retrieval-augmented large language models,

    W. Fan, Y . Ding, L. Ning, S. Wang, H. Li, D. Yin, T.-S. Chua, and Q. Li, “A survey on RAG meeting LLMs: Towards retrieval-augmented large language models,” inProc. 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 6491–6501

    2024

  32. [32]

    FactGraph: Evaluating factuality in summarization with semantic graph representations,

    L. F. R. Ribeiro, M. Liu, I. Gurevych, M. Dreyer, and M. Bansal, “FactGraph: Evaluating factuality in summarization with semantic graph representations,” inProc. 2022 Conf. North American Chapter of the Association for Computational Linguistics: Human Language Technolo- gies, Seattle, W A, USA, 2022, pp. 3238–3253

    2022

  33. [33]

    GraphEval: A knowledge-graph based LLM hallucination evaluation framework,

    H. Sansford, N. Richardson, H. Petric Maretic, and J. Nait Saada, “GraphEval: A knowledge-graph based LLM hallucination evaluation framework,” inProc. KiL’24: Workshop on Knowledge-infused Learn- ing, co-located with the 30th ACM SIGKDD Conf., Barcelona, Spain, 2024

    2024

  34. [34]

    SummaC: Re-visiting NLI-based models for inconsistency detection in summa- rization,

    P. Laban, T. Schnabel, P. N. Bennett, and M. A. Hearst, “SummaC: Re-visiting NLI-based models for inconsistency detection in summa- rization,”Transactions of the Association for Computational Linguistics, vol. 10, pp. 163–177, 2022

    2022

  35. [35]

    Knowledge-centric hallucination detection,

    X. Hu, D. Ru, L. Qiu, Q. Guo, T. Zhang, Y . Xu, Y . Luo, P. Liu, Y . Zhang, and Z. Zhang, “Knowledge-centric hallucination detection,” in Proc. 2024 Conf. Empirical Methods in Natural Language Processing (EMNLP), Miami, FL, USA, 2024, pp. 6953–6975

    2024

  36. [36]

    BERTScore: Evaluating text generation with BERT,

    T. Zhang, V . Kishore, F. Wu, K. Q. Weinberger, and Y . Artzi, “BERTScore: Evaluating text generation with BERT,” inProc. Inter- national Conference on Learning Representations (ICLR), 2020

    2020