pith. machine review for the scientific record. sign in

arxiv: 2604.04863 · v1 · submitted 2026-04-06 · 💻 cs.CV

Recognition: no theorem link

Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations

Tuan Dung Nguyen , Minh Khoi Ho , Qi Chen , Yutong Xie , Nguyen Cam-Tu , Minh Khoi Nguyen , Dang Huy Pham Nguyen , Anton van den Hengel
show 3 more authors
Johan W. Verjans Phi Le Nguyen Vu Minh Hieu Phan
Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:51 UTC · model grok-4.3

classification 💻 cs.CV
keywords hallucination detectionlarge vision-language modelstoken groundingattention patternspatch-level analysissemantic alignment
0
0 comments X

The pith

Hallucinated tokens in vision-language models show diffuse attention and weak patch alignment, enabling fine-grained detection at up to 90 percent accuracy.

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

The paper argues that current hallucination detectors for large vision-language models fail because they only check overall image relevance. A hallucinated object token can still score high globally if its weak connections scatter across many image patches. The authors instead examine token interactions at the patch level across layers and find two reliable signs of hallucination: attention spreads out instead of focusing on one region, and the token lacks semantic match with any visual patch. They build a simple detector from these patch statistics plus hidden states that reaches 90 percent token-level accuracy. This matters because it shifts detection from coarse scores to local grounding, which can flag errors that global methods miss.

Core claim

A faithful object token must be strongly grounded in a specific image region. Hallucinated tokens instead produce diffuse, non-localized attention patterns and fail to exhibit meaningful semantic alignment with any visual region. These two signatures, observed through patch-level analysis across model layers, allow a lightweight detector that combines patch statistical features with hidden-layer representations to identify hallucinations at the token level.

What carries the argument

Patch-level hallucination detection framework that extracts statistical features from token-to-patch attention maps and hidden representations to measure attention locality and semantic alignment.

If this is right

  • Token-level hallucination detection can outperform global relevance scoring on the same models and data.
  • Attention locality and semantic alignment serve as interpretable signals that separate faithful from hallucinated outputs.
  • The method requires no extra training and works with existing model internals.
  • Detection remains accurate even when global scores are misleading due to scattered low-level correlations.

Where Pith is reading between the lines

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

  • The same patch signatures could be used to guide correction steps that replace hallucinated tokens with better-grounded alternatives.
  • Extending the analysis to video or multi-frame inputs might reveal temporal versions of the same diffuse-attention pattern.
  • If the signatures hold across architectures, they offer a way to audit new models without task-specific fine-tuning.

Load-bearing premise

The two observed signatures of diffuse attention and missing semantic alignment are assumed to be consistent indicators of hallucination in any LVLM, task, or dataset.

What would settle it

Finding a token that is verifiably hallucinated yet shows compact localized attention and strong semantic alignment with at least one image patch, or a faithful token that shows diffuse attention, would disprove the detection rule.

Figures

Figures reproduced from arXiv: 2604.04863 by Anton van den Hengel, Dang Huy Pham Nguyen, Johan W. Verjans, Minh Khoi Ho, Minh Khoi Nguyen, Nguyen Cam-Tu, Phi Le Nguyen, Qi Chen, Tuan Dung Nguyen, Vu Minh Hieu Phan, Yutong Xie.

Figure 1
Figure 1. Figure 1: Left: SVAR [12] captures image-level statistics by globally summing the attention across the image, ignoring local attention structures. This triggers a false alarm when the local noisy attention appears. Right: Our proposed “attention dispersion score” encoded token-level attention distribution, quantifying how much the model allocates attention to various local regions simultaneously. Spreading the focus… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our token-level hallucination detection framework. We reveal two key indicators of hallucination: (1) Attention Dispersion Score (ADS), hallucinated tokens exhibit highly diffused, non-localized attention across local image patches, while faithful tokens show concentrated focus; and (2) Cross-modal Grounding Consistency (CGC), hallucinated text tokens exhibit low alignment with any object regio… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the Attention Dispersion Score (ADS) computation. After predicting an object token, we extract the text-to￾patch cross-modal attention map. The top k% activations are kept to isolate highly focused regions. Then, we form the attended object regions and suppress attention sinks by applying N-connected component. Finally, the proposed ADS score is computed by quantifying the entropy of the re… view at source ↗
Figure 4
Figure 4. Figure 4: The visual attention maps for or a true token (“camera”, [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Layer-wise attention entropy of true vs. hallucinated tokens across LVLMs (lower is better; more focused). Reported p-values indicate strong separation in early/mid layers. \bar {\mathbf {A}}^{(n)}_{t} = \frac {1}{H}\sum _{h=1}^H \mathbf {A}^{(n,h)}_{t}. (1) To isolate the most salient activations, we retain only the top-x% of patch responses (empirically x = 10), defining a foreground set F (n) t and its … view at source ↗
Figure 6
Figure 6. Figure 6: Attention dispersion distribution for true vs. halluci￾nated tokens for LLaVA-1.5-7B. calculation of our ADS metric in [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Layer-wise patch-level similarity scores for true vs. hallucinated tokens across LVLMs. Early/mid layers are most discriminative; [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Cross-modal grounding consistency heatmap for a true token (“television”, top) vs. a hallucinated token (“table”, bot￾tom) across layers (10, 17, 22). True tokens display similarity clusters aligned with the object’s local structure, while hallucina￾tion tokens have low alignment with all patches across the image. 0.2 0.3 0.4 0.5 0.6 Real Hallucinated Object token type CGC Score [PITH_FULL_IMAGE:figures/f… view at source ↗
Figure 10
Figure 10. Figure 10: Top–5% similarities for true vs. hallucinated tokens for LLaVA-1.5-7B. 4.2. Experimental Setup We benchmark our detector against state-of-the-art baselines for token-level hallucination detection, namely MetaToken [8], HalLoc [21], DHCP [29], ProjectAway [11] and SVAR [12]. Implementation details are presented in the Supplementary. For our method, the concatenated feature vector ft (which combines per-lay… view at source ↗
Figure 11
Figure 11. Figure 11: Mean absolute SHAP values for ADS and CGC across [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
read the original abstract

Large vision-language models (LVLMs) achieve strong performance on visual reasoning tasks but remain highly susceptible to hallucination. Existing detection methods predominantly rely on coarse, whole-image measures of how an object token relates to the input image. This global strategy is limited: hallucinated tokens may exhibit weak but widely scattered correlations across many local regions, which aggregate into deceptively high overall relevance, thus evading the current global hallucination detectors. We begin with a simple yet critical observation: a faithful object token must be strongly grounded in a specific image region. Building on this insight, we introduce a patch-level hallucination detection framework that examines fine-grained token-level interactions across model layers. Our analysis uncovers two characteristic signatures of hallucinated tokens: (i) they yield diffuse, non-localized attention patterns, in contrast to the compact, well-focused attention seen in faithful tokens; and (ii) they fail to exhibit meaningful semantic alignment with any visual region. Guided by these findings, we develop a lightweight and interpretable detection method that leverages patch-level statistical features, combined with hidden-layer representations. Our approach achieves up to 90% accuracy in token-level hallucination detection, demonstrating the superiority of fine-grained structural analysis for detecting hallucinations.

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 / 0 minor

Summary. The paper claims that existing global-score methods for detecting hallucinations in large vision-language models (LVLMs) are limited because hallucinated tokens can produce weak but scattered correlations that aggregate into misleadingly high relevance scores. It introduces a patch-level framework that analyzes fine-grained token interactions across layers, identifies two signatures of hallucinated tokens (diffuse/non-localized attention patterns and absence of semantic alignment with any image patch), and builds a lightweight detector from patch-level statistical features plus hidden representations. The approach is reported to reach up to 90% token-level accuracy and to demonstrate the superiority of fine-grained structural analysis.

Significance. If the signatures prove general and the accuracy holds under rigorous cross-model and cross-dataset testing, the work would offer a practical, interpretable alternative to coarse global metrics, directly addressing a key reliability bottleneck in multimodal reasoning systems. The emphasis on lightweight, feature-based detection without heavy additional training is a potential strength for deployment.

major comments (2)
  1. [Abstract] Abstract: The central claim of up to 90% token-level accuracy and superiority over global methods is presented without any reference to datasets, number of LVLMs evaluated, baselines, evaluation protocol, or statistical measures. This information is load-bearing for assessing whether the two signatures are reliable and general rather than artifacts of the specific test distribution.
  2. [Abstract] Abstract: The argument that diffuse attention and lack of semantic alignment are characteristic signatures of hallucination assumes these patterns hold across models, tasks, and datasets. No cross-validation results or diversity details are supplied, leaving open the possibility that the reported performance reflects overfitting rather than a robust, fine-grained advantage.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. The feedback highlights important aspects of how the abstract presents our claims, and we address each point below with a commitment to revision where appropriate.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim of up to 90% token-level accuracy and superiority over global methods is presented without any reference to datasets, number of LVLMs evaluated, baselines, evaluation protocol, or statistical measures. This information is load-bearing for assessing whether the two signatures are reliable and general rather than artifacts of the specific test distribution.

    Authors: We agree that the abstract's brevity leaves the central claim without immediate context on scope. The full manuscript details evaluations across multiple LVLMs, standard hallucination benchmarks, global-score baselines, and protocols with accuracy metrics. To strengthen the abstract, we will add a concise clause referencing the evaluation breadth (e.g., 'evaluated on multiple LVLMs and datasets with comparisons to global baselines'). This revision will make the claim more self-contained without expanding the abstract's length substantially. revision: yes

  2. Referee: [Abstract] Abstract: The argument that diffuse attention and lack of semantic alignment are characteristic signatures of hallucination assumes these patterns hold across models, tasks, and datasets. No cross-validation results or diversity details are supplied, leaving open the possibility that the reported performance reflects overfitting rather than a robust, fine-grained advantage.

    Authors: The signatures emerge from layer-wise patch-level analysis conducted on diverse LVLMs and tasks in the manuscript, with the detector validated across datasets to show consistency. Our approach relies on lightweight, non-trained statistical features rather than complex models prone to overfitting. We will revise the abstract to explicitly note the cross-model and cross-dataset scope of the evaluation, thereby underscoring the generality of the signatures while preserving conciseness. revision: yes

Circularity Check

0 steps flagged

No circularity; purely empirical observation and lightweight detector

full rationale

The paper describes an observational study: patch-level analysis reveals two attention signatures for hallucinated tokens, which are then used to construct a statistical-feature detector. No equations, parameter fits, derivations, or self-citations appear in the provided text that would reduce any claimed result to its own inputs by construction. The 90% accuracy is reported as an empirical outcome on the examined models and data, not a prediction forced by prior fitting or definitional equivalence. This satisfies the default expectation of no circularity for an empirical method without load-bearing self-referential steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No mathematical axioms, free parameters, or invented entities are introduced; the work rests on empirical observations of attention behavior in existing transformer architectures.

pith-pipeline@v0.9.0 · 5568 in / 1032 out tokens · 47814 ms · 2026-05-10T19:51:06.499465+00:00 · methodology

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 · 36 canonical work pages · 3 internal anchors

  1. [1]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, et al. Flamingo: a visual language model for few-shot learning. InProceedings of the 36th International Confer- ence on Neural Information Processing Systems, Red Hook, NY , USA, 2022. Curran Associates Inc. 2

    work page 2022

  2. [2]

    Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond, 2023

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond, 2023. 1, 2

    work page 2023

  3. [3]

    Qwen3-VL Technical Report

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025. 6

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    Attentropy: On the generalization ability of supervised semantic segmen- tation transformers to new objects in new domains

    Krzysztof Baron-Lis, Matthias Rottmann, Annika Mütze, Sina Honari, Pascal Fua, and Mathieu Salzmann. Attentropy: On the generalization ability of supervised semantic segmen- tation transformers to new objects in new domains. In35th British Machine Vision Conference 2024, BMVC 2024, Glas- gow, UK, November 25-28, 2024. BMV A, 2024. 5

    work page 2024

  5. [5]

    Multi-object hallucination in vision language models

    Xuweiyi Chen, Ziqiao Ma, Xuejun Zhang, Sihan Xu, Shengyi Qian, Jianing Yang, David Fouhey, and Joyce Chai. Multi-object hallucination in vision language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3

    work page 2024

  6. [6]

    Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhang- wei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test- time scaling.arXiv preprint arXiv:2412.05271, 2024. 2, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    InstructBLIP: Towards general-purpose vision-language models with instruction tuning

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. InThirty- seventh Conference on Neural Information Processing Sys- tems, 2023. 1, 2

    work page 2023

  8. [8]

    Metatoken: Detecting hallucination in image descriptions by meta classification

    Laura Fieback, Jakob Spiegelberg, and Hanno Gottschalk. Metatoken: Detecting hallucination in image descriptions by meta classification. InVISIGRAPP : VISAPP, 2024. 1, 3, 7, 8

    work page 2024

  9. [9]

    Detecting and pre- venting hallucinations in large vision language models

    Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and pre- venting hallucinations in large vision language models. In Proceedings of the Thirty-Eighth AAAI Conference on Artifi- cial Intelligence and Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Sym- posium on Educational Advances in Artificial Intelligence....

    work page 2024

  10. [10]

    Yixiao He, Haifeng Sun, Pengfei Ren, Jingyu Wang, Huazheng Wang, Qi Qi, Zirui Zhuang, and Jing Wang. Eval- uating and mitigating object hallucination in large vision- language models: Can they still see removed objects? In Proceedings of the 2025 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech...

    work page 2025

  11. [11]

    Interpreting and editing vision-language representations to mitigate hallucinations

    Nicholas Jiang, Anish Kachinthaya, Suzanne Petryk, and Yossi Gandelsman. Interpreting and editing vision-language representations to mitigate hallucinations. InThe Thir- teenth International Conference on Learning Representa- tions, 2025. 7, 8

    work page 2025

  12. [12]

    Devils in middle layers of large vision- language models: Interpreting, detecting and mitigating ob- ject hallucinations via attention lens

    Zhangqi Jiang, Junkai Chen, Beier Zhu, Tingjin Luo, Yankun Shen, and Xu Yang. Devils in middle layers of large vision- language models: Interpreting, detecting and mitigating ob- ject hallucinations via attention lens. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 25004–25014, 2025. 1, 2, 3, 4, 7, 8

    work page 2025

  13. [13]

    Your large vision-language model only needs a few attention heads for visual grounding

    Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. Your large vision-language model only needs a few attention heads for visual grounding. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 9339–9350, 2025. 5

    work page 2025

  14. [14]

    Mitigating object hallucinations in large vision-language models through vi- sual contrastive decoding

    Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shi- jian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through vi- sual contrastive decoding. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13872–13882, 2024. 4

    work page 2024

  15. [15]

    Otter: A multi-modal model with in-context instruction tuning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(9):7543–7557, 2025

    Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Joshua Adrian Cahyono, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(9):7543–7557, 2025. 1

    work page 2025

  16. [16]

    VisualBERT: A Simple and Performant Baseline for Vision and Language

    Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language.ArXiv, abs/1908.03557,

    work page internal anchor Pith review arXiv 1908

  17. [17]

    Evaluating object hallucination in large vision-language models

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. InProceedings of the 2023 conference on empirical methods in natural language pro- cessing, pages 292–305, 2023. 1, 2, 3, 7

    work page 2023

  18. [18]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 4

    work page 2014

  19. [19]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. InAdvances in Neural Information Processing Systems, pages 34892–34916. Curran Associates, Inc., 2023. 1, 2, 6

    work page 2023

  20. [20]

    Self- CheckGPT: Zero-resource black-box hallucination detection for generative large language models

    Potsawee Manakul, Adian Liusie, and Mark Gales. Self- CheckGPT: Zero-resource black-box hallucination detection for generative large language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Lan- guage Processing, pages 9004–9017, Singapore, 2023. As- sociation for Computational Linguistics. 3

    work page 2023

  21. [21]

    Hal- loc: Token-level localization of hallucinations for vision lan- guage models

    Eunkyu Park, Minyeong Kim, and Gunhee Kim. Hal- loc: Token-level localization of hallucinations for vision lan- guage models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 29893–29903, 2025. 1, 3, 6, 7, 8

    work page 2025

  22. [22]

    Object hallucination in image cap- tioning

    Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image cap- tioning. InProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035–4045,

    work page 2018

  23. [23]

    From pixels to tokens: Revisiting object hallucina- tions in large vision-language models

    Yuying Shang, Xinyi Zeng, Yutao Zhu, Xiao Yang, Zheng- wei Fang, Jingyuan Zhang, Jiawei Chen, Zinan Liu, and Yu Tian. From pixels to tokens: Revisiting object hallucina- tions in large vision-language models. InProceedings of the 33rd ACM International Conference on Multimedia, pages 10496–10505, 2025. 3

    work page 2025

  24. [24]

    Gemma 3 technical report, 2025

    Gemma Team et al. Gemma 3 technical report, 2025. 2

    work page 2025

  25. [25]

    Logical closed loop: Uncover- ing object hallucinations in large vision-language models

    Junfei Wu, Qiang Liu, Ding Wang, Jinghao Zhang, Shu Wu, Liang Wang, and Tieniu Tan. Logical closed loop: Uncover- ing object hallucinations in large vision-language models. In Findings of ACL 2024, 2024. 3

    work page 2024

  26. [26]

    Detecting and mitigating hallucination in large vision language models via fine-grained ai feedback

    Wenyi Xiao, Ziwei Huang, Leilei Gan, Wanggui He, Haoyuan Li, Zhelun Yu, Fangxun Shu, Hao Jiang, and Lin- chao Zhu. Detecting and mitigating hallucination in large vision language models via fine-grained ai feedback. InPro- ceedings of the AAAI Conference on Artificial Intelligence, pages 25543–25551, 2025. 3

    work page 2025

  27. [27]

    mPLUG- owl3: Towards long image-sequence understanding in multi- modal large language models

    Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mPLUG- owl3: Towards long image-sequence understanding in multi- modal large language models. InThe Thirteenth Interna- tional Conference on Learning Representations, 2025. 1

    work page 2025

  28. [28]

    Woodpecker: hallucination correction for multimodal large language models.Science China Information Sciences, 67(12), 2024

    Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. Woodpecker: hallucination correction for multimodal large language models.Science China Information Sciences, 67(12), 2024. 3

    work page 2024

  29. [29]

    Dhcp: Detecting hallucinations by cross-modal attention pat- tern in large vision-language models

    Yudong Zhang, Ruobing Xie, Xingwu Sun, Yiqing Huang, Jiansheng Chen, Zhanhui Kang, Di Wang, and Yu Wang. Dhcp: Detecting hallucinations by cross-modal attention pat- tern in large vision-language models. InProceedings of the 33rd ACM International Conference on Multimedia, pages 3555–3564, 2025. 3, 7, 8

    work page 2025

  30. [30]

    MiniGPT-4: Enhancing vision-language understanding with advanced large language models

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mo- hamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. InThe Twelfth International Conference on Learning Representa- tions, 2024. 1, 2 Beyond the Global Scores: Fine-Grained Token Grounding as a Robust Detector of LVLM Hallucinations Supplementary Material

    work page 2024

  31. [31]

    11 of the original paper is always zero; we instead use token probability directly

    Experimental configurations For hallucination detection, our configurations are as fol- lows: •MetaToken: For our greedy decoding setup, the proba- bility difference term in Eq. 11 of the original paper is always zero; we instead use token probability directly. We implement two binary classifiers, Logistic Regression (LR) withlbfgssolver and Gradient Boos...

    work page

  32. [32]

    This is be- cause the CHAIR toolkit often misses or return excessive words in case of ambiguity

    Dataset curation In order to construct a hallucination detection dataset, in- stead of following previous papers to use CHAIR [22], we used GPT-4o API to extract hallucinated words. This is be- cause the CHAIR toolkit often misses or return excessive words in case of ambiguity. The prompt structure we used is as follows 11: It is possible that an object w...

    work page

  33. [33]

    The results for our two features are shown in Tab

    Ablation Studies We keep the classifier fixed (LLaV A features + MLP) and vary which layer combinations are included. The results for our two features are shown in Tab. 6. We observe that, typ- ically the middle layers (around 12-24) are the layers with richest semantic features and best alignment between text and images, yielding the best results. Meanwh...

    work page

  34. [34]

    The re- sult is displayed with Fig 11

    Feature Importance We compute the mean SHAP values of layer-wise ADS and CGC features, for LLaV A-1.5 + MLP classifier. The re- sult is displayed with Fig 11. We observe the findings is consistent with 8, where features from the middle layers are the most influential for cross-modal attention and align- ment, meanwhile for the final layers these informati...

    work page

  35. [35]

    Effect of adding CGC and ADS features to hallucination detectors across three VLMs

    Combining Our Features With SV AR Model Features AUC F1 LLaV A-1.5 SV AR 85.12 69.35 +CGC 87.72 73.52 +ADS 89.04 74.03 +CGC+ADS89.41 74.07 Qwen 2.5-VL-8B SV AR 87.85 76.38 +CGC 86.79 76.24 +ADS 89.11 79.12 +CGC+ADS90.28 81.09 InternVL-2.5-8B SV AR 86.21 76.31 +CGC 86.39 79.43 +ADS 88.02 81.03 +CGC+ADS89.54 81.56 Table 7. Effect of adding CGC and ADS featu...

    work page

  36. [36]

    bowl", "broccoli

    Ablation Studies We report the sensitivity of our detector to two important threshold parameter choices: Top-x% (Equation 1, page 5), which determines how many attention patches we recognize as object patches with 8-connected components for ADS, and Top-kfor CGC (Equation 7), which are the percentage ADS CGC Top-x% AUC F1 Top-k% AUC F1 5% 81.65 76.54 1% 8...

    work page