Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Preemptive Hallucination Reduction: An Input-Level Approach for Multimodal Language Model

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that adaptively choosing among three preprocessed versions of an input image can cut measured visual hallucination rates in a multimodal language model by 44.3%, with no retraining and no architectural changes.

desk verdict The headline 44.3% reduction is an oracle artifact: the ensemble picks the variant with the lowest NLI score, and NLI scores are computed against ground truth, so the result is circular and non-deployable. read the letter →

arxiv 2505.24007 v2 pith:VF2UJZ6C submitted 2025-05-29 cs.CV

Nokimul Hasan Arif , Shadman Rabby , Md Hefzul Hossain Papon , Sabbir Ahmed This is my paper
classification cs.CV
keywords visualhallucinationmultimodallargelanguagemodelsinput-levelpreprocessingensemblefilteringnoisereductionedgeenhancementSelfCheckGPTHaloQuestdataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that visual hallucinations in multimodal language models can be reduced at the input stage, before the model ever sees the image. It reports a 44.3% reduction in hallucination rates, measured by Natural Language Inference (NLI) scores, on 1000 visually challenging samples from the HaloQuest dataset using GPT-3.5. The method compares three image variants—original, noise-reduced, and edge-enhanced—and selects the variant whose answer best matches the ground-truth answer. The claim matters because, if true, it offers a lightweight, model-agnostic way to improve factual grounding without fine-tuning or post-hoc correction.

What carries the argument

The ensemble of three input variants, scored by SelfCheckGPT-NLI. The variants are the raw image, a noise-reduced version produced by a 15x15 median filter, and an edge-enhanced version produced by a channel-wise Laplacian operator with weighted blending (α=1.5, β=-0.5). SelfCheckGPT-NLI uses a DeBERTa-v3-large model to assign contradiction probabilities between the generated answer and the reference answer, and the ensemble selects the variant with the lowest average contradiction score.

What would settle it

Run the same three-variant pipeline on a held-out subset of HaloQuest where the image version is selected without access to the ground-truth answer, using question category, image statistics, or a learned router, and compare average NLI scores: if the router's score is close to 0.171 the selection is real, and if it stays near the best single variant (about 0.301) the reported gain depends on oracle access to the correct answer.

Watch

Extended reading notes

Core claim

The central claim is that an ensemble of preprocessing filters can substantially reduce hallucination in multimodal LLM responses. The pipeline applies a median filter for noise reduction and a Laplacian operator for edge enhancement, generating three image variants—original, noise-reduced, and edge-enhanced—and feeds each to GPT-3.5. The answer from the variant with the lowest SelfCheckGPT NLI score is selected, yielding an average NLI score of 0.171 compared to 0.307 for the original image, a 44.3% decrease. The paper also finds that the best variant varies by question type and even within a type, so no single fixed filter consistently wins.

Load-bearing premise

The load-bearing premise is that the system can choose the best image version during real use, because the reported reduction is obtained by picking the version whose answer matches the ground-truth answer, which is not available when the model is deployed.

Editorial extensions

If this is right

  • If the claim holds, visual hallucination can be reduced without accessing model internals or retraining.
  • The per-question-type analysis suggests that different visual degradations call for different input representations, so a single fixed preprocessing would miss gains.
  • Because the best variant varies within every question category, per-input selection is necessary rather than a single global filter.
  • The approach is lightweight and model-agnostic, so it could be combined with training-time alignment or post-hoc correction methods.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported 44.3% reduction is an upper bound: the ensemble chooses the variant whose answer best matches the ground-truth answer, so the gain is an oracle-selection result unless a deployable routing rule can predict the best variant before generating an answer.
  • A routing rule based only on the three coarse question categories (object identification, quantity, color) would likely underperform the oracle, since within each category the best variant varies from sample to sample.
  • The same ensemble could be extended with additional image transforms such as contrast normalization or deblurring, and evaluated on other benchmarks to test whether input-level conditioning transfers across models and domains.
  • The finding indirectly suggests that hallucination is partly a function of how the image is presented, not only of model competence, which may open a complementary axis for reliability improvements.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes an input-level preprocessing framework that reduces visual hallucination in multimodal LLMs by generating three variants of each input image—original, noise-reduced (median filter), and edge-enhanced (Laplacian)—and then selecting among them. Responses are generated by GPT-3.5 for each variant on 1000 HaloQuest samples, and a 'SelfCheckGPT NLI' score is computed for each generated response. The authors report average NLI scores of 0.334 (edge-enhanced), 0.307 (original), 0.301 (noise-reduced), and 0.171 (ensemble), claiming a 44.3% reduction because the ensemble selects the variant with the lowest NLI score. They also report per-category winner counts and provide qualitative examples.

Significance. If the headline claim were valid, it would be a notable practical result: cheap input transformations could reduce hallucination in multimodal LLMs without retraining or architectural changes. The paper has some credible elements: it evaluates on 1000 samples from the public HaloQuest benchmark, uses simple and transparent filters, provides qualitative examples, and explicitly acknowledges that validation on one benchmark and one model family is a limitation. However, the central contribution is not established. The ensembling rule uses ground-truth answers to choose the best variant, and the NLI protocol is a supervised comparison to the answer key rather than the cited SelfCheckGPT self-consistency check. The 44.3% reduction is therefore a minimum-of-three artifact rather than evidence for input-level preprocessing as a deployable mitigation. This is a correctness problem, not a presentation issue.

major comments (4)
  1. [§3.4, §4.2] The ensemble selection rule is an oracle and cannot be deployed. §3.4 states that each generated response is 'evaluated against the ground truth answer' to obtain an NLI score, and §4.2 defines ensembling as selecting 'the optimal NLI score among the three' (original, noise-reduced, edge-enhanced) variants. In real use the ground-truth answer is unavailable, so the lowest-NLI variant cannot be identified. The reported drop from 0.307 to 0.171 is therefore the average of the per-sample minima of the three scores, which is mathematically no larger than the average of any single fixed variant; it is not the performance of an implementable routing mechanism. The abstract's claim of question-type-adaptive selection is not realized by any rule: Table 2 only reports, after the fact, how often each variant happened to win. This is the load-bearing issue: the headline 44.3% reduction is an artifact of the evaluation protocol.
  2. [§3.3.2, §3.4] The NLI protocol is not SelfCheckGPT as defined by the cited work. SelfCheckGPT (Ref. [41]) measures hallucination by sampling multiple responses to the same prompt and checking self-consistency; it is a zero-resource, black-box method that does not use reference answers. In §3.3.2 the equations are written in that framework (premise Sn, hypothesis ri), but §3.4 explicitly says the NLI score is obtained by comparing the LLM-generated response to the ground-truth answer. This replaces the intended self-consistency check with supervised oracle scoring, so the paper's claim that the method is 'zero-resource, black-box' and requires no external reference data is inaccurate. The reported hallucination rates are not comparable to standard SelfCheckGPT scores.
  3. [Table 2, §3.1] The per-category statistics are internally inconsistent. §3.1 reports 593 object-identification, 415 quantity, and 289 color questions and notes that the categories overlap because some 'what' questions are counted in both object-identification and color. Table 2 nevertheless presents these as three separate columns with fixed totals and counts such as 150, 111, and 75 in one row, as if they were disjoint. Moreover, the first three rows of the 'All Ques' column sum to 861 rather than 1000, with no explanation for the missing 139 cases. These inconsistencies affect the paper's secondary claim that preprocessing benefits differ by question type.
  4. [§3.2, §3.4] The model used for image-conditioned response generation is not clearly identified and appears to be text-only. The paper repeatedly states that images were presented to 'GPT-3.5' (e.g., §3.2: 'we used OpenAI's GPT-3.5 model for its image processing compatibility'; §3.4: 'the original image and its corresponding question were presented to GPT-3.5'). GPT-3.5 is not a vision-capable model. If a different model or API (e.g., a vision variant) was used, it must be named; without this, the reported image-conditioned responses cannot be reproduced or interpreted.
minor comments (5)
  1. [Tables 3–4] There are many typos in the qualitative tables, including 'N LIorg = 0.0.998', 'N LIN R= 0.265', and 'Holoquest' in the Table 1 caption; these should be corrected.
  2. [Figure 4] Figure 4 plots 1000 NLI scores as lines and is likely unreadable; consider showing aggregate statistics or a small multiple by category.
  3. [Eq. (5), §3.4] The description of the consistency score in §3.4 ('proportion of supporting (entailment) versus contradicting or neutral responses') is inconsistent with Eq. (5), which ignores the neutral class and normalizes over entailment and contradiction only.
  4. [§3.1] The paper does not specify how the 'first 1000 samples' were selected from HaloQuest or whether they are a random subset; this matters because the dataset's difficulty is not uniform.
  5. [References] Reference [43] contains a typo ('media filtering' should be 'median filtering') and several references use placeholder arXiv IDs (e.g., [28], [29], [32], [34], [37]), making them hard to verify.

Circularity Check

1 steps flagged · score 8.0 of 10

Headline reduction is an oracle artifact: the ensemble is defined as the per-sample minimum of three ground-truth-scored NLI values, so the 44.3% gain is forced by the evaluation protocol.

  1. self definitional [Section 4.2 (Implication of Ensembling Module) with Section 3.4 (Experiment)]
    "To leverage the strengths of each approach, the ensembling method selects the optimal NLI score among the three (original, noise reduced, and edge enhanced) for the final output. ... ensembling reduces the average NLI score from 0.307 (original) to 0.171, representing a 44.3% decrease. ... This generated response was then compared against the ground truth answer to determine the initial NLI score."

    Each variant's NLI score is computed by comparing the generated answer to the ground-truth answer (Section 3.4). The ensemble output is then defined as the variant with the lowest such NLI score (Section 4.2). Therefore the reported ensemble NLI (0.171) is, by construction, the per-sample minimum of NLI_org, NLI_NR, and NLI_EE. Selecting that minimum requires access to the ground-truth answer key, which is unavailable in deployment. The 44.3% reduction is the gap between the original score and the per-sample minimum, not the performance of any implementable question-type routing rule. The abstract's promised selection 'based on the type of question' is never operationalized; Table 2 only post-hoc tabulates which variant happened to win.

full rationale

The paper's preprocessing filters, the HaloQuest dataset, and the NLI model are external and not circular. The circularity is confined to the ensembling claim. Section 3.4 states that each variant's NLI score is obtained by comparing the generated response to the ground-truth answer. Section 4.2 then defines the ensemble output as the variant with the lowest such score. Consequently, the reported ensemble average (0.171) is, sample by sample, min(NLI_org, NLI_NR, NLI_EE); the 44.3% reduction over the original (0.307) is the difference between the original score and the per-sample minimum. That selection cannot be made at deployment, because the ground-truth answer is unavailable; the paper provides no learned or heuristic routing rule tied to question type. The abstract's promise of selection 'based on the type of question' is never operationalized; Table 2 only post-hoc reports which variant won in each category. The headline improvement is therefore an oracle bound imposed by the evaluation protocol, not the performance of an implementable input-level method. No self-citation chain is involved; the circularity is definitional, so the score is 8 rather than 10.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities. Its central claim rests on free parameters for filtering and blending, on the assumption that ground-truth-informed NLI scores can be used for selection in a deployed system, and on questionable assumptions about GPT-3.5's image capabilities and NLI transfer. The free parameters are not tuned, but the oracle selection is the main circular element.

free parameters (3)
  • kernel_size = 15x15
    The median and Laplacian filters use a 15x15 kernel chosen by hand; no ablation or sensitivity analysis is provided.
  • blending_weights = alpha=1.5, beta=-0.5, gamma=0
    The unsharp-mask-style blending weights are stated without justification or parameter search.
  • sample_subset = first 1000 HaloQuest samples
    The evaluation uses the first 1000 samples from HaloQuest with no justification, and question categories overlap, summing to 1297 rather than 1000.
assumptions (4)
  • domain assumption NLI scores computed against ground truth answers measure hallucination.
    The paper uses SelfCheckGPT NLI comparing generated answers to ground truth answers as the hallucination metric, assuming semantic similarity to ground truth equals factual grounding.
  • domain assumption SelfCheckGPT consistency indicates factuality.
    The paper relies on the hypothesis that consistent sampled responses are factual, as proposed in the SelfCheckGPT reference, without validating this for visual question answering.
  • domain assumption GPT-3.5 can accept image inputs.
    Section 3.2 uses OpenAI's GPT-3.5 for image processing, but GPT-3.5 is text-only in standard documentation and no image encoder or API modality is described.
  • domain assumption An MNLI-trained DeBERTa-v3 NLI model transfers to visual question-answer pairs.
    The NLI model is trained on text MNLI data and is applied here to compare VQA answers; the transfer is assumed without validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Preemptive Hallucination Reduction: An Input-Level Approach for Multimodal Language Model." pith.science (2026). https://pith.science/paper/VF2UJZ6C

@misc{pith2026250524007,
  author       = {Pith},
  title        = {Pith review of: Preemptive Hallucination Reduction: An Input-Level Approach for Multimodal Language Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VF2UJZ6C}},
  note         = {Machine review of arXiv:2505.24007}
}
read the original abstract

Visual hallucinations in Large Language Models (LLMs), where the model generates responses that are inconsistent with the visual input, pose a significant challenge to their reliability, particularly in contexts where precise and trustworthy outputs are critical. Current research largely emphasizes post-hoc correction or model-specific fine-tuning strategies, with limited exploration of preprocessing techniques to address hallucination issues at the input stage. This study presents a novel ensemble-based preprocessing framework that adaptively selects the most appropriate filtering approach -- noise reduced (NR), edge enhanced (EE), or unaltered input (org) based on the type of question posed, resulting into reduced hallucination without requiring any modifications to the underlying model architecture or training pipeline. Evaluated on the `HaloQuest' dataset -- a benchmark designed to test multimodal reasoning on visually complex inputs, our method achieves a 44.3% reduction in hallucination rates, as measured by Natural Language Inference (NLI) scores using SelfCheckGPT. This demonstrates that intelligent input conditioning alone can significantly enhance factual grounding in LLM responses. The findings highlight the importance of adaptive preprocessing techniques in mitigating hallucinations, paving the way for more reliable multimodal systems capable of addressing real-world challenges.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 11 canonical work pages

  1. [41]

    https: //doi.org/10.48550/arXiv.2303.08896

    Manakul, P., Liusie, A., Gales, M.J.F.: SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models (2023). https: //doi.org/10.48550/arXiv.2303.08896 . https://arxiv.org/abs/2303.08896

    arXiv doi:10.48550/arxiv.2303.08896 2023

  2. [1]

    Hamadi, R.: Large language models meet computer vision: A brief survey (2023) https://doi.org/10.48550/arXiv.2311.16673 arXiv:2311.16673 [cs.CV]

    work page Pith review arXiv doi:10.48550/arxiv.2311.16673 2023

  3. [2]

    Neural Computing and Applications 37(4), 1973–1997 (2025) https://doi.org/10.1007/s00521-024-10827-6

    Budnikov, M., Bykova, A., Yamshchikov, I.P.: Generalization potential of large language models. Neural Computing and Applications 37(4), 1973–1997 (2025) https://doi.org/10.1007/s00521-024-10827-6

    doi:10.1007/s00521-024-10827-6 2025

  4. [3]

    In: 2024 27th International Conference on Computer and Information Technology (ICCIT), pp

    Abrar, A., Tabassum, F., Ahmed, S.: Performance evaluation of large lan- guage models in bangla consumer health query summarization. In: 2024 27th International Conference on Computer and Information Technology (ICCIT), pp. 2748–2753 (2024). https://doi.org/10.1109/ICCIT64611.2024.11022034 . https://ieeexplore.ieee.org/abstract/document/11022034

    arXiv 2024

  5. [4]

    Neural Computing and Applications (2025) https://doi.org/10.1007/ s00521-024-10495-6

    Lee, J., Stevens, N., Han, S.C.: Large language models in finance (fin- llms). Neural Computing and Applications (2025) https://doi.org/10.1007/ s00521-024-10495-6

    work page 2025

  6. [5]

    In: Proceedings of the First Workshop on Bangla Language Processing (BLP-2023), pp

    Khan, A., Kamal, F., Chowdhury, M.A., Ahmed, T., Laskar, M.T.R., Ahmed, S.: BanglaCHQ-summ: An abstractive summarization dataset for medical queries in Bangla conversational speech. In: Proceedings of the First Workshop on Bangla Language Processing (BLP-2023), pp. 85–93. Association for Computational Linguistics, Singapore (2023). https://doi.org/10.1865...

    work page doi:10.18653/v1/2023.banglalp-1.10 2023

  7. [6]

    Natural Language Processing Journal 7, 100079 (2024) https: //doi.org/10.1016/j.nlp.2024.100079

    Ahmed, T., Ivan, S., Munir, A., Ahmed, S.: Decoding depression: Analyzing social network insights for depression severity assessment with transformers and explainable ai. Natural Language Processing Journal 7, 100079 (2024) https: //doi.org/10.1016/j.nlp.2024.100079

    work page arXiv 2024

  8. [7]

    In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp

    Mahbub, R., Khan, I., Anuva, S., Shahriar, M.S., Laskar, M.T.R., Ahmed, S.: Unveiling the essence of poetry: Introducing a comprehensive dataset and benchmark for poem summarization. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 14878–14886. Associa- tion for Computational Linguistics, Singapore (2023). ht...

    work page 2023

Show all 45 references
  1. [8]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 46(8), 5625–5644 (2024) https://doi.org/10.1109/TPAMI.2024.3369699 18

    Zhang, J., Huang, J., Jin, S., Lu, S.: Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence 46(8), 5625–5644 (2024) https://doi.org/10.1109/TPAMI.2024.3369699 18

    2024

  2. [9]

    In: 2024 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE), pp

    Ivan, S., Ahmed, T., Ahmed, S., Kabir, M.H.: A vision-language mul- timodal framework for detecting hate speech in memes. In: 2024 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE), pp. 464–468 (2024). https://doi.org/10.1109/CCECE59415.2024.10667078 . ht...

    2024

  3. [10]

    Neural Computing and Applications 36(33), 20849–20861 (2024) https://doi.org/10.1007/s00521-024-10310-2

    Wei, J., Tan, C., Gao, Z., Sun, L., Li, S., Yu, B., Guo, R., Li, S.Z.: Enhancing human-like multimodal reasoning: a new challenging dataset and comprehen- sive framework. Neural Computing and Applications 36(33), 20849–20861 (2024) https://doi.org/10.1007/s00521-024-10310-2

    2024 doi

  4. [11]

    ACM Transactions on Information Systems 43(2), 1–55 (2025) https://doi.org/10.1145/3703155

    Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., Liu, T.: A survey on hallucination in large language mod- els: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems 43(2), 1–55 (2025) http...

    2025 doi

  5. [12]

    arXiv preprint arXiv:2501.02189 1 (2025) https://doi.org/10.48550/arXiv.2501.02189

    Li, Z., Wu, X., Du, H., Nghiem, H., Shi, G.: Benchmark evaluations, applica- tions, and challenges of large vision language models: A survey. arXiv preprint arXiv:2501.02189 1 (2025) https://doi.org/10.48550/arXiv.2501.02189

    2025 doi arXiv

  6. [13]

    https://doi.org/10.48550/arXiv.2401

    Xu, Z., Jain, S., Kankanhalli, M.: Hallucination is Inevitable: An Innate Lim- itation of Large Language Models (2025). https://doi.org/10.48550/arXiv.2401. 11817 . https://arxiv.org/abs/2401.11817

    2025 doi arXiv

  7. [14]

    https://cdn.openai.com/papers/ GPTV System Card.pdf

    OpenAI: GPT-4V(ision) System Card (2023). https://cdn.openai.com/papers/ GPTV System Card.pdf

    2023

  8. [15]

    arXiv preprint arXiv:2304.10592 (2023) https://doi.org/10.48550/arXiv.2304.10592

    Zhu, D., Chen, J., Shen, X., Li, X., Elhoseiny, M.: Minigpt-4: Enhancing vision- language understanding with advanced large language models. arXiv preprint arXiv:2304.10592 (2023) https://doi.org/10.48550/arXiv.2304.10592

    2023 doi arXiv

  9. [16]

    arXiv preprint arXiv:2402.14683 (2024) https://doi.org/ 10.48550/arXiv.2402.14683

    Huang, W., Liu, H., Guo, M., Gong, N.Z.: Visual hallucinations of multi-modal large language models. arXiv preprint arXiv:2402.14683 (2024) https://doi.org/ 10.48550/arXiv.2402.14683

    2024 doi arXiv

  10. [17]

    Neural Computing and Applications (2025) https: //doi.org/10.1007/s00521-025-11229-y

    Mohammed, M.Y., Ali, S.A., Ali, S.K., Majeed, A.A., Mohamed, E.H.: Aftina: enhancing stability and preventing hallucination in ai-based islamic fatwa gen- eration using llms and rag. Neural Computing and Applications (2025) https: //doi.org/10.1007/s00521-025-11229-y

    2025 doi

  11. [18]

    arXiv e-prints, 2403 (2024) https://doi.org/10.48550/arXiv

    Miyai, A., Yang, J., Zhang, J., Ming, Y., Yu, Q., Irie, G., Li, Y., Li, H., Liu, Z., Aizawa, K.: Unsolvable problem detection: Evaluating trustworthiness of vision language models. arXiv e-prints, 2403 (2024) https://doi.org/10.48550/arXiv. 2403.20331

    2024 doi arXiv

  12. [19]

    Neural Computing and 19 Applications (2025) https://doi.org/10.1007/s00521-024-10895-8

    Rafi, S., Laitonjam, L., Das, R.: Reducing extrinsic hallucination in multimodal abstractive summaries with post-processing technique. Neural Computing and 19 Applications (2025) https://doi.org/10.1007/s00521-024-10895-8

    2025 doi

  13. [20]

    https://doi.org/10.48550/arXiv.2406.10185

    Chen, J., Yang, D., Wu, T., Jiang, Y., Hou, X., Li, M., Wang, S., Xiao, D., Li, K., Zhang, L.: Detecting and Evaluating Medical Hallucinations in Large Vision Language Models (2024). https://doi.org/10.48550/arXiv.2406.10185 . https:// arxiv.org/abs/2406.10185

    2024 doi arXiv

  14. [21]

    arXiv preprint arXiv:2404.18930 (2024) https://doi.org/10.13140/RG.2.2.27797.13280

    Bai, Z., Wang, P., Xiao, T., He, T., Han, Z., Zhang, Z., Shou, M.Z.: Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930 (2024) https://doi.org/10.13140/RG.2.2.27797.13280

    2024 arXiv

  15. [22]

    Advances in Neural Information Processing Systems 37, 133571–133599 (2024) https://doi.org/10

    Kim, J., Kim, H., Yeonju, K., Ro, Y.M.: Code: Contrasting self-generated descrip- tion to combat hallucination in large multi-modal models. Advances in Neural Information Processing Systems 37, 133571–133599 (2024) https://doi.org/10. 48550/arXiv.2406.01920

    2024 doi arXiv

  16. [23]

    arXiv preprint arXiv:2305.10355 (2023) https://doi.org/10.48550/arXiv.2305.10355

    Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W.X., Wen, J.-R.: Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355 (2023) https://doi.org/10.48550/arXiv.2305.10355

    2023 doi arXiv

  17. [24]

    arXiv preprint arXiv:2304.08485 (2023) https://doi.org/10.48550/arXiv.2304.08485

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. arXiv preprint arXiv:2304.08485 (2023) https://doi.org/10.48550/arXiv.2304.08485

    2023 doi arXiv

  18. [25]

    : Autohallusion: Automatic genera- tion of hallucination benchmarks for vision-language models

    Wu, X., Guan, T., Li, D., Huang, S., Liu, X., Wang, X., Xian, R., Shrivastava, A., Huang, F., Boyd-Graber, J.L., et al. : Autohallusion: Automatic genera- tion of hallucination benchmarks for vision-language models. arXiv preprint arXiv:2406.10900 (2024) https://doi.org/10.485...

    2024 doi arXiv

  19. [26]

    https://doi.org/10.1109/CVPR52733.2024.01230

    Yu, Q., Li, J., Wei, L., Pang, L., Ye, W., Qin, B., Tang, S., Tian, Q., Zhuang, Y.: HalluciDoctor: Mitigating Hallucinatory Toxicity in Visual Instruction Data (2024). https://doi.org/10.1109/CVPR52733.2024.01230 . https://arxiv.org/abs/ 2311.13614

    2024

  20. [27]

    arXiv preprint arXiv:1809.02156 (2018) https: //doi.org/10.48550/arXiv.1809.02156

    Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object hal- lucination in image captioning. arXiv preprint arXiv:1809.02156 (2018) https: //doi.org/10.48550/arXiv.1809.02156

    2018 doi arXiv

  21. [28]

    arXiv preprint arXiv:2304.08485 (2023) https://doi.org/10.48550/arXiv.2306.14565

    Liu, F., Lin, K., Li, L., Wang, J., Yacoob, Y., Wang, L.: Mitigating hallucina- tion in large multi-modal models via robust instruction tuning. arXiv preprint arXiv:2304.08485 (2023) https://doi.org/10.48550/arXiv.2306.14565

    2023 doi arXiv

  22. [29]

    arXiv preprint arXiv:2304.10592 (2023) https: //doi.org/10.48550/arXiv.2312.06968

    Jiang, C., Xu, H., Dong, M., Chen, J., Ye, W., Yan, M., Ye, Q., Zhang, J., Huang, F., Zhang, S.: Hallucination augmented contrastive learning for mul- timodal large language models. arXiv preprint arXiv:2304.10592 (2023) https: //doi.org/10.48550/arXiv.2312.06968

    2023 doi arXiv

  23. [30]

    : Aligning large multimodal models with factually aug- mented rlhf

    Sun, Z., Shen, S., Cao, S., Liu, H., Li, C., Shen, Y., Gan, C., Gui, L.-Y., Wang, 20 Y.-X., Yang, Y., et al. : Aligning large multimodal models with factually aug- mented rlhf. arXiv preprint arXiv:2309.14525 (2023) https://doi.org/10.48550/ arXiv.2309.14525

    2023 doi arXiv

  24. [31]

    arXiv preprint arXiv:2403.14401 (2024) https://doi

    Yang, D., Cao, B., Chen, G., Jiang, C.: Pensieve: Retrospect-then-compare mit- igates visual hallucination. arXiv preprint arXiv:2403.14401 (2024) https://doi. org/10.48550/arXiv.2403.14401

    2024 doi arXiv

  25. [32]

    arXiv preprint arXiv:2304.08485 (2023) https://doi.org/10.48550/arXiv.2311.07362

    Lee, S., Park, S.H., Jo, Y., Seo, M.: Volcano: Mitigating multimodal hallucination through self-feedback guided revision. arXiv preprint arXiv:2304.08485 (2023) https://doi.org/10.48550/arXiv.2311.07362

    2023 doi arXiv

  26. [34]

    arXiv preprint arXiv:2304.10592 (2023) https://doi.org/10.48550/arXiv.2412.15650

    Tan, W., Cao, Q., Zhan, Y., Xue, C., Ding, C.: Beyond human data: Align- ing multimodal large language models by iterative self-evolution. arXiv preprint arXiv:2304.10592 (2023) https://doi.org/10.48550/arXiv.2412.15650

    2023 doi arXiv

  27. [36]

    arXiv preprint arXiv:2304.10592 (2023) https://doi.org/10.48550/arXiv.2310

    Zhou, Y., Cui, C., Yoon, J., Zhang, L., Deng, Z., Finn, C., Bansal, M., Yao, H.: Analyzing and mitigating object hallucination in large vision-language mod- els. arXiv preprint arXiv:2304.10592 (2023) https://doi.org/10.48550/arXiv.2310. 00754

    2023 doi arXiv

  28. [37]

    arXiv preprint arXiv:2304.10592 (2023) https://doi.org/10.48550/arXiv

    Yin, S., Fu, C., Zhao, S., Xu, T., Wang, H., Sui, D., Shen, Y., Li, K., Sun, X., Chen, E.: Woodpecker: Hallucination correction for multimodal large language models. arXiv preprint arXiv:2304.10592 (2023) https://doi.org/10.48550/arXiv. 2310.16045

    2023 doi arXiv

  29. [38]

    arXiv preprint arXiv:2304.14178 (2023) https://doi.org/10.48550/ arXiv.2304.14178

    Ye, Q., Xu, H., Xu, G., Ye, J., Yan, M., Zhou, Y., Wang, J., Hu, A., Shi, P., Shi, Y., et al.: Mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178 (2023) https://doi.org/10.48550/ arXiv.2304.14178

    2023 doi arXiv

  30. [39]

    arXiv preprint arXiv:2503.10602 (2025) https: //doi.org/10.48550/arXiv.2503.10602

    Duan, J., Kong, F., Cheng, H., Diffenderfer, J., Kailkhura, B., Sun, L., Zhu, X., Shi, X., Xu, K.: Truthprint: Mitigating lvlm object hallucination via latent truthful-guided pre-intervention. arXiv preprint arXiv:2503.10602 (2025) https: //doi.org/10.48550/arXiv.2503.10602

    2025 doi

  31. [40]

    21 In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. 21 In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 26296–26306 (2024). https://doi.org/10.48550/arXiv.2310.03744 . https://arxiv.org/abs/2310.03744

    2024 doi arXiv

  32. [42]

    In: European Conference on Computer Vision, pp

    Wang, Z., Bingham, G., Yu, A.W., Le, Q.V., Luong, T., Ghiasi, G.: Haloquest: A visual hallucination dataset for advancing multimodal reasoning. In: European Conference on Computer Vision, pp. 288–304 (2024). https://doi.org/10.48550/ arXiv.2407.15680 . Springer. https://arxiv....

    2024 doi arXiv

  33. [43]

    Acoustics, Speech and Signal Processing, IEEE Transactions on 27, 13–18 (1979) https: //doi.org/10.1109/TASSP.1979.1163188

    Yang, G., Tang, G.: A fast two-dimensional media filtering algorithm. Acoustics, Speech and Signal Processing, IEEE Transactions on 27, 13–18 (1979) https: //doi.org/10.1109/TASSP.1979.1163188

    1979

  34. [44]

    Prentice- Hall, Inc., USA (2006)

    Gonzalez, R.C., Woods, R.E.: Digital Image Processing (3rd Edition). Prentice- Hall, Inc., USA (2006). https://www.amazon.com/Digital-Image-Processing- Rafael-Gonzalez/dp/0133356728

    2006

  35. [45]

    In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J

    Maynez, J., Narayan, S., Bohnet, B., McDonald, R.: On faithfulness and fac- tuality in abstractive summarization. In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J. (eds.) Proceedings of the 58th Annual Meeting of the Associa- tion for Computational Linguistics, pp. 1906–...

    2020 doi

  36. [46]

    In: Walker, M., Ji, H., Stent, A

    Williams, A., Nangia, N., Bowman, S.: A broad-coverage challenge corpus for sentence understanding through inference. In: Walker, M., Ji, H., Stent, A. (eds.) Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Hum...

    2018 doi

  37. [47]

    In: The Eleventh International Conference on Learning Representations (2023)

    He, P., Gao, J., Chen, W.: DeBERTav3: Improving deBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing. In: The Eleventh International Conference on Learning Representations (2023). https://doi.org/10.48550/arXiv.2111.09543 . https://openreview.n...

    2023 doi arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.