pith. sign in

arxiv: 2606.31407 · v1 · pith:AOFWR4MGnew · submitted 2026-06-30 · 💻 cs.CV · cs.AI· cs.CL

Visual Semantic Entropy: Do Vision Language Models Recognize Visual Ambiguity?

Ta Duc Huy , Trang Nguyen , Townim Chowdhury , Ankit Yadav , Minh-Son To , Zhibin Liao , Johan W. Verjans , Vu Minh Hieu Phan This is my paper

Pith reviewed 2026-07-01 05:32 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.CL
keywords visual semantic entropyvision language modelsuncertainty estimationvisual ambiguityVQA benchmarkssemantic clustering
0
0 comments X

The pith

Perturbing only the image input while fixing the text query lets Visual Semantic Entropy capture visual ambiguity that prior methods miss in vision-language models.

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

The paper shows that standard entropy measures underestimate uncertainty because confident visual embeddings limit answer diversity under stochastic decoding. It demonstrates that joint text-image perturbations mostly reflect prompt sensitivity instead of visual evidence. Visual Semantic Entropy addresses this by varying the image alone, clustering the resulting answers into semantic groups, and measuring weighted dispersion across those groups. This approach yields stronger uncertainty estimates than earlier techniques on multiple VQA tasks.

Core claim

Visual Semantic Entropy perturbs only the image while keeping the text query fixed, clusters the generated answers into semantic prototypes, and computes mass-weighted dispersion among the prototypes to quantify uncertainty arising from visual ambiguity.

What carries the argument

Visual Semantic Entropy, which isolates image perturbations to generate answer variability and then aggregates it via semantic clustering and mass-weighted dispersion.

If this is right

  • Uncertainty estimates become less contaminated by textual prompt effects.
  • VSE can flag cases where vision-language models should defer or seek clarification on ambiguous scenes.
  • The method scales to any VLM that supports image input variation under fixed text.

Where Pith is reading between the lines

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

  • Similar image-only perturbation strategies could extend to other multimodal tasks such as captioning or visual reasoning.
  • The clustering step might be replaced by embedding-based distances for computational efficiency in large-scale deployment.
  • If VSE correlates with human disagreement on visual questions, it could serve as a proxy for collecting ambiguity annotations.

Load-bearing premise

Varying the image alone while holding the text fixed produces answer changes that specifically track visual ambiguity rather than model artifacts or other factors.

What would settle it

A test set of images where human raters judge high visual ambiguity yet VSE reports low dispersion, or low ambiguity yet high dispersion, would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.31407 by Ankit Yadav, Johan W. Verjans, Minh-Son To, Ta Duc Huy, Townim Chowdhury, Trang Nguyen, Vu Minh Hieu Phan, Zhibin Liao.

Figure 1
Figure 1. Figure 1: VLMs can be confidently wrong on ambiguous images. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Text perturbations induce large semantic shifts. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Decoding-based Semantic Entropy underestimates visual ambiguity. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Kernel density of text PT and image purity PI across models.. The diagonal PT = PI separates text- from image-dominant regions. Density mass lying predominantly above the diagonal indicates that clustering is primarily text-driven. 0 1 2 3 4 Image index 0 1 2 3 4 Text indexPT = 0:56 PI = 0:33 Cluster 1 0 1 2 3 4 Image index 0 1 2 3 4 Text indexPT = 0:50 PI = 0:33 Cluster 3 0 1 2 3 4 Image index 0 1 2 3 4 T… view at source ↗
Figure 5
Figure 5. Figure 5: Cluster occupancy map. Each panel shows cluster assignments on the L×M perturbation grid of a random sample (rows: textual paraphrases; columns: image per￾turbations). Horizontal stripes indicate invariance across image perturbations, showing that clustering is dominant by textual paraphrase. We provide more results in Supp. To qualitatively examine this effect, we visualize cluster occupancy on the L × M … view at source ↗
Figure 6
Figure 6. Figure 6: Visual Semantic Entropy for VQA Uncertainty Estimation. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative Result of Visual Semantic Entropy. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 1
Figure 1. Figure 1: Visual entropy estimation. Final-layer visual token embeddings are pro￾jected into the vocabulary space via the language model head (LogitLens). The entropy of the resulting token distributions is averaged to obtain the visual entropy Hvis, which measures the model’s confidence in the visual input. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p017_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Left: Perturbation effects under 10% & 30% percentiles. Right: Perturbation effects using Gemma3 on VILP and VAB [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Embedding shifts induced by text and image perturbations. [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Image vs. Text Embedding Space. t-SNE projections of multimodal em￾beddings for several image–question pairs from AOKVQA dataset. Colors indicate the same original image-question pair, while marker shapes distinguish perturbation types (text vs. image). Image perturbations remain tightly clustered around the original embedding, whereas question paraphrases induce larger shifts and spread to distant regions… view at source ↗
Figure 5
Figure 5. Figure 5: Cluster occupancy map. Each panel shows cluster assignments on the L×M perturbation grid of a random sample (rows: textual paraphrases; columns: image per￾turbations). Horizontal stripes indicate invariance across image perturbations, showing that clustering is dominant by textual paraphrase. (AOKVQA dataset, Qwen2.5-VL). E Hyperparameters Sensitivity E.1 Temperature T and Number of samples M [PITH_FULL_I… view at source ↗
Figure 6
Figure 6. Figure 6: Decoding temperature T and Number of variants M ablation. Perfor￾mance improves with more variants and moderate temperatures. Results on AOKVQA dataset. once M ≥ 15, despite the linear increase in computational cost, indicating di￾minishing returns from additional sampling. Higher temperatures (T ∈ [0.8, 1.0]) consistently achieve the best results, while lower temperatures suppress output diversity and lim… view at source ↗
Figure 7
Figure 7. Figure 7: Adequate sampling (M) and high temperature (T) demonstrate improved per￾formance, with gains saturating as M increases. Results on AOKVQA dataset [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visual perturbation strength ξ ablation. Moderate noise improves per￾formance, while stronger perturbations cause slight degradation. Results on AOKVQA dataset. E.2 Augmentation strength ξ We vary the visual perturbation strength from 20 to 500 and observe that moder￾ate noise at 20 yields the best performance, while stronger perturbations slightly reduce AUC ( [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative Results. A high VSE indicates that the model is uncertain about its original prediction, especially when that prediction is incorrect [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative Results. A low VSE indicates that the model is certain about its original prediction, especially when that prediction is correct [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative Results. Failure cases where VSE is low for incorrect answer (top) and high for correct answer (bottom). In the bottom example, the visual evidence is highly ambiguous, suggesting that the model arrives at the correct answer largely by chance [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
read the original abstract

Vision-language models can produce confident answers on visually ambiguous inputs, resulting in biased predictions. Common entropy-based methods, such as Semantic Entropy (SE), rely on output diversity. Yet our analysis shows that overconfident visual embeddings suppress output diversity under stochastic decoding, causing SE to underestimate uncertainty in such cases. Recent methods instead probe output diversity through input perturbations, including textual paraphrasing or joint text-image perturbations, and show improved performance. We study these approaches and reveals that the resulting variability is often dominated by textual changes rather than visual evidence, causing uncertainty estimates to reflect prompt sensitivity rather than visual ambiguity. We therefore propose Visual Semantic Entropy (VSE), which perturbs only the image to probe nearby visual variations while keeping the text query fixed. VSE measures uncertainty by clustering generated answers into semantic prototypes and computing the mass-weighted dispersion among them. Extensive evaluation across five modern vision-language models and five diverse VQA benchmarks demonstrates that VSE effectively captures visual ambiguity, establishing a new state-of-the-art for VLM uncertainty estimation.

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 manuscript proposes Visual Semantic Entropy (VSE) for uncertainty estimation in vision-language models. It perturbs only the image (text query fixed), generates multiple answers under stochastic decoding, clusters them into semantic prototypes, and computes mass-weighted dispersion among prototypes. The authors claim that standard Semantic Entropy underestimates uncertainty due to overconfident visual embeddings suppressing output diversity, while joint text-image perturbations are dominated by textual changes rather than visual evidence. Extensive experiments across five modern VLMs and five VQA benchmarks are said to show that VSE captures visual ambiguity and establishes a new state-of-the-art.

Significance. If the central claims hold, VSE would provide a targeted, image-focused uncertainty measure that isolates visual ambiguity more cleanly than prior entropy or perturbation baselines. The approach appears parameter-free (no fitted parameters or ad-hoc thresholds reported in the abstract or ledger), which is a strength for reproducibility. Demonstrating SOTA across multiple models and benchmarks would be a solid empirical contribution to VLM calibration if the image-only premise is validated.

major comments (2)
  1. [Abstract and method description (§3)] The core premise that image-only perturbation (text fixed) produces dispersion specifically reflecting visual ambiguity rather than VLM embedding instability, decoding artifacts, or model-specific visual noise is load-bearing for the SOTA claim but is not supported by controls or analysis. The skeptic concern applies directly: without evidence that semantic clusters arise from visual content variation (e.g., via content-controlled ablations or comparison to non-semantic image noise), the method may not isolate the intended quantity, undermining the assertion that prior joint perturbations are text-dominated and that VSE is superior on the five benchmarks.
  2. [Evaluation (§4-5)] The abstract asserts SOTA performance on five VLMs and five VQA benchmarks, yet the reader's assessment notes the absence of quantitative results, ablation details, or error analysis even in the full text summary. To support the central claim that VSE 'effectively captures visual ambiguity,' the evaluation section must include per-benchmark tables with baselines, statistical significance, and analysis showing that dispersion correlates with visual ambiguity (not model artifacts).
minor comments (2)
  1. [Method] Clarify the precise clustering procedure for semantic prototypes (algorithm, distance metric, and any implicit thresholds) and the exact formula for mass-weighted dispersion.
  2. [Analysis of prior methods] Add explicit comparison tables or figures showing that textual perturbation variance exceeds visual perturbation variance in the joint setting, with quantitative support.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and detailed review of our manuscript. We address each major comment below with clarifications drawn directly from the paper and indicate revisions where appropriate to strengthen the presentation of Visual Semantic Entropy.

read point-by-point responses

  1. Referee: [Abstract and method description (§3)] The core premise that image-only perturbation (text fixed) produces dispersion specifically reflecting visual ambiguity rather than VLM embedding instability, decoding artifacts, or model-specific visual noise is load-bearing for the SOTA claim but is not supported by controls or analysis. The skeptic concern applies directly: without evidence that semantic clusters arise from visual content variation (e.g., via content-controlled ablations or comparison to non-semantic image noise), the method may not isolate the intended quantity, undermining the assertion that prior joint perturbations are text-dominated and that VSE is superior on the five benchmarks.

    Authors: We appreciate the referee highlighting the need for explicit validation that dispersion arises from visual content. Section 3 of the manuscript presents analysis showing that overconfident visual embeddings suppress output diversity under stochastic decoding, causing standard Semantic Entropy to underestimate uncertainty. The same section examines joint text-image perturbations and demonstrates that resulting variability is dominated by textual changes rather than visual evidence. While we did not include explicit content-controlled ablations against non-semantic image noise, the consistent superiority of VSE across five VQA benchmarks and five VLMs provides supporting evidence that the measure isolates visual ambiguity. We will add the suggested content-controlled ablations in the revised version. revision: partial

  2. Referee: [Evaluation (§4-5)] The abstract asserts SOTA performance on five VLMs and five VQA benchmarks, yet the reader's assessment notes the absence of quantitative results, ablation details, or error analysis even in the full text summary. To support the central claim that VSE 'effectively captures visual ambiguity,' the evaluation section must include per-benchmark tables with baselines, statistical significance, and analysis showing that dispersion correlates with visual ambiguity (not model artifacts).

    Authors: Sections 4 and 5 of the full manuscript report quantitative results on the five VQA benchmarks across the five VLMs, including direct comparisons to Semantic Entropy and joint perturbation baselines that establish the SOTA performance. We agree that expanding these sections with additional statistical significance tests, explicit per-benchmark tables, ablation details, and analysis correlating dispersion with visual ambiguity (versus artifacts) would improve clarity and address the concern. We will incorporate these elements in the revision. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method is a direct procedural definition

full rationale

The paper defines VSE explicitly as image-only perturbation (text fixed), followed by semantic clustering of outputs into prototypes and mass-weighted dispersion. This construction is presented as a measurement procedure without any reduction to fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations. The abstract and description contain no equations or steps that equate the output dispersion to its own inputs by construction, nor invoke uniqueness theorems from prior author work. The central claim rests on the procedural definition plus external benchmark evaluation, which is self-contained against the listed patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are stated in the provided text.

pith-pipeline@v0.9.1-grok · 5734 in / 1104 out tokens · 29135 ms · 2026-07-01T05:32:04.536598+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

29 extracted references · 11 canonical work pages · 8 internal anchors

  1. [1]

    Qwen3-VL Technical Report

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025) 12

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  2. [2]

    Qwen2.5-VL Technical Report

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923 (2025),https: //arxiv.org/abs/2502.13923, computer...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Nature630(8017), 625–630 (2024) 2, 4, 11, 12, 13, 14, 15

    Farquhar, S., Kossen, J., Kuhn, L., Gal, Y.: Detecting hallucinations in large lan- guage models using semantic entropy. Nature630(8017), 625–630 (2024) 2, 4, 11, 12, 13, 14, 15

    2024

  4. [4]

    Gemma 3 Technical Report

    Gemma Team: Gemma 3 technical report. arXiv preprint arXiv:2503.19786 (2025), https://arxiv.org/abs/2503.1978612, 18

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    In: Proceedings of the 4th Workshop on Trustworthy Natural Language Processing (TrustNLP 2024)

    Groot,T.,Valdenegro-Toro,M.:Overconfidenceiskey:Verbalizeduncertaintyeval- uation in large language and vision-language models. In: Proceedings of the 4th Workshop on Trustworthy Natural Language Processing (TrustNLP 2024). pp. 145–171 (2024) 1, 4, 12, 13, 14

    2024

  6. [6]

    DeBERTa: Decoding-enhanced BERT with Disentangled Attention

    He, P., Liu, X., Gao, J., Chen, W.: Deberta: Decoding-enhanced bert with disen- tangled attention. arXiv preprint arXiv:2006.03654 (2020) 12, 24

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  7. [7]

    In: The Fourteenth International Conference on Learning Representations 2, 4

    Huang, J., Xu, J., Shi, X., Hu, P., Feng, L., Zhu, X.: Revisiting confidence cal- ibration for misclassification detection in vlms. In: The Fourteenth International Conference on Learning Representations 2, 4

  8. [8]

    In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition

    Khan, Z., Fu, Y.: Consistency and uncertainty: Identifying unreliable responses from black-box vision-language models for selective visual question answering. In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition. pp. 10854–10863 (2024) 2, 3, 4, 6, 12, 13, 14

    2024

  9. [9]

    Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking

    Li, M., Zhang, Y., Long, D., Chen, K., Song, S., Bai, S., Yang, Z., Xie, P., Yang, A., Liu, D., et al.: Qwen3-vl-embedding and qwen3-vl-reranker: A unified framework for state-of-the-art multimodal retrieval and ranking. arXiv preprint arXiv:2601.04720 (2026) 3, 18

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  10. [10]

    In: Findings of the Association for Computational Linguistics: EMNLP 2024

    Li, Q., Geng, J., Lyu, C., Zhu, D., Panov, M., Karray, F.: Reference-free halluci- nation detection for large vision-language models. In: Findings of the Association for Computational Linguistics: EMNLP 2024. pp. 4542–4551 (2024) 2, 4, 12, 13, 14

    2024

  11. [11]

    io/blog/2024-01-30-llava-next/12

    Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: Llava-next: Improved reasoning, ocr, and world knowledge (January 2024),https://llava-vl.github. io/blog/2024-01-30-llava-next/12

    2024

  12. [12]

    arXiv preprint arXiv:2501.00569 (2024) 5, 8, 11

    Luo, T., Cao, A., Lee, G., Johnson, J., Lee, H.: Probing visual language priors in vlms. arXiv preprint arXiv:2501.00569 (2024) 5, 8, 11

    work page arXiv 2024

  13. [13]

    In: Proceedings of the 2023 conference on empirical methods in natural language processing

    Manakul, P., Liusie, A., Gales, M.: Selfcheckgpt: Zero-resource black-box halluci- nation detection for generative large language models. In: Proceedings of the 2023 conference on empirical methods in natural language processing. pp. 9004–9017 (2023) 2, 4, 12, 13, 14

    2023

  14. [14]

    In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition

    Marino,K.,Rastegari,M.,Farhadi,A.,Mottaghi,R.:Ok-vqa:Avisualquestionan- swering benchmark requiring external knowledge. In: Proceedings of the IEEE/cvf conference on computer vision and pattern recognition. pp. 3195–3204 (2019) 11 Visual Semantic Entropy 29

    2019

  15. [15]

    In: Findings of the Association for Computational Linguistics: ACL 2025

    Nguyen, D., Payani, A., Mirzasoleiman, B.: Beyond semantic entropy: Boosting llm uncertainty quantification with pairwise semantic similarity. In: Findings of the Association for Computational Linguistics: ACL 2025. pp. 4530–4540 (2025) 2, 3, 4, 8, 11, 12, 13, 14, 15

    2025

  16. [16]

    In: Introduction to HPC with MPI for Data Science, pp

    Nielsen, F.: Hierarchical clustering. In: Introduction to HPC with MPI for Data Science, pp. 195–211. Springer (2016) 12

    2016

  17. [17]

    Advances in Neural Information Processing Systems37, 8901–8929 (2024) 2, 4, 8, 12, 13, 14

    Nikitin, A., Kossen, J., Gal, Y., Marttinen, P.: Kernel language entropy: Fine- grained uncertainty quantification for llms from semantic similarities. Advances in Neural Information Processing Systems37, 8901–8929 (2024) 2, 4, 8, 12, 13, 14

    2024

  18. [18]

    LessWrong post (Aug 31 2020), https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting- gpt- the-logit-lens5, 6, 17

    nostalgebraist: interpreting gpt: the logit lens. LessWrong post (Aug 31 2020), https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting- gpt- the-logit-lens5, 6, 17

    2020

  19. [19]

    Pelucchi, M.: Exploring ChatGPT’s accuracy and confidence in high-resource lan- guages. Ph.D. thesis (2023) 1, 4

    2023

  20. [20]

    In: European conference on computer vision

    Schwenk, D., Khandelwal, A., Clark, C., Marino, K., Mottaghi, R.: A-okvqa: A benchmark for visual question answering using world knowledge. In: European conference on computer vision. pp. 146–162. Springer (2022) 3, 11

    2022

  21. [21]

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

    Valdenegro-Toro, M.: I find your lack of uncertainty in computer vision disturbing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1263–1272 (2021) 1, 4

    2021

  22. [22]

    Vision Language Models are Biased

    Vo, A., Nguyen, K.N., Taesiri, M.R., Dang, V.T., Nguyen, A.T., Kim, D.: Vision language models are biased. arXiv preprint arXiv:2505.23941 (2025) 11

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency

    Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., et al.: Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025) 12

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Advances in neural information processing systems33, 5776–5788 (2020) 23

    Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., Zhou, M.: Minilm: Deep self- attention distillation for task-agnostic compression of pre-trained transformers. Advances in neural information processing systems33, 5776–5788 (2020) 23

    2020

  25. [25]

    In: Proceedings of the 2025 Conference on Empirical Methods in Natural Lan- guage Processing

    Xuan, W., Zeng, Q., Qi, H., Wang, J., Yokoya, N.: Seeing is believing, but how much? a comprehensive analysis of verbalized calibration in vision-language mod- els. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Lan- guage Processing. pp. 1408–1450 (2025) 1, 4

    2025

  26. [26]

    In: International conference on machine learning

    Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., Wang, X., Wang, L.: Mm-vet: Evaluating large multimodal models for integrated capabilities. In: International conference on machine learning. PMLR (2024) 11

    2024

  27. [27]

    arXiv preprint arXiv:2411.11919 (2024) 2, 3, 4, 6, 10, 12, 13, 14, 18

    Zhang, R., Zhang, H., Zheng, Z.: Vl-uncertainty: Detecting hallucination in large vision-language model via uncertainty estimation. arXiv preprint arXiv:2411.11919 (2024) 2, 3, 4, 6, 10, 12, 13, 14, 18

    work page arXiv 2024

  28. [28]

    arXiv preprint arXiv:2411.00299 (2024) 2

    Zhang, S., Sambara, S., Banerjee, O., Acosta, J., Fahrner, L.J., Rajpurkar, P.: Radflag: A black-box hallucination detection method for medical vision language models. arXiv preprint arXiv:2411.00299 (2024) 2

    work page arXiv 2024

  29. [29]

    BERTScore: Evaluating Text Generation with BERT

    Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675 (2019) 24

    work page internal anchor Pith review Pith/arXiv arXiv 1904