Pith. sign in

REVIEW 3 major objections 6 minor 53 references

TruthLens: Object Hallucination Detection via Self-Evaluating Truthfulness Scores in LVLMs

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read TruthLens turns a rarely-used special token's log-probability into a per-object honesty score, beating prior object-hallucination detectors by more than 17 AUROC points on Qwen2.5-VL-7B.

desk verdict Worth engaging: a consistently strong empirical detector, but the truthfulness interpretation and the SOTA claim both need tightening before I fully trust them. read the letter →

arxiv 2608.05616 v1 pith:QLIYJDAQ submitted 2026-08-06 cs.CV

Yanqi Wu , Runhe Lai , Xinhua Lu , Qichao Chen , Zhiping Zhou , Jia-Xin Zhuang , Weijiang Yu , Ruixuan Wang This is my paper · ORCID
classification cs.CV
keywords objecthallucinationdetectionlargevision-languagemodeltruthfulnessscorespecialtokenlog-probabilityself-evaluationhidden-stateseparabilityMSEfine-tuningmitigation
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 claims that large vision-language models (LVLMs) carry a hidden truthfulness signal: in the last layer's hidden states, tokens for objects that actually appear in the image are separable from tokens for hallucinated objects, but the language-model head's output probabilities largely wash that separation out. TruthLens fine-tunes the model so that the log-probability it assigns to a designated low-probability special token at each object-token position becomes a per-object truthfulness score, regressed toward 1 for real objects and 0 for hallucinated ones, with a KL term to preserve generation quality. On MS-COCO and Object365, the resulting scores outperform earlier detectors across five open LVLMs, for example raising AUROC on Qwen2.5-VL-7B from 74.05 (the prior best GLSIM) to 91.46. The authors argue this is self-evaluation with no auxiliary model and no extra inference cost, since the score is just a log-probability the model already emits while generating.

What carries the argument

The load-bearing mechanism is the token-level truthfulness score S(o, X, I) = log π_θ(z_c | X, I, y_<o) − c_ref, applied at the position of the object token o, where z_c is a low-probability special token that the model almost never emits. Fine-tuning pushes this score toward 1 for grounded objects and 0 for hallucinated ones via a re-weighted MSE loss, while a KL-divergence term against the frozen original model preserves generation behavior and class re-weighting balances the real-versus-hallucinated token ratio. The paper also uses linear discriminant analysis (LDA) on last-layer hidden states to establish that real and hallucinated tokens are separable in feature space, and an output-null-space ratio defined as |Wv| / (|W|_F |v|) with v = m_r − m_h, which measures how weakly the discriminative direction aligns with the LM head's readout and thereby explains why the separability is lost in output probabilities.

What would settle it

Take a held-out set of object tokens, match real and hallucinated tokens for word frequency, category prior, and caption position, and rerun the LDA separator; if AUROC drops to near chance, the claimed truthfulness signal is a confound of lexical correlates. A complementary check is to compute per-category AUROC for rare versus common categories: if hallucinated tokens concentrate in rare categories, control for category frequency and see whether the fine-tuned score still separates.

Watch

Extended reading notes

Core claim

The central claim is that the LM head can be taught to expose a truthfulness signal that is present in hidden representations but absent from its usual next-token probabilities. The discovery is operational: define S(o, X, I) = log π_θ(z_c | X, I, y_<o) − c_ref, the difference between the fine-tuned model's log-probability of a rarely-used special token (such as <unk> or <|image_pad|>) at the position of object token o and a predefined constant; fine-tune with an MSE loss toward 1 for real objects and 0 for hallucinated ones, plus a KL divergence constraint against a frozen copy of the original model; then S(o, X, I) becomes a per-object truthfulness score. The paper reports consistent AUROC and AUPR gains over negative-log-likelihood, entropy, internal-confidence, attention-based, and similarity baselines on MS-COCO and Object365, and generalization to attribute-hallucination benchmarks CLEVR and SpatialMQA with negligible loss of general capability.

Load-bearing premise

TruthLens assumes that the separability its LDA probe measures really encodes visual truthfulness, not side information such as category rarity, word frequency, or caption position; if the separator exploits those correlates, the fine-tuned score may be fitting lexical statistics rather than groundedness.

Editorial extensions

If this is right

  • Object-hallucination detection becomes a zero-overhead readout: at inference, a single log-probability from the already-running model replaces auxiliary detectors or post-hoc similarity computations.
  • Because training uses only MS-COCO object labels, the method's transfer to Object365's 365 categories suggests the score captures a general truthfulness dimension rather than memorized category statistics.
  • The same score can drive a lightweight mitigation loop: thresholding the score and asking the model to revise its caption removes hallucinated mentions and lowers CHAIRi and CHAIRs while keeping recall nearly unchanged.
  • The representation–projection mismatch finding implies that other latent truthfulness signals in LVLMs might be surfaced by similar special-token probing across tasks and modalities.

Reading between the lines

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

  • An implication the authors leave implicit is that the same fine-tuning recipe could be applied to factual-consistency detection in text-only LLMs, where hidden-state separability of true versus false claims has been observed, yielding a self-evaluative check without extra machinery.
  • Because training data are generated by the model's own sampling, the method's ceiling is set by the model's capability: a model that cannot recognize an object as present cannot be made honest by this regression alone, so TruthLens inherits rather than repairs perceptual limits.
  • A direct test of the truthfulness interpretation would be to measure whether the calibration of c_ref transfers across datasets without retuning; if the same reference constant works on Object365 and MS-COCO, that supports a grounded semantics rather than dataset-specific fitting.
  • The near-invariance to the choice of special token (AUROC varies less than 0.4 points across candidates) suggests the output distribution contains a generic low-probability channel that can carry supervised truthfulness information, which could be exploited for other self-evaluation targets.
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

3 major / 6 minor

Summary. The paper proposes TruthLens, a fine-tuning framework for object hallucination detection in LVLMs. The authors report that real and hallucinated object tokens are linearly separable in last-layer hidden states but that this separability is largely absent in the token probabilities produced by the LM head. They repurpose a rarely used special token as a reference and define a per-object truthfulness score as the log-probability of that token at the object position minus a constant. The model is fine-tuned with an MSE loss that pushes truthfulness scores of real tokens toward 1 and hallucinated tokens toward 0, with a KL divergence constraint to preserve generation capabilities. Experiments on MSCOCO, Objects365, CLEVR, and SpatialMQA across five LVLMs report large AUROC gains over prior baselines, and the method is also applied to hallucination mitigation via caption revision. The paper additionally analyzes layer-wise separability and an output-null ratio to explain the LM-head failure.

Significance. If the empirical results are taken at face value, TruthLens is a strong and practical OH detector: it adds no inference-time modules, preserves general capabilities on a broad benchmark suite (Table 5), and shows large gains over existing methods, with standard deviations over three seeds reported in Appendix A.2. The cross-dataset transfer from a COCO-trained model to Objects365 is impressive, and the release of code is a concrete asset. However, the central interpretative claim—that the score captures intrinsic truthfulness rather than lexical, categorical, or positional confounds—is not yet supported by the evidence presented, and the motivation experiment in Section 3.3 compares a supervised probe with an unsupervised scalar. The paper is therefore a potentially strong engineering contribution whose scientific interpretation and mechanistic claims require additional substantiation.

major comments (3)
  1. [Section 3.3, Fig. 3a] The comparison between LDA on hidden states and NLL on output probabilities is confounded: LDA is a supervised classifier trained on 80% of the tokens and evaluated on the remaining 20%, whereas NLL is an unsupervised pointwise score. This design cannot establish that the LM head loses the separability present in hidden states. A fair comparison would use a supervised probe on the output logits or pre-softmax activations, or an unsupervised measure of hidden-state separability. As written, the claim that the discriminative signal is 'not fully propagated or preserved in the final output probabilities' is overstated.
  2. [Sections 4.2 and 5] The claim that TruthLens captures 'intrinsic truthfulness-related signals rather than dataset-specific object statistics' is not supported by the reported aggregate AUROC numbers. The LDA separability in Section 3.3 and the Objects365 transfer could be driven by confounds such as object-category frequency, token position in the caption, or token surface form. The authors should report per-category AUROC for both categories seen and unseen in training, and/or provide confound-controlled evaluations (e.g., stratified by position and word-frequency bands, or matched-frequency analysis). Without such controls, the truthfulness interpretation of the score is not established.
  3. [Section 5, Eq. (6)] The output-null ratio analysis is not convincing as stated. The ratio |Wv|/(|W|_F |v|) is a crude measure of alignment; small values may simply reflect the high dimensionality of the output space, and the authors do not compare against random directions or against the ratio for other discriminative directions. As a result, the conclusion that the LM head is 'functionally silent' is not supported. In addition, Eq. (6) as printed is missing the division operator, which should be corrected.
minor comments (6)
  1. [Throughout] The word 'Table' is consistently misspelled as 'T able' in table captions and in-text references; please correct.
  2. [Table 2] The entry '82,60' in the GLSIM row under LLaVA-OneVision on CLEVR should be '82.60'.
  3. [Section 2 and Table 1] The baseline is referred to as 'ContextLens' in the related work but as 'Contextual Lens' in tables and elsewhere; please unify the naming.
  4. [Eq. (3)] The symbol z_c is not explicitly defined as the designated special token in the main text; please define it when first used.
  5. [Section 3.3] The sentence 'using 80% of the samples for training' should say '80% of the object tokens' rather than 'samples', since the units are tokens.
  6. [Section 4.3, hyperparameter sensitivity] The text says performance saturates 'beyond 23 rollouts', but the x-axis of Fig. 4 appears to show powers of 2 (2^1 to 2^4), so the intended meaning is likely 'beyond 2^3 rollouts' or 'beyond 8 rollouts'; please clarify.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: TruthLens is a supervised detector trained on external labels and tested on held-out data; the sole self-citation is a related-work baseline, not load-bearing.

full rationale

The derivation chain is empirical and self-contained. The truthfulness score in Eq. 3 is defined as the log-probability of a special token minus a constant, and Eq. 5 fits this score to CHAIR-derived ground-truth labels via MSE; no equation reduces a target result to its input by construction. The LDA separability in Section 3.3 is measured on held-out tokens (80/20 split) and motivates, but does not define, the score; the final AUROC is evaluated on held-out MSCOCO test instances and on Object365/CLEVR/SpatialMQA, so the fitted score is not being 'predicted' back onto its own training labels. The only self-citation is [19] (InsLen), a related-work baseline by the same group; it is not used as a premise for TruthLens and does not carry the argument. Concerns about possible confounds (category priors, token frequency, position) are validity threats, not circularity: they question whether the learned score captures truthfulness rather than corpus statistics, but they do not show that any claim is true by definition or by fitted-input renaming. The borrow from LASER [49] is explicitly attributed to external work, so no ansatz is smuggled via self-citation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or architectural entity; it repurposes an existing special token. Its central dependencies are empirical: label quality from CHAIR matching, the interpretation of LDA separability as truthfulness, and the assumption that special-token log-probability can be reshaped without hurting generation. The main fitted quantities are c_ref, beta, the decision threshold, the rollout count, and LoRA hyperparameters.

free parameters (5)
  • c_ref = -28.0
    Additive baseline in Eq. 3, tuned to the special-token log-probability scale. AUROC at inference is invariant to c_ref, but during training it shifts the MSE regression target and affects optimization; sensitivity is shown in Fig. 4.
  • beta = 0.1
    Scales the MSE gradient in Eq. 4-5; set by sensitivity analysis and controls the trade-off between score alignment and KL preservation.
  • detection threshold mu = 0.3 (case study)
    Decision threshold in Eq. 1 is set on a 500-sample validation set in Appendix A.4. It is not needed for AUROC evaluation but is used for the binary detection claim.
  • number of sampled responses n = 8
    Number of stochastic captions sampled per training image; Fig. 4 shows performance saturates near 2^3 rollouts, so this is an empirical hyperparameter.
  • LoRA rank and alpha = r=8, alpha=16
    Standard parameter-efficient fine-tuning choices; no sensitivity analysis is reported for these values.
assumptions (4)
  • domain assumption Generated object tokens matched to ground-truth annotations via CHAIR and finite synonym lists yield correct real and hallucinated labels.
    Entered in Section 4.1 Implementation Details. Noisy labels would bias both training targets and evaluation.
  • domain assumption The separability measured by supervised LDA is a truthfulness signal, not a confound of object identity, word frequency, or token position.
    Section 3.3 uses LDA trained on labels; no per-category or position counterbalancing is provided.
  • domain assumption Adjusting the log-probability of a rarely used special token does not materially change generation, and the KL penalty preserves original capability.
    Section 3.4 and Table 5; supported by general benchmark results for LLaVA-1.5-7B only.
  • domain assumption Fine-tuning on MSCOCO categories transfers because the score captures intrinsic truthfulness rather than dataset-specific statistics.
    Used to interpret Object365 generalization; supported empirically but not proven by a controlled analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TruthLens: Object Hallucination Detection via Self-Evaluating Truthfulness Scores in LVLMs." pith.science (2026). https://pith.science/paper/QLIYJDAQ

@misc{pith2026260805616,
  author       = {Pith},
  title        = {Pith review of: TruthLens: Object Hallucination Detection via Self-Evaluating Truthfulness Scores in LVLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QLIYJDAQ}},
  note         = {Machine review of arXiv:2608.05616}
}
read the original abstract

Despite the remarkable progress of large vision language models (LVLMs), object hallucination remains a fundamental challenge that hinders their trustworthy deployment. A key finding motivates our work: real and hallucinated object tokens are clearly separable in hidden representations, yet this separability is largely lost at the language-modeling (LM) head. We propose TruthLens, a self-evaluation framework that teaches the LM head to expose a per-object truthfulness signal without any auxiliary model or additional inference cost. Concretely, a rarely-used special token is repurposed as a reference token. For each object-token position, we extract the log-probability assigned to this special token by the LM head, and define its difference from a predefined constant as the truthfulness score. The model is then fine-tuned with an MSE objective that drives scores toward 1 for real objects and 0 for hallucinated ones, while a divergence constraint preserves the original generation capability. Despite being trained on only a limited set of object categories, TruthLens generalizes effectively to benchmarks with substantially larger label spaces. Extensive experiments across multiple LVLMs demonstrate state-of-the-art performance; notably, on Qwen2.5-VL-7B, TruthLens outperforms the previous best method on MS-COCO by over 17\% in AUROC. Our code is available at https://github.com/wyqstan/TruthLens.

Figures

Figures reproduced from arXiv: 2608.05616 by the authors.

Figure 1
Figure 1. Hallucination detection performance of five open-source LVLMs [1, 2, 26, 27] using six different scoring metrics on MSCOCO [24] and Object365 [41] datasets. reasoning by integrating visual perception with the language understanding ca￾pabilities of large language models (LLMs). Despite their impressive perfor￾mance, LVLMs frequently suffer from object hallucination (OH) [23, 40]– gen￾erating mentions of objects abse… view at source ↗
Figure 2
Figure 2. Overview of the proposed self-evaluation framework. The truthfulness score is formulated as the difference between the log-probability of a designated special token, predicted at the corresponding token position, and a constant. During training, an MSE loss aligns this score with binary rewards (1 for real, 0 for hallucinated). At inference, the score is directly leveraged for hallucination detection without auxilia… view at source ↗
Figure 3
Figure 3. Comparison of hallucination detection performance and LDA-based visualiza￾tion on MSCOCO across different LVLM architectures. encode substantial information about answer correctness. By analogy, in OH detection, the token positions corresponding to generated object mentions can be viewed as critical decision points that determine the factual consistency of the output. We therefore hypothesize that the hidden represe… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Sensitivity studies of hyperparameters on LLaVA-1.5-7B and Qwen2.5-VL-7B. Dashed lines indicate the strongest baselines. tings—even with a single rollout—demonstrating that the improvements pri￾marily stem from our reward modeling design rather than increased sampling.…
Figure 5
Figure 5. Figure 5: Evolution of token separability. by extracting hidden states from each Transformer layer and applying LDA in￾dependently. As shown in [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of InsLen and GLSIM for detecting object hallucina￾tions, using the LLaVA-1.5-7B model. In the generated responses, ground-truth objects are highlighted in green, while hallucinated objects appear in red. Detection outcomes are color-coded: green…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 23 canonical work pages

  1. [1]

    arXiv preprint arXiv:2509.23661 (2025)

    An, X., Xie, Y., Yang, K., Zhang, W., Zhao, X., Cheng, Z., Wang, Y., Xu, S., Chen, C., Zhu, D., Wu, C., Tan, H., Li, C., Yang, J., Yu, J., Wang, X., Qin, B., Wang, Y., Yan, Z., Feng, Z., Liu, Z., Li, B., Deng, J.: LLaVA-OneVision- 1.5: Fully Open Framework for Democratized Multimodal Training. arXiv preprint arXiv:2509.23661 (2025)

    arXiv 2025

  2. [2]

    arXiv preprint arXiv:2502.13923 (2025)

    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., Wang, P., 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)

    arXiv 2025

  3. [3]

    Journal of Systems and Software (2025)

    Bui, T.D., Vu, T.T., Nguyen, T.T., Nguyen, S., Vo, H.D.: Correctness Assessment of Code Generated by Large Language Models using Internal Representations. Journal of Systems and Software (2025)

    work page 2025

  4. [4]

    arXiv preprint arXiv:2402.03744 (2024)

    Chen, C., Liu, K., Chen, Z., Gu, Y., Wu, Y., Tao, M., Fu, Z., Ye, J.: INSIDE: LLMs’ Internal States Retain the Power of Hallucination Detection. arXiv preprint arXiv:2402.03744 (2024)

    arXiv 2024

  5. [5]

    Chiang, W.L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J.E., Stoica, I., Xing, E.P.: Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality (2023)

    work page 2023

  6. [6]

    arXiv preprint arXiv:2510.17205 (2025)

    Fan, Y., Zhao, A., Fu, J., Tong, J., Su, H., Pan, Y., Zhang, W., Shen, X.: VisiPruner: Decoding Discontinuous Cross-Modal Dynamics for Efficient Mul- timodal LLMs. arXiv preprint arXiv:2510.17205 (2025)

    arXiv 2025

  7. [7]

    Annals of Eugenics pp

    Fisher, R.A.: The Use of Multiple Measurements in Taxonomic Problems. Annals of Eugenics pp. 179–188 (1936)

    work page 1936

  8. [8]

    Fu, C., Chen, P., Shen, Y., Qin, Y., Zhang, M., Lin, X., Yang, J., Zheng, X., Li, K., Sun, X., Wu, Y., Ji, R., Shan, C., He, R.: Mme: A comprehensive evaluation benchmark for multimodal large language models (2025)

    work page 2025

Show all 53 references
  1. [9]

    arXiv preprint arXiv:1612.00837 (2017)

    Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in vqa matter: Elevating the role of image understanding in visual question answering. arXiv preprint arXiv:1612.00837 (2017)

    2017 arXiv

  2. [10]

    arXiv preprint arXiv:1802.08218 (2018)

    Gurari, D., Li, Q., Stangl, A.J., Guo, A., Lin, C., Grauman, K., Luo, J., Bigham, J.P.: Vizwiz grand challenge: Answering visual questions from blind people. arXiv preprint arXiv:1802.08218 (2018)

    2018 arXiv

  3. [11]

    ICLR (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: LoRA: Low-rank Adaptation of Large Language Models. ICLR (2022)

    2022

  4. [12]

    arXiv preprint arXiv:1902.09506 (2019)

    Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. arXiv preprint arXiv:1902.09506 (2019)

    2019 arXiv

  5. [13]

    arXiv preprint arXiv:2410.02762 (2025)

    Jiang, N., Kachinthaya, A., Petryk, S., Gandelsman, Y.: Interpreting and Edit- ing Vision-Language Representations to Mitigate Hallucinations. arXiv preprint arXiv:2410.02762 (2025)

    2025 arXiv

  6. [14]

    In: CVPR (2025)

    Jiang, Z., Chen, J., Zhu, B., Luo, T., Shen, Y., Yang, X.: Devils in middle lay- ers of large vision-language models: Interpreting, detecting and mitigating object hallucinations via attention lens. In: CVPR (2025)

    2025

  7. [15]

    In: CVPR (2017)

    Johnson, J., Hariharan, B., Van Der Maaten, L., Fei-Fei, L., Lawrence Zitnick, C., Girshick, R.: Clevr: A Diagnostic Dataset for Compositional Language and Elementary Visual Reasoning. In: CVPR (2017)

    2017

  8. [16]

    arXiv preprint arXiv:2207.05221 (2022)

    Kadavath, S., Conerly, T., Askell, A., Henighan, T., Drain, D., Perez, E., Schiefer, N., Hatfield-Dodds, Z., DasSarma, N., Tran-Johnson, E., Johnston, S., El-Showk, TruthLens 17 S., Jones, A., Elhage, N., Hume, T., Chen, A., Bai, Y., Bowman, S., Fort, S., Gan- guli, D., Hernan...

    2022 arXiv

  9. [17]

    arXiv preprint arXiv:2509.04664 (2025)

    Kalai, A.T., Nachum, O., Vempala, S.S., Zhang, E.: Why Language Models Hal- lucinate. arXiv preprint arXiv:2509.04664 (2025)

    2025 arXiv

  10. [18]

    Nature Neuroscience (2014)

    Kaufman, M., Churchland, M., Ryu, S., Shenoy, K.: Cortical Activity in the Null Space: Permitting Preparation without Movement. Nature Neuroscience (2014)

    2014

  11. [19]

    arXiv preprint arXiv:2605.12258 (2026)

    Lai, R., Lu, X., Wu, Y., Ye, J., Yu, W., Wang, R.: Instruction lens score: Your instruction contributes a powerful object hallucination detector for multimodal large language models. arXiv preprint arXiv:2605.12258 (2026)

    2026 arXiv

  12. [20]

    arXiv preprint arXiv:2307.16125 (2023)

    Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., Shan, Y.: Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125 (2023)

    2023 arXiv

  13. [21]

    Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping Language-Image Pre- trainingwithFrozenImageEncodersandLargeLanguageModels.In:ICML(2023)

    2023

  14. [22]

    NeurIPS (2023)

    Li, K., Patel, O., Viégas, F., Pfister, H., Wattenberg, M.: Inference-time Interven- tion: Eliciting Truthful Answers from a Language Model. NeurIPS (2023)

    2023

  15. [23]

    In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (2023)

    Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, X., Wen, J.R.: Evaluating Object Hallu- cination in Large Vision-Language Models. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (2023)

    2023

  16. [24]

    In: ECCV (2014)

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common Objects in Context. In: ECCV (2014)

    2014

  17. [25]

    arXiv preprint arXiv:2306.14565 (2024)

    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:2306.14565 (2024)

    2024 arXiv

  18. [26]

    In: CVPR (2024)

    Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: CVPR (2024)

    2024

  19. [27]

    Liu, H., Li, C., Li, Y., Li, B., Zhang, Y., Shen, S., Lee, Y.J.: LLaVA-NeXT: Im- proved reasoning, OCR, and world knowledge (2024)

    2024

  20. [28]

    NeurIPS (2023)

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. NeurIPS (2023)

    2023

  21. [29]

    Liu, J., Liu, Z., Cen, Z., Zhou, Y., Zou, Y., Zhang, W., Jiang, H., Ruan, T.: Can Multimodal Large Language Models Understand Spatial Relations? In: Proceed- ings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (2025)

    2025

  22. [30]

    arXiv preprint arXiv:2503.15850 (2025)

    Liu, X., Chen, T., Da, L., Chen, C., Lin, Z., Wei, H.: Uncertainty Quantification and Confidence Calibration in Large Language Models: A Survey. arXiv preprint arXiv:2503.15850 (2025)

    2025 arXiv

  23. [31]

    Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., et al.: Mmbench: Is your multi-modal model an all-around player? In: ECCV (2024)

    2024

  24. [32]

    NeurIPS (2022)

    Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.W., Zhu, S.C., Tafjord, O., Clark, P., Kalyan, A.: Learn to explain: Multimodal reasoning via thought chains for science question answering. NeurIPS (2022)

    2022

  25. [33]

    arXiv preprint arXiv:2002.07650 (2021)

    Malinin, A., Gales, M.: Uncertainty estimation in autoregressive structured pre- diction. arXiv preprint arXiv:2002.07650 (2021)

    2021 arXiv

  26. [34]

    NeurIPS (2024) 18 Wu et al

    Meng, L., Yang, J., Tian, R., Dai, X., Wu, Z., Gao, J., Jiang, Y.G.: Deepstack: Deeply Stacking Visual Tokens is Surprisingly Simple and Effective for LMMs. NeurIPS (2024) 18 Wu et al

    2024

  27. [35]

    arXiv preprint arXiv:2410.02707 (2025)

    Orgad, H., Toker, M., Gekhman, Z., Reichart, R., Szpektor, I., Kotek, H., Belinkov, Y.: LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations. arXiv preprint arXiv:2410.02707 (2025)

    2025 arXiv

  28. [36]

    arXiv preprint arXiv:2508.19972 (2025)

    Park, S., Li, S.: GLSIM: Detecting Object Hallucinations in LVLMs via Global- Local Similarity. arXiv preprint arXiv:2508.19972 (2025)

    2025

  29. [37]

    Phukan, A., Divyansh, D., Morj, H.K., Vaishnavi, V., Saxena, A., Goswami, K.: Beyond Logit Lens: Contextual Embeddings for Robust Hallucination Detection & Grounding in VLMs. In: Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for C...

    2025

  30. [38]

    arXiv preprint arXiv:2412.15115 (2025)

    Qwen, Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Wei, H., Lin, H., Yang, J., Tu, J., Zhang, J., Yang, J., Yang, J., Zhou, J., Lin, J., Dang, K., Lu, K., Bao, K., Yang, K., Yu, L., Li, M., Xue, M., Zhang, P., Zhu, Q., Men, R., Lin, R....

    2025 arXiv

  31. [39]

    In: ICML (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: ICML (2021)

    2021

  32. [40]

    In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (2018)

    Rohrbach, A., Hendricks, L.A., Burns, K., Darrell, T., Saenko, K.: Object Halluci- nation in Image Captioning. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (2018)

    2018

  33. [41]

    In: ICCV (2019)

    Shao, S., Li, Z., Zhang, T., Peng, C., Yu, G., Zhang, X., Li, J., Sun, J.: Objects365: A Large-Scale, High-Quality Dataset for Object Detection. In: ICCV (2019)

    2019

  34. [42]

    arXiv preprint arXiv:2502.12964 (2025)

    Simhi, A., Itzhak, I., Barez, F., Stanovsky, G., Belinkov, Y.: Trust Me, I’m Wrong: LLMs Hallucinate with Certainty Despite Knowing the Answer. arXiv preprint arXiv:2502.12964 (2025)

    2025 arXiv

  35. [43]

    In: CVPR (2019)

    Singh, A., Natarajan, V., Shah, M., Jiang, Y., Chen, X., Batra, D., Parikh, D., Rohrbach, M.: Towards vqa models that can read. In: CVPR (2019)

    2019

  36. [44]

    arXiv preprint arXiv:2302.13971 (2023)

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  37. [45]

    arXiv preprint arXiv:2307.09288 (2023)

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bash- lykov, N., Batra, S., Bhargava, P., Bhosale, S., Bikel, D., Blecher, L., Ferrer, C.C., Chen, M., Cucurull, G., Esiobu, D., Fernandes, J., Fu, J., Fu, W., Fuller, B., Gao, C., Goswami, V., Goyal, N...

    2023 arXiv

  38. [46]

    arXiv preprint arXiv:2308.15126 (2023)

    Wang, J., Zhou, Y., Xu, G., Shi, P., Zhao, C., Xu, H., Ye, Q., Yan, M., Zhang, J., Zhu, J., Sang, J., Tang, H.: Evaluation and Analysis of Hallucination in Large Vision-Language Models. arXiv preprint arXiv:2308.15126 (2023)

    2023 arXiv

  39. [47]

    Blog post (Nov 2024), accessed: 2025-05-12 TruthLens 19

    X.AI: RealWorldQA. Blog post (Nov 2024), accessed: 2025-05-12 TruthLens 19

    2024

  40. [48]

    arXiv preprint arXiv:2407.10671 (2024)

    Yang, A., Yang, B., Hui, B., Zheng, B., Yu, B., Zhou, C., Li, C., Li, C., Liu, D., Huang, F., Dong, G., Wei, H., Lin, H., Tang, J., Wang, J., Yang, J., Tu, J., Zhang, J., Ma, J., Yang, J., Xu, J., Zhou, J., Bai, J., He, J., Lin, J., Dang, K., Lu, K., Chen, K., Yang, K., Li, M....

    2024 arXiv

  41. [49]

    arXiv preprint arXiv:2510.14943 (2025)

    Yang, W., Liu, W., Xie, R., Guo, Y., Wu, L., Yang, S., Lin, Y.: LaSeR: Reinforce- ment Learning with Last-Token Self-Rewarding. arXiv preprint arXiv:2510.14943 (2025)

    2025

  42. [50]

    arXiv preprint arXiv:2408.04840 (2024)

    Ye, J., Xu, H., Liu, H., Hu, A., Yan, M., Qian, Q., Zhang, J., Huang, F., Zhou, J.: mPLUG-Owl3: Towards Long Image-Sequence Understanding in Multi-Modal Large Language Models. arXiv preprint arXiv:2408.04840 (2024)

    2024 arXiv

  43. [51]

    arXiv preprint arXiv:2308.02490 (2023)

    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. arXiv preprint arXiv:2308.02490 (2023)

    2023 arXiv

  44. [52]

    In: CVPR (2025)

    Zhang, Z., Yadav, S., Han, F., Shutova, E.: Cross-modal information flow in mul- timodal large language models. In: CVPR (2025)

    2025

  45. [53]

    arXiv preprint arXiv:2310.00754 (2024) 20 Wu et al

    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 Models. arXiv preprint arXiv:2310.00754 (2024) 20 Wu et al. A Appendix A.1 Baselines Negative Log-likelihood[53]. To quantify...

    2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.