Pith. sign in

REVIEW 4 major objections 5 minor 15 references

When Visual Signals Mislead: A Mechanistic Study of Attribute Hallucination in Vision-Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Attribute hallucination in vision-language models is driven by weak or misaligned visual signals, not by a dominant language prior.

desk verdict Useful diagnostic, but the headline causal claim is partly a tautology; the scoped material/color asymmetry is the real contribution. read the letter →

arxiv 2608.11024 v1 pith:RE3OQ3DO submitted 2026-08-11 cs.CV

Yufei Zhang , Chenlu Zhan , Hongwei Wang This is my paper · ORCID
classification cs.CV
keywords attributehallucinationvision-languagemodelsvisualsignalqualitylanguagepriornull-imagediagnosismechanisticanalysiscalibrationadaptation
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

This paper claims that attribute hallucination in vision-language models—saying a visibly present object has the wrong color, material, or state—is not primarily caused by a dominant language prior, the model's tendency to lean on word co-occurrence statistics. Instead, the raw visual signal itself predicts false "yes" answers, either because it is too weak to cross the decision threshold or because it is misaligned in the final output layer. The authors measure, for each attribute query, the logit difference between "yes" and "no" on the real image and on a blank gray image that should carry no visual content. Across 10,791 negative examples from three model families, the visual logit difference correlates with hallucination while the language-prior difference is near chance. If the finding holds, hallucination mitigation should start by diagnosing the visual signal rather than uniformly suppressing language priors.

What carries the argument

The load-bearing object is the VSNR diagnostic, which decomposes each yes/no attribute prediction into a visual logit difference, $\delta_{\mathrm{vis}} = \mathrm{logit}(\mathrm{yes}|I,q_w) - \mathrm{logit}(\mathrm{no}|I,q_w)$, and a prior logit difference, $\delta_{\mathrm{prior}} = \mathrm{logit}(\mathrm{yes}|I_\emptyset,q_w) - \mathrm{logit}(\mathrm{no}|I_\emptyset,q_w)$, computed with a uniform gray null image. The decision rule $\mathrm{sign}(\delta_{\mathrm{vis}})$ matches the hallucination label on all 10,791 samples, and a Bayesian decomposition writes the model's log-odds as a visual likelihood ratio plus a language prior, so $\delta_{\mathrm{vis}} - \delta_{\mathrm{prior}}$ approximates the visual contribution. The same coordinate yields a per-layer signal-to-noise ratio that localizes material failures to the final decoder layers and routes each query to calibration, abstention, or per-word visual adaptation.

What would settle it

Construct an attribute word whose blank-image prior is strongly positive yet whose real-image visual margin is negative, and show that false positives occur at a high rate despite the negative visual margin; alternatively, show that varying the gray level of the null image changes $\delta_{\mathrm{prior}}$ in a way that is correlated with hallucination, indicating the null image itself carries a visual prior.

Watch

Extended reading notes

Core claim

The central discovery is that attribute false positives are better explained by visual-signal quality than by language-prior dominance. For negative-ground-truth queries, the Spearman rank correlation between the visual logit margin and the hallucination outcome is high (0.755–0.835 across models), while the correlation between the null-image prior logit margin and hallucination is near zero (0.066–0.177), and a prior-only classifier barely beats the majority baseline. Two failure modes emerge: color and state attributes have directionally correct but low-margin visual signals, correctable by moving the decision threshold; material attributes have low signal-to-noise ratio or wrong-direction visual signals in late decoder layers, which threshold moves cannot fix. The paper concludes that language-prior suppression is misaligned with the actual mechanism for attribute hallucination.

Load-bearing premise

The conclusion depends on the blank gray image having no visual content relevant to the attribute, so the logit difference it produces truly measures the language prior, and on the residual $\delta_{\mathrm{vis}} - \delta_{\mathrm{prior}}$ being a valid measure of the visual likelihood contribution rather than just a definition of the leftover.

Editorial extensions

If this is right

  • Prior-suppression decoding methods cannot fix attribute false positives where the visual margin is already wrong; the paper shows VCD and ICD produce zero color-FPR improvement while VISOR-Calib cuts color FPR by 11–17 percentage points.
  • Material attribute hallucination requires either abstaining on low-SNR queries or adapting the visual projection per word; VISOR-Abstain lowers Qwen material FPR from 15.7% to 10.2% at a 14.2% abstention rate, and VISOR-Adapt cuts the six-word mean from 38.1% to 23.9%.
  • The diagnosis generalizes beyond the VAW format: the $\delta_{\mathrm{vis}} \ge 0$ criterion yields zero false-positive violations on the GQA color subset across all three models.
  • Late-layer SNR collapse in layers 28–36 implies that material attribute errors are a projection problem, not a missing-visual-evidence problem, so targeted visual adaptation is more appropriate than output-level correction.

Reading between the lines

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

  • A testable extension is to apply the same null-image decomposition to other hallucination types—object existence, spatial relations, counting—to see whether they also follow visual-signal quality rather than language prior.
  • If the bottleneck is late-layer visual-to-logit projection, interventions at intermediate layers, such as early-exit or layer-wise adapters, might recover material accuracy at lower cost than per-word LoRAs.
  • The two-forward-pass VSNR coordinate is cheap enough to serve as a runtime abstention signal in high-stakes deployments, since it requires no training and no extra data.
  • The paper leaves open whether material failures stem from pretraining coverage or from projection, so a natural follow-up is to train on material-rich captions and check whether final-layer SNR rises.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 introduces VISOR, a framework that diagnoses attribute hallucination in vision-language models by comparing real-image and blank-image forward passes. It defines a visual logit difference δvis, a null-image prior logit difference δprior, and a residual interpreted as the visual likelihood contribution. Across 10,791 negative-ground-truth attribute probes on the VAW benchmark with Qwen2.5-VL-3B, InternVL3.5-4B, and LLaVA-1.5-7B, the paper reports that δvis strongly predicts false positives while δprior is near chance. It then proposes a taxonomy of two failure modes, low-margin color/state errors and low-SNR material errors, and evaluates three routed interventions: calibration, abstention, and per-word visual LoRA adaptation.

Significance. If the mechanistic claim were established, the paper would challenge the prevailing language-prior-dominance account of attribute hallucination and reframe mitigation around visual-signal quality. The paper deserves credit for a clean controlled probing setup, layer-wise SNR tracing, and a targeted per-word LoRA adaptation scheme with MME stability checks; the late-layer SNR improvement after adaptation is a concrete, falsifiable prediction. However, the headline correlation is partially definitional, and the null-image decomposition that carries the causal claim is not independently validated. The significance is therefore conditional on substantial revision of the central evidence.

major comments (4)
  1. [Section III, Eq. (1); Table II] The manuscript defines hallucination for a negative sample as δvis ≥ 0 and then reports ρ(δvis, halluc.) in Table II. Under greedy decoding, the hallucination label is exactly sign(δvis), so the Spearman correlation is a rank-biserial correlation between a variable and its own thresholded sign; the text's statement that sign(δvis) matches the label on all 10,791 samples makes this explicit. This correlation provides no causal evidence that the visual signal 'strongly predicts false positives' and is at most a self-consistency check of the decoding rule. Please report a non-tautological predictive analysis, such as the sign of δvis − δprior against the dataset label, and also report the δvis-only accuracy, which the current text implies is 100%.
  2. [Section III, Eq. (4)] Equation (4) is a formal Bayes identity, not a measurement of a visual likelihood. The identification of δprior with the language prior requires that the blank image I0 has zero visual likelihood ratio, and the identification of δvis − δprior with the visual contribution requires that the residual is not contaminated by image-independent biases. The stability checks across gray, black, white, and random-noise images do not validate either identification, and the text reports text-only agreement only for Qwen and InternVL, not LLaVA. Without such validation, the conclusion that the language prior does not drive attribute hallucination does not follow; the authors should compare null-image priors with true text-only priors on all three models and demonstrate that δvis − δprior tracks a known visual signal in a controlled setting.
  3. [Section IV, Table III] The prior-positive conflict analysis is a restatement of the decision rule. The 278 cases with δprior > 0 and δvis < 0 are, by construction, cases where the model says 'no', so the 0.0% hallucination rate in Table III is entailed by the definition of hallucination rather than evidence that the image overrides the prior. The argument would need to show that δprior, measured independently of the model's final logits, predicts behavior differently; as it stands, Table III cannot support the claim that real image content 'actively modifies model behavior' in a causal sense.
  4. [Section IV, Correlational analysis; Table V] The sample-level Spearman correlation for δprior is not a well-specified test of language-prior dominance because the null-image prior contains no image-specific variation and is defined from the same forward pass as δvis; the word-level results in Section IV are more appropriate but mixed (color ρ = +0.420, material ρ = +0.034). In addition, the SNR–FPR correlation of −0.916 in Table V is approximately mechanical under a Gaussian model: if FPR is the mass of the δvis distribution above zero, it is essentially a function of SNR, so this correlation does not by itself establish SNR as a causal driver of hallucination severity.
minor comments (5)
  1. [Throughout] The name 'LLaV A' appears repeatedly in the abstract and in Table II; it should read 'LLaVA' consistently.
  2. [Section IV, Correlational analysis] The main text announces the correlational analysis without citing Table II at that point; add a citation and consider presenting the image-sensitivity analysis table in the main text rather than referring to it implicitly.
  3. [Section III, Eq. (2)] Equation (2) writes δprior(M, w), but the query prompt contains an object token; clarify whether the prior is defined for the object–attribute pair and update the notation accordingly.
  4. [Section III, Eq. (4)] The phrase 'carries no semantic content' should be replaced by an explicit mathematical assumption about the visual likelihood ratio of the null image, since 'semantic content' is not a formal condition.
  5. [Section IV, Word-level correlations] The per-word Spearman correlations in Section IV are reported with p-values but without effect sizes or confidence intervals; include these for the main comparisons.

Circularity Check

3 steps flagged · score 6.0 of 10

The headline correlation is definitional: hallucination is defined as δvis ≥ 0, so δvis must 'predict' it; Table III and the SNR–FPR link are restatements of the same decision rule, though the remediation and word-level results are independent.

  1. self definitional [Section III, 'Formulation' (Eq. 1); Section IV, 'Correlational analysis' and Table II]
    "For a negative sample (ground truth: No), hallucination occurs when δvis ≥ 0. ... Empirically, sign(δvis) matches the hallucination label across all 10,791 negative-ground-truth samples. δvis strongly predicts hallucination across all three models."

    The hallucination label is defined as 1{δvis ≥ 0}, so sign(δvis) matching the label is true by construction. Table II's ρ(δvis, halluc.) = 0.755–0.835 is a rank-biserial correlation between a variable and its own thresholded sign, not evidence about what caused that sign. The abstract's claim that 'the visual signal strongly predicts false positives' therefore restates the decision rule rather than testing a mechanistic hypothesis.

  2. self definitional [Section IV, 'Image-sensitivity analysis (additional evidence)' and Table III]
    "In all 278 cases where blank-image prediction favors 'yes' but real image flips the prediction to 'no', no hallucination occurs, confirming that real image content actively modifies model behavior."

    The subset is selected by δprior > 0 and δvis < 0, while hallucination is defined as δvis ≥ 0. A 0.0% hallucination rate is therefore guaranteed by the definition of the subset. The table's non-tautological content is only that such δprior > 0, δvis < 0 cases exist; it cannot by itself establish that image content 'actively modifies' behavior as an independently measured causal effect.

1 more flagged steps
  1. other [Section V, 'SNR predicts material severity' and Table V]
    "SNR(w) = |¯δvis(w)|/σδvis(w). Across 46 material words, SNR is the strongest predictor of FPR. ... Predictors of material FPR across 46 material words. SNR achieves ρ=−0.916, outperforming either component alone."

    FPR is the fraction of negative samples with δvis ≥ 0, and SNR is a standardized mean of the same δvis distribution. Under approximate normality, FPR ≈ Φ(−δvis/σδvis), so the near-perfect rank correlation is largely a re-expression of the same distributional summary rather than an independent predictor. This does not invalidate the layer-wise SNR analysis, but the 'SNR predicts FPR' framing is partly mechanical.

full rationale

The central correlational claim is circular by construction: Eq. 1 defines hallucination as δvis ≥ 0, so the statement that δvis predicts hallucination, the assertion that sign(δvis) matches all 10,791 labels, and the 0% hallucination rate in Table III's δvis < 0 subset are all consequences of the definition. The paper's causal language ('visual signal strongly predicts false positives', 'visual signal drives hallucination') goes beyond what this definition can support, especially because an equally definitional δvis-only classifier would achieve 100% accuracy on the negative probes while only prior-only accuracy (58.5–59.0%) is reported. The Bayesian decomposition in Eq. 4 is a formal identity, and treating δvis − δprior as the 'visual likelihood contribution' depends on the unvalidated assumption that the gray null image carries no learned visual prior; the paper's own limitation statements acknowledge this scope and note that text-only prior alignment is reported for Qwen and InternVL but not explicitly for LLaVA. On the other hand, several parts of the paper are not circular and contain independent content: the word-level asymmetry (color ρ(δprior, FPR) = +0.420 vs material ρ = +0.034), the late-layer SNR collapse in L28–L36, the cross-model remediation gains of Calib, Abstain, and Adapt, the MME drift within ±1.25 pp, and the explicit negative controls in Table IX. These results would stand even if the headline correlational claim were removed. Because the paper's strongest advertised finding reduces by definition, but the framework also contains substantial non-tautological empirical work, a score of 6 is appropriate rather than 0 or 10.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim rests on two constructs that are not independently measured: the visual likelihood ratio (defined as a residual) and the null-image prior (assumed content-free). The Calib threshold gamma, the Abstain threshold tau, LoRA hyperparameters, and the B1/B2 manual labels are fitted or hand-chosen and directly affect the reported FPR improvements.

free parameters (4)
  • gamma (Calib threshold) = not reported in main text
    The calibrated decision rule 1[delta_vis - gamma * delta_prior > 0] requires offline calibration of gamma per model and attribute; the reported color/state FPR gains are after fitting gamma, so the improvement is a fitted result.
  • tau (Abstain SNR threshold) = 0.3, 0.6, 1.0 explored
    Abstain rate and FPR reduction depend on the chosen SNR threshold tau; the reported operating point tau=0.6 is selected to trade off FPR and coverage.
  • LoRA hyperparameters (rank, learning rate) = unstated (supplementary)
    The Adapt operator trains per-word visual LoRA adapters; hyperparameters are deferred to the supplementary material, which is not included in the preprint.
  • B1/B2 word classification = manual labels (steel B1; rubber, paper, leather B2; ceramic mixed)
    The separation of annotation-granularity noise from true representation failure relies on manual relabeling of 20 steel negatives as visually metallic; the "true FPR" numbers depend on this judgment.
assumptions (4)
  • domain assumption The null image I_empty (uniform gray) carries no semantic content and isolates the language prior.
    Section III and IV define delta_prior from a blank gray image; the causal conclusion that language prior does not drive hallucination depends on this baseline being a clean prior measurement. The paper reports stability across null images (r>0.90) but cannot prove the null image is content-free for the model.
  • ad hoc to paper The Bayesian decomposition in Eq. 4 applies to the model's logits, i.e., log P(yes|I,q)/P(no|I,q) = visual likelihood ratio + language prior.
    This is an application of Bayes' rule to a neural network's outputs, assuming the model's log-odds follow a generative form with an identifiable visual likelihood. The visual likelihood term is defined as a residual, not measured independently.
  • domain assumption VAW negative_attributes annotations are correct ground truth for attribute absence.
    All FPR metrics rely on negative labels from VAW; if annotations are noisy, FPR is miscalibrated. The authors partially address this with the B1/B2 relabeling, which admits that VAW labels contain noise.
  • domain assumption The decision boundary for hallucination is delta_vis = 0 under greedy decoding.
    Section III states that the first answer token fixes the label, so delta_vis = 0 is the boundary; this holds only if the surface-form set for "yes"/"no" captures the actual generated token, which the reported Spearman rho < 1 suggests is not always the case.
invented entities (1)
  • Visual likelihood contribution (log P(I|yes,q)/P(I|no,q))
    purpose: Used as the causal interpretation of delta_vis minus delta_prior in Eq. 4 and as the basis for the claim that visual signal quality, not language prior, drives hallucination.
    This term is defined as the residual of a Bayesian decomposition applied to VLM logits; no independent measurement isolates it from the model's decision rule, so it is a construction rather than an observed quantity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Visual Signals Mislead: A Mechanistic Study of Attribute Hallucination in Vision-Language Models." pith.science (2026). https://pith.science/paper/RE3OQ3DO

@misc{pith2026260811024,
  author       = {Pith},
  title        = {Pith review of: When Visual Signals Mislead: A Mechanistic Study of Attribute Hallucination in Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RE3OQ3DO}},
  note         = {Machine review of arXiv:2608.11024}
}
read the original abstract

Attribute hallucination---where vision-language models (VLMs) correctly identify an object but mischaracterize its properties---is prevalent yet mechanistically poorly understood. The dominant explanation, language-prior dominance, has motivated prior-suppression methods, but this explanation has not been directly tested at the attribute level. We present VISOR (Visual-Operational Remediation), a unified framework that couples null-image-based diagnosis with routed remediation. Its VSNR diagnostic decomposes each prediction into a visual logit signal and a language-prior signal. Across 10,791 negative-ground-truth samples from three VLM families and three attribute types, the visual signal strongly predicts false positives, whereas the language-prior signal is near chance. VISOR uses this diagnosis to separate two failure modes: low-margin but directionally correct visual signals in color/state attributes, and low-SNR or misaligned visual signals in material attributes. The same diagnosis routes each query to the appropriate operator: calibration for threshold-placement errors, abstention for training-free low-SNR handling, or targeted visual adaptation for material failures that prior suppression cannot correct. Across Qwen, InternVL, and LLaVA, VISOR reduces attribute false positives without relying on the prior-dominance assumption.

Figures

Figures reproduced from arXiv: 2608.11024 by the authors.

Figure 1
Figure 1. Framework overview. VISOR first uses VSNR to decompose each attribute query into visual and null-image prior signals. The same diagnostic signal supports mechanism analysis, layer-wise back-tracing, and routing to a mechanism￾specific operator: Calib for low-margin visual signals, and Abstain or Adapt for low-SNR material signals. summarized in recent surveys (Yin et al., 2023; Zhang et al., 2024; Bai et al., 2024; … view at source ↗
Figure 2
Figure 2. SNR vs. FPR and δvis distributions for material words. B2 words have broad δvis distributions straddling zero, while high-SNR words concentrate below zero. case studies are in the supplementary material [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Per-layer SNR for low-SNR material words vs. high [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 6
Figure 6. Figure 6: VISOR-Adapt raises final-layer SNR for five B2 mate [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 5
Figure 5. Figure 5: VISOR-Abstain traces the FPR–coverage trade-off as [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 6 canonical work pages

  1. [3]

    Hallucination of multimodal large language models: A survey.arXiv preprint arXiv:2404.18930,

    Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey.arXiv preprint arXiv:2404.18930,

    arXiv

  2. [6]

    DAMRO: Dive into the attention mechanism of LVLM to reduce object hallucination

    Xuan Gong, Tianshi Ming, Xinpeng Wang, and Zhihua Wei. DAMRO: Dive into the attention mechanism of LVLM to reduce object hallucination. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP),

    work page 2024

  3. [8]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024b. Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiu- tian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large v...

    arXiv 2023

  4. [9]

    Object hallucination in image captioning

    Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. InProceedings of the 2018 Conference on Em- pirical Methods in Natural Language Processing (EMNLP), pages 4534–4545,

    work page 2018

  5. [11]

    V-DPO: Mitigating hallucination in large vision language models via vision-guided direct preference optimization

    Yuxi Xie, Guanzhen Li, Xiao Xu, and Min-Yen Kan. V-DPO: Mitigating hallucination in large vision language models via vision-guided direct preference optimization. InFindings of the Association for Computational Linguistics: EMNLP 2024,

    work page 2024

  6. [12]

    DeCo: Decoupling token compression from semantic abstraction in multimodal large language models.arXiv preprint arXiv:2405.20985,

    Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. DeCo: Decoupling token compression from semantic abstraction in multimodal large language models.arXiv preprint arXiv:2405.20985,

    arXiv

  7. [13]

    A survey on multimodal large language models.arXiv preprint arXiv:2306.13549,

    Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models.arXiv preprint arXiv:2306.13549,

    arXiv

  8. [14]

    Beyond hallucinations: Enhanc- ing LVLMs through hallucination-aware direct preference optimization.arXiv preprint arXiv:2311.16839,

    Zhiyuan Zhao, Bin Wang, Linke Ouyang, Xiaoyi Dong, Jiaqi Wang, and Conghui He. Beyond hallucinations: Enhanc- ing LVLMs through hallucination-aware direct preference optimization.arXiv preprint arXiv:2311.16839,

    arXiv

Show all 15 references
  1. [15]

    MiniGPT-4: Enhancing vision-language understanding with advanced large language models.arXiv preprint arXiv:2304.10592, 2023

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mo- hamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models.arXiv preprint arXiv:2304.10592, 2023

    2023 arXiv

  2. [2018]

    Mitigating hallucinations in large vision-language models with instruction contrastive decoding

    Xintong Wang, Jingheng Pan, Liang Ding, and Chris Biemann. Mitigating hallucinations in large vision-language models with instruction contrastive decoding. InFindings of the As- sociation for Computational Linguistics: ACL 2024,

    2024

  3. [2021]

    MME: A comprehensive evaluation benchmark for multimodal large language models.arXiv preprint arXiv:2306.13394,

    Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Meng- dan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. MME: A comprehensive evaluation benchmark for multimodal large language models.arXiv preprint arXiv:2306.13394,

    arXiv

  4. [2022]

    O’Sullivan, Fang Cao, Tahoura Nedaee, Kamyar Rajabalifardi, Fei-Fei Li, Ehsan Adeli, and Euan Ashley

    Mohammad Asadi, Jack W. O’Sullivan, Fang Cao, Tahoura Nedaee, Kamyar Rajabalifardi, Fei-Fei Li, Ehsan Adeli, and Euan Ashley. Mirage: The illusion of visual understanding. arXiv preprint arXiv:2501.09905,

    arXiv

  5. [2023]

    InternVL2.5: An open-source multimodal LLM for competitive performance with GPT-4o.arXiv preprint arXiv:2412.05271,

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Zhong Ji, Jiangning Song, Muyi Sun, Xiying Tan, Lewei Yao, Haodong Duan, Wenbo Li, Yu Qiao, Jifeng Dai, and Conghui He. InternVL2.5: An open-source multimodal LLM for competitive performance with GPT-4o.arXiv pr...

    arXiv

  6. [2024]

    BLIP- 2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP- 2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InProceedings of the 40th International Conference on Machine Learning (ICML), pages 19730–19742, 2023a. Yifan Li, Yi...

    2023

  7. [2025]

    Qwen2.5-VL technical report.arXiv preprint arXiv:2502.13923,

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhao- hai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, H...

    arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.