Pith. sign in

REVIEW 5 major objections 7 minor 39 references

UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations

T0 review · 5 major / 7 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 A lightweight probe that reads one mid-network layer's attention and hidden states can localize hallucinated tokens in frozen LVLMs and, when streamed during decoding, resample them away, cutting object hallucinations by up to 55% at…

desk verdict A credible, well-ablated extension to LVLMs; the SOTA numbers are plausible but need error bars and a direct test of the attention premise. read the letter →

arxiv 2608.10835 v1 pith:MEMJNWCR submitted 2026-08-11 cs.CV cs.LG

Dvir Samuel , Guy Bar-Shalom , Fabrizio Frasca , Ethan Fetaya , Yftah Ziser , Gal Chechik , Haggai Maron This is my paper · ORCID
classification cs.CVcs.LG
keywords token-levelhallucinationdetectionlargevision-languagemodelsattentiongraphinternalrepresentationshallucination-awaredecodingself-adaptationcomputationaltracefrozenbackbone
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

UniProbe claims that token-level hallucination detection in large vision-language models is best treated as a structured-data problem: one forward pass of a frozen model yields hidden states and attention weights, which UniProbe arranges into a directed graph over image patches, query tokens, and generated response tokens. Alternating GNN, ViT, and GRU blocks let relational, spatial, and sequential evidence interact, and a linear head scores each response token. On MHALO and HalLoc the detector surpasses trained baselines that fine-tune the backbone, and a streaming version used during decoding rejects and resamples flagged tokens, cutting object hallucinations by up to 55% at 1.06x vanilla latency. The paper further shows that adapting the detector to the host model's own free-form generations closes a train-test distribution shift that otherwise degrades precision.

What carries the argument

The computational-trace graph is the load-bearing object: each response token keeps its top attention edges to image patches, query tokens, and earlier response tokens, with the attention weight as edge weight and a per-modality attention-mass summary c_i (Eq. 3), so the detector reads 'attends to image' and 'attends to query' as distinct evidence. L alternating blocks apply a GNN (Eq. 2) to mix evidence across modalities, a ViT (Eq. 4) over the 2-D image grid, and a BiGRU (Eq. 5) along response order; a linear head (Eq. 6) gives per-token probabilities. The streaming variant swaps the BiGRU for a unidirectional GRU so scores depend only on the prefix, and reject-and-resample decoding bans flagged tokens. The paper identifies a mid-network 'grounding band' (layer 20 of GLM-4.1V, layer 14 of LLaVA-1.5) where detection peaks, and prunes the graph to at most 200 image and 64 query nodes, keeping the probe at about 16M parameters.

What would settle it

Take a backbone and construct a test set where hallucinated tokens are deliberately generated to attend to image patches exactly as strongly as grounded tokens, for instance by conditioning the model on an image that contains the mentioned object but in a different attribute, so the object is present yet the mention is still wrong. If UniProbe's token-level F1 on this set drops to near the flat-probe level, the attention-signature assumption is falsified; if it stays high, the detector is reading a different cue.

Watch

Extended reading notes

Core claim

The central claim is that a lightweight probe over a single mid-network layer's attention and hidden states can localize hallucinated tokens more accurately than full-model fine-tuning or external verifiers, and that the same probe, made causal, can act as a decoding guardrail. Concretely, UniProbe builds a typed directed graph in which response tokens draw edges to the most-attended image patches, query tokens, and earlier response tokens, with attention weights as edge features; three interleaved structure-aware modules update a shared per-token representation, and a linear head outputs per-token hallucination probabilities. The paper reports token-level F1 of 63.2/52.9 (F1M/F1IoU) on MHALO with GLM-4V, surpassing HaloDet's 59.1/49.8; object-hallucination F1 of 63.1 on POPE versus 41.0 for the strongest prior; and, with self-adaptation, a 55% reduction in CHAIRi at 1.06x latency while preserving caption quality and output distribution.

Load-bearing premise

Hallucinated tokens leave a distinguishable trace in the attention a single mid-network layer pays to image patches, query tokens, and earlier response tokens; if a hallucinated token attends to the image as strongly as a grounded one, the graph message passing loses its primary signal and the detector's reported gains would not survive.

Editorial extensions

If this is right

  • Backbone-freezing detection: a single forward pass of the frozen LVLM provides enough signal to beat fine-tuned taggers, so safeguards no longer require retraining or degrading the protected model.
  • Decoding-time prevention: the same detector, streamed causally, turns detection into prevention by rejecting and resampling hallucinated tokens before they enter the final response.
  • Distribution-shift correction: self-adaptation on the host model's own free-form captions, labeled automatically with CHAIR, lifts streaming detection F1 from 32.6 to 63.8 and cuts CHAIRi by 55%.
  • Latency budget: scoring a full response costs 1.15x vanilla latency and the streaming guardrail 1.06x, below all trained baselines, so the accuracy gain does not come at prohibitive compute.
  • Per-backbone readouts: internal representations differ across models, so the paper trains a separate readout per backbone and peaks the read-out layer by validation F1.

Reading between the lines

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

  • If the mid-layer grounding band is a general property of LVLMs, the read-out layer could be chosen automatically by a lightweight probing sweep, and the same graph construction might transfer to other fine-grained faithfulness tasks such as evidence attribution or chain-of-thought verification.
  • The ablation result that flat MLP and Transformer probes trail the graph-based probe by at least 30 F1M suggests that the signal is relational: hallucinated tokens are not identifiable from isolated hidden states but from how they route attention to context.
  • A testable extension: applying the same reject-and-resample guardrail to attribute and relationship hallucinations, not just object mentions, may require richer labels than CHAIR, since CHAIR only captures object presence.
  • The paper's threshold analysis implies an operating curve: tightening tau below 0.70 removes more hallucinations but starts deleting correct content, so deployment needs a content-preservation constraint, not just a hallucination target.
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

5 major / 7 minor

Summary. The paper introduces UniProbe, a lightweight token-level hallucination detector for large vision-language models (LVLMs) that operates on a frozen backbone. From a single forward pass, it reads the hidden states and attention weights at a mid-network layer, builds a directed graph over image patches, query tokens, and response tokens, and processes this graph with alternating GNN, ViT, and GRU modules. A streaming variant with a unidirectional GRU enables hallucination-aware decoding that rejects and resamples flagged tokens. A self-adaptation procedure fine-tunes the detector on the target model's own CHAIR-labeled generations. Experiments on MHALO, HalLoc, POPE, and COCO captions across five backbones report state-of-the-art token-level and object-level hallucination detection, and a 55% reduction in object hallucinations at 1.06× vanilla latency in streaming decoding.

Significance. If the reported results hold, UniProbe is a significant contribution to trustworthy multimodal AI. It shows that a single lightweight probe over a frozen model's internal trace can localize hallucinated tokens with higher accuracy than full-model fine-tuning (HaloDet) and external verifiers (HalLocalizer), and that the same detector can drive decoding-time mitigation that reduces object hallucinations by more than half at a small latency cost. The paper is commendable for its breadth: five backbones, three tasks, extensive architecture ablations (Table 5 and Appendix D), and a human preference study. The method is reference-free, preserves the backbone, and the self-adaptation strategy directly addresses an important deployment shift. The systematic ablation isolating the contributions of the GNN, ViT, and GRU is a particular strength. However, the empirical claims rest entirely on point estimates with no error bars or significance tests, and the core mechanistic premise—that hallucinated tokens leave a distinguishable attention signature at the chosen mid-layer—is never directly verified. These gaps prevent the results from being fully convincing at this stage.

major comments (5)
  1. [Sec. 4, Eq. (2) and graph-construction paragraph] The method's central premise is that hallucinated response tokens have a distinguishable attention signature at the chosen mid-layer, particularly in how they attend to image patches, query tokens, and earlier response tokens. The paper never directly measures this separation. The ablation in Table 5b shows that removing the graph costs 10.1 F1M, but it does not establish that the attention weights themselves, rather than the GNN's inductive bias over hidden-state features, are the source of the gain. Please report a direct diagnostic, for example the distribution of total attention mass to image patches, the entropy of the per-modality attention masses c_i in Eq. (3), or the AUROC of a simple attention-based statistic for hallucinated versus grounded tokens at the read-out layer, across backbones and token positions. Without such evidence, the central mechanism is plausible but unverified, and the layer sweep in Fig. 4 could be selecting a layer for reasons unrelated to attention.
  2. [Tables 1–5 and Appendix D] All main results are presented as point estimates with no standard deviations, confidence intervals, or significance tests, despite the appendix mentioning 'seed noise' in Appendix D. Given that headline margins (for example, 4–6 F1M over HaloDet in Table 1 and 22.1 F1 over Token Grounding in Table 3) support the state-of-the-art claim, and that the read-out layer and decoding threshold are selected on validation data, the paper should report mean ± standard deviation over at least three seeds for the main comparisons and a paired significance test (for example, bootstrap or paired t-test) for the differences against the strongest baseline. This is especially important for the 55% mitigation claim in Table 4, where a single CHAIRi value could be within run-to-run variance.
  3. [Section 5 (Evaluation) and Table 1] The evaluation does not specify whether the trained baselines (HaloDet, HalLocalizer, Whitehead et al., PAS, and the 'FT on self-generation' variants) were retrained by the authors under matched conditions or whether numbers were taken from prior publications. For example, Table 1 reports HaloDet at 55.4/59.1 F1M on Qwen-3-VL and GLM-4V backbones, but HaloDet was originally trained on a different backbone; the fine-tuning data, epochs, learning rate, and any adaptation to the new backbone are not described. Similarly, Table 3's POPE protocol is only referenced to Nguyen et al. (2026) and Jiang et al. (2025), and Table 4's 'FT on self-generation' is a single line without details. Please state for each baseline whether it was re-run, with which hyperparameters, and provide the training protocol or an exact reference to the implementation used.
  4. [Sec. 5, 'Mitigating Distribution Shift' and Table 4] The self-adaptation procedure labels hallucinated object mentions using CHAIR on 500 Objects365 images, and the subsequent evaluation in Table 4 uses CHAIR on COCO captions as the metric. Training on the same heuristic used for evaluation risks overfitting to CHAIR's specific false-positive and false-negative patterns (for example, synonym mismatches or unannotated objects). In addition, the paper does not explain how the object-level CHAIR labels are converted to token-level training targets for the token-level detector. Please specify the token-labeling procedure, and include an analysis of the agreement between CHAIR labels and human judgments, or at least a discussion of how robust the reported reductions are to CHAIR's labeling noise.
  5. [Sec. 6 and Fig. 5] The read-out layer ℓ and decoding threshold τ are selected on validation data, which is standard, but the paper should report the sensitivity of the final headline metrics to these choices. Fig. 4 shows the layer curve for GLM-4.1V and LLaVA-1.5, but not the corresponding detection F1 at layers near the peak on the actual test sets; Fig. 5 shows CHAIRi versus τ but not the detection F1 or content-preservation trade-off around τ=0.70. The reported 55% reduction may depend strongly on the validation-based τ choice; please provide a small sensitivity table (for example, F1 obj, CHAIRi, and correct-object coverage for τ = 0.60, 0.70, and 0.80).
minor comments (7)
  1. [Abstract and Section 1] The text mentions a 'Project Page' but provides no URL or linked identifier; please add the complete URL or a footnote.
  2. [Tables 2 and 3] Please use 'LLaVA-1.5' consistently instead of 'LLaV A-1.5' in the table headings and text.
  3. [Section 5 (Datasets and metrics)] The definitions of F1M and F1IoU are brief; please clarify that F1M is a word-overlap F1 at the token level and specify the exact matching rules (for example, case-insensitive matching, stemming, and how partial overlaps are handled).
  4. [Section 4, Eq. (4)] The ViT module is described as operating 'over the image grid'; please clarify that it operates on the selected image patch nodes arranged by their 2-D coordinates, rather than on the full image grid.
  5. [Appendix C] The claim that 'the backbone verbally denies an object that its own internal state encodes as present' is interesting but not quantitatively supported; consider adding a small analysis of such cases.
  6. [Appendix B] The user study reports preference percentages without confidence intervals or inter-rater agreement statistics; please add 95% confidence intervals and a measure such as Cohen's kappa.
  7. [Appendix A] The paper states that all hyperparameters were selected through ablations on held-out validation data, but several important budgets (N_img, N_qry, k_img, k_qry, k_resp) are only ablated in Appendix D; consider summarizing these results in the main text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: UniProbe's predictions are supervised by external labels, and its main gains are measured on held-out data.

full rationale

UniProbe is a supervised detector: its per-token probabilities (Eq. 6) are trained with class-balanced binary cross-entropy against external span labels from MHALO and HalLoc, and its object-level results are evaluated on POPE and COCO CHAIR labels that are independent of the detector's parameters. The self-adaptation procedure uses ground-truth object annotations from Objects365 with CHAIR matching, then evaluates on COCO captions, so the training labels are not derived from the detector's own outputs. The mitigation results in Table 4 apply the same reject-and-resample guardrail to every detector, isolating detection quality; the 55% CHAIRi reduction is measured on final generated captions against COCO ground truth, not against the detector's flags. The only self-citations (Bar-Shalom et al. 2025/2026; Frasca et al. 2026; and Chuang et al. 2024 for the decoding rule) motivate the architecture and the attention-based pruning heuristic, but the architecture's contribution is validated internally by ablations (Tables 5a/5b), and the pruning budgets are shown to be non-critical (Appendix D). No equation in the paper reduces to its own inputs, and no fitted parameter is renamed as a prediction. The lack of error bars and the fact that the attention-signature separation is not measured directly are correctness and robustness concerns, not circularity.

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

The central claim rests on the empirical sufficiency of a single-layer attention and hidden-state trace, plus standard supervised learning assumptions. The only free parameters are architecture and decoding hyperparameters selected on validation data; none are fitted to the test results. No new physical or conceptual entities are postulated.

free parameters (5)
  • Read-out layer index ℓ = 20 for GLM-4.1V and InstructBLIP-7B, 14 for LLaVA-1.5 and InternVL2
    Selected per backbone by validation F1 during a coarse layer sweep (Sec. 6, Fig. 4). Different backbones peak at different depths.
  • Decoding threshold τ = 0.70
    Selected on a held-out development set to maximize hallucination reduction while preserving correct content (Sec. 6, Fig. 5). Lower thresholds remove more hallucinations but also delete correct mentions.
  • Graph node budgets (Nimg, Nqry) = 200 image nodes, 64 query nodes
    Chosen via ablation (Appendix D). The paper reports F1 is flat for image caps {100,200,400,600}, so the value is not critical.
  • Per-token edge budgets (k_img, k_qry, k_resp) = 16, 12, 8
    Chosen via ablation (Appendix D). F1 varies within seed noise across tested values, so defaults are adequate.
  • Architecture size (L, hidden width h) = L=2 blocks, h=256
    Fixed across backbones, about 16M parameters; selected via validation ablations (Appendix A).
assumptions (5)
  • domain assumption Averaged attention weights from one mid-layer of the frozen LVLM encode the grounding relations between response tokens, query tokens, and image patches.
    The detector constructs its graph from mean-over-heads attention at a single layer (Sec. 4, Eq. 2 and graph construction). If attention does not discriminate grounded from hallucinated tokens, the GNN loses its primary signal. Table 5 shows the graph module is the largest single contributor, but that is evidence, not proof.
  • domain assumption Hidden states from a single mid-network layer retain enough grounding information for token-level classification.
    UniProbe reads layer-l residual-stream activations only (Eq. 1). The layer sweep (Fig. 4) shows a mid-layer peak, but the choice is empirical and backbone-specific. The paper does not provide a mechanistic reason why this layer, rather than the full trajectory, is sufficient.
  • domain assumption CHAIR matching against Objects365 ground-truth boxes is a reliable automatic labeler for self-adaptation.
    Self-adaptation labels the model's own captions by checking whether mentioned objects appear in the ground-truth list (Sec. 4, self-adaptation paragraph). Any object present in the image but absent from the annotations would be mislabeled as hallucinated, injecting noise into adaptation.
  • domain assumption Standard supervised learning assumptions hold: training data (MHALO, HalLoc) and self-adaptation data (Objects365 captions) are representative enough that the detector transfers to deployment.
    The detector is trained offline and then applied to freely sampled generations. Self-adaptation reduces but does not eliminate distribution shift; the paper provides no theoretical guarantee of transfer.
  • standard math Backpropagation and Adam optimization will find a good detector on the training objective.
    The method relies on standard supervised training (Appendix A); no convergence or generalization guarantees are provided. This is a routine background assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations." pith.science (2026). https://pith.science/paper/MEMJNWCR

@misc{pith2026260810835,
  author       = {Pith},
  title        = {Pith review of: UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MEMJNWCR}},
  note         = {Machine review of arXiv:2608.10835}
}
abstract

Large Vision-Language Models (LVLMs) achieve impressive visual reasoning and dialogue capabilities, yet frequently hallucinate content unsupported by the visual input. Effective mitigation requires token-level localization, enabling targeted intervention without discarding the entire response. Existing detectors require expensive full-model fine-tuning, rely on external verifiers that ignore the model's generation process, or reduce internal signals to isolated features and hand-crafted statistics, discarding spatial, sequential, and relational structure. We introduce \textbf{UniProbe}, a lightweight, unified, learnable detector that models a frozen LVLM's heterogeneous computational trace from a single forward pass. UniProbe constructs a directed graph over image patches, query tokens, and generated tokens, with attention weights encoding their relations. It processes this trace with alternating structure-aware modules: a GNN for relational evidence, a ViT for 2-D visual geometry, and a GRU for response order. Interleaving them allows spatial, relational, and sequential evidence to interact throughout the detector. We further develop a streaming variant for hallucination-aware decoding, which detects and resamples hallucinated tokens during generation, and a self-adaptation strategy aligning the detector with the LVLM's own generations. Across diverse LVLM backbones, UniProbe achieves state-of-the-art token-level and object-hallucination detection. During decoding, it reduces object hallucinations by up to 55\% at $1.06\times$ the latency of standard generation.

Figures

Figures reproduced from arXiv: 2608.10835 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. UniProbe architecture. From a single for￾ward pass of the frozen LVLM M, UniProbe reads layer-ℓ hidden states and attention to construct a computational-trace graph over image, query, and re￾sponse tokens. L alternating blocks of GNN, ViT, and GRU, integrate cross-modal, spatial, and sequential evi￾dence, followed by a linear head that predicts a halluci￾nation probability for each response token. Computational-trac… view at source ↗
Figure 3
Figure 3. Qualitative token-level detection and mitigation across different LVLM backbones. Existing SOTA methods exhibit different failure modes: they incorrectly flag visually grounded text, detect only part of the hallucinated span, or miss it entirely. In contrast, UniProbe accurately localizes the hallucinated tokens in both examples and uses its predictions during hallucination-aware decoding to produce the grounded out… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: The threshold controls how much halluci￾nation the guardrail removes. CHAIRi on GLM-4.1V falls as τ tightens; we operate at τ=0.70 (marked), the most aggressive setting that still leaves the correctly mentioned objects intact (lower τ removes more but starts deleting c…
Figure 6
Figure 6. Figure 6: Additional detection and mitigation exam￾ples across backbones and hallucination types. For each example we show the backbone response with the tokens flagged as hallucinated by HaloDet, HalLocal￾izer, and UniProbe (Detection), followed by UniProbe’s corrected generati…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 30 canonical work pages

  1. [1]

    MHALO : Evaluating MLLM s as Fine-grained Hallucination Detectors

    Cai, Yishuo and Gu, Renjie and Li, Jiaxu and Huang, Xuancheng and Chen, Junzhe and Gu, Xiaotao and Huang, Minlie. MHALO : Evaluating MLLM s as Fine-grained Hallucination Detectors. Findings of the Association for Computational Linguistics: ACL 2025. 2025

    work page 2025

  2. [2]

    CVPR , year=

    HalLoc: Token-level Hallucination Localization for Vision-Language Models , author=. CVPR , year=

    work page

  3. [3]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    Evaluating Object Hallucination in Large Vision-Language Models , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    work page 2023

  4. [4]

    Object Hallucination in Image Captioning

    Rohrbach, Anna and Hendricks, Lisa Anne and Burns, Kaylee and Darrell, Trevor and Saenko, Kate. Object Hallucination in Image Captioning. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2018

    work page 2018

  5. [5]

    HALP : Detecting Hallucinations in Vision-Language Models without Generating a Single Token

    Kogilathota, Sai Akhil and G, Sripadha Vallabha E and Sun, Luzhe and Zhou, Jiawei. HALP : Detecting Hallucinations in Vision-Language Models without Generating a Single Token. Proceedings of the 19th Conference of the E uropean Chapter of the A ssociation for C omputational L inguistics (Volume 1: Long Papers). 2026

    work page 2026

  6. [6]

    Wenyi Hong and Wenmeng Yu and Xiaotao Gu and Guo Wang and Guobing Gan and Haomiao Tang and Jiale Cheng and Ji Qi and Junhui Ji and Lihang Pan and Shuaiqi Duan and Weihan Wang and Yan Wang and Yean Cheng and Zehai He and Zhe Su and Zhen Yang and Ziyang Pan and Aohan Zeng and Baoxu Wang and Bin Chen and Boyan Shi and Changyu Pang and Chenhui Zhang and Da Yi...

    work page

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Improved Baselines with Visual Instruction Tuning , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    work page

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

Show all 39 references
  1. [9]

    Dai, Wenliang and Li, Junnan and Li, Dongxu and Tiong, Anthony Meng Huat and Zhao, Junqi and Wang, Weisheng and Li, Boyang and Fung, Pascale and Hoi, Steven , booktitle =

  2. [10]

    The Internal State of an LLM Knows When It`s Lying

    Azaria, Amos and Mitchell, Tom. The Internal State of an LLM Knows When It`s Lying. Findings of the Association for Computational Linguistics: EMNLP 2023. 2023

    2023

  3. [11]

    Advances in Neural Information Processing Systems , year=

    Inference-time intervention: Eliciting truthful answers from a language model , author=. Advances in Neural Information Processing Systems , year=

  4. [12]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

    Lookback Lens: Detecting and Mitigating Contextual Hallucinations in Large Language Models Using Only Attention Maps , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

    2024

  5. [13]

    Gaurang Sriramanan and Siddhant Bharti and Vinu Sankar Sadasivan and Shoumik Saha and Priyatham Kattakinda and Soheil Feizi , booktitle=

  6. [14]

    Beyond Token Probes: Hallucination Detection via Activation Tensors with

    Bar-Shalom, Guy and Frasca, Fabrizio and Galron, Yaniv and Ziser, Yftah and Maron, Haggai , booktitle =. Beyond Token Probes: Hallucination Detection via Activation Tensors with

  7. [15]

    International Conference on Learning Representations (ICLR) , year =

    Neural Message-Passing on Attention Graphs for Hallucination Detection , author =. International Conference on Learning Representations (ICLR) , year =

  8. [16]

    Beyond Next Token Probabilities: Learnable, Fast Detection of Hallucinations and Data Contamination on

    Bar-Shalom, Guy and Frasca, Fabrizio and Lim, Derek and Gelberg, Yoav and Ziser, Yftah and El-Yaniv, Ran and Chechik, Gal and Maron, Haggai , booktitle =. Beyond Next Token Probabilities: Learnable, Fast Detection of Hallucinations and Data Contamination on

  9. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

    Opera: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , year=

  10. [18]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  11. [19]

    Science China Information Sciences , year=

    Woodpecker: Hallucination correction for multimodal large language models , author=. Science China Information Sciences , year=

  12. [20]

    arXiv preprint arXiv:2409.00238 , year =

    Pre-Training Multimodal Hallucination Detectors with Corrupted Grounding Data , author =. arXiv preprint arXiv:2409.00238 , year =

    arXiv

  13. [21]

    and Nguyen, Phi Le and Phan, Vu Minh Hieu , booktitle =

    Nguyen, Tuan Dung and Ho, Minh Khoi and Chen, Qi and Xie, Yutong and Cam-Tu, Nguyen and Nguyen, Minh Khoi and Nguyen, Dang Huy Pham and van den Hengel, Anton and Verjans, Johan W. and Nguyen, Phi Le and Phan, Vu Minh Hieu , booktitle =. Beyond the Global Scores: Fine-Grained T...

  14. [22]

    Fieback, Laura and Spiegelberg, Jakob and Gottschalk, Hanno , journal =

  15. [23]

    Proceedings of the 33rd ACM International Conference on Multimedia , year=

    DHCP: Detecting Hallucinations by Cross-modal Attention Pattern in Large Vision-Language Models , author=. Proceedings of the 33rd ACM International Conference on Multimedia , year=

  16. [24]

    arXiv preprint arXiv:2411.16724 , year=

    Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens , author=. arXiv preprint arXiv:2411.16724 , year=

    arXiv

  17. [25]

    The Thirteenth International Conference on Learning Representations (ICLR) , year=

    Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations , author=. The Thirteenth International Conference on Learning Representations (ICLR) , year=

  18. [26]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    PAS: Prelim Attention Score for Detecting Object Hallucinations in Large Vision-Language Models , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  19. [27]

    Wada, Yuiga and Matsuda, Kazuki and Sugiura, Komei and Neubig, Graham , year =

  20. [28]

    Proceedings of the 34th International Conference on Machine Learning (ICML) , year =

    Neural Message Passing for Quantum Chemistry , author =. Proceedings of the 34th International Conference on Machine Learning (ICML) , year =

  21. [29]

    International Conference on Learning Representations (ICLR) , year =

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author =. International Conference on Learning Representations (ICLR) , year =

  22. [30]

    Learning Phrase Representations using

    Cho, Kyunghyun and van Merri. Learning Phrase Representations using. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    2014

  23. [31]

    Transactions of the Association for Computational Linguistics (TACL) , volume =

    From Image Descriptions to Visual Denotations: New Similarity Metrics for Semantic Inference over Event Descriptions , author =. Transactions of the Association for Computational Linguistics (TACL) , volume =

  24. [32]

    Microsoft

    Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll. Microsoft. European Conference on Computer Vision (ECCV) , year =

  25. [33]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    arXiv

  26. [34]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year=

    Objects365: A Large-Scale, High-Quality Dataset for Object Detection , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year=

  27. [35]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year=

    TruthPrInt: Mitigating Large Vision-Language Models Object Hallucination Via Latent Truthful-Guided Pre-Intervention , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year=

  28. [36]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-VL Technical Report , author=. arXiv preprint arXiv:2511.21631 , year=

    arXiv

  29. [37]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Where's Waldo: Diffusion Features for Personalized Segmentation and Retrieval , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  30. [38]

    International Conference on Learning Representations (ICLR) , year=

    EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition , author=. International Conference on Learning Representations (ICLR) , year=

  31. [39]

    arXiv preprint arXiv:2602.08099 , year=

    VidVec: Unlocking Video MLLM Embeddings for Video-Text Retrieval , author=. arXiv preprint arXiv:2602.08099 , year=

Pith tools

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