Pith. sign in

REVIEW 3 major objections 6 minor 44 references

Hallucinations Leave a Grounding Signature:Verifier-Guided Decoding for Selective Object Correction

T0 review · 3 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read This paper claims that when a vision-language model begins to mention an object that is not in the image, its attention reorganizes into a consistent signed pattern — image attention falls and sink attention rises in visual-grounding heads

desk verdict Signed cross-head attention is a genuinely new diagnostic and the causal experiment is careful, but the CHAIR-label circularity and the suspiciously small coverage drop at @rec90 keep this from being an immediate accept. read the letter →

arxiv 2607.27823 v1 pith:YVYJVH4R submitted 2026-07-30 cs.CV

classification cs.CV
keywords objecthallucinationvision-languagemodelsattentionsignaturesvisualgroundingselectivedecodingverifier-guidedKV-cacherollbackconfidentungrounding
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 tries to establish that object hallucination is not an invisible output error but a generation-time event with an internal signature: as a model starts to mention an object that is not in the image, attention to image tokens drops and attention to the sequence-start sink rises in visual-grounding heads, while language-continuation heads compensate by raising attention to the mention itself. Because these two moves have opposite signs, averaged attention cancels them; the paper defines the Intrinsic Grounding Signature (IGS) as the full signed, head-specific attention pattern. A lightweight sparse logistic verifier reads IGS for each emerging object mention and, when risk is high, rolls back the KV cache to the mention start, suppresses the object and synonyms, and regenerates just that span. The stated result is that IGS reaches 0.927–0.939 AUROC across six vision-language models, and this selective decoding cuts hallucination by roughly 37–44% while retaining near-full grounded-object coverage and caption length. If right, this turns object hallucination from a post-hoc output error into a locatable, correctable event without broadly suppressing generation.

What carries the argument

The central object is the Intrinsic Grounding Signature (IGS): for an object-token position, the vector of image-attention mass and sink-attention values across all layers and heads, plus a three-dimensional confidence vector. The key identity behind its signed structure is Δ log α_{o,s} = Δ z_s − Δ Z_o: whether sink attention rises or falls when evidence is removed depends on whether the BOS score stays roughly constant (denominator-driven sink-up) or local-language routing strengthens mention and history keys while weakening BOS (query-driven sink-down). IGS preserves head-specific directions so complementary responses add instead of cancel. A sparse ℓ1-regularized logistic readout maps IG

What would settle it

On a held-out set of object mentions, replace the reference-caption oracle with human annotations of whether each mentioned object is actually present and recompute the verifier's AUROC and the decoding improvement; if the IGS advantage over output confidence shrinks or vanishes, the grounding signature was partly an artifact of the label source.

Watch

Extended reading notes

Core claim

The central claim is that a frozen vision-language model, while confidently generating an object absent from the image, leaves a distributed signed attention pattern — IGS — that is more informative than output confidence. The paper demonstrates 'confident visual ungrounding': standard confidence cues reach only about 0.70–0.73 AUROC, while signed IGS reaches about 0.935. The mechanism: visual-grounding heads show sink-up, denominator-driven attention fallback to the BOS sink when evidence is removed, while linguistic-continuation heads show sink-down, query-driven compensation to the forming mention and recent context. This handoff lets the model remain fluent and confident after visual gro

Load-bearing premise

The entire pipeline is fitted and evaluated against reference-caption-derived labels of which objects are present; if those labels misclassify what is actually in the image, both the IGS verifier and the reported hallucination reductions inherit the bias.

Editorial extensions

If this is right

  • Object hallucination can be detected during generation, mention by mention, using attention the frozen model already computes — no auxiliary model or retraining is needed.
  • Selective correction preserves what a coarse intervention would destroy: at the high-recall operating point, hallucination drops by roughly 37–44% while grounded-object coverage stays at 99.6% and caption length is essentially unchanged.
  • A single threshold exposes an explicit hallucination–coverage frontier, so the cost of stricter suppression is visible to the user rather than hidden.
  • The grounding diagnosis transfers to a new dataset without refitting the verifier, with verified-negative mentions decreasing and caption quality near baseline.
  • Because most mentions are accepted as-is, over 85% of generations remain on the greedy path and the decoding overhead is about 1.6× greedy.

Reading between the lines

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

  • The same signed fallback/compensation mechanism is not object-specific in principle, so IGS may extend to attribute, relation, or whole-claim hallucinations; that is an extrapolation, not a result in the paper.
  • The identity relating sink-logit changes to denominator changes suggests a direct manipulation test: perturbing the BOS sink key during decoding should shift hallucination rates in the predicted direction — the paper observes the signature but does not perform that intervention.
  • Because the verifier is fitted and evaluated on reference-caption-derived labels, an independent human-annotated oracle would give a cleaner estimate of the true reduction; this is a natural next experiment, not a claim made here.
  • VGD's rollback regenerates greedily; combining the risk signal with sampling or beam search could produce a different frontier, which the paper does not explore.
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 claims that object hallucinations in LVLMs can be detected at generation time via an Intrinsic Grounding Signature (IGS): the distributed, signed pattern of per-head image-attention and BOS/sink-attention masses. A sparse logistic verifier over IGS assigns grounding risk to each emerging object mention, and the proposed Verifier-Guided Decoding (VGD) rolls back the KV cache, suppresses the object and its synonyms, and regenerates only high-risk spans. The authors report 0.927–0.939 AUROC for IGS across six LVLM configurations, and at @rec90 claim a 43.6% reduction in AMBER-G CHAIR, a 37.0%/30.4% reduction in CHAIR-MSCOCO CHAIRi/CHAIRs, near-unchanged caption length and grounded-object coverage, and cross-dataset transfer to NoCaps/OI-VN without verifier refitting.

Significance. If the central claim holds, this is a valuable step: instead of coarse generation-wide suppression, object hallucination becomes a local, generation-time event that can be corrected selectively. The paper's strengths include a cleanly designed causal masking experiment (180 paired target-mask versus sham interventions, cross-fitting, exact identity in Eq. 3), a threshold-controlled intervention frontier, and transfer to an external Open Images-based evaluation. The paper also reports computational cost transparently (1.61× greedy) and avoids shortening captions at the primary operating point. The main risk is that the diagnostic and the headline metric share the same label construct: the verifier is fitted and evaluated on CHAIR labels, which are an imperfect proxy for true visual grounding; this limits the current support for the paper's strongest claim.

major comments (3)
  1. [§3.1, §4.1, §5.3] The verifier is trained and evaluated against CHAIR labels, which are derived from reference captions rather than direct visual evidence. The reported 0.927–0.939 AUROC in §5.3 therefore measures agreement with CHAIR's definition of grounding, not necessarily with visual grounding. CHAIR covers only 80 COCO categories and can be systematically biased for objects that human annotators happen not to list. The NoCaps/OI-VN transfer in §5.4 uses human-verified Open Images labels but only at the caption level (OI-VNs/OI-VNi), not as mention-level grounding labels, so it cannot validate individual verifier decisions. To support the central 'IGS detects visual ungrounding' claim, the authors should validate the verifier on mention-level labels that are independent of CHAIR, or at least quantify the verifier's behavior as a function of CHAIR's known failure modes (e.g., categories, reference-cap
  2. [§5.2, Table 1, Abstract] There is an unexplained discrepancy between the @rec90 threshold and the observed coverage retention. The threshold is set to retain 90% of CHAIR-grounded mentions, so 10% of grounded mentions are flagged and rolled back. If most flagged mentions were truly grounded, AMBER-G Cover should drop by a noticeable fraction, yet for LLaVA it drops only from 51.0 to 50.8 (0.4% relative). The abstract credits VGD with 'retaining 99.6% of grounded-object coverage,' but this number does not follow from the 90% recall guarantee. It implies either that CHAIR-grounded mentions and AMBER-G ground-truth objects are largely disjoint, or that rollback regenerates the same grounded objects. The authors should report, among rolled-back mentions, the fraction that were CHAIR-grounded, the fraction that reappear after regeneration, and the resulting per-object recall for AMBER-G. Without this decomposition, t
  3. [§3.3, Figure 2] The causal anatomy analysis identifies two specific heads (L31H16 and L15H26) as visual-grounding and linguistic-continuation heads, with opposite sink-attention responses. The text says heads are cross-fitted before held-out intervention analysis, but it does not state whether these particular heads were selected a priori or on the basis of the same intervention data. If selection was data-dependent, the paired 95% intervals in Figure 2 do not reflect selection bias. The authors should describe the full head-selection rule and report the distribution of sign consistency across all heads, including how many heads show the sink-up, sink-down, or neither pattern. This is important because the signed-IGS design rests on the claim that opposing head responses coexist and must be aggregated without cancellation.
minor comments (6)
  1. [§4.2, §5.5] The '40-retry cap' is mentioned in the implementation details and in §5.5, but the rollback loop description in §4.2 does not define how retries are counted or what happens when the cap is reached. Please state this explicitly in the method section.
  2. [Table 3] The rows in Table 3 are described as 'Components,' but the final row replaces L2 with L1 regularization rather than adding a feature component. The AUROC jump from 0.892 to 0.935 is therefore attributed to the regularizer, not to a new signal. The table headers and text should clarify this distinction.
  3. [§5.2, Table 1] For the main CHAIR/Cover deltas no confidence intervals are reported. Since decoding is deterministic per image, bootstrap over images would be simple and would help assess whether the reported differences are stable, as done for the OI-VN results in §5.4.
  4. [Supplementary material] Several details are deferred to the supplement: the noun recognizer, synonym sets, Open Images hierarchy processing, head-selection procedure, and bootstrap protocols. These are necessary for reproducibility and should either be included in the main paper or made explicitly available with the submission.
  5. [Eq. (3)] The notation Δ and the meaning of 'target-mask minus sham' are introduced somewhat abruptly. Please define Δ just before Eq. (3) and clarify that j_sink uses the same one-based indexing established in §3.1.
  6. [Abstract] The abstract's 'retaining 99.6% of grounded-object coverage' is confusing because @rec90 implies that 10% of grounded mentions are flagged. Please rephrase to state that this is the observed AMBER-G Cover ratio (50.8/51.0), not a verifier recall guarantee, and address the relationship in the main text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the verifier is fit on disjoint CHAIR calibration data and the central claims are corroborated by external AMBER-G and Open Images transfer; the CHAIR-label overlap is a validity concern, not a circular step.

full rationale

The claimed derivation chain is: raw attention masses g,s and confidence c (Eq. 1) are assembled into IGS features (Eq. 4); a sparse logistic readout (Eqs. 5-6) is fit once per backbone to CHAIR labels from calibration images that are explicitly disjoint from final evaluation; the threshold τ is selected on the same calibration data by grounded-object recall, and VGD uses this fixed verifier during decoding. The 0.927-0.939 AUROC is a held-out classification result against the same CHAIR label construct used for training, but because the calibration and evaluation sets are disjoint and the threshold is not tuned on test labels, this is supervised evaluation, not a prediction forced by construction. The headline CHAIR-MSCOCO reductions are measured on held-out CHAIR data, and the AMBER-G and NoCaps/OI-VN results use a fixed verifier with external labels, so the central claim does not reduce to its training inputs. The paper's dependence on CHAIR as the oracle for mention-level grounding is a measurement-validity limitation (CHAIR is reference-caption based and covers 80 COCO categories), but that is a correctness/robustness concern, not circularity: no equation or fitted parameter is renamed as an independent prediction. No load-bearing self-citation or imported uniqueness theorem is present. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The core diagnostic is a fitted classifier over attention features; the only 'first-principles' component is Eq 3. The main load-bearing assumption is CHAIR-label validity. IGS itself is not an independent physical entity but has falsifiable empirical handles.

free parameters (4)
  • Verifier weights w and bias b (2LH+3 dimensional)
    Sparse logistic readout fitted to CHAIR labels via Eq 6; values not reported. Central to IGS risk scoring.
  • L1 penalty lambda
    Regularization strength in Eq 6; no numeric value given; tuned during calibration.
  • Decision threshold tau = @rec90 primary; @rec70 stricter
    Chosen on calibration data to retain q% of grounded mentions (§4.1, §5.1); controls intervention aggressiveness.
  • Rollback cap = 40
    Manual safeguard (§5.1); hit in 13/4500 @rec90 captions and 68/4500 @rec70, so not load-bearing.
assumptions (5)
  • domain assumption CHAIR hallucination labels are a valid object-level visual-grounding oracle.
    The verifier is trained on 'automatically CHAIR-labeled mentions' (§4.1) and the headline metrics are CHAIR/AMBER-G; CHAIR labels groundedness from reference captions over 80 COCO categories, so label noise would corrupt both the diagnosis and the reductions.
  • domain assumption The BOS token is a stable, image-invariant attention sink whose key and value are unaffected by masking image evidence.
    Used in §3.3 to interpret sink-up vs sink-down responses: 'BOS precedes all image tokens, so masking leaves its key and value fixed.'
  • domain assumption There are stable visual-grounding and linguistic-continuation head classes with opposite signed responses to evidence removal.
    §3.3 attributes the two sign directions to head families (L31H16 vs L15H26) and verifies recurrence in 4/5 cross-fitted splits; the mechanistic narrative assumes these classes generalize.
  • domain assumption The noun recognizer and synonym sets correctly isolate object mentions during decoding.
    §4.2: VGD suppresses B={o}∪Syn(o); if the recognizer or synonym mapping is wrong, rollback and suppression can hit grounded content. Implementation is not shipped.
  • standard math Log-sum-exp identity for attention probabilities (Eq 3).
    Δ log alpha_o,s = Δ z_s − Δ Z_o is exact given the softmax definition; this is the only formal derivation component.
invented entities (1)
  • Intrinsic Grounding Signature (IGS) independent evidence
    purpose: A distributed signed attention pattern (image-mass vs sink-mass shifts across heads) used to score whether an emerging object mention is visually grounded.
    Defined in §3.2; falsifiable through causal masking (Fig 2), cross-architecture AUROC (0.927–0.939), and NoCaps/OI-VN transfer. It is a representational construct, not a physical entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hallucinations Leave a Grounding Signature:Verifier-Guided Decoding for Selective Object Correction." pith.science (2026). https://pith.science/paper/YVYJVH4R

@misc{pith2026260727823,
  author       = {Pith},
  title        = {Pith review of: Hallucinations Leave a Grounding Signature:Verifier-Guided Decoding for Selective Object Correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YVYJVH4R}},
  note         = {Machine review of arXiv:2607.27823}
}
abstract

Large vision-language models (LVLMs) often hallucinate objects that are absent from an image. Despite recent progress, existing mitigation methods still lack reliable object-level grounding diagnostics and therefore tend to apply coarse-grained interventions, which can impair visual understanding, shorten responses, and reduce coverage of genuinely grounded objects. The key challenge is thus to detect, during generation, whether each emerging object mention is supported by reliable visual evidence, so that hallucination can be mitigated selectively. Yet output confidence reflects next-token plausibility rather than visual support, allowing language priors to make absent objects appear certain. We show that the missing diagnostic evidence is encoded in an Intrinsic Grounding Signature (IGS), a distributed signed attention pattern that remains informative for such confident hallucinations. Based on IGS, we propose Verifier-Guided Decoding (VGD), a decoding framework in which a lightweight verifier examines each emerging object mention, rolls back the KV cache when the mention is identified as high risk, suppresses the object and its synonyms, and regenerates the affected continuation. Because VGD intervenes only on object mentions identified as high risk, it reduces object hallucination while preserving the model's original visual understanding and grounded object coverage. Experiments on CHAIR and AMBER-G show that VGD achieves state-of-the-art object hallucination reduction: at @rec90, it cuts AMBER-G CHAIR by 43.6\% while retaining 99.6\% of grounded-object coverage, and reduces CHAIR-MSCOCO CHAIR$_i$/CHAIR$_s$ by 37.0\%/30.4\% without shortening captions.

Figures

Figures reproduced from arXiv: 2607.27823 by the authors.

Figure 1
Figure 1. Intervention scope and retained utility. Prior [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Causal and structural evidence for signed IGS. Ev [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. From IGS to verifier-guided decoding. Top: distributed image/sink attention forms per-object IGS, which an offline [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: IGS diagnosis, operating points, and efficiency. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 2 linked inside Pith

  1. [1]

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , year=

    Object Hallucination in Image Captioning , author=. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , year=. doi:10.18653/v1/D18-1437 , pages=

  2. [2]

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

    Evaluating Object Hallucination in Large Vision-Language Models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year=. doi:10.18653/v1/2023.emnlp-main.20 , pages=

  3. [3]

    2024 , eprint=

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution , author=. 2024 , eprint=

  4. [4]

    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=

  5. [5]

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

  6. [6]

    Computer Vision -- ECCV 2024 , year=

    Paying More Attention to Image: A Training-Free Method for Alleviating Hallucination in LVLMs , author=. Computer Vision -- ECCV 2024 , year=

  7. [7]

    2311.07397 , archivePrefix=

    Wang, Junyang and Wang, Yuhang and Xu, Guohai and Zhang, Jing and Gu, Yukai and Jia, Haitao and Wang, Jiaqi and Xu, Haiyang and Yan, Ming and Zhang, Ji and Sang, Jitao , year=. 2311.07397 , archivePrefix=

  8. [8]

    Science China Information Sciences , year=

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

Show all 44 references
  1. [9]

    International Conference on Learning Representations , pages=

    Analyzing and Mitigating Object Hallucination in Large Vision-Language Models , author=. International Conference on Learning Representations , pages=

  2. [10]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Causally-Grounded Dual-Path Attention Intervention for Object Hallucination Mitigation in LVLMs , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  3. [11]

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

    Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  4. [12]

    International Conference on Learning Representations , pages=

    Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models , author=. International Conference on Learning Representations , pages=

  5. [13]

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

    Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  6. [14]

    Advances in Neural Information Processing Systems , pages=

    Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling , author=. Advances in Neural Information Processing Systems , pages=

  7. [15]

    2026 , eprint=

    SAVAA: Mitigating Hallucinations in LVLMs via Step-wise Adaptive Visual Attention Amplification , author=. 2026 , eprint=

  8. [16]

    2026 , eprint=

    SAGE: Sink-Aware Grounded Decoding for Multimodal Hallucination Mitigation , author=. 2026 , eprint=

  9. [17]

    34th USENIX Security Symposium (USENIX Security 25) , year=

    Mirage in the Eyes: Hallucination Attack on Multi-modal Large Language Models with Only Attention Sink , author=. 34th USENIX Security Symposium (USENIX Security 25) , year=

  10. [18]

    2026 , doi=

    Kogilathota, Sai Akhil and G, Sripadha Vallabha E and Sun, Luzhe and Zhou, Jiawei , booktitle=. 2026 , doi=

  11. [19]

    Proceedings of the 38th International Conference on Machine Learning , pages=

    Learning Transferable Visual Models From Natural Language Supervision , author=. Proceedings of the 38th International Conference on Machine Learning , pages=. 2021 , volume=

  12. [20]

    Advances in Neural Information Processing Systems , doi=

    Flamingo: a Visual Language Model for Few-Shot Learning , author=. Advances in Neural Information Processing Systems , doi=

  13. [21]

    2023 , volume=

    Li, Junnan and Li, Dongxu and Savarese, Silvio and Hoi, Steven , booktitle=. 2023 , volume=

  14. [22]

    Advances in Neural Information Processing Systems , doi=

    Visual Instruction Tuning , author=. Advances in Neural Information Processing Systems , doi=

  15. [23]

    International Conference on Learning Representations , pages=

    Efficient Streaming Language Models with Attention Sinks , author=. International Conference on Learning Representations , pages=

  16. [24]

    International Conference on Learning Representations , pages=

    DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models , author=. International Conference on Learning Representations , pages=

  17. [25]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , 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 , year=. doi:10.18653/v1/2024.emnlp-main.84 , pages=

  18. [26]

    International Conference on Learning Representations , pages=

    Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning , author=. International Conference on Learning Representations , pages=

  19. [27]

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

    RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-grained Correctional Human Feedback , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  20. [28]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Detecting and Preventing Hallucinations in Large Vision Language Models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  21. [29]

    2025 , eprint=

    Qwen2.5-VL Technical Report , author=. 2025 , eprint=

  22. [30]

    International Conference on Learning Representations , year=

    A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks , author=. International Conference on Learning Representations , year=

  23. [31]

    Proceedings of the 34th International Conference on Machine Learning , pages=

    On Calibration of Modern Neural Networks , author=. Proceedings of the 34th International Conference on Machine Learning , pages=. 2017 , volume=

  24. [32]

    International Conference on Learning Representations , year=

    Uncertainty Estimation in Autoregressive Structured Prediction , author=. International Conference on Learning Representations , year=

  25. [33]

    First Conference on Language Modeling , year=

    Massive Activations in Large Language Models , author=. First Conference on Language Modeling , year=

  26. [34]

    , booktitle=

    Clark, Kevin and Khandelwal, Urvashi and Levy, Omer and Manning, Christopher D. , booktitle=. What Does. 2019 , doi=

  27. [35]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , year=

    Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned , author=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , year=. doi:10.18653/v1/P19-1580 , pages=

  28. [36]

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

    nocaps: novel object captioning at scale , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year=

  29. [37]

    International Journal of Computer Vision , year=

    The Open Images Dataset V4 , author=. International Journal of Computer Vision , year=

  30. [38]

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

    Mitigating Object Hallucinations in Large Vision-Language Models with Assembly of Global and Local Attention , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  31. [39]

    Findings of the Association for Computational Linguistics: EMNLP 2025 , year=

    Mitigating Hallucination in Large Vision-Language Models through Aligning Attention Distribution to Information Flow , author=. Findings of the Association for Computational Linguistics: EMNLP 2025 , year=

  32. [40]

    Vision-Language Introspection: Mitigating Overconfident Hallucinations in

    Liu, Shuliang and Yang, Songbo and Fang, Dong and Jia, Sihang and Tang, Yuqi and Su, Lingfeng and Peng, Ruoshui and Yan, Yibo and Zou, Xin and Hu, Xuming , booktitle=. Vision-Language Introspection: Mitigating Overconfident Hallucinations in. 2026 , pages=

  33. [41]

    2025 , volume=

    Park, Seongheon and Li, Sharon , booktitle=. 2025 , volume=

  34. [42]

    and Bhattarai, Manish , booktitle=

    Hoang, Nhat and Vu, Minh and Thai, My T. and Bhattarai, Manish , booktitle=. 2026 , pages=

  35. [43]

    2604.06165 , archivePrefix=

    Zohrabi, Reihaneh and Hasani, Hosein and Gupta, Akshita and Baghshah, Mahdieh Soleymani and Rohrbach, Anna and Rohrbach, Marcus , year=. 2604.06165 , archivePrefix=

  36. [44]

    International Conference on Learning Representations , year=

    See What You Are Told: Visual Attention Sink in Large Multimodal Models , author=. International Conference on Learning Representations , year=

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.