Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

The paper proposes SAVAA (called AdaVBoost in the manuscript body), a training-free method that adaptively scales visual attention per token based on a hallucination-risk score, and reports that it outperforms fixed-strength visual boosting

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 23:28 UTC pith:FQGFWFD4

load-bearing objection A useful, cheap adaptive-boosting method for LVLM hallucinations, but the grounding-risk component is less validated than the headline claims, and the name/code inconsistencies need fixing. the 3 major comments →

arxiv 2602.13600 v2 pith:FQGFWFD4 submitted 2026-02-14 cs.CV

SAVAA: Mitigating Hallucinations in LVLMs via Step-wise Adaptive Visual Attention Amplification

Jiacheng Zhang , Feng Liu , Chao Du , Tianyu Pang This is my paper
classification cs.CV
keywords visual attention amplificationhallucination mitigationlarge vision-language modelsvisual grounding entropytraining-free inferencetoken-level risk estimationCHAIR benchmarkAMBER benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that existing training-free methods for reducing hallucinations in large vision-language models, which amplify attention to visual tokens with a fixed scaling factor, are fundamentally limited: a single strength is too weak at some generation steps, leaving hallucinations unresolved, and too strong at others, introducing new hallucinations. To fix this, the paper proposes SAVAA, which estimates per-token hallucination risk with Visual Grounding Entropy (VGE), a signal that combines predictive entropy with a visual grounding score computed once from prefill logits. That risk score then controls a step-wise modulation factor that scales how much visual attention is amplified and how much input-text attention is suppressed at each step. Across LLaVA-NeXT-7B, Qwen3-VL-8B, and InternVL3.5-8B, the adaptive scheme consistently lowers hallucination metrics on CHAIR, SHR, and AMBER while preserving generation quality, suggesting that 'how much to boost' matters as much as 'where to boost'.

Core claim

The central claim is that visual attention reliance in LVLMs is uneven across tokens during autoregressive generation, so hallucination mitigation should determine per-token boost strength rather than applying a uniform scaling. The paper first shows that a fixed boost factor leaves some hallucinated tokens unfixed (under-boosted) while creating new ones (over-boosted), with both types coexisting in the same response. It then defines a risk score from Visual Grounding Entropy, where VGE_t = alpha * normalized predictive entropy + (1 - alpha) * (1 - G_t), with G_t being the visual grounding score of the predicted token read from a prefill-computed vocabulary-level vector. The boost factor m_t

What carries the argument

Visual Grounding Entropy (VGE), defined as a weighted combination of normalized predictive entropy and the complement of a visual grounding score, serves as the hallucination-risk estimator governing all intervention strength. The grounding score comes from a static vocabulary-level vector G[v] = max over visual tokens of Softmax(visual-token logits)[v], computed during prefill, and is assigned to each generated token after decoding. A risk score r_t = min(VGE_t / gamma, 1) is mapped through m_t = 1 + (m_max_vis - 1) r_t to scale pre-softmax attention scores for visual tokens in a chosen layer range, while a companion suppression factor divides the pre-softmax scores of input text tokens. Th

Load-bearing premise

The whole adaptive scaling chain relies on the prefill-time visual grounding vector G[v] = max over visual tokens of the softmaxed LM-head logits staying a faithful and stable measure of whether a generated token is supported by the image throughout decoding; if that static max is miscalibrated, high-risk tokens will not get the boost they need and low-risk tokens will be over-boosted.

What would settle it

On a fresh LVLM, measure whether tokens with high VGE are actually more likely to be judged hallucinated than tokens with low VGE; if the near-monotone VGE-hallucination correlation shown in Figure 4 does not reproduce, the risk signal loses its load-bearing role. Alternatively, run the method against a fixed boost factor tuned to the best average strength on the same benchmarks; if fixed boosting matches or beats the adaptive scheme, the paper's central trade-off is not decisive.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Fixed-strength visual attention boosting methods have a dual failure mode, and any single scaling factor will leave some hallucinations unresolved while creating others; adaptive per-token scaling is necessary to avoid both under-boosting and over-boosting.
  • Guided by VGE, adaptive scaling consistently reduces hallucination rates on CHAIR, SHR, and AMBER across three different LVLM architectures, without sacrificing F1 or discriminative accuracy on POPE and AMBER discriminative tasks.
  • The grounding vector is computed once during prefill and VGE uses the model's own output logits, so the risk signal adds negligible inference-time overhead, making the method practical for deployment.
  • Suppressing attention to input text tokens acts as a complementary mechanism to visual boosting, converting a substantial fraction of otherwise unfixable hallucinations into correct tokens, as shown in the paper's proof-of-concept analysis.

Where Pith is reading between the lines

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

  • A natural next experiment is to test whether VGE can serve as a general token-level risk filter outside the attention-boosting framework, for example by gating beam search or candidate rejection; the near-monotone VGE-hallucination relationship suggests it would transfer.
  • Because the grounding vector G takes a max over all visual tokens and is computed only once, high-resolution models with many visual tokens may yield optimistic grounding scores; recomputing or normalizing G could keep the risk signal calibrated as resolutions grow.
  • The paper's use of an LLM judge for token-level hallucination labels highlights a methodological issue: rule-based metrics can misclassify verbs like 'building' and benign narrative elaborations, so benchmark results on open-ended descriptions may depend heavily on the choice of judge.
  • A direct stress test would compare the adaptive scheme against a fixed boost factor tuned to the best average strength on the same benchmark; if adaptive still wins, the improvement comes from the modulation itself rather than from choosing a better global strength.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a training-free, inference-time method (SAVAA in the abstract, AdaVBoost in the body) for mitigating hallucinations in LVLMs. It identifies a trade-off in fixed visual-attention-boosting methods: a constant boost factor is too weak for some tokens and too strong for others. The method introduces Visual Grounding Entropy (VGE), a token-level hallucination risk estimator that combines predictive entropy with a static visual grounding score computed once at prefill. VGE is used to modulate the visual attention boost strength at each generation step, with additional suppression of text-input attention. Experiments on LLaVA-NeXT-7B, Qwen3-VL-8B, and InternVL3.5-8B on CHAIR, SHR, and AMBER show consistent improvements over fixed-boosting baselines (PAI, VAF, VGA).

Significance. If the results hold, the paper makes a useful contribution to training-free hallucination mitigation by showing that the appropriate visual attention boost strength varies per token and can be estimated by a lightweight signal computed from the model's own logits. The method is simple, efficient, and code is provided. The main caveat is that the specific contribution of the visual grounding component over entropy-only risk is not rigorously established; the empirical support for VGE's advantage is mixed and partly based on an unvalidated LLM judge. Nevertheless, the adaptive boosting mechanism itself appears to deliver substantial improvements over uniform boosting, which is valuable regardless of the precise risk estimator.

major comments (3)
  1. [§3.1, Eq. (1)] The grounding vector G[v] is computed once at prefill as a max over all visual tokens and is never updated during generation. This makes VGE_t blind to the generated prefix and to which visual region the current query attends. For example, a color word that appears somewhere in the image receives a high grounding score even when applied to the wrong object. The motivating example in Figure 3 covers only an object entirely absent from the image, not the common case of contextually misplaced grounding. Since VGE is the core risk estimator driving the adaptive factor m_t (Eq. 5), the validity of this static score is load-bearing for the method's mechanism. Please provide evidence that G is well-calibrated (e.g., compare risk quantiles against fine-grained human/rule-based annotations) or evaluate a dynamic variant that recomputes grounding using current-step attention.
  2. [§4.4, Table 8; §3.1, Figure 4] The sensitivity analysis over α in Table 8 is the only evidence that the visual-grounding term contributes beyond entropy, since α=1.0 reduces VGE to normalized entropy. The differences on LLaVA-NeXT (28.80 vs 29.80 CHAIR_s) and InternVL3.5 (34.40 vs 34.20) are small and likely within sampling noise; only Qwen3-VL shows a clear gap (46.00 vs 50.20). The paper does not frame this as an ablation, report multiple seeds, or compute significance. Moreover, Figure 4's correlation comparison (r=0.96 vs 0.87) is based on GPT-5-mini judgments with no inter-annotator agreement or error bars. Thus the paper's central claim that visual grounding, rather than adaptive boosting itself, drives the improvements is not currently supported. Please add an explicit ablation: adaptive boosting with entropy-only vs. with VGE on all benchmarks, with seeds and significance tests.
  3. [§4.1, Appendix B.1] All main results in Tables 1–3 are single-run with sampling-based decoding (Algorithm 1, line 18: y_t ∼ p_t), and no variances or confidence intervals are reported. Since some margins are small (e.g., SHR HSR for LLaVA-NeXT: 30.7 vs. 31.1 for VGA), the consistency claim across models/benchmarks needs repeated-seed experiments. Additionally, the hyperparameters are tuned on a 500-image CHAIR subset and the main CHAIR numbers come from a different 500-image subset; while the split is disjoint, the same benchmark is used for selection and reporting. Please clarify the exact split and report variance; ideally, tune all baselines on the same validation set for a fair comparison.
minor comments (5)
  1. [Title/Abstract] The abstract refers to 'SAVAA' and the GitHub URL is SAVVA, while the body title and code link are 'AdaVBoost'. This inconsistency must be fixed before submission.
  2. [§3.2, Algorithm 1] Equation (5) defines m_t in terms of r_t, but Algorithm 1 line 9 uses r_{t-1}. The text implies the boost for step t is based on the previous token's risk; please reconcile the notation.
  3. [§2, References] The related work names 'Visual Amplification Fusion (VAF)' but the cited Yin et al. (2025) reference is 'Clearsight'. Please reconcile the method name and reference.
  4. [§4.1] Typo: 'ending layer for 16' should be 'ending layer to 16'.
  5. [§3.2] The pre-softmax attention scores Z^{(t)}_ℓ are not explicitly defined; please refer back to the attention equation in Section 2 or define them locally.

Circularity Check

0 steps flagged

No significant circularity: the adaptive VGE-guided boosting loop is empirically evaluated against independent benchmarks, not forced by construction.

full rationale

The paper's derivation chain is self-contained and non-circular. The risk estimator VGE (Eq. 3) combines predictive entropy with a static prefill visual-grounding vector G[v]=max_i Softmax(h_i)[v] (Eq. 1). Neither VGE nor the adaptive modulation factor m_t=1+(m_max_vis-1)r_t (Eq. 5) is defined in terms of the hallucination benchmark labels or the final CHAIR/SHR/AMBER metrics. The claim that VGE correlates with hallucinated tokens is tested empirically in Figure 4, not assumed. The boosting intervention (Eq. 6) is a concrete modification of pre-softmax attention scores, and its effect is measured against external benchmarks. Hyperparameters (alpha, gamma, m_max_vis, m_max_txt) are tuned on a separate 500-image validation subset, as explicitly stated: 'Following Zhao et al. (2025), we additionally select another 500 samples for hyperparameter tuning.' This is standard model selection, not a fitted parameter renamed as a prediction. GPT-5-mini is used as a judge for motivation and SHR, but the main CHAIR and AMBER results are rule-based and do not depend on that judge. No load-bearing self-citation appears; prior work is cited for baselines and benchmarks. The paper's central contribution is an empirical improvement, and no equation reduces to its own inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The method contributes an adaptive control policy on top of existing VAA interventions, but its performance depends on five per-model hyperparameters tuned on a CHAIR validation subset, as well as the unvalidated grounding-score proxy G and an LLM-judge-based signal for the motivation analysis.

free parameters (5)
  • α (balance coefficient) = LLaVA-NeXT: 0.5; Qwen3-VL: 0.6; InternVL3.5: 0.8
    Controls the trade-off between normalized predictive entropy and visual grounding in VGE (Eq. 3). Tuned per model on a 500-image CHAIR validation subset.
  • γ (risk scale) = LLaVA-NeXT: 0.5; Qwen3-VL: 0.6; InternVL3.5: 0.7
    Scales VGE into a risk score r_t via Eq. 4; affects how quickly boosting saturates. Tuned per model.
  • m_max_vis (max visual boost factor) = LLaVA-NeXT: 1.1; Qwen3-VL: 1.3; InternVL3.5: 1.3
    Upper bound on visual attention multiplier in Eq. 5; tuned per model.
  • m_max_txt (max text suppression factor) = LLaVA-NeXT: 1.7; Qwen3-VL: 1.3; InternVL3.5: 1.6
    Maximum divisor for text token attention scores in Eq. 7; tuned per model.
  • layer range [L_s, L_e) = LLaVA-NeXT: [0,16); Qwen3-VL: [4,16); InternVL3.5: [4,16)
    Which Transformer layers receive the attention modulation; chosen to match model architecture.
axioms (5)
  • domain assumption Amplifying pre-softmax attention scores of visual tokens during autoregressive decoding reduces hallucinations in LVLMs.
    Inherited from PAI (Liu et al., 2024b) and subsequent VAA works; this is the base intervention AdaVBoost builds on.
  • ad hoc to paper G[v] = max_i Softmax(h_i)[v] computed once at prefill is a valid measure of visual grounding for vocabulary token v.
    Introduced specifically for VGE (Eq. 1); no calibration or robustness analysis is provided, and the max over all visual tokens may be optimistic.
  • domain assumption VGE correlates monotonically with token-level hallucination risk.
    Supported by Figure 4 on 200 AMBER samples with GPT-5-mini as judge; not established on other benchmarks or with other judges.
  • domain assumption Suppressing attention to text input tokens helps resolve hallucinations that visual boosting cannot fix.
    Based on Figure 1c, where 22 of 27 unfixable tokens were corrected by text suppression on a 200-sample AMBER subset.
  • ad hoc to paper Hyperparameters tuned on a 500-image CHAIR validation subset generalize to other benchmarks and models without re-tuning.
    The authors select another 500 CHAIR samples for tuning and fix hyperparameters across SHR, AMBER, POPE and all models; no separate validation sets for those benchmarks are reported.
invented entities (1)
  • Visual Grounding Entropy (VGE) no independent evidence
    purpose: Token-level hallucination risk estimator combining normalized predictive entropy and a static vocabulary-level visual grounding score.
    Defined in Eq. 3 and used to control step-wise visual attention boosting. Its utility is demonstrated only through this paper's internal experiments; no external benchmark or independent validation is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 21783 in / 15448 out tokens · 128313 ms · 2026-08-02T23:28:18.430303+00:00 · methodology

0 comments
read the original abstract

A line of recent training-free methods for mitigating hallucinations in large vision-language models (LVLMs) operates by amplifying attention to visual tokens during autoregressive generation within a single forward pass. We refer to this paradigm as visual attention amplification (VAA). In this paper, we identify a dual failure pattern in existing VAA methods caused by their use of a fixed amplification factor across generation steps: it can be too weak at some steps, leaving hallucinations unresolved, while too strong at others, introducing new hallucinations. Motivated by this finding, we propose Step-wise Adaptive Visual Attention Amplification (SAVAA), a new VAA framework that estimates hallucination risk for each generated token and uses the estimated risk to adaptively amplify visual attention at the next generation step. Specifically, we introduce Visual Grounding Entropy (VGE), a lightweight hallucination-risk estimator that augments predictive entropy with visual grounding, assigning higher risk to tokens that are uncertain, weakly grounded in the image, or both. Guided by VGE, SAVAA uses the estimated risk to calibrate the VAA factor for the next generation step, applying stronger amplification to higher-risk steps and weaker amplification to lower-risk steps. Across LLaVA-NeXT-7B, Qwen3-VL-8B, and InternVL3.5-8B, SAVAA significantly outperforms baseline methods on generative hallucination benchmarks such as CHAIR, SHR and AMBER. Code is available at: https://github.com/JiachengZ01/SAVVA.

Figures

Figures reproduced from arXiv: 2602.13600 by Chao Du, Feng Liu, Jiacheng Zhang, Tianyu Pang.

Figure 1
Figure 1. Figure 1: Proof-of-concept experiments. We randomly sample 200 examples from the AMBER benchmark (Wang et al., 2023) and use GPT-5-mini (OpenAI, 2025) as a judge to determine whether a token is hallucinated. The full prompts and detailed justifications are provided in Appendix A. Following the experimental setting of Liu et al. (2024b), we set the scaling factor to 1.2. All experiments are conducted on LLaVA-NeXT-7B… view at source ↗
Figure 2
Figure 2. Figure 2: Examples of under-boosted and over-boosted halluci￾nated tokens in LLaVA-NeXT, Qwen3-VL and InternVL3.5. The success of AdaVBoost can be summarized to three key factors: (1) The reliance on visual attention is inher￾ently uneven across tokens during autoregressive gener￾ation, implying that effective hallucination mitigation re￾quires token-specific visual interventions. (2) Existing vi￾sual boosting metho… view at source ↗
Figure 3
Figure 3. Figure 3: Left: An example where LLaVA-NeXT generates a hal￾lucinated attribute (i.e., camera) with extremely low token entropy (i.e., extremely high confidence), indicating that using entropy alone sometimes fails to provide reliable risk signals. We use GPT-5-mini as a judge to decide if a token is hallucinated. This observation highlights a structural blind spot for confidence-based metrics (e.g., entropy) and mo… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of entropy and VGE as hallucination risk estimators on LLaVA-NeXT. Each bar represents the number of hallucinated tokens within a signal quantile (Q1 = lowest, Q10 = highest). We use GPT-5-mini as a judge to decide if a token is hallucinated. Notably, VGE achieves a stronger correlation with the number of hallucinated tokens compared to entropy alone, demonstrating that incorporating visual grou… view at source ↗
Figure 5
Figure 5. Figure 5: Inference time comparison of different methods on the CHAIR benchmark using LLaVA-NeXT-7B. is consistent with observations reported in Liu et al. (2024b). Notably, within a moderate range of values, our method consistently achieves strong hallucination reduction while maintaining stable F1 scores, indicating that the method is effective and robust under reasonable boosting strengths. 4.5. Inference Efficie… view at source ↗
Figure 6
Figure 6. Figure 6: An example of GPT-5-mini as a judge vs. rule-based metrics for hallucination detection. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: An example of GPT-5-mini as a judge vs. rule-based metrics for hallucination detection. A.2. Definition of Under-boosted Hallucinated Tokens vs. Over-boosted Hallucinated Tokens In the motivation analysis (see [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Prompt for GPT-5-mini to decide whether a word is hallucinated. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Prompt for GPT-5-mini to decide whether a hallucinated word is over-boosted or derived from a previous hallucinated token. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Attending to Multimodal Generation One Token at a Time

    cs.CV 2026-07 accept novelty 7.0

    MLLMs attend to the relevant modality at the exact generation steps that need it; causal blocking of those peaks breaks answers while timed boosting improves multimodal performance.

  2. Listening makes Vision Clear for VLMs

    cs.CV 2026-06 unverdicted novelty 6.0

    PV-TAM uses prompt-side semantics and a bias filter to improve attention-based and IoU localization metrics for vision-language models over answer-side baselines.

Reference graph

Works this paper leans on

15 extracted references · 10 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Bai, S., Cai, Y ., Chen, R., Chen, K., Chen, X.-H., Cheng, Z., Deng, L., Ding, W., Fang, R., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y ., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L. Y ., Ren, X., yi Ren, X., Son...

    Pith/arXiv arXiv

  2. [8]

    An llm-free multi-dimensional benchmark for mllms hallucination evaluation.ArXiv, abs/2311.07397,

    Wang, J., Wang, Y ., Xu, G., Zhang, J., Gu, Y ., Jia, H., Xu, H., Yan, M., Zhang, J., and Sang, J. An llm-free multi-dimensional benchmark for mllms hallucination evaluation.ArXiv, abs/2311.07397,

    Pith/arXiv arXiv

  3. [9]

    Internvl3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency.ArXiv, abs/2508.18265,

    Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., Wang, Z., Chen, Z., Zhang, H., Yang, G., Wang, H., Wei, Q., Yin, J., Li, W., Cui, E., Chen, G., Ding, Z., Tian, C., Wu, Z., Xie, J., Li, Z., Yang, B., Duan, Y ., Wang, X., Hao, H., Li, S., Zhao, X., Duan, H., Deng, N., Fu, B., He, Y ., Wang, Y ., He, C., Shi, B., He,...

    Pith/arXiv arXiv

  4. [10]

    Looking beyond text: Reducing lan- guage bias in large vision-language models via multi- modal dual-attention and soft-image guidance.ArXiv, abs/2411.14279,

    Zhao, H., Si, S., Chen, L., Zhang, Y ., Sun, M., Zhang, M., and Chang, B. Looking beyond text: Reducing lan- guage bias in large vision-language models via multi- modal dual-attention and soft-image guidance.ArXiv, abs/2411.14279,

    Pith/arXiv arXiv

  5. [11]

    Tell model where to look: Mitigating hallucinations in mllms by vision-guided attention.ArXiv, abs/2511.20032,

    Zhao, J., Zhang, F., Sun, X., Feng, C., and Tan, Z. Tell model where to look: Mitigating hallucinations in mllms by vision-guided attention.ArXiv, abs/2511.20032,

    Pith/arXiv arXiv

  6. [12]

    Describe the image

    Zhao, Y ., Yan, L., Sun, W., Xing, G., Meng, C., Wang, S., Cheng, Z., Ren, Z., and Yin, D. Knowing what llms do not know: A simple yet effective self-detection method. ArXiv, abs/2310.17918, 2023a. Zhao, Z., Wang, B., Ouyang, L., wen Dong, X., Wang, J., and He, C. Beyond hallucinations: Enhancing lvlms through hallucination-aware direct preference optimiz...

    Pith/arXiv arXiv

  7. [13]

    Please describe the image in detail

    and measures hallucination from two complementary perspectives: CHAIRi= |Ohall| |Oall| , and CHAIRs= |Chall| |Call| . Here, Ohall denotes the set of hallucinated object mentions that appear in the generated captions but are not present in the ground-truth annotations, and Oall denotes the set of all object mentions in the generated captions. Chall denotes...

    2025

  8. [14]

    Is there a chair in the image?

    provides a discriminative evaluation protocol for object hallucinations in multimodal large language models. Instead of relying on free-form image captions, POPE reformulates hallucination assessment as a binary visual question answering task, where models are asked to answer simple yes-or-no questions about the existence of specific objects in an image (...

    2025

  9. [15]

    15 AdaVBoost: Mitigating Hallucinations in LVLMs via Token-Level Adaptive Visual Attention Boosting C

    for Qwen3-VL and InternVL3.5. 15 AdaVBoost: Mitigating Hallucinations in LVLMs via Token-Level Adaptive Visual Attention Boosting C. Additional Experiments C.1. Full Experiment Results on AMBER Table 4.Results on AMBER benchmark. The AMBER metric is calculated as(1−CHAIR+F1)/2. MLLM Method CHAIR↓Cover↑Hal↓Cog↓ Acc.↑Prec.↑Rec.↑F1↑AMBER↑ LLaV A-NeXT Vanilla...

    arXiv 1975

  10. [2014]

    Y ., Wang, X

    Liu, C., Xu, Z., Wei, Q., Wu, J., Zou, J. Y ., Wang, X. E., Zhou, Y ., and Liu, S. More thinking, less seeing? as- sessing amplified hallucination in multimodal reasoning models.ArXiv, abs/2505.21523,

    Pith/arXiv arXiv

  11. [2019]

    Self-introspective decoding: Alleviating hallucinations for large vision-language models.ArXiv, abs/2408.02032,

    Huo, F., Xu, W., Zhang, Z., Wang, H., Chen, Z., and Zhao, P. Self-introspective decoding: Alleviating hallucinations for large vision-language models.ArXiv, abs/2408.02032,

    Pith/arXiv arXiv

  12. [2021]

    Ker- nel language entropy: Fine-grained uncertainty quan- tification for llms from semantic similarities.ArXiv, abs/2405.20003,

    Nikitin, A., Kossen, J., Gal, Y ., and Marttinen, P. Ker- nel language entropy: Fine-grained uncertainty quan- tification for llms from semantic similarities.ArXiv, abs/2405.20003,

    Pith/arXiv arXiv

  13. [2023]

    Liu, H., Li, C., Li, Y ., Li, B., Zhang, Y ., Shen, S., and Lee, Y . J. Llava-next: Improved reason- ing, ocr, and world knowledge, 2024a. URL https://llava-vl.github.io/blog/ 2024-01-30-llava-next/. 9 AdaVBoost: Mitigating Hallucinations in LVLMs via Token-Level Adaptive Visual Attention Boosting Liu, S., Zheng, K., and Chen, W. Paying more attention to ...

    2024

  14. [2024]

    F., Yaldiz, D

    Kang, S., Bakman, Y . F., Yaldiz, D. N., Buyukates, B., and Avestimehr, A. S. Uncertainty quantification for hallucination detection in large language models: Foun- dations, methodology, and future directions.ArXiv, abs/2510.12040,

    arXiv

  15. [2025]

    Bai, Z., Wang, P., Xiao, T., He, T., Han, Z., Zhang, Z., and Shou, M. Z. Hallucination of multimodal large language models: A survey.ArXiv, abs/2404.18930,

    Pith/arXiv arXiv