acceptodds
Under review as a conference paper at ICLR 2027

SAM: Rethinking Object Hallucination Detection through Visual-Evidence Reliability and Semantic-Update Consistency

Abstract

Large vision-language models remain prone to object hallucination, describing objects unsupported by the input image. Existing attention or semantics-based detectors mainly characterize visual support, leaving the relation between selected evidence and subsequent semantic updates less explored. We revisit object hallucination from the joint perspective of visual-evidence reliability and update consistency. Our layer-wise analysis shows that grounded objects exhibit stronger semantically calibrated visual writes, whereas hallucinated objects exhibit higher evidence-mismatch scores over broad layer ranges. Grounded in these observations, we introduce Semantic-Attention and evidence Mismatch (SAM), a dual-signal framework comprising Semantic-Attention Reliability (SAR) and Evidence Mismatch Score (EMS).SAR recalibrates source-level attention writes with target-semantic support, while EMS compares the FFN-induced visual preference with target-conditioned attention evidence through a state-calibrated optimal-transport score. A lightweight detector combines their layer-wise trajectories without modifying the LVLM parameters or decoding procedure. Experiments across four LVLMs on three benchmarks covering MS COCO captioning and CLEVR and AMBER question answering show consistent gains over existing detectors, achieving the highest AUROC among the compared methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Large Vision-Language Models, Object Hallucination, Hallucination Detection, Model Interpretability

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_hV9SOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents achieves competitive results across multiple LVLMs and benchmarks. However, the methodological novelty is somewhat limited, and the relationship between the proposed geometric measurements and actual semantic grounding is not sufficiently established. Furthermore, inconsistencies in the detection targets across benchmarks make the broader generalization claims difficult to assess.

Critical strengths
  1. The paper providing a potentially useful perspective beyond conventional attention-based hallucination detection.
  2. The paper provides component ablations, layer-wise analyses, and visualizations that help characterize the complementary contributions of SAR and EMS.
Critical weaknesses
  1. The paper does not sufficiently demonstrate why these specific formulations are necessary compared with simpler combinations of attention, semantic similarity, and FFN-based features.
  2. The COCO experiments detect hallucinated object mentions, whereas the CLEVR and AMBER experiments classify whether generated answers match ground-truth answers. In the latter settings, the extracted target is the first answer token, making it unclear whether the method consistently measures object-specific grounding.
Questions for authors
  1. Can the authors compare SAM against simpler alternatives that directly concatenate attention, semantic-support, and FFN-alignment features under the same downstream classifier?
  2. How do the authors justify using first-answer-token correctness as a proxy for hallucination detection on CLEVR and AMBER? Would the results remain consistent when targeting the queried object or relation explicitly?
  3. Can the authors provide significance tests for the relatively small improvements over the strongest baselines?
5 comments
AuthorsOpenReview ↗

Response to Reviewer hV9S (1)

We thank Reviewer hV9S for the careful review and constructive suggestions on feature design, VQA targets, and statistical reliability. We will include these results and clarifications in the revised manuscript.

W1.1 & Q1: On SAR/EMS Design and Simple Feature Concatenation

How does SAM compare with simply concatenating attention, semantic support, and FFN alignment features under the same downstream classifier?

We have added simple concatenation and source-level pairing ablations to examine feature fusion, the role of EMS relative to mean FFN alignment, and the contributions of SAR's individual components.

Setup. We conduct experiments on COCO using LLaVA-1.5-7B and Qwen3-VL-8B. Let I=∑mImI=\sum_m I_m, S=∑mβmS=\sum_m\beta_m, and Fmean=1M∑mcos⁡(u,vm)F_{\mathrm{mean}}=\frac{1}{M}\sum_m\cos(\boldsymbol u,\boldsymbol v_m), where FmeanF_{\mathrm{mean}} measures the mean alignment between the FFN update and visual reference representations. To examine source-level pairing, we replace SAR with I(1+sˉ)I(1+\bar s), where sˉ=1M∑msm\bar s=\frac{1}{M}\sum_m s_m and sm=log⁡(1+Cβm)s_m=\log(1+C\beta_m). We also evaluate a variant retaining only the pairing term B=∑mImsmB=\sum_m I_m s_m to examine the role of the original WRITE magnitude in SAR=I+B\mathrm{SAR}=I+B.(The differences from the original manuscript arise from randomness: the original three seeds varied only MLP initialization with a fixed image split, while the additional experiments vary both.)

Features per layerLLaVA AUROCLLaVA AUPRQwen3 AUROCQwen3 AUPR
[I,S,Fmean][I,S,F_{\mathrm{mean}}]91.4597.1089.1396.52
[I,S,EMS][I,S,\mathrm{EMS}] (SAM-V2)91.7997.2089.3896.55
[I(1+sˉ),EMS][I(1+\bar s),\mathrm{EMS}]90.8996.8888.4396.12
[B,EMS][B,\mathrm{EMS}]91.1396.9687.7795.94
[SAR,EMS][\mathrm{SAR},\mathrm{EMS}] (SAM)91.6197.1288.8296.34

Observations. Simple concatenation performs well: the fully concatenated baseline is competitive with SAM and achieves higher AUROC on Qwen3, while SAM-V2 obtains the highest mean performance in the table on both models.

The controlled comparisons further show that replacing mean FFN alignment with EMS while keeping (I,S)(I,S) fixed improves detection performance on both models. With EMS fixed, full SAR outperforms variants that remove either source-level pairing or the original WRITE magnitude. These trends support target-conditioned evidence–update comparison, source-level pairing, and the retention of the original WRITE magnitude, respectively. The rationale for these modeling choices and our choice of the default implementation are discussed further in W1.2.


Q3: On the Statistical Reliability of Gains over Strong Baselines

Could the authors provide statistical significance tests for the small improvements over the strongest baseline?

Response. We thank the Reviewer for this helpful suggestion. We have conducted a paired bootstrap analysis comparing SAM with the strong baseline SVAR.

Setup. We use three random seeds, each controlling both the image split and MLP initialization. For each bootstrap iteration, we resample images from the test set of each seed, using the same sampled images for both methods. We then compute the AUROC difference for each seed and average the differences across the three seeds. We repeat this process 3,000 times to obtain the 95% confidence interval (CI) of the mean difference.(The differences from the original manuscript arise from randomness: the original three seeds varied only MLP initialization with a fixed image split, while the additional experiments vary both.)

ModelSAM AUROCSVAR AUROCΔ\Delta95% CI
LLaVA-1.5-7B91.6190.49+1.12[+0.28, +1.96]
Qwen3-VL-8B88.8287.55+1.27[+0.32, +1.87]

Observations. SAM improves mean AUROC by 1.12 and 1.27 percentage points on LLaVA and Qwen3, respectively. Both 95% confidence intervals are above zero, supporting the average performance gains over SVAR under this evaluation setting.

AuthorsOpenReview ↗

Response to Reviewer hV9S (2)

W1.2: On the Modeling Choices of SAR and EMS

demonstrate why these specific formulations are necessary

Response. Our core intuition goes beyond jointly using three types of information: it concerns two relationships—whether the visual sources contributing to attention writes support the current target, and how much the visual preferences of subsequent FFN updates deviate from this target-conditioned evidence. SAR and EMS translate these relationships into computable readouts. The theoretical analysis characterizes their properties, while the controlled comparisons in W1.1 & Q1 examine the roles of the individual modeling choices.

SAR: Preserving source-level attention–semantic correspondence. Separately aggregating WRITE magnitude and semantic support captures their overall strengths, but does not directly indicate whether strong writes originate from visual sources that support the target. For example, permuting semantic scores across visual sources leaves total semantic support SS unchanged, yet can alter the correspondence between semantic support and WRITE magnitude. SAR pairs the two before aggregation and therefore captures this distinction. As derived in Appendix A.2, SAR=I+∑mImsm=I(1+sˉ)+MCov⁡V(I,s).\mathrm{SAR}=I+\sum_m I_m s_m=I(1+\bar s)+M\operatorname{Cov}_{\mathcal V}(I,s). Here, I(1+sˉ)I(1+\bar s) uses only total WRITE magnitude and mean semantic support, whereas the covariance term captures whether strong writes and strong semantic support occur at the same visual sources. The I(1+sˉ)I(1+\bar s) ablation therefore examines source-level pairing, while the variant retaining only BB examines the contribution of the original WRITE magnitude. Full SAR performs better in both comparisons, linking the observed benefits to specific modeling components.

EMS: Comparing target-conditioned evidence with update preferences, rather than merely measuring mean visual alignment. FmeanF_{\mathrm{mean}} measures how similar the FFN update is to visual representations on average, without explicitly assessing whether those representations support the current target. For example, two updates can have the same mean visual alignment while favoring target-supporting regions and irrelevant background regions, respectively. Concatenating this mean with separately aggregated I,SI,S allows the MLP to learn statistical relationships among the aggregate features, but the omitted source-level correspondences cannot be determined from these aggregates alone. EMS instead retains the distribution of FFN preferences over visual sources and compares it with target-conditioned visual evidence, directly examining which visual sources the update favors and how these preferences deviate from the evidence supporting the current target. The geometric term further accounts for representation distances between visual sources, while the state term incorporates their compatibility with the current prediction state. The decomposition and geometric lower bound in the original manuscript characterize these properties. EMS improves detection performance on both models, supporting the value of this target-conditioned comparison over mean FFN alignment.

Default implementation: Cross-task performance and interpretability. SAM-V2 retains EMS and changes only the fusion of attention and semantic support. It therefore primarily examines the choice between SAR and the fusion of separately aggregated features. SAM-V2 performs better on COCO, but this advantage does not consistently hold across the evaluated VQA settings. As shown in W2 & Q2, SAM outperforms SAM-V2 in the All Answer, Pred-Yes, and Pred-No settings on Qwen3. On LLaVA, SAM performs better in Pred-Yes and Pred-No, while SAM-V2 performs better in All Answer.

The original visualizations further illustrate the meanings of the two readouts. The heatmaps show target-conditioned visual evidence and its spatial relationship with FFN preference. Figure 3 shows generally higher SAR for real objects and larger EMS for hallucinated objects across a broad range of layers. In contrast, the uncalibrated WRITE magnitude in the appendix exhibits substantial class overlap and local reversals. These observations align with our intuition: examining both whether visual writes receive target-semantic support and how subsequent updates deviate from that evidence.

We therefore retain SAR and EMS as our default: SAR explicitly preserves source-level attention–semantic correspondence, while EMS explicitly compares target-conditioned evidence with update preferences, and the corresponding controlled experiments support these modeling choices. We appreciate the reviewer's suggestion and will report these comparisons in full in the revised manuscript, explaining our default choice through cross-task performance and interpretability. We will also report SAM-V2, which retains EMS, as another effective feature-fusion implementation.

AuthorsOpenReview ↗

Response to Reviewer hV9S (3)

W2 & Q2.1: On Answer Correctness as a Proxy for Hallucination Detection

Why is the correctness of the first answer token used as a proxy for hallucination detection?

Response. We thank the Reviewer for this helpful question.

This choice follows prior work evaluating object hallucinations through existence questions. InsLen applies CLS and CCS to Yes/No tokens for object-existence hallucination detection. TokenGrounding assigns labels based on answer correctness, using the "Yes" token for ADS and the object term in the question for CGC.

For object-existence questions, Yes/No represents the model's judgment about whether the queried object exists, conditioned on the image and the full question. Thus, answer-targeted SAR evaluates the semantically calibrated visual evidence supporting this judgment, while EMS measures the mismatch between the corresponding evidence and FFN preference.


W2 & Q2.2: On Object-Targeted Evaluation

Do the results still hold when explicitly targeting the object or relation in the question?

Response. Following the Reviewer's suggestion, we further use the queried object as the EMS target while keeping SAR answer-targeted, to directly examine the role of object-related visual evidence.

Setup. To further connect the QA evaluation with our proposed object-level interpretation, we conduct additional experiments on CLEVR. We use the object term in the question as the EMS target to measure object-related evidence–update mismatch, while SAR uses the generated Yes/No answer as its target to evaluate the semantically calibrated visual evidence for that answer. For SAM-V2, the two separate features replacing SAR also use the answer token.(The differences from the original manuscript arise from randomness: the original three seeds varied only MLP initialization with a fixed image split, while the additional experiments vary both.)

We evaluate three settings:

  • All answers: All samples, with labels determined by agreement between the generated and ground-truth answers.
  • Pred-Yes: Among samples predicted as Yes, we distinguish correct Yes answers (GT = Yes) from unsupported Yes answers (GT = No).
  • Pred-No: Among samples predicted as No, we distinguish correct No answers (GT = No) from incorrect No answers (GT = Yes).

Each entry in the following tables reports AUROC / AUPR.

LLaVA-1.5-7B

MethodAll answersPred-YesPred-No
SVAR61.42 / 63.8363.11 / 64.3870.82 / 75.41
DHCP56.30 / 58.3561.86 / 62.8663.92 / 67.68
ProjectAway47.04 / 50.8545.57 / 49.4253.50 / 59.47
TokenGrounding54.56 / 55.5958.69 / 60.3564.58 / 70.43
InsLen46.37 / 50.2045.19 / 49.1152.78 / 59.00
SAM65.36 / 67.8865.39 / 67.8171.38 / 75.62
SAM-V266.12 / 68.7364.12 / 67.2470.70 / 73.68

Qwen3-VL-8B

MethodAll answersPred-YesPred-No
SVAR66.49 / 98.1768.03 / 98.4475.64 / 98.44
DHCP55.72 / 96.8964.74 / 98.1661.47 / 97.07
ProjectAway44.28 / 95.9646.39 / 97.2644.26 / 94.80
TokenGrounding69.79 / 98.3162.07 / 98.3982.78 / 98.94
InsLen51.20 / 96.6249.21 / 97.6152.34 / 95.64
SAM91.31 / 99.5990.17 / 99.6192.98 / 99.62
SAM-V290.30 / 99.4185.77 / 99.3892.44 / 99.59

Observations. On LLaVA, SAM improves overall AUROC over SVAR by 3.94 percentage points and achieves better results in both answer subsets. Compared with SAM-V2, SAM improves AUROC by 1.27 and 0.68 points on Pred-Yes and Pred-No, while SAM-V2 performs better overall.

On Qwen3, SAM achieves the highest AUROC in all three settings, outperforming SAM-V2 by 1.01, 4.40, and 0.54 points.

These results directly support combining object-related evidence–update mismatch with answer-targeted visual evidence to detect unsupported Yes answers and incorrect No answers. Together with the source-level pairing ablation and interpretability analysis in W1 & Q1, these results further support our SAR/EMS design and show that the method maintains its detection performance when explicitly targeting the queried object.

Remarks. We again thank the Reviewer for the careful comments, which help us better clarify our method design, QA evaluation, and the statistical evidence for the performance differences.

Reviewer_d4dCOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SAM (Semantic-Attention and evidence Mismatch), a framework for detecting object hallucinations in LVLMs. Motivated by the limitations of attention-based detection and the underexplored relationship between visual evidence and subsequent semantic updates, SAM introduces two complementary signals: Semantic-Attention Reliability (SAR) measures source-level attention-write strength, recalibrated by target-semantic support obtained through a Visual Logit Lens; Evidence Mismatch Score (EMS) measures the discrepancy between target-conditioned attention evidence and FFN-induced visual preferences using state-calibrated optimal transport. The layer-wise SAR and EMS scores are combined through a lightweight MLP without modifying the underlying LVLM. Experiments across four LVLMs and three benchmarks demonstrate consistent improvements over existing detectors. The authors further provide theoretical analyses, ablation studies, and qualitative visualizations.

Critical strengths
  1. An interesting perspective on object hallucination detection. The paper extends conventional attention-based detection by jointly considering visual-evidence reliability and subsequent FFN updates. In particular, SAR incorporates source-level attention writes and target-semantic support rather than relying solely on raw attention weights. The distinction between evidence strength and evidence-update consistency provides a meaningful perspective for studying object hallucinations.
  2. Consistent empirical improvements across multiple models and tasks. SAM achieves the highest AUROC on COCO and CLEVR across all four evaluated LVLMs, demonstrating its effectiveness across different architectures and tasks. The ablation studies further support the complementarity of SAR and EMS, while the layer-wise analyses highlight the benefits of integrating signals across decoder layers.
Critical weaknesses
  1. Insufficient evidence supporting the proposed interpretation of FFN-update consistency. A central contribution of this paper is the interpretation of EMS as measuring the consistency between visual evidence and subsequent FFN semantic updates. However, the current experiments do not sufficiently establish that its discriminative power specifically originates from FFN-induced semantic changes. The FFN visual-preference distribution is constructed from cosine similarities between the FFN update and visual-token representations. It remains unclear whether this captures distinctive properties of FFN updates or merely general hidden-state alignment.
  2. Limited practical gains relative to computational overhead, with insufficient evidence of generalizability. On COCO, the AUROC improvements over SVAR are below one percentage point for three of the four LVLMs, while feature extraction is substantially slower (e.g., 0.501 vs. 0.116 seconds per image on LLaVA-1.5-7B). Moreover, the evaluation relies on separately trained detectors for each LVLM, using in-distribution train/test splits. The robustness of the proposed signals to distribution shifts remains unclear.
Questions for authors
  1. How to verify that EMS specifically captures FFN-update consistency rather than generic representation alignment? Could the authors compare FFN-derived visual preferences with those constructed from pre-FFN residual states, post-FFN residual states, or attention outputs, while keeping the remaining computation unchanged?
  2. How does SAM generalize to unseen distributions, and are its performance gains worth the additional computational cost? Could the authors evaluate SAM under distribution shifts and provide a clearer analysis of the trade-off between detection performance and computational efficiency?
  3. Can the proposed SAR and EMS signals be used to mitigate object hallucinations? Could the authors explore whether these signals can guide generation-time interventions to reduce hallucinations, rather than serving solely as detection features?
5 comments
AuthorsOpenReview ↗

Response to Reviewer d4dC (1)

We sincerely thank Reviewer d4dC for the positive evaluation and the insightful suggestions on further scientific investigation, which we find highly valuable for improving our work. We address the comments below and will incorporate the corresponding revisions into a revised manuscript, which we will upload as soon as it is complete.

W1 & Q1.1: On the FFN-Update Interpretation

How can we verify that EMS specifically captures FFN-update consistency rather than generic representation alignment?

We further clarify the FFN-update interpretation of EMS based on the role of FFN outputs in residual updates and the readout relation in the original manuscript. The FFN output u\boldsymbol u represents the directional update added to the residual stream. For a fixed visual reference v^m\widehat{\boldsymbol v}_m, let gm(z)=⟨z,v^m⟩g_m(\boldsymbol z)=\langle\boldsymbol z,\widehat{\boldsymbol v}_m\rangle. The linear readout change introduced by the FFN update is

Δgm=gm(r+u)−gm(r)=⟨u,v^m⟩\Delta g_m=g_m(\boldsymbol r+\boldsymbol u)-g_m(\boldsymbol r)=\langle\boldsymbol u,\widehat{\boldsymbol v}_m\rangle.

For a nonzero FFN update u\boldsymbol u, we have

cos⁡(u,vm)=Δgm∥u∥2\cos(\boldsymbol u,\boldsymbol v_m)=\frac{\Delta g_m}{\|\boldsymbol u\|_2}.

Thus, the cosine score measures the normalized linear-readout increment introduced by the actual FFN update along a fixed visual reference, and its softmax defines the relative preference distribution PP over visual sources. EMS compares this preference with target-conditioned visual evidence TT to measure evidence–update mismatch.

Qualitative support. The heatmaps in Figure 1 and Appendix D further illustrate the spatial meaning of this readout: locations with higher weights in PP correspond to visual references with larger normalized FFN-induced readout increments. In the illustrated grounded-object cases, TT concentrates on the corresponding visible regions, while PP generally preserves similar spatial support. In hallucinated cases, spatial shifts of PP relative to TT can be observed. These visualizations make the FFN update's directional preferences and their relationship with target-conditioned evidence explicit, providing qualitative support for the readout interpretation.


W1 & Q1.2: On Controlled Comparisons of Visual Preference Sources

Could the authors compare visual preferences constructed from FFN updates, pre-FFN residual states, post-FFN residual states, and attention outputs, while keeping the remaining computation unchanged?

Response. We thank the Reviewer for this constructive suggestion.

We have conducted controlled experiments to further verify the role of FFN updates in constructing visual preferences.

Setup. On LLaVA-1.5-7B and Qwen3-VL-8B, we replace the FFN output u\boldsymbol u used to construct the visual preference distribution PP with the pre-FFN residual r\boldsymbol r, post-FFN residual r+u\boldsymbol r+\boldsymbol u, and attention output oA\boldsymbol o_A. All other computations and experimental settings remain unchanged. Each entry reports AUROC / AUPR. ( The differences from the original manuscript arise from randomness: the original three seeds varied only MLP initialization with a fixed image split, while the additional experiments vary both. )

Preference sourceLLaVA-1.5-7BQwen3-VL-8B
FFN output u\boldsymbol u (SAM)91.61 / 97.1288.82 / 96.34
Pre-FFN residual r\boldsymbol r91.31 / 96.9588.16 / 95.97
Post-FFN residual r+u\boldsymbol r+\boldsymbol u91.16 / 96.9388.18 / 95.99
Attention output oA\boldsymbol o_A90.97 / 96.8788.21 / 96.28

Observations. The FFN-output version achieves the highest AUROC and Real AUPR on both models. Using the pre-FFN residual, post-FFN residual, and attention output reduces AUROC by 0.30, 0.45, and 0.64 percentage points on LLaVA, and by 0.66, 0.64, and 0.61 points on Qwen3. Real AUPR follows the same trend. These results support the value of explicitly extracting FFN updates for visual preference construction. The FFN output also achieves better detection performance than the attention output, which is another sublayer update. Together, these comparisons support our choice of FFN updates as the default source for visual preferences.

Remarks. The readout relation in W1 & Q1.1 clarifies the update-level meaning of FFN preference, while the controlled experiments further support its detection value. Together, they support our design of measuring the mismatch between target-conditioned visual evidence and subsequent FFN updates.

AuthorsOpenReview ↗

Response to Reviewer d4dC (2)

W2 & Q2.1: On Generalization under Distribution Shifts

How does SAM generalize to unseen distributions? Could the authors evaluate SAM under distribution shifts?

Setup. To mitigate the impact of limited source-domain sample sizes on detector training and reduce the potential confounding effect of insufficient training on cross-dataset generalization evaluation, we combine CLEVR and AMBER into a shared source training set. We compare the methods' transfer performance using the same source-domain training data. The trained detectors are directly applied to POPE, with results obtained by training and testing within POPE serving as in-domain references. All results are reported in the order AUROC / AUPR.

POPE train/test splitModelMethodIn-domain AUROC / AUPR ↑Transfer AUROC / AUPR ↑Generalization gap ↓
5:5LLaVASVAR68.50 / 96.3936.15 / 90.2332.35 / 6.16
SAM80.49 / 97.8565.31 / 95.9615.18 / 1.89
Qwen3SVAR62.98 / 95.3741.88 / 91.8121.10 / 3.56
SAM86.48 / 98.7770.59 / 96.8015.89 / 1.97
8:2LLaVASVAR64.08 / 94.3838.70 / 88.7625.38 / 5.62
SAM86.73 / 98.5565.17 / 95.4321.56 / 3.12
Qwen3SVAR66.67 / 95.1441.85 / 89.8424.82 / 5.30
SAM90.67 / 99.0067.83 / 95.4722.84 / 3.53

Observations. Across both backbones and both splits, SAM consistently achieves better transfer performance than SVAR, with AUROC gains of approximately 26–29 percentage points. More importantly, in three of the four settings, SAM trained without POPE data outperforms POPE-trained SVAR in both AUROC and AUPR. SAM also exhibits smaller gaps from its own in-domain performance, while its advantage over SVAR becomes larger under transfer. These results show that SAM retains its performance advantage over SVAR without using POPE data for detector training, achieving higher transfer performance and smaller in-domain-to-transfer performance drops under both target-domain splits.

Changing the target-domain split further examines robustness to variations in the size and composition of the evaluation samples. This shows that SAM not only achieves better transfer performance within each setting but also retains its advantage as the target evaluation samples change.

Remarks. These results support the cross-dataset effectiveness of SAM's detection signals: their discriminative power extends beyond the datasets used for detector training. SAM achieves stronger direct-transfer performance and consistently maintains its advantage across both backbones and both target splits, demonstrating its practical value when target-domain detector retraining is unavailable.

AuthorsOpenReview ↗

Response to Reviewer d4dC (3)

W2 & Q2.2: On the Performance–Efficiency Trade-off

Are SAM's performance gains worth the additional computational cost? Could the authors provide a clearer analysis of the trade-off between detection performance and computational efficiency?

Response. We thank the Reviewer for raising this important question. We have further optimized our implementation and provide additional analysis of the computational cost and detection gains.

Setup. Our analysis identifies the vocabulary projection of visual representations as the main computational bottleneck. We therefore batch the projection of visual tokens and remove additional computations used only for visualization. The table below reports feature-extraction time after optimization. GPU memory denotes peak usage.

MethodLLaVA GPU (GiB)LLaVA Time (ms/img)Qwen3 GPU (GiB)Qwen3 Time (ms/img)
SAM14.15307.8217.69429.04
SVAR14.52118.2317.44358.60
TokenGrounding14.891,150.0417.71886.09

Observations. After optimization, SAM takes approximately 1.2–2.6 times the feature extraction time of SVAR. Compared with TokenGrounding, which also uses all layers, SAM reduces extraction time by 73.2% on LLaVA and 51.6% on Qwen3, with comparable peak GPU memory usage. The generation time is 2,217.99 ms per image on LLaVA and 9,496.11 ms on Qwen3. SAM's feature extraction accounts for approximately 13.9% and 4.5% of these generation times.

In terms of detection performance, the additional VQA results reported in our response to Reviewer mZYU (W2 & Q3) show that SAM improves AUROC over SVAR by 3.94 and 24.82 percentage points on LLaVA and Qwen3, respectively, in the All Answer setting. SAM also achieves higher transfer AUROC on POPE, as shown in W2 & Q2.1 of this response.

Remarks. While the AUROC gains on COCO are relatively modest, the stronger VQA results and cross-dataset performance provide further evidence of SAM's practical value. Considering these detection gains and the optimized feature extraction cost, SAM provides a reasonable trade-off between performance and computational efficiency.

Reviewer_mZYUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SAM, a supervised hallucination detector for frozen vision-language models. It combines semantically calibrated attention writes (SAR) with an optimal-transport score comparing visual evidence and FFN-derived visual preferences (EMS). A small MLP integrates these signals across layers. Evaluation covers COCO captioning and CLEVR/AMBER question answering, with comparisons across four models and additional component, layer, and efficiency analyses.

Critical strengths

The COCO improvements are consistent across all four models, and the component ablations show that SAR and EMS provide complementary information. Image-level splits, validation-based tuning, and extensive sensitivity analyses strengthen the empirical evidence. These results support the usefulness of combining visual support with information from the subsequent FFN update.

Critical weaknesses
  1. I am not convinced that the distinctive formulation adds enough beyond simpler combinations of existing signals. Figure 5 already shows strong class separation from visual semantic support alone, but there is no comparison using layer-wise semantic support and write magnitude as separate inputs to the same MLP. In Table 4, replacing optimal transport with JS divergence reduces AUROC by only 0.38 points on LLaVA and 0.18 on Qwen3, without uncertainty estimates for these differences. Meanwhile, full SAM feature extraction takes approximately 2.3–4.3 times as long as SVAR. The additional complexity needs a clearer empirical justification.
  2. The QA evaluation changes what is being detected without adequately establishing the connection to the proposed object-level interpretation. Appendix B.3 labels samples by answer correctness and extracts features for the first answer token, typically “Yes” or “No.” This groups incorrect denials together with unsupported positive claims, while semantic calibration targets an answer token rather than an object word. The results are useful evidence for answer-error prediction, but less direct evidence for the claimed visual-evidence mechanism. A breakdown by answer polarity and error type is needed to assess this distinction.
Questions for authors
  1. Clarification: Do the three random seeds change the image splits or only detector training? The zero reported variation for the training-free baselines makes this unclear.
  2. Validation: With the same MLP and splits, how does SAM compare with simply combining layer-wise semantic support and write magnitude? Are the small OT-over-JS gains consistent across image splits?
  3. Clarification and validation: In QA, is the semantic target explicitly the generated Yes/No token? How does detection perform separately for unsupported positive answers and incorrect negative answers?
5 comments
AuthorsOpenReview ↗

Response to Reviewer mZYU (1)

We sincerely thank Reviewer mZYU for the careful review and for the comprehensive, insightful comments. These suggestions are highly valuable for improving the clarity of the paper and for enriching its empirical validation. We report the results and clarifications below and will incorporate them into the revised manuscript.

Q1: On random seeds and image splits

“Do the three random seeds change the image splits or only detector training? The zero reported variation for the training-free baselines makes this unclear.”

In the original submission, the image split was fixed, and the three random seeds affected detector training only.

Setup. To further examine the effects of image partitioning and detector training, we repeat the COCO experiments on LLaVA-1.5-7B and Qwen3-VL-8B with three random seeds, each controlling both the image split and detector training. Let I=∑mImI=\sum_m I_m and S=∑mβmS=\sum_m\beta_m. We construct a variant, termed SAM-V2, by replacing SAR with separate layer-wise WRITE magnitude and semantic support features while keeping EMS unchanged. Specifically, SAM-V2 concatenates [I,S,EMS][I,S,\mathrm{EMS}] across layers, whereas the original SAM uses [SAR,EMS][\mathrm{SAR},\mathrm{EMS}].

MethodLLaVA AUROCLLaVA AUPRQwen3 AUROCQwen3 AUPR
MetaToken89.78 ± 0.5796.37 ± 0.2486.27 ± 0.5295.69 ± 0.33
DHCP87.86 ± 0.2595.43 ± 0.1285.94 ± 0.3895.13 ± 0.44
SVAR90.49 ± 0.8396.53 ± 0.0287.55 ± 0.2196.11 ± 0.07
ProjectAway76.99 ± 1.3390.34 ± 0.6072.27 ± 1.2187.68 ± 0.94
TokenGrounding87.81 ± 0.3695.33 ± 0.5076.58 ± 1.4190.94 ± 0.69
InsLen86.86 ± 0.5295.34 ± 0.1870.84 ± 0.8288.88 ± 0.31
SAM91.61 ± 0.5297.12 ± 0.0888.82 ± 0.9696.34 ± 0.52
SAM-V291.79 ± 0.3497.20 ± 0.0289.38 ± 1.1996.55 ± 0.37

Observations. Under this additional protocol, SAM improves mean AUROC over SVAR by 1.12 and 1.27 points on LLaVA and Qwen3, respectively. SAM-V2 achieves further improvements in mean performance. We will include these results and distinguish the original fixed-split experiments from the additional experiments with varying splits.

AuthorsOpenReview ↗

Response to Reviewer mZYU (2)

W1.1 & Q2.1: Simple Concatenation vs. SAR Design

How does SAM compare with directly combining layer-wise semantic support and WRITE magnitude using the same MLP and data splits?

Response. We thank Reviewer mZYU for this constructive suggestion.

Simple concatenation also follows the core intuition of our method and represents a simpler modeling choice within the same framework. Our central contribution lies in proposing and investigating the following intuition: visual-evidence reliability should be jointly assessed through the strength of attention writes and target-semantic support, while further examining the mismatch between this evidence and subsequent FFN updates. This intuition connects visual-evidence formation with subsequent updates, providing two complementary perspectives for analyzing object hallucination. The heatmaps and layer-wise analyses in the original manuscript provide empirical support for this intuition. The heatmaps illustrate how visual evidence supports the generated target and its spatial relationship with FFN preference. The layer-wise analyses show substantial overlap between real and hallucinated samples in uncalibrated WRITE magnitude, whereas SAR exhibits clearer class separation. Meanwhile, hallucinated objects tend to exhibit larger EMS across a broad range of layers. Together, these observations support jointly examining attention writes, target-semantic support, and evidence–update mismatch.

Our formulation translates this intuition into computable readouts, while the theoretical analysis explains their meanings and characterizes the properties of the proposed constructions. SAR explicitly combines WRITE magnitude with target-semantic support at each visual source before aggregation. SAM-V2 instead adopts simple concatenation, treating separately aggregated WRITE magnitude and semantic support as independent features and combining them with EMS through an MLP. Therefore, SAM-V2 still jointly utilizes attention writes and target-semantic support while retaining the assessment of subsequent evidence–update mismatch. It embodies the same core intuition as the original SAM, differing only in how these signals are modeled. The effectiveness of simple concatenation demonstrates that this intuition can also be realized through an alternative formulation, providing another valuable modeling choice within our framework.

Observations. The COCO results are reported in the Q1, while the VQA results are provided in W2 & Q3 below. On COCO, SAM-V2 achieves a higher mean AUROC than SAM. However, this advantage does not consistently hold across VQA settings, where the original SAM achieves a higher mean AUROC than SAM-V2. These results suggest that the relative performance of the two modeling choices depends on the specific task and evaluation setting. Furthermore, SAR provides an explicit source-level interpretation:SAR=I+∑mImsm.\mathrm{SAR}=I+\sum_m I_m s_m. It explicitly preserves the correspondence between WRITE magnitude and target-semantic support at each visual source before aggregation, directly characterizing whether the visual sources contributing to the WRITE support the generated target. For fixed WRITE magnitudes, stronger semantic support from these contributing sources leads to a higher SAR value.

Considering the additional VQA results and this explicit source-level interpretability, we retain the original SAM as our default implementation. Meanwhile, we sincerely appreciate the reviewer's valuable suggestion and will incorporate SAM-V2 into the same framework as a simple yet effective alternative implementation. The two variants can be selected according to task-specific validation performance and interpretability requirements, providing a more comprehensive understanding of the value of different modeling choices under the same core intuition.

AuthorsOpenReview ↗

Response to Reviewer mZYU (3)

W1.2 & Q2.2: Gains of OT over JS

Are the small gains of OT over JS consistent across different image splits?

Response. We thank Reviewer mZYU for this helpful suggestion.

JS also follows the core intuition of our method and offers a simpler, more efficient modeling choice within the same framework. OT measures evidence–update mismatch by considering the geometry of visual representations. Replacing OT with JS retains the comparison between target-conditioned visual evidence and FFN preference, changing how their discrepancy is quantified. OT and JS therefore represent different modeling choices under the same intuition, with JS providing a more efficient realization of the framework.

Setup. We repeat the comparison between SAM (OT) and SAM (JS) using the same three matched random seeds described above. ( In the original submission, the image split was fixed and the seeds affected detector training only; the additional experiments vary both. )

ModelMethodAUROCAUPRms/img
LLaVA-1.5-7BSAM (OT)91.61 ± 0.5297.12 ± 0.08101.69
LLaVA-1.5-7BSAM (JS)91.05 ± 0.5396.92 ± 0.1150.58
Qwen3-VL-8BSAM (OT)88.82 ± 0.9696.34 ± 0.52108.82
Qwen3-VL-8BSAM (JS)88.63 ± 0.5196.30 ± 0.4355.36

Observations. Across the three runs varying both image splits and detector training, OT achieves higher mean AUROC on both models, outperforming JS by 0.56 and 0.19 percentage points on LLaVA and Qwen3, respectively.

Remarks. OT explicitly accounts for distances between visual representations, allowing probability mass assigned to similar visual sources to be matched at a lower cost. This provides a geometric interpretation of evidence–update mismatch. Together with its higher mean detection performance on both models, this motivates our default choice of OT. JS directly compares the two distributions without computing cross-source transport, making it a useful alternative when computational efficiency is prioritized.

We appreciate the reviewer’s suggestion, which helps clarify this performance–efficiency trade-off, and will explicitly incorporate JS into the revised framework as a simple and effective alternative. The same core intuition can be realized through different discrepancy measures, with the choice adapted to task-specific validation performance, computational budgets, and the need for a geometric interpretation. We therefore retain OT as our default while offering JS as an efficiency-oriented alternative.


W1.3: Feature Extraction Cost and Practical Gains

The full SAM feature extraction takes approximately 2.3–4.3 times longer than SVAR, and the additional complexity requires clearer empirical justification.

We have further optimized our implementation and provide a more detailed analysis of SAM's performance and efficiency in terms of feature extraction, generation time, and detection gains.

Setup. Our analysis shows that projecting visual representations into the vocabulary is the main computational bottleneck. We therefore batch the vocabulary projection of visual tokens and remove additional computations used only for visualization. The table below reports the full feature extraction time, GPU memory peak usage.

MethodLLaVA GPU (GiB)LLaVA (ms/img)Qwen3 GPU (GiB)Qwen3 (ms/img)
SAM14.15307.8217.69429.04
SVAR14.52118.2317.44358.60
TokenGrounding14.891,150.0417.71886.09

Observations. After optimization, SAM takes 307.82 ms per image on LLaVA and 429.04 ms on Qwen3 for feature extraction, about 1.2–2.6 times the feature extraction time of SVAR. Compared with TokenGrounding, which also uses all layers, SAM reduces extraction time by approximately 73.2% on LLaVA and 51.6% on Qwen3, with comparable peak GPU memory usage. The generation time is 2,217.99 ms per image for LLaVA and 9,496.11 ms for Qwen3. SAM's feature extraction accounts for approximately 13.9% and 4.5% of the generation time on these two models. Meanwhile, in the VQA experiments reported below, SAM improves the overall AUROC over SVAR by 3.94 percentage points on LLaVA and 24.82 percentage points on Qwen3, and achieves better results on both answer-polarity subsets.

Considering the average performance gains on COCO, the larger improvements on VQA, and the extraction cost relative to generation time, we believe that the optimized SAM provides a reasonable trade-off between detection performance and computational efficiency.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal and Revision Roadmap

We have conducted a comprehensive revision of the manuscript to address the constructive feedback from all three reviewers, with revisions highlighted in blue. Below, we provide a roadmap linking each reviewer's concerns to our detailed responses and the corresponding revisions in the manuscript.


W = Weakness; Q = Question; A = Answer.

Reviewer 1 (mZYU)

  • Q1. (On random seeds and image splits): [A1] Clarified the original fixed-split protocol and added three-seed experiments varying both image splits and detector training, with updated baseline comparisons (Sec. C.6, C.11, C.12).
  • W1.1 & Q2.1. (On simple feature concatenation): [A2] Introduced SAM-V2 and compared separate WRITE magnitude and semantic-support features with SAR under matched settings, including COCO and CLEVR results (Sec. 5.3, Sec. C.6).
  • W1.2 & Q2.2. (On OT versus JS): [A3] Repeated the OT–JS comparison across different image splits and analyzed their detection performance, computational costs, and modeling trade-offs (Sec. C.11).
  • W1.3. (On computational overhead and practical gains): [A4] Optimized feature extraction and reported runtime, peak GPU memory, generation-time overhead, and detection gains relative to strong baselines (Sec. 5.3, Sec. C.12).
  • W2 & Q3.1. (On the semantic target in QA): [A5] Clarified the use of the generated Yes/No token as the original semantic target and introduced an object-targeted EMS evaluation to examine object-related evidence–update mismatch (Sec. 5.1, Sec. B.3, Sec. C.6).
  • W2 & Q3.2. (On answer polarity and error types): [A6] Added CLEVR evaluations for All Answers, Pred-Yes, and Pred-No, separately examining unsupported positive answers and incorrect negative answers (Sec. C.6).

Reviewer 2 (d4dC)

  • W1 & Q1.1. (On the FFN-update interpretation): [A1] Clarified the relationship between FFN outputs and normalized linear-readout increments, with a corresponding proof and qualitative visualizations supporting the update-level interpretation (Sec. 3.3, Sec. 4.2, Sec. C.7).
  • W1 & Q1.2. (On FFN-specific versus generic alignment): [A2] Compared visual preferences constructed from FFN outputs, pre-FFN residuals, post-FFN residuals, and attention outputs under identical settings, showing the strongest mean detection performance with FFN outputs (Sec. 5.3, Sec. C.7).
  • W2 & Q2.1. (On generalization under distribution shifts): [A3] Added cross-dataset transfer experiments from CLEVR+AMBER to POPE, comparing SAM with SVAR under two target-domain splits and reporting in-domain performance, transfer performance, and generalization gaps (Sec. 5.4, Sec. C.10).
  • W2 & Q2.2. (On the performance–efficiency trade-off): [A4] Optimized feature extraction and compared runtime and peak GPU memory with SVAR and TokenGrounding, further relating the overhead to generation time and detection gains on QA and transfer tasks (Sec. 5.3, Sec. C.12).
  • Q3. (On generation-time hallucination mitigation): [A5] Explored SAR/EMS-guided attention redistribution with SADI and SVAR on Qwen3-VL-8B (Sec. E).

Reviewer 3 (hV9S)

  • W1.1 & Q1. (On simple alternatives to SAR and EMS): [A1] Added controlled comparisons of feature concatenation, mean FFN alignment, source-level semantic pairing, and the base WRITE contribution (Sec. 5.3, Sec. C.6).
  • W1.2. (On the justification of modeling choices): [A2] Expanded the analysis of SAR's source-level correspondence and EMS's target-conditioned evidence–update comparison, supported by theoretical properties, component ablations, and cross-task results (Sec. 4.1–4.2, Sec. C.6).
  • W2 & Q2.1. (On answer correctness as a detection target): [A3] Clarified the first-answer-token evaluation protocol, the use of answer correctness as the detection label, and the semantic interpretation of Yes/No predictions in object-existence questions (Sec. 5.1, Sec. B.3)
  • W2 & Q2.2. (On object-targeted evaluation): [A4] Added CLEVR experiments using the queried object as the EMS target while retaining answer-targeted SAR, with separate results for All Answers, Pred-Yes, and Pred-No (Sec. C.6).
  • Q3. (On statistical significance): [A5] Conducted paired image-level bootstrap tests with 3,000 resamples across three image-split and training seeds, reporting 95% confidence intervals for AUROC improvements over SVAR (Sec. C.12).
AuthorsOpenReview ↗

Author Remarks to the Area Chair

We sincerely thank the Area Chair for the time and effort devoted to reviewing and evaluating our work.

To facilitate your assessment, we briefly summarize the core storyline and contributions, the reviews and our responses, and provide a roadmap of revisions.


Core Storyline with Contributions

  • Problem Identification — Introducing a Joint Perspective on Visual Evidence and Subsequent Semantic Updates. Existing methods primarily focus on attention or semantic support, while paying limited attention to whether the visual sources contributing to attention WRITE support the target and whether subsequent FFN updates deviate from this evidence. We observe that real objects generally receive more reliable, semantically calibrated visual evidence, whereas hallucinated objects exhibit larger evidence–update mismatch across a broad range of layers.
  • Theoretical Analysis — Establishing an Interpretable Model of Evidence and Updates. Starting from source-level attention WRITE, we analyze the relationship between WRITE magnitude and target-specific semantic support. We further establish the relationship between FFN updates and linear-readout increments along visual reference directions, providing a theoretical basis for analyzing visual evidence reliability and evidence–update mismatch.
  • Practical Method — Proposing SAM, a Dual-Signal Hallucination Detection Framework. SAR jointly measures source-level WRITE magnitude and target-specific semantic support, while EMS employs optimal transport to quantify the deviation between target-conditioned visual evidence and FFN preferences. The two signals are integrated through a lightweight MLP to detect hallucinations without modifying the underlying model, translating our core research intuition into interpretable and computable detection signals.
  • Experimental Validation. Experiments across four LVLMs and multiple benchmarks demonstrate that SAM consistently outperforms existing detection methods. Component ablations, layer-wise analyses, and visualizations further support the complementarity of the two signals and the proposed core research intuition.

Summary of Reviews and Responses

All three reviewers recognized the value of our work and provided constructive feedback on the method design, mechanistic interpretation, and experimental validation.

The reviewers particularly acknowledged the following contributions:

  • A Meaningful Research Perspective. Connecting the reliability of visual evidence with subsequent FFN updates, extending beyond conventional attention-based detection methods.
  • Complementary and Interpretable Signals. SAR and EMS analyze object hallucinations from two complementary perspectives, namely evidence formation and subsequent semantic updates, supported by ablation studies and visualizations.
  • Effectiveness Across Models. Consistent performance improvements across different LVLMs and tasks, accompanied by comprehensive component-level and layer-wise analyses.

In our rebuttal, we addressed the main concerns through the following additional analyses and experiments:

  • Deepening Methodological and Mechanistic Analyses. We introduced comparisons involving simple feature concatenation, FFN alignment, and OT/JS alternatives to clarify the roles of different modeling choices. Furthermore, we provide a readout-based interpretation of FFN updates and conduct replacement experiments using different hidden states to further validate the rationale behind FFN preference modeling.
  • Strengthening Experimental Reliability and Evaluation Clarity. We conducted additional experiments with three random seeds controlling both image splits and detector training, and used paired bootstrap analysis to verify the statistically significant improvements over SVAR. We also clarified the detection targets in VQA and introduced object-targeted evaluations and Pred-Yes/Pred-No subgroup analyses to assess detection performance across different error types.
  • Extending Generalization, Efficiency, and Practical Applicability. We conducted additional cross-dataset transfer experiments, demonstrating stronger transfer performance for SAM compared with SVAR. We also optimized feature extraction and analyzed runtime, and detection gains to better characterize the trade-off between performance and computational efficiency. In addition, we explored generation-time hallucination mitigation guided by SAR/EMS, demonstrating their preliminary potential as intervention signals while acknowledging the limitations of this extension.

We have carefully addressed all comments and incorporated all corresponding revisions into the manuscript. We sincerely thank the reviewers for their constructive suggestions, which have helped strengthen the methodological justification, experimental reliability, and practical relevance of our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.