- Rating
- 2
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This work analyzes the limitations in existing attention-based dehallucination methods, highlighting potential confounding effects from generation position, semantic associations, and attention-induced response preferences. It proposes RASE, which employs MLPs to extract and integrate cross-head attention relations with native semantic evidence to detect object hallucinations and guide rollback and reselection.
Critical strengths
- I think the analysis of potentially misleading attention signals relatively novel and informative. Examining generation-position bias, object association priors, and directional response changes provides useful perspectives for future works on attention-based hallucination detection and mitigation.
- RASE demonstrates a substantial advantage over the evaluated baselines in object hallucination detection.
Critical weaknesses
- Several essential implementation details are insufficiently specified: how to identify object words during generation; how multi-word object and multi-token word are handled; labeling of training data; actual thresholds; how the required attention distributions are obtained when using FlashAttention, whether a different attention backend is required, or the resulting computational overhead. These omissions make the method difficult to reproduce and its practical applicability difficult to assess.
- I thought the analyses in Section 3 have limited support for generalization. The Object Association Prior analysis primarily relies on qualitative examples, without statistical analysis. For Directional Effects of Attention Intervention, the sample size and aggregation procedure underlying Figure 3b are insufficiently described, while the transition analysis is only reported for ClearSight on the POPE adversarial split.
Questions for authors
- The paper suggests that object hallucinations exhibit identifiable internal feature patterns. Can the trained MLP detectors generalize across backbones without retraining? Since the current experiments train a separate detector for each backbone, they do not directly establish detector transferability.
- Can rollback and reselection reduce text fluency or introduce hallucinations involving attributes, relations, or other non-object content? The reported object-level hallucination and coverage metrics do not directly assess these effects.
The authors’ responses to these questions and clarification of the missing implementation details could change my assessment.
4 comments
Response to Reviewer MmJ7 2/4
W2-B Details and Generalization of Attention Intervention
We thank the reviewer for pointing out the insufficient experimental details and limited generalization evidence. We provide further clarification and additional evaluations below.
Experimental Clarification. Figure 3(b) was intended as a qualitative illustration of the directional effects of attention scaling, rather than an aggregated statistical evaluation. To strengthen the empirical support, we extend this analysis across multiple LVLM backbones and generation settings.
Cross-Backbone Validation. We extend the analysis to three LVLM backbones, covering different versions of LLaVA and the Qwen3-VL architecture. The results consistently show that HSVR scaling increases affirmative response probabilities even when the queried object is absent, while LSVR scaling produces the opposite effect.
| Backbone | Intervention | Probability | 1.0 | 1.25 | 1.5 | 1.75 | 2.0 | 2.25 |
|---|---|---|---|---|---|---|---|---|
| LLaVA-1.5 | HSVR | P(Yes) | 0.714 | 0.722 | 0.725 | 0.731 | 0.736 | 0.747 |
| P(No) | 0.027 | 0.027 | 0.027 | 0.028 | 0.028 | 0.028 | ||
| LSVR | P(Yes) | 0.714 | 0.688 | 0.666 | 0.647 | 0.633 | 0.605 | |
| P(No) | 0.027 | 0.029 | 0.030 | 0.032 | 0.033 | 0.035 | ||
| LLaVA-1.6 | HSVR | P(Yes) | 0.170 | 0.171 | 0.173 | 0.173 | 0.174 | 0.177 |
| P(No) | 0.630 | 0.607 | 0.594 | 0.581 | 0.571 | 0.564 | ||
| LSVR | P(Yes) | 0.170 | 0.169 | 0.167 | 0.166 | 0.166 | 0.165 | |
| P(No) | 0.630 | 0.637 | 0.639 | 0.645 | 0.647 | 0.624 | ||
| Qwen3-VL | HSVR | P(Yes) | 0.086 | 0.121 | 0.184 | 0.189 | 0.195 | 0.231 |
| P(No) | 0.913 | 0.722 | 0.690 | 0.689 | 0.632 | 0.590 | ||
| LSVR | P(Yes) | 0.086 | 0.035 | 0.033 | 0.031 | 0.031 | 0.007 | |
| P(No) | 0.913 | 0.958 | 0.967 | 0.967 | 0.968 | 0.982 |
Object-Token Probability Analysis. We further examine the effects of attention scaling in generation contexts. The results show that HSVR enhancement increases the probabilities of both visually present objects and semantically related but absent objects, while unrelated object probabilities remain largely unchanged. This suggests that attention enhancement can amplify semantic associations alongside genuine visual evidence.
| Backbone | Object Token | 1.00 | 1.25 | 1.50 | 1.75 | 2.00 | 2.25 | relative change |
|---|---|---|---|---|---|---|---|---|
| LLaVA-1.5 | Present (motorcycle) | 0.0381 | 0.0393 | 0.0400 | 0.0405 | 0.0407 | 0.0415 | +9.0% |
| Related absent (traffic light) | 0.00284 | 0.00288 | 0.00292 | 0.00294 | 0.00302 | 0.00309 | +8.8% | |
| Unrelated absent (teddy bear) | 0.00014 | 0.00014 | 0.00014 | 0.00014 | 0.00014 | 0.00014 | −0.6% | |
| LLaVA-1.6 | Present (motorcycle) | 0.2083 | 0.2111 | 0.2125 | 0.2156 | 0.2166 | 0.2179 | +4.6% |
| Related absent (traffic light) | 0.00943 | 0.00963 | 0.00985 | 0.00999 | 0.01012 | 0.01017 | +7.8% | |
| Unrelated absent (teddy bear) | 0.00009 | 0.00009 | 0.00009 | 0.00009 | 0.00009 | 0.00009 | −4.1% |
Quantitative Transition Analysis. We selected ClearSight for the original transition analysis because its pre-softmax attention scaling sharpens visual attention distributions, emphasizing highly attended visual tokens while suppressing weakly attended ones. To further validate the directional effects illustrated in Figure 3(b), extend the transition analysis to the remaining two POPE splits (Random and Popular) on LLaVA-1.5 and Qwen3-VL.
The results show that ClearSight consistently increases both TPR and FPR across all settings, indicating a shift toward affirmative predictions. Notably, the increase in FPR exceeds that in TPR on the adversarial splits, resulting in decreased accuracy. These findings further demonstrate that attention enhancement can induce directional response bias without consistently improving object-specific verification.
| Backbone | Split | Δ Yes Rate | Δ TPR | Δ FPR | Δ Acc |
|---|---|---|---|---|---|
| LLaVA-1.5 | Random | +2.97 | +3.54 | +2.40 | +0.57 |
| Popular | +3.44 | +3.54 | +3.34 | +0.10 | |
| Qwen3-VL | Random | +2.40 | +3.60 | +1.20 | +1.20 |
| Popular | +3.10 | +3.60 | +2.60 | +0.50 |
Response to Reviewer MmJ7 3/4
Q1 Cross-Backbone Generalization
We would like to clarify that RASE demonstrates framework-level generalization across different LVLM architectures, rather than the transferability of a single trained detector.
Different LVLM backbones vary in their architectures, including the number of decoder layers and attention heads, resulting in different feature dimensions and internal representations. Moreover, differences in training data and training procedures can lead to distinct internal representations, even when the model architectures are similar. Therefore, RASE trains a separate lightweight detector for each backbone. Nevertheless, our experiments demonstrate that the proposed feature extraction and detection framework consistently achieves strong performance across diverse LVLM architectures.
Q2 Fluency and Non-Object Hallucination Analysis
We thank the reviewer for raising this important concern regarding the potential effects of rollback on text fluency and non-object hallucinations. We provide further clarification and additional evaluation below.
Clarification of Rollback Mechanism. RASE performs hallucination detection only on object mentions, leaving non-object tokens unaffected. Moreover, reselection operates on the backbone's original prediction distribution, preserving its native language modeling behavior without modifying internal representations.
Additional Evaluation. To assess whether rollback affects text fluency or introduces non-object hallucinations, we conduct additional evaluations across three LVLM backbones. Specifically, we use GRUEN [2] to measure linguistic quality and FaithScore [1] to evaluate fine-grained visual faithfulness, including attribute and relation consistency. We compare RASE with vanilla decoding on the same COCO images under identical generation settings. The results show that RASE maintains largely comparable fluency and fine-grained visual faithfulness across all three backbones.
| Model | Metric | Vanilla | RASE |
|---|---|---|---|
| LLaVA-1.5 | GRUEN ↑ | 0.8109 | 0.8142 |
| FaithScore (Attribute) ↑ | 0.9065 | 0.8970 | |
| FaithScore (Relation) ↑ | 0.8655 | 0.8817 | |
| FaithScore (Overall) ↑ | 0.8810 | 0.8882 | |
| LLaVA-1.6 | GRUEN ↑ | 0.7620 | 0.7656 |
| FaithScore (Attribute) ↑ | 0.9036 | 0.9054 | |
| FaithScore (Relation) ↑ | 0.8881 | 0.8889 | |
| FaithScore (Overall) ↑ | 0.9003 | 0.8981 | |
| Qwen3-VL | GRUEN ↑ | 0.6414 | 0.6384 |
| FaithScore (Attribute) ↑ | 0.8781 | 0.8699 | |
| FaithScore (Relation) ↑ | 0.8888 | 0.8820 | |
| FaithScore (Overall) ↑ | 0.8845 | 0.8874 |
[1] FaithScore: Fine-grained Evaluations of Hallucinations in Large Vision-Language Models, Findings of EMNLP 2024
[2] GRUEN for Evaluating Linguistic Quality of Generated Text, Findings of EMNLP 2020
Response to Reviewer MmJ7 4/4
We sincerely thank the reviewer for the constructive feedback and the opportunity to clarify our work. We hope our responses have addressed the concerns and would be happy to provide further clarification if any questions remain.
Response to Reviewer MmJ7 1/4
W1 Implementation details
We thank the reviewer for pointing out the insufficiently implementation details and providing an opportunity for further clarification. We have incorporated these clarifications into the revised Appendix A.
Multi-Token Object Handling. For multi-token objects, detection is triggered once the complete object mention is generated and identified. We extract diagnostic features from the last subtoken, whose attention patterns and hidden states incorporate the context of the entire object mention through causal attention. If the mention is classified as hallucinated, RASE rolls back to the beginning of the object span and performs reselection.
FlashAttention Compatibility. Replacing FlashAttention with eager attention throughout decoding would introduce considerable latency, but this is unnecessary for RASE. A straightforward solution is to selectively use eager attention for object-token detection while retaining FlashAttention for all other tokens. A more efficient alternative is to keep FlashAttention enabled throughout decoding and independently compute the required attention distributions using the current object-token queries and visual keys stored in the KV cache. Since object tokens constitute only a small fraction of generated tokens, this selective computation limits the additional overhead.
Object Mention Identification and Labeling. We follow the standard CHAIR-based protocol adopted in prior work to identify object mentions using the COCO object vocabulary, including synonym mappings and compound-word matching. During training, we generate captions for 20,000 COCO train2017 images using each backbone and label the identified object mentions as grounded or hallucinated based on the corresponding COCO ground-truth annotations. During inference, we use the same strategy to identify completed object mentions in generated text and trigger RASE detection, without requiring any ground-truth annotations.
Threshold Calibration. For each backbone, we select the detection threshold using five-fold image-level out-of-fold predictions on the training set, maximizing hallucination F1 subject. The resulting thresholds are 0.530, 0.517, 0.571, and 0.589 for LLaVA-1.5, LLaVA-1.6, Qwen2.5-VL, and Qwen3-VL, respectively.
W2-A Quantitative Validation of Object Association Prior
We thank the reviewer for pointing out the limited quantitative evidence supporting our analysis of object association priors. To address this concern, we provide additional statistical validation across multiple LVLM backbones.
To complement the qualitative examples in Figure 3(a), we conduct a quantitative analysis across three LVLM backbones. For each image, we compare the attention allocated to the same visual region under semantically related and unrelated absent-object queries. We consider two types of semantic associations, contextual co-occurrence (car → traffic light) and category similarity (orange → apple), using teddy bear as the unrelated control. Across all three backbones, semantically related queries consistently attract more attention to the same visual regions than unrelated queries. These results provide quantitative evidence that visual attention is influenced by semantic associations even when the target object is absent, reinforcing our finding that attention allocation alone does not establish object-specific grounding.