acceptodds
Under review as a conference paper at ICLR 2027

HSS-DR: From Hidden-State Spectral Signatures to Object Hallucination Detection and Repair in LVLMs

16 views

Abstract

Large vision-language models (LVLMs) have demonstrated strong capabilities in multimodal understanding, yet they still hallucinate objects that are absent from the input image. We investigate this problem by analyzing the spatial spectra of visual representations computed before response generation. In late decoder layers, these spectra show a lower proportion of low-frequency energy and a higher spectral centroid for subsequent hallucinated mentions than for grounded mentions. Based on this finding, we propose HSS-DR, a framework for object hallucination detection and repair. We train a lightweight detector that uses spectral features to identify object mentions at risk of hallucination. For each flagged object mention, HSS-DR uses constrained search to find a lower-risk continuation. The same detector evaluates candidate continuations over a short lookahead and guides their selection based on hallucination risk. Experiments across multiple LVLMs demonstrate that HSS-DR consistently reduces object hallucination while preserving general multimodal capabilities.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large vision-language models, object hallucination, hallucination mitigation, hidden representations, spatial frequency analysis, visual grounding, hallucination detection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_UdfvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HSS-DR, a framework for detecting and mitigating object hallucinations in large vision-language models (LVLMs). The authors observe that grounded and hallucinated object mentions exhibit different spatial spectral patterns in visual hidden states, particularly in late decoder layers. Based on this observation, they train a lightweight detector using spectral features, response-map statistics, and generation context to identify potentially hallucinated object mentions. When a high-risk mention is detected, the method performs constrained local regeneration and selects a lower-risk continuation using short lookahead evaluation. Experiments across six LVLMs show consistent hallucination reductions relative to vanilla decoding, while largely preserving general multimodal capabilities.

Critical strengths
  • An interesting observation about spectral signatures of object hallucinations. The paper identifies differences in the spatial frequency distributions of hidden visual representations between grounded and hallucinated object mentions. The authors further validate these observations across data splits, within-image comparisons, and different model architectures. This provides an interesting perspective on hallucination detection using internal representations.

  • A practical detection-and-repair framework evaluated across multiple architectures. HSS-DR combines lightweight hallucination detection with localized regeneration while keeping the LVLM backbone frozen. The experiments demonstrate consistent reductions in object hallucination across six models, suggesting that the framework can be adapted to different visual-token architectures.

Critical weaknesses
  • Weak connection between the hypothesis and the proposed mechanism. The paper hypothesizes that grounded objects exhibit more spatially coherent responses, but the connection between spatial coherence and the observed spectral differences is not directly validated. Moreover, the lack of an end-to-end ablation removing spectral features leaves their unique contribution to hallucination reduction unclear.

  • Limited evidence for comparative cross-model advantages. While HSS-DR consistently reduces hallucinations across six LVLMs, the evaluation lacks comparisons with established methods such as PAI and OPERA across different architectures. Since cross-model effectiveness is a key contribution, it is important to investigate whether these existing methods exhibit similar improvements or face limitations in cross-model generalization. Without such comparisons, the claimed cross-model advantage remains insufficiently substantiated.

Questions for authors
  • Can the authors provide direct evidence that the observed spectral differences reflect spatial coherence rather than global response statistics (e.g., the DC component)? Additionally, how does removing spectral features affect the final hallucination reduction when the repair procedure remains unchanged?

  • How does HSS-DR's cross-model effectiveness compare with established methods such as PAI and OPERA under matched evaluation settings? Do these methods exhibit similarly consistent improvements across different architectures, and if not, what limitations distinguish them from HSS-DR? If direct comparisons are infeasible, please clarify the reasons.

3 comments

We sincerely thank the reviewer for the careful evaluation and thoughtful comments. We appreciate the valuable feedback and would like to provide the following clarifications and additional experimental results.

W1. Spectral Hypothesis and End to End Ablation

DC component and spatial arrangement. We conducted additional analyses on 299 images with 1,915 grounded and 353 hallucinated mentions. The table reports AUROC using the low frequency energy ratio, with the DC energy ratio evaluated separately.

ConditionAUROC
Original0.7187
DC Energy Ratio0.6789
Without DC0.5952
Without DC + Shuffling0.5220

After removing DC, the remaining frequency components still provide useful information for distinguishing grounded from hallucinated mentions. Their discriminative ability decreases further when spatial positions are shuffled, highlighting the importance of spatial arrangement.

We also evaluated the frequency features separately across 20 object categories. After removing DC, the average AUROC reaches 0.6666, compared with 0.5052 after spatial shuffling. These results further support the value of spatial frequency information beyond global response statistics.

End to end ablation. We conducted an additional ablation on LLaVA-1.5 using the same 500 CHAIR images. The repair procedure, search budget, and training settings were kept unchanged.

MethodCHAIRs ↓CHAIRi ↓Recall ↑
Without Spectral Features37.2010.2775.99
Spectral Features Only33.809.4975.34
Full HSS-DR32.409.1176.45

Notably, the spectral feature block alone achieves a CHAIRs of 33.80%, approaching the performance of the full method. Removing spectral features increases CHAIRs from 32.40% to 37.20% and CHAIRi from 9.11% to 10.27%, while Recall slightly decreases from 76.45% to 75.99%.

These results demonstrate that spectral features contribute to hallucination mitigation beyond the repair procedure itself.

W2. Cross Model Comparisons with PAI and OPERA

We conducted additional comparisons with PAI and OPERA on LLaVA-1.5 and MiniGPT-4, using the same 500 CHAIR images and evaluation metrics.

ModelMethodCHAIRs ↓CHAIRi ↓Recall ↑
LLaVA-1.5PAI36.8011.3077.56
OPERA44.6012.7079.19
HSS-DR32.409.1176.45
MiniGPT-4PAI34.0010.9163.34
OPERA29.2010.2460.86
HSS-DR26.808.9158.64

HSS-DR achieves the lowest CHAIRs and CHAIRi on both models. On LLaVA-1.5, CHAIRs is 32.40%, compared with 36.80% for PAI and 44.60% for OPERA. On MiniGPT-4, HSS-DR further reduces CHAIRs to 26.80%, outperforming PAI (34.00%) and OPERA (29.20%).

These results provide additional evidence that HSS-DR achieves effective hallucination reduction across different model architectures.

Thank you for the additional experiments and clarifications. While I still have some reservations about the interpretation of the spectral signatures, the new end-to-end ablation and comparisons with PAI/OPERA substantially address my main concerns. I have therefore updated my rating to 3.

Thank you very much for your thoughtful follow-up and for taking the time to carefully consider our additional experiments and clarifications. We sincerely appreciate your constructive feedback and recognition of our efforts. Your comments have been very valuable in helping us improve our work, and we are grateful for your time and consideration throughout the review process.

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

This paper introduces HSS-DR, a framework designed to detect and mitigate "object hallucinations" in Large Vision-Language Models (LVLMs). The method projects visual hidden states onto object-specific directions and extracts spatial spectral features from the resulting response maps. A lightweight supervised detector integrates these features with response map statistics and the generation context. During the generation process, if a high-risk object mention occurs, the system triggers a constrained local search, using the same detector to evaluate the generated candidate continuations. Experiments cover six LVLM backbone models, with additional evaluations conducted on the AMBER and MM-Vet benchmarks.

Critical strengths
  1. The observations regarding spectral features are supported by several complementary analyses, including validation on an independent dataset, intra-image comparisons, and further analysis specifically for the Qwen-VL-Chat model.
  2. Ablation studies on the detector demonstrate the predictive value of spectral features. Combining these spectral features with response map statistics and generation context improves the AUPRC from 0.6743 to 0.7071.
  3. The framework shows significant effectiveness in mitigating hallucinations. Notably, for the Qwen-VL-Chat model, HSS-DR reduces the CHAIRₛ metric from 40.0 to 31.0—outperforming the FLB method—while maintaining an object recall of 74.82.
Critical weaknesses
  1. This construction method explicitly preserves the direct current (DC) component. For a response map with NN spatial locations, the DC energy—when using an orthogonal Discrete Cosine Transform (DCT)—is given by Nmean⁡(M)2N\operatorname{mean}(M)^2. Consequently, by introducing a spatially constant offset—without altering the relative differences between spatial locations—one can increase the proportion of low-frequency energy and lower the spectral centroid. This suggests that the "discriminability" mentioned in the paper may, to some extent, reflect the contrast between the "mean response" and "spatial variation," rather than differences in "spatial coherence." Figure 2 shows that most of this discriminability is concentrated in the lowest frequency band, though this does not in itself prove that the DC component is the dominant factor.

  2. While this method outperforms "vanilla decoding" across various metrics, its advantage over FLB is not robust. On LLaVA-1.5, the CHAIRₛ metric shifts only from 32.6 to 32.4 (implying a difference in just one description across 500 images), yet the reported generation time increases from 2.498 seconds to 3.777 seconds per image. On MiniGPT-4, although HSS-DR yields a lower CHAIRₛ score than FLB, its CHAIRᵢ score is significantly higher, and its recall is lower. On the LLaVA AMBER model, FLB achieves a lower hallucination score, whereas HSS-DR maintains a higher coverage rate.

Questions for authors
  1. Does this "spectral separation" phenomenon persist if the DC component is removed or if the response maps are centered? Furthermore, how does the method perform when relying solely on mean/DC component features, or when using spatially shuffled response maps?
  2. Given identical supervisory signals, search budgets, and acceptance rules, how would the final hallucination scores and recall rates change if spectral features were removed or replaced with simpler detection signals?
5 comments

We sincerely thank the reviewer for the careful evaluation and thoughtful comments. We appreciate the valuable feedback and would like to provide the following clarifications and additional experimental results.

W1. DC Component and Spectral Discrimination

We conducted additional analyses on 299 images with 1,915 grounded and 353 hallucinated mentions to examine the effects of DC and spatial arrangement. The table reports AUROC using the low frequency energy ratio, with the DC energy ratio evaluated separately.

ConditionAUROC
Original0.7187
DC Energy Ratio0.6789
Without DC0.5952
Without DC + Shuffling0.5220

The DC energy ratio achieves an AUROC of 0.6789, indicating that DC contributes to the observed differences. After removing DC and normalizing the remaining energy, the frequency features still retain discriminative ability. Shuffling the spatial positions further reduces AUROC to nearly 0.5. These results highlight the value of spatial frequency information in distinguishing grounded from hallucinated mentions, beyond the contribution of the DC component.

We also evaluated the frequency features separately across 20 object categories. After removing DC, the average AUROC reaches 0.6666, compared with 0.5052 after spatial shuffling. Moreover, 19 of the 20 categories achieve an AUROC above 0.5, and 18 show better discrimination with the original spatial arrangement.

These results indicate that the observed spectral differences are not solely explained by DC or global response statistics. Spatial arrangement also provides useful information for distinguishing grounded from hallucinated mentions.

W2. Comparison with FLB and End to End Ablation

End to end ablation. We conducted an additional ablation on LLaVA-1.5 using the same 500 CHAIR images. The repair procedure, search budget, and training settings were kept unchanged, with each detector trained and calibrated under the same protocol.

MethodCHAIRs ↓CHAIRi ↓Recall ↑
Without Spectral Features37.2010.2775.99
Spectral Features Only33.809.4975.34
Full HSS-DR32.409.1176.45

Notably, the spectral feature block alone achieves a CHAIRs of 33.80%, approaching the performance of the full method. Removing spectral features increases CHAIRs from 32.40% to 37.20% and CHAIRi from 9.11% to 10.27%, while Recall slightly decreases from 76.45% to 75.99%.

These results demonstrate that spectral features contribute to hallucination mitigation beyond the repair procedure itself.

Comparison with FLB. On LLaVA-1.5, HSS-DR achieves a similar CHAIRs to FLB (32.4% vs. 32.6%) but requires more inference time (3.777 vs. 2.498 seconds per image). Nevertheless, HSS-DR achieves lower CHAIRs across all six models, with a particularly clear improvement on Qwen-VL-Chat (31.0% vs. 40.0%) at the same Recall of 74.82%. Overall, HSS-DR consistently improves CHAIRs across the evaluated models, with its relative advantages varying across metrics and architectures.

We sincerely appreciate your careful review and constructive feedback.

We would like to kindly follow up on our previous response and the additional experiments. After removing DC, the average AUROC across 20 object categories remains 0.6666, compared with 0.5052 after spatial shuffling. Furthermore, under the same repair procedure, incorporating spectral features reduces CHAIRs from 37.20% to 32.40%, while slightly improving object Recall.

We would be very grateful if you could share any further thoughts on whether these additional results help address your concerns. Thank you again for your valuable time and consideration throughout the review process.

Thank you for the additional experiments. The end-to-end ablation addresses part of my concern: the spectral feature block improves CHAIRₛ from 37.2 to 32.4 under the same repair procedure. The shuffling result also shows that non-DC spatial structure carries predictive information. However, the DC energy ratio alone achieves an AUROC of 0.6789, compared with 0.7187 for the original low-frequency ratio, and the spectral-only ablation is not a DC-only control. What remains unclear is how much the non-DC spatial information contributes to the final repair results beyond simpler response statistics.

I appreciate the clear improvement over FLB on Qwen-VL-Chat at matched object recall. On LLaVA-1.5, however, the CHAIRₛ improvement is small, while inference takes about 51% longer. The practical advantage over FLB therefore appears to be model-dependent. I therefore maintain my rating.

Reviewer_4fLmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the spatial spectral characteristics of visual hidden states associated with object hallucinations in large vision-language models. The authors find that hallucinated object mentions tend to exhibit lower low-frequency energy and higher spectral centroids than grounded mentions. They propose HSS-DR, which uses a lightweight hallucination detector and risk-guided local decoding to identify and repair potentially hallucinated object mentions.

Critical strengths
  1. The paper investigates object hallucination from the spatial spectral characteristics of visual hidden representations, providing an additional perspective on hallucination detection.

  2. The proposed spectral features demonstrate reasonable predictive performance, and the experiments show consistent hallucination reduction across six LVLMs with different architectures.

  3. The paper includes ablation studies on detector features and repair strategies, along with additional analyses across data splits and within individual images.

Critical weaknesses
  1. Limited novelty and insufficient positioning. Prior works such as HALP [1] have already explored hallucination detection using internal representations before generation, while FLASH [2] investigates spectral characteristics of visual information flow and uses spectral modulation to mitigate hallucinations. In comparison, the main contribution of this paper appears to be identifying another detection signal and incorporating it into an existing detection-and-repair framework. The distinction from these closely related works is not sufficiently discussed.

  2. Insufficient justification of the spectral hypothesis. The central hypothesis assumes that grounded objects produce more spatially coherent responses with stronger low-frequency components. However, spatial frequency characteristics are not necessarily determined by grounding quality. For example, a correctly grounded small object may produce a highly localized response with substantial high-frequency energy, while multiple spatially separated instances of the same object category may produce discontinuous response patterns. Object shapes and category-specific spatial distributions can also systematically influence the DCT spectrum. Therefore, the observed spectral differences may partly reflect variations in object geometry or category distributions rather than hallucination itself. Although the authors provide within-image comparisons, these do not adequately control for such factors. More controlled analyses are needed to establish whether the proposed spectral signatures reliably distinguish hallucination from these potential confounders.

  3. Limited mechanistic understanding. The paper establishes correlations between spectral characteristics and hallucination but does not investigate whether these characteristics play a causal role. In particular, unlike FLASH [2], which explores spectral modulation as an intervention, this work does not examine whether directly modifying the spectral components of hidden representations can reduce hallucination. Such experiments could provide stronger evidence for the proposed hypothesis and potentially lead to a more distinctive mitigation approach.

  4. Incomplete ablation of the proposed detection signal. Although Table 4 demonstrates the predictive value of spectral features, the paper does not evaluate their contribution to final hallucination mitigation performance under an identical repair pipeline. It remains unclear how much of the improvement comes from spectral features rather than the existing decoding and repair mechanisms.

  5. Potential annotation noise and evaluation trade-offs. Grounded and hallucinated mentions are automatically labeled using CHAIR object annotations without additional human verification. Furthermore, hallucination reduction is consistently accompanied by lower object recall, making it difficult to distinguish improved visual grounding from more conservative generation.

References

[1] Kogilathota SA, G SVE, Sun L, Zhou J. HALP: Detecting Hallucinations in Vision-Language Models without Generating a Single Token. In: Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics; 2026:6067-6085. doi:10.18653/v1/2026.eacl-long.287

[2] Lu S, Suo W, Zheng Y, Yao J, Xu W, Wang P. Beyond Attention Imbalance: Mitigating Hallucinations via Spectral Surgery. In: Proceedings of the 43rd International Conference on Machine Learning. Vol 306. Proceedings of Machine Learning Research; 2026:81550-81568.

Questions for authors

See above.

5 comments

We sincerely thank the reviewer for the careful evaluation and thoughtful comments. We appreciate the valuable feedback and would like to provide the following clarifications and additional experimental results in response to the concerns raised.

W1. Novelty and Positioning

We would like to clarify how HSS-DR differs from HALP and FLASH.

HALP predicts hallucination risk before generation using internal representations. In contrast, HSS-DR analyzes spatial frequency patterns in visual hidden states to assess the risk of individual object mentions during generation and guide local repair.

FLASH investigates the frequency characteristics of visual information flow through attention heads and directly modifies them to mitigate hallucination. HSS-DR instead extracts frequency features from visual hidden states for hallucination detection, without modifying the model's internal representations.

The main contribution lies in identifying spatial frequency patterns associated with object hallucination and demonstrating their usefulness for detection and repair. Our new end-to-end ablation (W4) further shows that these features improve hallucination mitigation beyond the existing repair mechanism.

W2. Spectral Hypothesis and Potential Confounders

We conducted additional analyses on 299 images with 1,915 grounded and 353 hallucinated mentions to examine the effects of DC, spatial arrangement, and object properties.

DC component and spatial arrangement. We compared the original spectra with those obtained after removing DC and shuffling spatial positions:

ConditionOverall AUROCAverage AUROC by Category
Original0.71870.6860
Without DC0.59520.6666
Without DC + Shuffling0.52200.5052

For each of the 20 object categories, we separately evaluated the ability of the spectral features to distinguish grounded from hallucinated mentions. After removing DC, the average AUROC was 0.6666, with 19 of 20 categories scoring above 0.5. Shuffling the spatial positions reduced the average AUROC to 0.5052, with lower performance in 18 of the 20 categories. These results show that the spectral differences cannot be explained solely by DC or differences in object categories. Spatial arrangement also carries useful information.

Object geometry. We also examined the effects of object size and instance count. In our stratified analysis, smaller grounded objects showed a lower proportion of low-frequency energy after removing DC (58.67% vs. 63.08%), indicating that object geometry can influence the spectral patterns. Nevertheless, the category comparisons and spatial shuffling results above show that these features remain useful for distinguishing grounded from hallucinated mentions.

W3. Mechanistic Understanding

The observed spectral differences may arise from multiple factors rather than a single underlying mechanism. To investigate this, we explored several interventions on visual hidden states in earlier experiments, including frequency filtering using FFT, frequency modulation using DCT, and adjusting the spectra of later layers toward those associated with grounded objects.

However, these interventions did not yield consistent improvements in hallucination reduction. This suggests that the relationship between spectral patterns and hallucination is more complex than can be addressed by directly modifying individual frequency components.

These findings motivate our approach of using spectral features as predictive signals for hallucination detection. As demonstrated by the additional experiments in W2 and W4, these features capture useful information about hallucination risk and contribute to more effective detection and local repair.

W4. End to End Ablation of Spectral Features

We conducted an additional ablation on LLaVA-1.5 using the same 500 images. The repair procedure, search budget, and training settings were kept unchanged.

MethodCHAIRs ↓CHAIRi ↓Recall ↑
Without Spectral Features37.2010.2775.99
Spectral Features Only33.809.4975.34
Full HSS-DR32.409.1176.45

Notably, using spectral features alone achieves a CHAIRs of 33.80%, approaching the performance of the full method. Removing spectral features increases CHAIRs from 32.40% to 37.20% and CHAIRi from 9.11% to 10.27%, while Recall decreases slightly from 76.45% to 75.99%.

These results demonstrate that spectral features make a substantial contribution to hallucination mitigation beyond the repair procedure itself.

W5. Annotation Reliability and Evaluation Trade-offs

Annotation reliability. CHAIR is a widely used benchmark for evaluating object hallucination. Beyond CHAIR, we also evaluated HSS-DR on AMBER Generative, providing additional evidence of its effectiveness under a different evaluation protocol.

Hallucination reduction and object recall. On AMBER Generative, HSS-DR reduces CHAIR from 7.39 to 5.91 on LLaVA-1.5 and from 6.71 to 5.55 on Qwen-VL-Chat, while Cover slightly improves from 51.74 to 51.85 and from 53.55 to 53.81, respectively. On Qwen-VL-Chat, HSS-DR also achieves a lower CHAIRs than FLB (31.0% vs. 40.0%) at the same object Recall (74.82%). Furthermore, the average response length on LLaVA-1.5 remains nearly unchanged (118.57 vs. 118.89 tokens for Vanilla and HSS-DR, respectively).

These results show that HSS-DR can reduce hallucinations while maintaining object coverage, rather than simply generating shorter or less informative responses.

Thank you for the detailed response and the additional experiments. I appreciate the end-to-end ablation, which provides useful evidence. This addresses my W4.

However, I remain concerned about the interpretation of the proposed spectral signatures. In particular, the DC-only feature achieves an AUROC of 0.6789, relatively close to the original low-frequency ratio (0.7187), whereas removing the DC component reduces the overall AUROC to 0.5952. It does not seems like the observed spatial patterns reflect grounding quality. Instead, it could be object geometry or other properties of the response maps. Indeed, the additional object-size analysis confirms that geometric properties systematically influence the spectral statistics even among grounded objects. The category-wise comparison and spatial shuffling experiments do not directly control for this confounding factor.

Regarding the novelty concern, I acknowledge the differences between HSS-DR and HALP/FLASH. However, my concern is that the main methodological contribution is introducing a new feature into an existing detection-and-repair paradigm. While the additional ablation demonstrates that this feature is useful, the DC analysis also suggests that much of its predictive power may come from simple response statistics. Therefore, I remain unconvinced that the spectral analysis provides sufficiently novel insights into object hallucination to justify the overall contribution.

I therefore maintain my original rating.

Comments on the paper

1 comment

Summary of Additional Evidence for the Area Chair

We sincerely thank the Area Chair and reviewers for their careful evaluation and constructive discussion. We would like to briefly summarize the key experimental evidence added during the rebuttal.

1. Practical Contribution of Spectral Features

Our new end-to-end ablation demonstrates that spectral features contribute directly to hallucination reduction, beyond their predictive performance as detection signals.

On the same 500 CHAIR images, with the repair procedure and search budget unchanged, removing spectral features increases CHAIRs from 32.40% to 37.20%, while object recall slightly decreases from 76.45% to 75.99%. This confirms that spectral information improves the final generation results under an identical repair pipeline.

We further examined whether a simpler DC statistic could achieve comparable results. Although the DC energy ratio alone achieves an AUROC of 0.6789, using it as the sole detection feature yields substantially worse CHAIRs (41.20% vs. 32.40%) and object recall (73.45% vs. 76.45%) than the full method.

These results establish the practical value of spectral features in hallucination detection and repair, which cannot be explained by the DC ratio alone.

2. Further evidence connecting spectral patterns to visual grounding.

After removing DC, the spectral features retain discriminative ability, while spatial shuffling reduces their AUROC from 0.5952 to 0.5220.

We also conducted a new analysis comparing grounded and hallucinated object responses over identical visual patches and equally sized reference regions. Across 70 images, the spectral difference is 7.55 percentage points stronger at real object locations, providing further evidence of its connection to visual grounding.

3. Effectiveness across different architectures.

HSS-DR consistently reduces CHAIRs across six LVLM architectures. Additional comparisons with PAI and OPERA on LLaVA-1.5 and MiniGPT-4 further support its effectiveness. Results on AMBER also demonstrate hallucination reduction while maintaining object coverage.

Together, these additional experiments connect our spectral findings to measurable improvements in final generation. HSS-DR uses these signals to guide local repair during generation, demonstrating their practical value in reducing object hallucinations across different LVLM architectures.

We sincerely appreciate the reviewers' time, thoughtful feedback, and consideration of these additional results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.