AVCD: Mitigating Hallucinations in Audio-Visual Large Language Models through Contrastive Decoding

We appreciate the reviewer’s insightful comments. The reviewer’s experimental suggestions enabled us to offer a more in-depth analysis of our method. Based on the comments, we have:

(1) Categorized hallucination cases to demonstrate which types our model effectively mitigates,
(2) Conducted broader performance validation using the OmniBench dataset,
(3) Added clarification on the motivation behind Eq. (11),
(4) Provided per-component ablation results.

Our detailed responses are provided below.

[W1 & Q1] What types of hallucinations does AVCD primarily mitigate, such as cross-modal inconsistencies, unimodal fabrications, or spurious alignments?

Thank you for your insightful suggestion regarding the taxonomy of hallucinations. In response, we provide a deeper analysis of the types of hallucinations mitigated by AVCD, using a subset of the AVHBench dataset. Following your guidance, we categorize hallucinations into three representative types defined in AVHBench:

• Case 1. Audio-driven video hallucination: Vision-centric questions misled by irrelevant or misleading audio cues.

  Example: Q: Is the ship visible in the video? Video: No ship visible, Audio: Ship sounds present.

• Case 2. Video-driven audio hallucination: Audio-centric questions misled by visual content.

  Example: Q: Is the lion making sound? Video: Lion visible, Audio: No lion sound.

• Case 3. Audio-visual alignment hallucination: Failures to correctly judge the consistency between audio and visual modalities.

  Example: Q: Are the contexts of audio and visual content matching?

We evaluate AVCD on each category using VideoLLaMA2 as the base model. The results are as follows:

While AVCD maintains the already high performance in Case 1, it shows significant improvements in Cases 2 and 3. These results support our claim that AVCD balances modality contributions more effectively.

We will include this analysis and the above categorization in the revised version to better characterize the nature of hallucinations and the targeted efficacy of AVCD.

[W2 & Q2] Could you evaluate AVCD on more challenging benchmarks like OmniBench to better assess its effectiveness in complex hallucination scenarios?

In response to the reviewer’s suggestion, we additionally evaluate various decoding methods (Base, VCD, OPERA, VCD*, and AVCD) on the OmniBench dataset, which is known to require more complex reasoning and better tests hallucination mitigation in AV-LLMs.

Following the OmniBench [1], we use video-SALMONN as the backbone model, where they report an overall accuracy of 35.6%. We reproduce this setting and obtain 33.5% accuracy with Base decoding, which we use as our baseline.

As shown in the table below, AVCD is the only method that consistently outperforms the Base model, whereas other methods often underperform or show comparable results. This highlights AVCD’s strong generalization ability, even on more challenging benchmarks.

[Abbreviation in the table]
Action: Action & Activity, Story: Story Description, Plot: Plot Inference, Object: Object Identification & Description, Context: Contextual & Environmental, Identity: Identity & Relationship, Text: Text & Symbols, Count: Count & Quantity

[Reference]
[1] OmniBench: Towards The Future of Universal Omni-Language Models, arXiv 2025.

[W3 & Q3] What is the motivation behind Equation (11), and what role does it play in the proposed method?

Eq. (11) shows an intermediate step to verify whether hallucinations caused by masking audio, video, and both modalities could be treated as independent. The equation combines the three masked variants in a single formulation, assuming that their effects on hallucination could be simply additive.

However, as shown in Table 3, the performance of Eq. (11) was lower than applying Eq. (2). We interpret this result as evidence that hallucinations caused by different modalities are not independent, and that masking multiple modalities simultaneously can lead to redundant or overlapping corrections.

This finding became a key motivation for the design of AVCD, which explicitly considers interactions between modalities rather than naively aggregating all masked logits—thereby avoiding interference, as explained in L204 of the main paper.

As a result, AVCD can treat hallucinations from different modalities as interacting factors, and improves decoding performance through more sophisticated formulation.

[W4 & Q4] Could you provide per-component ablations to show how much each of the three components contributes to performance?

Following reviewer’s suggestion, we conducted an ablation study on AVHBench using VideoLLaMA2 as the base model. The results are summarized below:

Baseline refers to a contrastive decoding (CD) setup with randomly selected dominant modality and Eq. (11).

Dominance-aware masking improves accuracy by approximately +5%.

Trimodal CD with Eq. (10), our proposed extension of CD to trimodal alignment, further improves performance by +2.9%.

Entropy-guided adaptive decoding (EAD) maintains accuracy while reducing inference latency by over 30%.

These results validate the complementary roles of all three AVCD components as each contributes meaningfully to either accuracy or efficiency.

We appreciate reviewer's insightful comments. We were able to add crucial details that enhance the clarity of our work. Based on the comments, we have:

(1) elaborated on the attentive masking mechanism,
(2) provided details of the inference speed experiments,
(3) clarified how AV-LLMs suffer from hallucination,
(4) highlighted differences from the SID model,
(5) demonstrated the reliability of attention via Attention-Guided Masking,
(6) explained the motivation for using the final query token,
(7) compared performance against a non-CD method (OPERA),
(8) clarified how our entropy-based decoding differs from existing methods.

Our detailed responses are provided below.

[W1 & Q6 & Q7] Could you clarify how "masking out" and "corrupted states" are implemented in your method?

We apologize for the confusion. In our paper, terms such as “corrupted” or “masked out” consistently refer to the same operation: zeroing out the corresponding embeddings/features to reduce their influence during decoding.

For the corruption operation, we first compute attention values across all layers to derive a global attention distribution, which determines both the dominance score and the masking threshold. We then identify tokens that exceed this threshold and apply zero-out masking to those associated with less dominant modalities in the attention weights.

We agree that this clarification is important and will explicitly state this process in the final version of the paper.

[W2 & Q8] Have you conducted a direct speed comparison on AVHBench? Additionally, can the multiple forward passes for masked inputs share parts of the computation graph, or are they executed independently?

In Figure 6, we measured inference time on the AVHBench dataset using the VideoLLaMA2 model. Specifically, we averaged the response time over 100 samples to compute the results. As suggested by the reviewer, we will include the inference speed scale in the final version, as shown in the table below.

While base contrastive decoding typically requires up to four separate full forward passes in AV-LLMs, our Entropy-Guided Adaptive Decoding (EAD) effectively reduces unnecessary passes by leveraging entropy-based filtering.

Although, like other existing CD methods, our implementation does not employ shared computation across passes, AVCD still achieves notable improvements. For example, with τ = 0.8, AVCD not only surpasses the Base method in accuracy but also runs faster than VCD, which generally requires two separate decoding passes.

[W3 & Q1] Could you clarify how hallucination is defined in the audio-visual context, and how it relates to hallucinations in other multimodal models?

Hallucination in AV-LLMs differs fundamentally from that in VLMs due to the presence of three modalities (audio, vision, and language), as opposed to two in VLMs. In AV-LLMs, hallucinations often arise when the model imagines non-existent audio from visual input or visual content from audio input, as noted in AVHBench [1]. This is further elaborated by CMM [2], which attributes such hallucinations to over-reliance on unimodal priors and spurious inter-modality correlations.

This tri-modal structure introduces 7 types of modality interactions (3: unimodal, 3: bimodal, 1: trimodal), making hallucination more complex than in VLMs.

Our proposed AVCD framework is designed specifically to handle this tri-modal challenge by applying modality-aware CD. Despite this challenge, AVCD shows strong generalizability to video-LLMs and image-LLMs, outperforming prior CD methods (see Table 2, Supp. Sec. C, and Supp. Sec. F).

In particular, Table 2 shows that AVCD brings significant performance gains over the base model even on video-LLM settings, highlighting its broad applicability beyond the tri-modal setup.

[Reference]
[1] AVHBench: A Cross-Modal Hallucination Benchmark for Audio-Visual Large Language Models, ICLR 2025.
[2] The Curse of Multi-Modalities: Evaluating Hallucinations of Large Multimodal Models across Language, Visual, and Audio, arXiv, 2024.

[W4] Could you clarify how your method differs from SID?

While we acknowledge that attention-guided perturbations have been explored in prior work (e.g., SID), our contribution lies in extending this idea to the tri-modal setting of AV-LLMs. Specifically, AVCD adaptively identifies less dominant modalities and selectively masks their most informative tokens, introducing a novel application of CD tailored to multimodal interactions.

To our best knowledge, this is the first systematic exploration of CD in AV-LLMs, and we further demonstrate its generalizability across AV-LLMs, video-LLMs, and image-LLMs, underscoring its broad applicability beyond the original trimodal context.

[Q2] Have you analyzed attention visualizations before and after applying AVCD?

While we cannot provide explicit attention maps, we conducted an indirect yet rigorous evaluation to verify whether the model's attention captures semantically meaningful auditory-visual information.

Specifically, we compared two masking strategies during decoding: a random zero-out masking strategy, which randomly masks 50% of tokens, and our attentive zero-out masking strategy, which masks the top 50% most attended tokens as identified by the attention scores.

As shown in Table above, random masking significantly degrades decoding performance, while our proposed attentive masking improves performance by 3.9% over the base decoding.

These results suggest that the attention mechanism in AVCD is indeed focusing on informative and modality-relevant tokens.

[Q3] Is it appropriate to use the last query token's attention distribution for identifying dominant modality?

We use the attention distribution from the final query token because it is directly responsible for next-token prediction in autoregressive decoding. This design choice is supported by prior works [3,4], where attention from the final token is commonly used to interpret token-level importance in LLMs.

To validate this decision, we compare our strategy against an alternative approach that averages attention across all query tokens. As shown in the table below, using final query token achieves higher accuracy, indicating that the final token provides a more reliable signal for analyzing modality dominance during decoding.

[Reference]
[3] Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models, ICLR 2025.
[4] Hierarchical Context Merging: Better Long Context Understanding for Pre-trained LLMs, ICLR 2024.

[Q4] Could you provide a comparison with OPERA to demonstrate the effectiveness of AVCD?

we attempted to apply OPERA [5] to VideoLLaMA2; however, due to its beam search with multiple attention rollbacks, we encountered out-of-memory (OOM) issues when handling its long input sequences (2K+ tokens). To circumvent this, we evaluated OPERA on a lighter AV-LLM—video-SALMONN, which processes only a few hundred tokens, to assess its feasibility in the AV domain.

As shown in the table below, both OPERA and VCD led to noticeable performance degradation on AVHBench and AVHBench-Captioning. In contrast, our trimodal extension of VCD (VCD*) improved over Base decoding, though it still lagged behind AVCD.

Overall, AVCD achieves the most significant performance gain over Base across both benchmarks.

[Reference]
[5] OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation, CVPR 2024.

[Q5] How does your entropy-based decoding differ from existing approaches?

While [6] utilizes entropy to set a hyperparameter (e.g., α) that controls the degree of contrasting in CD, simply adapting this approach requires up to four full forward passes—making it impractical for AV-LLMs.

In contrast, we treat entropy as a threshold signal to decide whether to apply CD at all, rather than how strongly to contrast. This design choice enables us to maintain reasonable inference speed even in the tri-modal setting. We believe this approach is more suitable for real-world deployment scenarios, where efficiency and scalability are critical.

[Reference]
[6] Adaptive Contrastive Decoding in Retrieval-Augmented Generation for Handling Noisy Contexts, EMNLP 2024.

Thanks to the reviewer’s insightful comments, we were able to add detailed information that enhances reader understanding. Based on the comments, we have:

(1) re-emphasized the broad validation scope,
(2) referred to the sections describing derivations and experimental setups,
(3) supplemented the discussion with failure cases,
(4) added experiments to verify attention reliability,
(5) included comparisons with non-CD methods and added explanations on ethical considerations and potential real-world applications,
(6) provided more detailed descriptions of the datasets used,
(7) explained the role of the softmax function in Eq. (1),
(8) supplemented the explanation on modality distortion.

Our detailed responses are provided below.

[W1] Limited validation scope with respect to models and datasets.

We would like to clarify that AVCD is evaluated on four multimodal large language models (VideoLLaMA2, video-SALMONN, Video-LLaVA, and LLaVA-1.5) across five datasets (MUSIC-AVQA, AVHBench, MSVD-QA, ActivityNet-QA, and POPE). These evaluations cover a diverse range of tasks and modalities, demonstrating the robustness and generalizability of our approach.

[W2] Insufficient details on derivations and experimental setups.

We would like to clarify that our paper does not involve OCR in any part of the method or evaluation. Additionally, we provide detailed information about the datasets, evaluation protocols, baselines, and implementation details in Sec. 4.1 of the main paper. The complete mathematical derivation is also included in Supp. Sec. A

[W3] Absence of discussion on limitations and failure cases.

We would like to note that we include a discussion of our method’s limitations in Supp. Sec. H. To further support readers' understanding, we will include additional failure cases in the final version of the paper.

[W4] Concerns on attention reliability.

We understand the reviewer’s concern regarding the reliability of attention. To address this, we conducted additional experiments on AVHBench dataset using VideoLLaMA2 model comparing attention-guided masking with random masking strategies.

When random masking is applied, performance drops significantly to 71.71%, compared to the Base decoding at 78.05%. In contrast, our proposed attention-guided masking selectively distorts the modality that truly affects the model output, resulting in improved performance of 81.95%. These results indicate that attention-based masking effectively identifies influential content and contributes to reducing hallucinations.

Moreover, attention mechanisms have been extensively studied in prior works [1,2] for identifying important tokens in LLMs, further supporting their reliability.

[Reference]
[1] Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models, ICLR 2025.
[2] Hierarchical Context Merging: Better Long Context Understanding for Pre-trained LLMs, ICLR 2024.

[W5] Limited discussion on non-CD methods, ethical considerations, and potential real-world applications.

• Non-CD method comparisons

To address comparisons with non-CD methods, we include experiments with the OPERA [3] method as a representative baseline using video-SALMONN model.

Our results show that OPERA underperforms in the AV-LLM setting, with lower scores on both AVHBench and AVHBench_cap. In contrast, our method achieves the best performance, demonstrating its effectiveness in handling complex audio-visual inputs and reducing hallucinations where OPERA struggles.

• Ethical implications

We acknowledge that CD can produce more persuasive and coherent text, which raises ethical concerns. Specifically, it may be misused to generate fake news, misinformation, or manipulated content by making false information appear more credible. Responsible deployment and mitigation strategies should be considered in future work.

• Real-world applications

Our proposed method is model-agnostic and can be applied broadly in any environment using AV-LLMs to reduce hallucinations. Considering the trade-off between model performance and inference time, users can adapt the method flexibly according to their application needs, making it practical for real-world deployment.

[Reference]
[3] OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation, CVPR 2024.

[Q1] What are the total numbers of samples in the three datasets?

Thank you for your question. Our evaluation covers diverse test sets across different model types. For AV-LLMs, we use AVQA (9,185 pairs), AVHBench (5,302 pairs), and AVHBench-Captioning (1,000 pairs). For Video-LLMs, we evaluate on MSVD-QA (1,000 pairs) and ActivityNet-QA (6,760 pairs). Lastly, for image-LLMs, we use the POPE dataset, which contains 9,000 samples across three categories: Random, Popular, and Adversarial (3,000 each). We will include detailed dataset descriptions in the revised paper.

[L1] What is the purpose of applying softmax in Eq. (1)?

In Eq. (1), the softmax function is used in the standard way to convert the LLM’s output logits into a probability distribution over the vocabulary. This is an essential step for enabling word prediction, as it allows the model to assign interpretable likelihoods to each possible token.

[L2] What does the term "damaged and biased visual information" refer to?

We would like to clarify that the phrase “damaged and biased visual information” specifically refers to intentionally corrupted inputs introduced during inference, such as noise addition, data augmentation, or masking, as mentioned in L144 of the main paper. It is important to note that this does not refer to any issues arising from data collection.

We sincerely thank the reviewer for the thorough and insightful feedback. The reviewer’s detailed critique greatly helped us deepen our explanation regarding KL divergence and improve the clarity of the paper. Based on the comments, we have:

(1) provided an analysis of the relationship between low KL divergence and the strength of the contrastive signal,
(2) explained how our implementation captures local token-level changes,
(3) compared our approach to scenarios using FlashAttention,
(4) revised the notation to a more general form reflecting the adaptive modality selection.

Our detailed responses are provided below.

[W1.1 & Q1] Can low KL divergence still lead to sufficient contrastive effectiveness?

Contrary to reviewer’s concern, low KL divergence does not mean the contrastive signal is ineffective or hallucinations are absent. To support this, we applied AVCD and VCD on the AVHBench test set using VideoLLaMA2. AVCD changed 13.1% of answers (8.5% improved, 4.6% degraded), while VCD changed only 7.4% (4.4% improved, 3.0% degraded), despite VCD showing higher KL divergence. This demonstrates that semantically coherent perturbations from attention-based masking (used in AVCD) provide a more effective contrastive signal than simply increasing KL. Therefore, the two conditions, low divergence and effective contrast, are not mutually exclusive.

[W1.2 & Q1] Can KL divergence effectively capture local token-level changes that are critical to the final answer?

Thank you for raising this important point. We clarify our use of KL divergence in the proposed framework.

In our implementation, KL is computed after applying the Adaptive Plausibility Constraint, as detailed in Supp. Sec. B. This constraint, commonly used in the CD literature, ensures that the final output is restricted to tokens with sufficient confidence in the original logit distribution.

As a result, the KL divergence is calculated only over the top-K logits that are most likely to influence output generation, addressing the concern that KL fails to capture local behavior.

We will revise the main paper to explicitly describe this procedure.

[W2 & Q2] Is AVCD computationally efficient and practically applicable, for example in settings that leverage FlashAttention?

We acknowledge the reviewer’s valid concern regarding efficiency. Indeed, our current AVCD implementation is incompatible with FlashAttention. To clarify the efficiency gap, we provide the following inference speed comparison (lower is faster):

While FlashAttention is not currently applicable due to the need to access attention scores, our goal is fundamentally different: this paper is the first to design a contrastive decoding (CD) strategy tailored for AV-LLMs. Simply extending existing CD methods designed for unimodal or vision-language models often focused on a fixed modality, which leads to suboptimal or unsuitable behavior in AV settings. Our work identifies this issue and proposes a solution via modality-aware masking and new mathematically extended CD formula for three modalities.

Moreover, instead of naively applying CD across all tokens with multiple forward passes, we introduce Entropy-Guided Adaptive Decoding (EAD) to selectively apply AVCD only when necessary, reducing the computational overhead.

Although FlashAttention cannot be used at this stage, we are optimistic. Recent work [1] on optimizing CD for LLMs demonstrates that more efficient CD implementations are possible, and we believe future work can extend these techniques to AV-LLMs once modality-aware CD becomes more mature. Our contribution lays the foundation for this line of research, and we ask for your generous understanding of the current trade-off in efficiency in light of the broader methodological contribution.

[Reference]
[1] Fast Large Language Model Collaborative Decoding via Speculation, ICML 2025.

[W3 & Q3] On the generality of the framework.

Thank you for pointing out this confusion. We apologize for the lack of clarity. In the final version, we will ensure that Equation (10) reflects the generalized case, allowing any modality to be adaptively masked based on its dominance score.

Questions relating to the mathematical framework still apply after the author's response.

1. Your response has not addressed the question I asked about the validity of the mathematical approximation you use. Even if you restrict the KL-divergence calculation to top-k logits in the original distribution, this doesn't constrain the logit values of the secondary distribution to be close to the top-k logits, you could still have a large difference in per-token logits and therefore the approximation you use would have large error in these cases. The experimental result you have provided is an empirical justification amounting to "this works in practice" which is valid, but isn't mathematically rigorous as the submission suggests. Once again, you are adjusting per token logits and so the logit difference needs to be small for the approximation to be valid. KL divergence does not tell a reader what the logit difference of the tokens which would've been originally selected and are eventually selected are. If your argument is really "it works empirically" you should say so.
2. On the speed argument, I would say >2x inference cost is pretty substantial, but given this is an initial work in this direction it is a minor weakness as other reviewers have pointed out. When authors have a baseline which do not require attention matrix materialization, they should use FlashAttn or whatever SotA attention method is applicable as no-one should run a base model without FlashAttn or similar if the attention matrix does not need to be materialized. So the AVCD method is really 70/140% slower than the baseline rather than 30/80% slower as you have outlined in other reviews.

We sincerely thank the reviewer for the constructive and detailed feedback. Your comments helped us clarify the limitations of our current explanation and better position our empirical evidence.

[Q1] Validity of the Approximation

We understand the reviewer’s concern that KL divergence does not directly justify the error approximation in Eq. (7), and we appreciate the opportunity to clarify this point. To address this, we additionally computed the exact approximation error described in Supp. A.1 ($\sigma^2 / 2M^2$), using logits selected via the adaptive plausibility constraint.

This offers a more direct and rigorous quantification of the approximation error. We evaluated this error on 100 examples from the AVBench dataset. As shown in the tables below, AVCD consistently yields smaller or comparable approximation errors than VCD across all modality masking settings. Furthermore, the absolute error values are small, providing strong empirical support for the validity of our approximation.

[Table] Approximation Error with Adaptive Plausiblity Constraint | Method | Video Masked | Audio Masked | Audio & Video Masked | |---|:---:|:---:|:---:| | VCD | 0.015 | 0.083 | 0.073 | | AVCD (Ours) | 0.015 | 0.032 | 0.037 |

We will revise the final version of the paper to replace the KL divergence-based explanation with this analysis, which more directly supports the validity of the approximation. As such, we will further revise L224–225 to clarify that it is found to work well empirically, rather than for its theoretical implications.

[Q2] Inference Speed under FlashAttention

We fully agree that reporting inference speed under FlashAttention provides a more realistic estimate of practical deployment costs. Accordingly, we have shared FlashAttention-based inference-time comparisons in our response to another reviewer and will include these results in the final version of the paper to more accurately reflect the runtime overhead of AVCD in modern setups.

We thank the reviewer for the valuable feedback. The comments helped us create a clearer and more understandable final version for readers. Based on the comments, we have:

(1) conducted comparison experiments between random masking and attention-guided masking,
(2) examined the reliability of analyses based on attention weights,
(3) re-emphasized our efforts to improve inference speed,
(4) supplemented the explanation that our masking strategy is based on global attention.

Our detailed responses are provided below.

[W1] Validation of attention-guided masking against random masking or noise injection.

Thank you for the insightful comments. In response, we conducted additional experiments to compare Attention-Guided Masking with Random Masking strategies.

When we apply random masking, the performance (71.71%) drops significantly compared to the base decoding (78.05%). In contrast, our proposed attention-guided masking selectively distorts the modality that truly affects the model output, resulting in improved performance (81.95%).

Additionally, as shown in Table 1 of the main paper, AVCD consistently outperforms VCD, which perturbs the input by injecting noise, across all benchmarks. This suggests that simply corrupting the input with random noise is not as effective as selectively masking semantically meaningful tokens. Therefore, our method provides evidence that attention can serve as a useful signal for locating critical information, helping reduce hallucinations more reliably.

[W2] How reliable are the attention weights in identifying key modalities?

In Table 4, we validate the effectiveness of using attention weights to identify the dominant modality. Our adaptive strategy frequently selects “language” as the dominant modality, and this choice consistently yields better performance compared to manually assigning video or audio as dominant.

Moreover, attention weights have been extensively studied in prior works [1,2] for identifying important tokens in LLMs, further supporting their reliability in modality attribution.

[Reference]
[1] Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models, ICLR 2025.
[2] Hierarchical Context Merging: Better Long Context Understanding for Pre-trained LLMs, ICLR 2024.

[W3] A minor concern regarding the 2× slower inference speed.

Thank you for recognizing that the speed gap is not a major weakness in the context of our overall contribution. We would like to emphasize that, despite handling three modalities in AV-LLMs, our method accelerates inference speed by incorporating an entropy-guided adaptive decoding strategy that effectively skips unnecessary steps. This mechanism helps reduce computational overhead and maintains a balance between richer multimodal reasoning and decoding efficiency.

[Q1] Have you considered varying the masking probability (P) across different layers or models?

We apologize for any confusion caused.

To clarify, we compute attention values across all layers and mask the top 50% of tokens based on the global distribution. As a result, the number of masked tokens varies per layer, effectively leading to a layer-wise adaptive masking scheme.

To validate this, we conducted a comparison against a setting where we mask the top 50% attention-value tokens separately within each layer. The results are shown below:

While fixed masking is already more effective than Base decoding, our adaptive strategy guided by global attention achieves significantly better performance. This supports the idea that masking based on global attention better identifies influential tokens across the entire model.

Thanks for the response. The authors have answered all my questions and I currently have no concern. I am not worried about the scoping of 'what is hallucination' here since general performance is still an indicator of levels hallucination. But it did seem that the scoping of this paper can be slightly more general by focusing on this method improves the general performance. Otherwise, the authors can include more analysis on hallucination, which fortunately is included in other reviewers' rebuttals.