Mitigating Object Hallucination in Large Vision-Language Models via Image-Grounded Guidance

Q1: Reasons to balance the original branch.

Thank you for your valuable suggestion. While increasing the guidance strength leads to a notable decrease in the CHAIR score, we have observed that overly strong guidance can negatively affect the model's ability to follow instructions, as illustrated in Figure 12.

Q1.1: Additional experimental validation for CFG with different $\gamma$.

In response to your suggestion, we conducted additional experiments to further validate our findings. Figure 8 of the updated revision illustrates how varying guidance strength affects the quality of LLaVA's output text in both the LLaVA-QA90 task and the image captioning task (maximum generation length = 256), evaluated using metrics such as BLEU, ROUGE-L, and CIDEr. Due to space constraints, we summarize a subset of the results in the table below for the LLaVA-QA90 task.

Our experiments demonstrate that a guidance strength of 1 does not yield the best image captioning performance (i.e., quality of the generated text). In LLaVA-QA90, a guidance strength $\gamma$ between 0.5 and 0.7 provides the most enhanced text quality. This aligns with a common concern in methods employing classifier-free guidance, where excessively strong guidance can divert the generation process.

Additionally, we conducted experiments comparing the overall generation quality of LLaVA using GPT-4V as a judge, scoring outputs on a scale of 10 for accuracy and detail, which further supports our findings. The comparison between models with (✅) and without (❌) balancing the original LVLM branch is summarized below and Table 25 of our revision:

In summary, our additional experiments confirm that setting γ = 1 results in decreased text quality, both quantitatively across several evaluation metrics and qualitatively as assessed by GPT-4V evaluation.

Q1.2: Presentation of specific cases with and without balancing the original branch.

We analyzed specific cases to illustrate the effect of overly strong guidance. Figure 9 of the updated revision demonstrates how overly strong guidance can reduce instruction adherence by inducing the model to include unnecessary details from the image.

• Question: How might the art installation encourage interaction among people?
• LLaVA with overly strong guidance: "The art installation, which features a large, colorful tiger sculpture and a car parked next to it, can encourage interaction among people..."

This case highlights that overly strong guidance can induce the model to prioritize providing exhaustive visual details from the image, even when such details are irrelevant to the specific instruction (e.g., "a car parked next to it"). In contrast, balanced guidance enables the model to maintain better adherence to the instruction while still utilizing the visual information effectively.

Thank you for your constructive feedback. Please find our detailed response below for the questions raised in the review.

W1: Discussions on other relevant methods using CFG

Thank you for mentioning these references; while they are very relevant, we note that they are concurrent works to ours. We have included them in the concurrent work discussion in our revised manuscript on lines 127-135 and a detailed discussion in Appendix B lines 825-829. Specifically:

• Similarity: The methodologies of both MARINE and these concurrent works [1-3] can be formulated as approaches using Classifier-Free Guidance (CFG) aimed at improving the factuality of LVLMs.
• Differences: These concurrent approaches differ in the focus of their method designs within the larger framework of CFG. Specifically:
  • [1] addresses the issue of “Text Inertia”, where LVLMs produce the same hallucinated object descriptions regardless of whether an image is actually provided.
  • [2] focuses on fine-grained guidance and especially operates at the sub-image level.
  • [3] adjusts the attention weights of the visual tokens within the attention mechanism to control the focus of the generation on the highlighted parts.

We highlight that MARINE operates at the image level and uniquely employs a vision toolbox to ensemble information from multiple vision models, producing guidance that reaches a consensus among the models for more accurate results.

Lastly, we emphasize that these approaches are complementary to each other, and integrating them could pave the way for a more effective multimodal prompt-following strategy in future research.

[1] Paying more attention to image: A training-free method for alleviating hallucination in lvlms

[2] Contrastive region guidance: Improving grounding in vision-language models without training

[3] Prompt Highlighter: Interactive Control for Multi-Modal LLMs

Thank you for the detailed feedback. Please find our responses below. We hope that our clarifications and additional experiments resolve the concerns and any potential misunderstandings.

Q1: Explore the reason why MARINE-Truth’s recall score are significantly lower than MARINE

Thank you for highlighting the numerical errors in Tables 1 and 2; we have corrected them in the revised paper. Specifically, we updated the recall values for LLaVA and MiniGPT-4v in Table 1 and fixed the swapped F1 and Yes Ratio values in Table 2. With these corrections, MARINE-Truth generally achieves better performance, aligning with our intuition and the statement in Line 361.

Q2: MARINE’s performance depends on the quality of the external image-grounded models.

Regarding the question, we’d like to first emphasize that the performance of the current LVLMs significantly depends on the quality of the vision models integrated into their architectures. This dependency motivated us to propose MARINE, which introduces additional vision guidance to enhance the LVLM's ability to accurately recognize visual information. Our approach thus involves leveraging strong external guidance by ensembling information from multiple advanced vision models specialized in information extraction. The commonly used CLIP model in LVLM, on the other hand, focuses on vision-language alignment.

The importance of the external vision model’s quality is indeed significant, as confirmed by our extended ablation studies (Figures 10, 12, and 13). By examining the impact of noise intensity in object grounding features, we demonstrate the critical role of precise object grounding information and illustrate how noise can degrade performance.

This underscores:

1. Why our method works: These vision models are highly specialized and outperform CLIP in object detection, providing precise object grounding information that helps mitigate hallucinations.
2. Why we adopt ensembling: By aggregating visual information from multiple image-grounded models, we can reduce the impact of noisy or inaccurate guidance, thereby improving overall robustness.

Q3: Using a single model leads to a significant performance drop, even worse than greedy decoding

We would like to first point out that using a single model effectively reduces hallucinations in most cases. In our ablation study of Table 6, for example, MARINE-RAM successfully reduces $C_S$ and $C_I$ for LLaVA-v1.5, achieving the second-best performance while MARINE achieves the best. Although using a single model may slightly underperform compared to greedy decoding in some instances, this inconsistent improvement is also observed in many of the compared baselines for certain models and metrics. This motivates our approach of ensemble multiple vision models to reach a consensus in guidance, thereby achieving the best performance.

This finding was aimed to highlight the following two key points:

1. The importance of ensembling in MARINE: Single models lack the robustness provided by ensembling, making them more susceptible to inaccurate or misleading guidance. On the other hand, ensembling provides a more precise object grounding in guidance.
2. The limitation of the early LLaVA model: The relatively weak robustness of this early version makes it more vulnerable to noisy guidance, leading to degraded performance.

To gain a deeper understanding of this phenomenon, we conducted an extended ablation study on yes/no questions using POPE metrics. The results below include subscripts indicating performance gains compared to greedy decoding:

Note: The ideal Yes Ratio for a non-biased LVLM is 50%.

These results demonstrate that while single models can provide valuable visual grounding information, our ensemble method is more robust and yields stronger overall improvements.

We're grateful for your strong support and helpful suggestions on our work, for which we have accordingly made modifications to our manuscript. Please find our detailed response below for the questions raised in the review.

W1 Minor errors in writing.

Thank you for pointing out these minor errors in our writing. We have corrected them in line 236 of the revised version of the manuscript.

W2: Discussions on the difference with direct prompting methods

Thank you for the suggestion. We conducted additional experiments to quantitatively compare with direct prompting and present the results in the table below:

Specifically, we evaluated:

• Direct Prompting: The original input query was replaced with the following prompt: Describe the visible contents of this image in as much detail as possible without adding any information not clearly visible. Only mention objects, colors, shapes, and textures that can be directly observed in the image, avoiding assumptions about materials, functions, or contexts. If there are any uncertainties about what an object is, describe its visual characteristics (e.g., 'a circular object with a smooth surface') without inferring its purpose or identity. Avoid creative or hypothetical descriptions, and focus on observable details only.
• Prompts as Additional Guidance: We incorporated the prompt as supplemental context to guide the models in generating outputs. The results demonstrate that incorporating the prompt can effectively enhance the recall performance for some models (e.g., LLaVA-v1.5 improves from 41.1 to 52.3 in recall). However, it did not consistently reduce hallucinations across all metrics and the performance on CHAIR scores (e.g., $C_s$, $C_i$) dropped. Meanwhile, MARINE significantly outperforms the prompting baseline approaches on CHAIR.

We note the following differences between our method and prompting method:

• Prompting method depends heavily on the instruction-following ability of the model. While it might mitigate the hallucination to a mild extent for strong models (e.g., LLaVA-v1.5), it may cause a weak model to hallucinate even more (e.g., LLaVA). Models also require more sophisticated fine-tuning approaches to generate better and more precise responses conditioned on the prompts, as discussed in [1]. In contrast, our method directly addresses deficiencies in the model’s vision capabilities by introducing stronger vision guidance. This makes our approach more effective even for weaker models and more cost-efficient.
• Unlike prompting methods, which need to be tailored to specific tasks or datasets, our method generalizes effectively across models and datasets, reducing hallucinations while maintaining competitive recall.

[1] Enhancing Large Vision Language Models with Self-Training on Image Comprehension

W1b: This closed vocabulary nature of detectors is different from the LLM.

Thank you for highlighting this point. For clarification, we indeed aimed to address the detector’s fixed set of vocabularies by incorporating RAM++ [1], a state-of-the-art open-set image tagging model to effectively mitigate object hallucinations in more general scenarios. The MARINE-RAM variant (Table 6) and MARINE-union variant (Table 7) demonstrate the effectiveness of open-vocabulary approaches. While these methods reduce hallucinations, the best performance is still achieved with the ensemble that includes DETR. However, MARINE’s flexibility allows for seamless integration of advances in open-vocabulary detection models and holds promise for further improvement through ensembling multiple open-vocabulary models.

For closed-vocabulary detectors, we emphasize that their closed vocabulary set does not constrain the LVLM’s free-form responses. Instead, these detectors help reduce hallucination during LVLM decoding by providing more grounded guidance within the predefined set. The set, moreover, sufficiently encompasses frequent objects commonly observed in natural images in many datasets other than MSCOCO. This generalizability is supported by the significant improvement observed in POPE results on the A-OKVQA [2] shown in Table 15, which demonstrates the ability of our method to generalize beyond MSCOCO.

[1] Open-Set Image Tagging with Multi-Grained Text Supervision

[2] A-OKVQA: A Benchmark for Visual Question Answering using World Knowledge

W2: Comparing with related specialized non-LLM models in image captioning.

Thank you for your suggestion. We extended our captioning quality evaluation (Table 16) to include the Generative Image-to-text Transformer (GIT) [1], a highly specialized non-LLM-based model (~0.8B) pretrained on 4M images and fine-tuned on the MSCOCO dataset (GIT-Large-COCO [2]), as a baseline. GIT notably outperforms many traditional image captioning methods across multiple benchmarks and, at the time, achieved state-of-the-art results with a significant margin.

The results are as follows and also in Table 20 Appendix D.1.3 of our updated revision:

We make the following key notes:

1. While GIT-Large-COCO outperforms the older LLaVA model as expected given its specialized training, newer LVLMs like mPLUG-Owl2 achieve better performance (21.70 vs 19.94 average score) even compared to specialized captioning models. This demonstrates the rapid advancement of LVLMs (based on powerful LLM backbone Llama) in exceeding task-specific models.
2. As we previously pointed out in the paper (line 414-416), image captioning evaluation serves to demonstrate that our method, while primarily targeting hallucination mitigation, maintains LVLMs' performance on broader tasks without significant trade-offs. For the LLaVA model, our method even further improves LLaVA's performance across all metrics.
3. For LLaVA, MARINE improves performance across all metrics. For mPLUG-Owl2 with MARINE, performance remains largely stable and continues to outperform the specialized model.

The competitive performance of LVLMs, combined with their flexibility across multiple tasks, supports our focus on enhancing these models rather than specialized ones. Lastly, we fully agree that including specialized methods could better contextualize the results, and have included them in Table 20 of our revision.

[1] GIT: A Generative Image-to-text Transformer for Vision and Language

[2] https://huggingface.co/microsoft/git-large-coco

Thank you for the constructive feedback. Please find our responses below. We hope that our clarifications and additional experiments resolve the concerns.

W1: Further discussion on contributions.

Thank you for your constructive suggestions. In response to the specified questions, we a) added experiments introducing a component that dynamically adjusts $\gamma$ based on confidence scores and b) clarified that the inclusion of RAM++ in our method does provide a flexible open-set vocabulary. We begin with clarification of our contributions. In the following W1a and W1b, we provide detailed results and explanations.

Clarification of our contributions

The novelty of MARINE lies in its training-free and API-free guidance framework with an ensemble mechanism that integrates any open-source vision models. This offers an efficient alternative to training LVLMs with different vision backbones from scratch, saving resources while enabling the ensemble of specialized vision models without excessive memory and computation costs. Recent works [1-4] have also explored training-free approaches to guide LVLMs, highlighting the importance of effective training-free methods. In lines 825-829 of our revision, we also provided detailed discussion on the concurrent training-free methods.

[1] Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding

[2] Paying more attention to image: A training-free method for alleviating hallucination in lvlms

[3] Contrastive region guidance: Improving grounding in vision-language models without training

[4] Prompt Highlighter: Interactive Control for Multi-Modal LLMs

W1a: Dynamically tuning the guidance strength ($\gamma$)

Thank you for your insightful suggestion. In response, we conducted additional experiments on dynamic guidance strength based on confidence scores, evaluating both CHAIR and POPE metrics, as shown in the tables below and included as Table 21 and 22 in appendix D.2 of our updated revision.

Specifically:

• Fix Guidance Strength uses a fixed guidance strength of 0.7, selected to balance hallucination reduction and instructions adherence.
• Dynamic Guidance Strength adjusts the guidance strength dynamically by mapping the mean confidence score ($s$) of the image-grounding models to a range of (0.4, 0.8) using the formula $\gamma' = 0.4 + \frac{(0.8 - 0.4)\cdot (s - s_{\text{min}})}{s_{\text{max}} - s_{\text{min}}}.$ A higher confidence score indicates more reliable guidance and corresponds to a stronger guidance strength.

The results show that dynamic guidance improves performance for the weaker model LLaVA, which is more susceptible to noisy guidance. For stronger models like mPLUG-Owl2, a fixed guidance strength sufficiently mitigates object hallucinations with a robust performance. Lastly, we consider setting the guidance strength as a tunable parameter a promising direction for future work.

Addressing specific follow-up comments

Q1: The method's effectiveness appears inversely correlated with model capability.

We respectfully disagree. We respectfully disagree. For example, using the POPE metric, MARINE significantly improves the performance of mPLUG-Owl2:

• Accuracy: 72.5% → 82.8%
• F1 score: 77.5% → 82.7%
• Yes ratio: Improved from 72.4% → 49.2%, aligning closely with the ideal ratio of 50%.

Q2. For LLaVA, while improvements are shown, the performance gap is substantial compared to specialized image captioning models.

As emphasized earlier, MARINE’s focus is not on image captioning but on reducing object hallucinations in LVLMs. Image captioning evaluations are provided to demonstrate MARINE's broad applicability without trade-offs, not to claim superiority over specialized models.

Q3. With stronger models (e.g., mPLUG-Owl2), more sophisticated approaches like dynamic guidance weight did not yield significant improvements.

Dynamic guidance weight is an optional enhancement, not the core of MARINE. The modest improvement compared to the vanilla MARINE approach highlights the robustness and effectiveness of our original method, which performs strongly even without additional complexity. For your clearer reference, we added the original performance of mPLUG-Owl2 below, with subscripts indicating the performance gain compared to greedy decoding:

We hope these clarifications address your concerns. Please reach out if further explanations are needed. We appreciate your efforts in reviewing our work and look forward to continued discussion.

[1] Analyzing and mitigating object hallucination in large vision-language models.

[2] Woodpecker: Hallucination correction for multimodal large language models.

[3] Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding.

Thank you very much for your prompt feedback and for acknowledging that our rebuttal has addressed several of your concerns. We appreciate your time in this discussion, and would like to respectfully address some significant misunderstandings regarding the scope and contribution of our work. Below, we provide detailed clarifications:

Scope and contribution of MARINE

In response to the limitation you mentioned, we must clarify that

• Primary focus: The focus of our method MARINE is to reduce object hallucinations of LVLMs for their diverse responses to a wide range of tasks, not to outperform specialized image captioning models on captioning. This focus is emphasized in lines 39-53 and further discussed in lines 112-135. Object hallucination reduction is a critical area of research in the LVLM domain, as demonstrated by numerous recent studies.
• Baselines (recent works): As outlined in our work, the appropriate comparison for MARINE is against the recent methods targeting hallucination reduction in LVLMs such as LURE [1], Woodpecker [2] and VCD [3]. Comparing MARINE to specialized image captioning models, which address fundamentally different tasks, misrepresents the scope and contribution of our work.
• State-of-the-art performance: As shown in both Table 1 and 2 of our manuscript, MARINE consistently achieves SOTA performance across multiple metrics when compared to other methods designed to mitigate hallucinations in LVLMs.

Clarifications on image captioning experiments (Bleu, ROUGE-L, CIDEr)

The evaluation of image captioning in our paper is not the focus of MARINE but serves a complementary purpose. These additional experiments demonstrate that MARINE can reduce hallucinations while maintaining overall task performance across diverse applications. This is consistent with related works [1-3], which also focus on hallucination reduction rather than specialized image captioning.

We have similarly pointed this out in our previous rebuttal:

“As we previously pointed out in the paper (line 414-416), image captioning evaluation serves to demonstrate that our method, while primarily targeting hallucination mitigation, maintains LVLMs' performance on broader tasks without significant trade-offs.”

“The competitive performance of LVLMs, combined with their flexibility across multiple tasks, supports our focus on enhancing these models rather than specialized ones.”

In addition, as demonstrated in Figure 4, 7, 9 and 13, our paper includes question examples of our targeted evaluation that are fundamentally different from image captioning tasks, such as:

• (Figure 4) Is there a chair in the image?
• (Figure 4) What is the position of the skateboard in the image?
• (Figure 7) What might be the purpose of this table arrangement?
• (Figure 9) How might the art installation encourage interaction among people?
• (Figure 13) What precautions has the little girl taken to stay dry in the rain?

While the question format of one metric CHAIR may resemble image captioning in format, they specifically assess the degree of object hallucination in LVLMs as compared the captioning metrics like CIDEr. The objective is not to have models generate captions that mimic ground truth but to reduce hallucinations in LVLMs across diverse tasks and response types.

Thank you very much for your support and the constructive feedback that helped us improve our work. Please see our detailed response with additional experiments below.

W1. Compare to a baseline using carefully engineered prompts.

Thank you for your insightful suggestion to compare our approach against a baseline using carefully engineered prompts aimed at reducing hallucination. Based on your recommendation, we conducted additional experiments and present the results in the table below and Table 18 of our revision:

Specifically, we evaluated:

• Direct Prompting: The original input query was replaced with the prompts as described.
• Prompts as Additional Guidance: We incorporated the prompt as supplemental context to guide the models in generating outputs. The results demonstrate that incorporating the prompt can effectively enhance the recall performance for some models (e.g., LLaVA-v1.5 improves from 41.1 to 52.3 in recall). However, it did not consistently reduce hallucinations across all metrics and the performance on CHAIR scores (e.g., $C_s$, $C_i$) dropped. Meanwhile, MARINE significantly outperforms the prompting baseline approaches on CHAIR.

We note the following differences between our method and prompting method:

• The Prompting method depends heavily on the instruction-following ability of the model. While it might mitigate the hallucination to a mild extent for strong models (e.g., LLaVA-v1.5), it may cause a weak model to hallucinate even more (e.g., LLaVA). Models also require more sophisticated fine-tuning approaches to generate better and more precise responses conditioned on the prompts, as discussed in [1]. In contrast, our method directly addresses deficiencies in the model’s vision capabilities by introducing stronger vision guidance. This makes our approach more effective even for weaker models and more cost-efficient.
• Unlike prompting methods, which need to be tailored to specific tasks or datasets, our method generalizes effectively across models and datasets, reducing hallucinations while maintaining competitive recall.

[1] Enhancing Large Vision Language Models with Self-Training on Image Comprehension