Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models

$**Concern 2:**$ A few more important benchmark comparisons are needed - MathVista, MMVet, LLaVA-Bench, MMMU.

$**Response 2:**$ Thanks for your valuable suggestions. We conduct experiments on MathVista (testmini), MMVet, LLaVA-Bench, and MMMU, which mainly focus on the general and complex reasoning abilities of LVLMs, to validate the effectiveness of SID. Due to time constraints, we do our best to compare SID with two representative decoding methods (i.e., VCD and OPERA) on LLaVA-1.5 7B.

$*$ denotes employing the greedy decoding strategy

The above table indicates that our SID improves the reasoning ability, especially in the LLaVA-Bench benchmark. However, VCD slightly degrades LVLM's complex reasoning ability, as reflected in MathVista, MMVet, and MMMU benchmarks. OPERA has little gains in terms of discrimination tasks. These results are consistent with Table 5 (MME and MMbench benchmarks), Table 3, and Table 4.

Note that our paper has been validated on various benchmarks and metrics, including POPE (Table 4), CHAIR (Table 3), GPT-4 assisted Sentence-level Hallucination Ratio (SHR) (Figure 7), 1/2-gram, the number of generated words/sentences per image (WPI/SPI) (Figure 7), GPT4-V assisted evaluation (Table 7), MME, and MMBench (Table 5) to evaluate the generated texts regarding $**hallucination**$, $**fluency**$, $**detailness**$, and various $**general**$ ability.

We add experimental results of MathVista (testmini), MMVet, LLaVA-Bench, and MMMU in Table 8 of Appendix A.5.

Line 905: APPENDIX A.5 MORE GENERAL BENCHMARKS EVALUATION...

[1] MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts, ICLR 2024.

[2] MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities, ICML 2024.

[3] Visual instruction tuning, NeurIPS 2023.

[4] Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi, CVPR 2024.

Hope these baseline comparisons and experiments could address your concerns. Please let me know if you have any concerns and we will be more than happy to answer them.

Best,

Authors

Thank you for the thoughtful concerns. We are thrilled that the reviewer considers our paper to have a clear presentation, extensive experiments and ablations, and well-supported motivations. We appreciate the opportunity to add more discussions and experiments on more baselines and benchmarks.

$**Concern 1:**$ An important baseline (Ghosh et al [1]) needs to be compared.

$**Response 1:**$ Thanks for your valuable comments. We compare with Ghosh et al [1] in Table 15 and analyze the results. Concretely, the simple but effective Visual Description Grounded Decoding (VDGD) [1] enhances visual perception and improves reasoning capabilities by first generating a detailed description of the image and appending it as a prefix to the instruction. We compare VDGD with our SID in terms of hallucination evaluation benchmarks( i.e., POPE and CHAIR (testing on the same 500 images)) and the general ability benchmark (i.e., MMbench) and add detailed analyses. We re-implement VDGD [1] (version: arXiv:2405.15683v2) based on official codes on LLaVA-1.5 7B.

• denotes employing the greedy decoding strategy

The results above demonstrate the effectiveness of VDGD [1] in alleviating hallucinations and enhancing general reasoning abilities in LVLMs. However, while the grounding visual descriptions generated by the LVLMs themselves effectively improve visual perception and reasoning capabilities, they may inevitably contain hallucinations. Consequently, VDGD is inferior in the POPE ($**adversial**$) subset, which prioritizes $**co-occurring**$ objects that are not present in the image. Meanwhile, VDGD shares somewhat similar motivations in enhancing vision information via Equation (6) as we analyzed. The experiments in Table 2 are consistent with the above results, indicating that boosting the vision information is effective in mitigating hallucinations but is less effective in adversarial environments.

We add experimental results and analysis of VDGD in Table 15 of Appendix A.7, marked in blue, as follows:

Line 1086: Visual Enhancing Decoding Strategy. Although LVLMs can accurately recognize visual elements,....

[1] Visual Description Grounding Reduces Hallucinations and Boosts Reasoning in LVLMs, Ghosh et al., 2024, arXiv:2405.15683v2.

Thank you for clarifying all my queries. I have increased my score from 5 to 6.

Thank you for the insightful reviews! We are thrilled that our paper is considered by the reviewer as having the novel and interesting methodology, valuable insights, comprehensive experiments and low computational cost. We appreciate the opportunity to clarify details of our motivations, statements, and experimental results. Here, we answer the comments point by point.

$**Concern 1:**$ Comparisons and Discussion with AVISC [1].

$**Response 1:**$ Thank you for the constructive suggestion. We want to clarify in terms of motivation, robustness, and experiments.

$**Motivation**$: AVISC [1], in contrast to ours, preserves outlier high attention tokens (named 'blind token') and substracts output logits to counteract the over-emphasis of 'blind token'. In this way, AVISC promotes balanced consideration of all tokens to alleviate hallucinations [1]. Differently, SID preserves spurious related tokens to amplify fine-grained hallucinations and mitigate them.

$**Robustness**$: Similar to 'artifact token' [2], 'blind token' aggregates global information and maintains global discrimination ability. Appendix Table 9 shows that the Top-100 vision tokens with high attention scores can largely maintain the original performance. Therefore, 'blind token' tends to have $**high probability**$ of target class logits, and contrastive decoding does not improve the target class's probability while might bring extra noise. AVISC heavily $**relies on**$ the adaptive plausibility constraint (Eq. 3) and often degrades the original greedy decoding performance, similar to VCD and ICD, as indicated by the experiments below.

$**Experiments**$: Due to time constraints, we compare AVISC on both greedy and sampling settings in POPE and CHAIR benchmarks with LLaVA-1.5 7B. Although AVISC outperforms vanilla sampling decoding under adaptive plausibility constraint Eq. (3), it still degrades the greedy decoding to some extent, which indicates that the attentional vision calibration (AVISC) strategy induces some annoying noise.

$*$ and ^ denote the greedy and sampling decoding strategies

We add the analyses and experimental results in Table 13 of Appendix A.7, marked in blue, as follows:

Line 1050: A APPENDIX A.7 Token Selection Strategies Analyses. ...Moreover, AVISC (Woo et al., 2024)...

[1] Don't Miss the Forest for the Trees: Attentional Vision Calibration for Large Vision Language Models. arXiv preprint arXiv:2405.17820, 2024.

[2] Vision Transformers Need Registers. ICLR 2024.

$**Concern 2:**$ Clarify Figure 6 and statement.

$**Response 2:**$ Thanks for your careful concerns.

1. Figure 6 illustrates analyses of varying i and preserved ratios of vision tokens. Note that when pruning all vision tokens, the remaining tokens (instructions) still pass through subsequent decoding layers. The multimodal knowledge absorbed in the early decoder layer induces contextual hallucinations during subsequent decoding.
2. The ideal induced hallucination distributions should be target-co-occurring while suppressing target logits. If we prune all vision tokens, the absorbed visual information diminishes during subsequent decoding. An extreme case is VIG [1], which directly ablates vision input. The induced distributions of VIG are noisy (i.e., aimless hallucinations) due to insufficient vision grounding. Therefore, we preserve vision tokens with low attention values to induce target-co-occurring hallucinations while suppressing target logits.

[1] Multi-modal hallucination control by visual information grounding. CVPR 2024.

We explain and rephrase the statement, marked in blue, as followings:

Line 329: ...aggressive pruning of vision tokens (i.e., 0%) after layer 𝑖 =1 is not optimal. As the ideal induced hallucination distributions are target-co-occurring but suppress target logits, the loss of visual information for subsequent decoding results in visual context diminishing, which can lead to aimless hallucinations due to insufficient grounding in visual information...

$**Concern 3:**$ Standard deviations of MME and MMbench.

$**Response 3:**$ We add the standard deviations using three different seeds (i.e., 41, 42, 43) in the updated paper. Please note that we use the $*greedy*$ decoding strategy, and the standard deviations are small. Our goal is to mitigate the hallucination issue, and results on MME and MMBench benchmarks demonstrate SID greatly $*preserves*$ and $*improves*$ LVLMs' various general abilities.

Table5: LVLM benchmark evaluations. DoLa, ICD, VCD,and SID employ the same greedy decoding.

$**Concern 4:**$ Lack interpretability of token selections and InstructBLIP, compared to HALC [1].

$**Response 4:**$ Thanks for your insightful comments. HALC [1] provides a more intuitive and interpretable approach by locating fine-grained tokens within the original image for each generated token. However, it is essential to clarify that HALC relies on $*extra*$ pre-trained networks like Grounding DINO [2] / OWLv2 [3] to identify vision regions and BLIP [4] for vision matching, while our method aims to avoid utilizing auxiliary analysis networks and instead focuses on introspectively eliminating hallucinations. In Sec. 7 Future Work, we consider training the external network to automatically determine optimal hyperparameters and also consider enhancing the interpretability by leveraging extra analysis networks. The discussions are marked in blue as follows:

Line 533: ...In addition, to enhance the interpretability of hallucination alleviations, we consider resorting to pre-trained analysis networks to intuitively locate spurious related vision regions....

[1] HALC: Object Hallucination Reduction via Adaptive Focal-Contrast Decoding. ICML 2024.

[2] Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ECCV 2024.

[3] Scaling open-vocabulary object detection. NeurIPS 2023.

[4] Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. ICML 20222

$**Concern 5:**$ Minor issues regarding inference time of Table 6, wrong citation format in Line 910, and concerns of deviations in Table 1.

$**Response 5:**$ Thanks for your careful concerns!

1. We calculate the inference time of the entire benchmark to avoid bias. We add the explanations in the revised paper.

2. We revise the wrong citations in the revised paper.

3. We directly adopt results of VCD from the original paper in the sampling setting and re-implement VCD in the greedy setting, which has small deviations; We conduct experiments using three random seeds (i.e., 41, 42, 43).

$**Concern 6:**$ Does the adaptive plausibility constraint affect CHAIR and MME?

$**Response 6:**$ Yes, the adaptive plausibility constraint also greatly affects CHAIR and MME. Due to space constraints, we do not show the results of CHAIR and MME in Table 1. In the updated paper, we add the experimental results of CHAIR and MME in the Appendix Table 11.

All methods adopt the sampling decoding strategy, and the results are the average of three running times with three random seeds.

We add the analysis of the adaptive plausibility constraint in Table 11 of Appendix A.7.

Line 990: APPENDIXA.7 More Analyses on Adaptive Plausibility Constraint...

$**Concern 7:**$ How to define vision-and-text association hallucination?

$**Response 7:**$ 'Vision-and-text association' means 'vision-and-text aware', as opposed to being 'vision-and-text agnostic' (i.e., VCD and ICD). Vision and text are both fed to the LLM decoder, enabling multimodal information interactions. We focus on amplifying the vision-and-text association (or vision-and-text awareness) hallucinations informed by low-attention vision tokens.

$**Concern 8:**$ The specific experimental setup of ID in Figure 5.

$**Response 8:**$ ID utilizes the negative prompt 'You are a confused image caption model.' following [1,2] as shown in the $*upper right*$ of Figure 5. All other parameters are set to their default values. For the open-end generation task, LVLMs tend to follow the negative prompt, leading them to refuse to respond.

[1] Mitigating Hallucinations in Large Vision-Language Models with Instruction Contrastive Decoding, ACL 2024 Findings.

[2] Instructive Decoding: Instruction-tuned Large Language Models are Self-refiner from Noisy Instructions, ICLR 2024.

$**Concern 9:**$ How do the performances on the subsets (i.e., Existence, Count, Position, and Color) of the MME benchmark?

$**Response 9:**$ Thanks for your insightful suggestion. Existence, Count, Position, and Color subsets of the MME benchmark cover both object-level and attribute-level hallucinations. Experimental results are illustrated below:

*denotes the greedy decoding strategies

We add the results of Existence, Count, Position, and Color subsets of the MME benchmark in Table 9 of Appendix A.5.

Line 945: APPENDIX A.5...We also illustrate experimental results of Existence...

$**Concern 10:**$ Explaining the interesting experimental results in Lines 346-362.

$**Response 10:**$ Thanks for appreciating our analysis! In the $*random*$ setting, enhancing vision information typically boosts the target logits of the original prediction. However, in the $*adversarial*$ setting, which is more challenging due to the emphasis on prioritizing co-occurring confusing objects, the enhanced vision information $*might not*$ boost target logits but instead boost the logits of co-occurring objects. It is $*easier*$ to prune important vision tokens to amplify the distribution of co-occurring objects and then subtract them.

Hope these explanations could address your concerns. Please let me know if you have any concerns and we will be more than happy to answer them.

Best,

Authors

Thank you for the authors' detailed follow-up responses. Please include these new results and discussions into the revised paper. I will maintain my positive rating and look forward to seeing this work accepted. Good luck!

$**Concern 5:**$ How does the performance of the proposed approach vary across different LVLM sizes?

$**Response 5:**$ Thanks for the thoughtful concern! We want to clarify that the SID has been validated across various scales ($**7B**$, $**8B**$, and $**13B**$) in Tables 4 and 14. Concretely, Table 4 illustrates the performance on LLaVA-NeXT, which utilizes LLaMA3 8B. Additionally, Table 14 (previously Table 10) in Appendix A.7 $**Larger-scale**$ $**LVLM**$ $**Backbone**$ shows the performance on LLaVA-1.5 13B and InstructBLIP 13B.

$**Concern 6:**$ Where does the proposed approach fail to mitigate hallucinations?

$**Response 6:**$ Thanks for raising this concern. Since our approach is a training-free decoding method that does not rely on auxiliary analysis networks, it inherently carries over the existing weaknesses of LVLMs. Intuitive case studies, as illustrated in Appendix Figures 13, 14, and 15, reveal that SID still generates some hallucinations, particularly in finer details such as eye color and vehicle identification specifics.

These issues may arise from the relatively limited visual perception capabilities of the vision encoder. For future work, it is promising to integrate SID with InternVL [1], which scales the vision encoder up to 6B, or consider leveraging auxiliary analysis networks like Grounding DINO [2] or OWLv2 [3] to address the internal weaknesses of LVLMs.

We include failure analysis in Appendix A.6 CASE STUDY, marked in blue:

Line 955 APPENDIX A.6 CASE STUDY. As we propose...

[1] InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks, CVPR 2024.

[2] Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ECCV 2024.

[3] Scaling open-vocabulary object detection. NeurIPS 2023.

Hope these explanations and experimental results could address your concerns. Please let me know if you have any concerns and we will be more than happy to answer them.

Best,

Authors

$**Concern 2:**$ The paper does not cover a comprehensive set of baselines - Woodpecker, LRV, LURE.

$**Response 2:**$ Thank you for the attentive concern. It’s essential to clarify SID is a $**training-free**$ plug-and-play decoding strategy that does not requires $**robust**$ $**instruction**$ $**tuning**$ with $**curated datasets**$ (as in LRV and HA-DPO) and $**post-hoc**$ utilizing $**auxiliary analysis**$ $**networks**$ (as in Woodpecker and LURE). This distinction is highlighted in bold in Section 2, "Related Work" (Lines 115-119: describing three different types). Therefore, it might be $**unfair**$ to directly compare SID with Woodpecker, LURE, LRV, and HA-DPO.

To further validate the effectiveness of SID, we compare SID with auxiliary analysis networks-based method (i.e., LURE [1]) and the robust instruction tuning-based method (i.e., HA-DPO [3]). We also apply the plug-and-play SID to LURE and HA-DPO on the POPE, CHAIR, and MME benchmarks following their official implementations.

Note that all methods adopt the greedy decoding strategy. LURE and HA-DPO adopt MiniGPT-4 13B and LLaVA-1.5 7B, respectively.

Experimental results indicate that SID outperforms LURE and HA-DPO in most cases. Additionally, seamlessly integrating SID with LURE and HA-DPO can significantly enhance performance.

We add experimental results of more baselines in Table 10 of Appendix A.7, as follows:

Line 905 APPENDIX A.5 More Analyses on Other Baselines...

[1] Analyzing and Mitigating Object Hallucination in Large Vision-language Models, ICLR 2024.

[2] Aligning Large Multi-Modal Model with Robust Instruction Tuning, ICLR 2024.

[3] Beyond Hallucinations: Enhancing Lvlms through Hallucination-aware Direct Preference Optimization, arXiv:2311.16839, 2024.

$**Concern 3:**$ The approach aims to tackle hallucinations in LVLMs as a whole but does not mention or tackle style hallucinations or biases introduced when LVLMs are instruction-tuned.

$**Response 3:**$ Thank you for the thoughtful concern!

Our method proposes the $*training-free*$ and $*plug-and-play*$ decoding strategy to mitigate hallucinations. The self-introspective mechanism automatically amplifies the contextual hallucinations and then alleviates them.

Although employing auxiliary analysis networks or instruction tuning with curated datasets can explicitly diagnose hallucination styles and biases, self-introspective decoding is free from external networks and instruction tuning and can also be seamlessly integrated with the above methods.

In the updated paper, we add the detailed quantitative results and analysis of fine-grained hallucination types, specifically focusing on object-level (existence and count) and attribute-level (position and color) hallucination subsets in Appendix A.5 Table 9.

$**Concern 4:**$ Two typos.

$**Response 4:**$ Thank you for pointing typos out! We update the figure and revise the typos.

Thank you for the thoughtful reviews. We appreciate the opportunity to clarify some misunderstandings and add more discussions and benchmarks evaluation. Here, we answer the concerns point by point.

$**Concern 1:**$ Different benchmarks to show the preservation of LVLM ability.

$**Response 1:**$ Thank you for the valuable comments. Kindly note that in addition to hallucination-related benchmarks and metrics (POPE, CHAIR, SHR, GPT4-V assisted correctness evaluation), our paper had been validated on $*various*$ benchmarks and metrics including 1/2-gram, the number of generated words/sentences per image (WPI/SPI) (in Figure 7), GPT4-V assisted detailedness evaluation (in Table 7), $**MME**$, and $**MMBench**$ (in Table 5) to evaluate the generated texts regarding $**fluency**$, $**detailness**$, and LVLMs' various $**general**$ abilities. MME comprises $**ten**$ sub-tasks to evaluate model’s perceptual capabilities and $**four**$ sub-tasks for assessing cognitive abilities. MMBench systematically evaluates $**twenty**$ ability dimensions of LVLMs. Experiments presented in $**Table 5**$ on MME and MMBench illustrate SID can maintain and improve various multimodal general abilities.

To further analyze the abilities of LVLMs from different perspectives, we conduct experiments on MathVista (testmini), MMVet, LLaVA-Bench, and MMMU, which mainly focus on the general and complex reasoning abilities of LVLMs, to validate the effectiveness of SID. Due to time constraints, we do our best to compare SID with two representative decoding methods (i.e., VCD and OPERA) on LLaVA-1.5 7B.

$*$ denotes employing the greedy decoding strategy

This table indicates that our SID enhances reason abilities, particularly in the LLaVA-Bench benchmark. However, VCD slightly degrades LVLM's complex reasoning ability, as reflected in MathVista, MMVet, and MMMU benchmarks. OPERA has little gains in discrimination tasks. These results are consistent with Table 5 (MME and MMbench benchmarks), Table 3, and Table 4.

We add the experimental results for these benchmarks in Table 8 of Appendix A.5, highlighted in blue, as follows:

Line 905: APPENDIX A.5 MORE GENERAL BENCHMARKS EVALUATION...

[1] MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts, ICLR 2024.

[2] MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities, ICML 2024.

[3] Visual instruction tuning, NeurIPS 2023.

[4] Mmmu: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert Agi, CVPR 2024.

Thank you for the insightful reviews! We appreciate the opportunity to add more explanations and discussions of the proposed method. Below, we address the concerns point by point.

$**Concern 1:**$ The proposed SID is performed on pre-trained LVLMs, so it’s possible that the performance of SID is limited by those pre-trained models.

$**Response 1:**$ Thanks for your insightful concern. Since our method is a training-free, plug-and-play decoding approach, it does indeed inherit the existing limitations of LVLMs. As demonstrated in the case studies in Appendix Figures 13, 14, and 15, SID still generates some hallucinations, particularly in finer details such as eye color and vehicle identification specifics. These issues may stem from the relatively limited visual perception capabilities of the vision encoder. For future work, it is promising to integrate SID with InternVL [1], which scales the vision encoder up to 6B, or consider leveraging auxiliary analysis networks like Grounding DINO [2] or OWLv2 [3] to mitigate LVLMs' internal weaknesses.

We add an analysis of SID faults, which inherit weaknesses from the pre-trained LVLM, in Appendix A.6 under CASE STUDY. These sections are highlighted in blue for clarity.

[1] InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks, CVPR 2024.

[2] Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ECCV 2024.

[3] Scaling open-vocabulary object detection. NeurIPS 2023.

$**Concern 2:**$ Currently, there’s no solution to specifically tune SID for different LVLMs.

$**Response 2:**$ Thank you for the careful concern. Since SID is designed to be parameter-free, we have not proposed tuning-based methods. However, in $**Section 7**$, we outline two potential future directions: $**1)**$ As the pruning ratios and layer are set manually, we consider training the external network to automatically determine optimal hyperparameters, inspired by [1]. In addition, to enhance the interpretability of hallucination alleviations, we consider resorting to pre-trained analysis networks to intuitively locate spurious related vision regions. $**2)**$ Moreover, given that SID amplifies fine-grained hallucinations, we consider leveraging the CT$^2$S strategy to automatically construct high-quality negative instruction for robust visual instruction tuning rather than relying on expensive GPT-4 [2,3]. Note that the self-generated hallucination dataset ensures $*style consistency*$, which is crucial for preference learning [3].

[1] Diffrate: Differentiable Compression Rate for Efficient Vision Transformers, CVPR 2024.

[2] Mitigating Hallucination in Large Multi-Modal Models via Robust Instruction Tuning, ICLR 2024.

[3] Beyond Hallucinations: Enhancing Lvlms through Hallucination-aware Direct Preference Optimization, arXiv:2311.16839, 2024

$**Concern 3:**$ Explain the results of greedy and sampling decoding in Table 1.

$**Response 3:**$ Thanks for your thoughtful comment. Greedy decoding consistently outperforms sampling decoding regarding hallucination metrics. $*However*$, greedy decoding generates repetitive text and often only reaches a local optimum [1,2]. In contrast, sampling decoding generates diverse and coherent output texts [2,3]. As a result, most open-source and closed-source LVLMs utilize the sampling decoding by default to facilitate diverse and coherent chat interactions. It is therefore practically meaningful to consider the constraints in Eq. 3 when analyzing sampling decoding.

More detailed introductions of decoding strategies are in APPENDIX A.1 MORE BACKGROUNDS Decoding Strategy in LLMs. We add further explanations in the revised paper in APPENDIX A.1, highlighted in blue.

[1] Neural text generation with unlikelihood training, ICLR 2020.

[2] On decoding strategies for neural text generators, Transactions of the Association for Computational Linguistics, 2022.

[3] The curious case of neural text degeneration, ICLR,2020.