Unveiling the Tapestry of Consistency in Large Vision-Language Models (2024)

1 Introduction

Recently, benefiting from notable advancements in large language models (LLMs)[1; 25; 2], the realm of large vision-language models (LVLMs) has undergone a revolutionary transformation. These novel LVLMs [18; 24; 3; 8; 15; 13] try to combine visual signals with textual semantics and spark cognitive brilliance across modalities.Although LVLMs can generate high-quality responses to task prompts, we discover that for correctly answered cases, simply modifying the prompt will result LVLMs in providing contradictory responses. In Figure 1 (a.2), LLaVA-7B [18] properly describes the picture as “It is a man wearing a dinosaur costume.”, but when prompted “Is the dinosaur played by humans? Please answer yes or no.”, it responds with “No, they are dinosaurs”. The above phenomenon of Inconsistency is widely observed across mainstream LVLMs, and a preliminary study was conducted only on LLMs [14]. In practice, in contrast to the fixed patterns of questions, designed in existing multimodal benchmarks, the users tend to pose questions in arbitrary ways. Therefore, it is necessary to ensure the LVLMs in predicting correct and consistent answers, even when faced with various formats of queries.

However, there are currently no benchmarks or research studies that specifically focus on evaluating the Consistency of LVLMs responses. These single-prompt type evaluation approaches [12; 10; 28; 21; 6] lead to a disconnect between benchmark accuracy and real-world user practical experience.

Based on the above observations, we systematically introduce a Consistency Benchmark dubbed ConBench, to estimate the capabilities of LVLMs more thoroughly via diverse question formats. It consists of $1,000$ public pictures, and each was manually selected from four multimodal benchmarks [10; 12; 28; 21]. Apart from the original discriminative prompt, we constructed two additional discriminative types of questions¹¹1e.g., Multiple-choice questions and limited VQA questions are generated for MME benchmark. by ChatGPT/GPT-4 [1]. Notably, three types of questions of each case are around the same knowledge point. Besides, every set is accompanied by a generative question without ground truth. Consequently, ConBench serves as an evaluation tool that observes the Consistency performance of LVLMs and surpasses the limitations of previous assessments.

Furthermore, grounded on the ConBench, we conduct an in-depth analysis and visualization of Consistency on $14$ popular LVLMs. In a nutshell, the conclusions of noteworthy insight are threefold:

C1 In the discriminative question-answering (QA) domain: (1) A decrease in LVLMs accuracy as the prompt’s solution space increases. (2) Instances of erroneous yet consistent answers are scarce.

C2 Extended to the generative domain, we establish a connection between discriminative and generative domains by the perspective of Consistency. (1) As the solution space of discriminative questions expands, the Consistency between its answer and caption grows stronger. (2) The accuracy of discriminative answer and its Consistency with the caption exhibit a positive correlation.

C3 Closed-source models exhibit a pronounced bias advantage in terms of Consistency, compared to open-source models. This provides an alternative perspective to demonstrate why closed-source models, despite sometimes having lower accuracy, offer a better user experience in practical applications.

Eventually, leveraging the insights gained from our theoretical discoveries, we enhance the caption performance of LVLMs without any additional costs associated with training. Specifically, we construct discriminative prompts based on the low-confidence words in the answers of LVLMs, forcing the LVLMs to introspect. Then, through iterative refinement in multiple rounds of question-answering, the quality of LVLMs’ captions gets an impressive achievement (e.g., our method improves the LLaVA-NeXT-34B [19] by $9.1\%$ and MiniGemini-34B [15] by $9.6\%$ on metric[C] in Sec. 4).

In summary, to the best of our knowledge, we are the first to propose a Consistency evaluation method and conduct a comprehensive analysis of Inconsistency in LVLMs. We hope this paper serves as a catalyst for further exploration, and look forward to the community applying the above findings to polish up the usability and practicality of large vision-language models.

2 Related Work

Inconsistency in LLMs

A amount of prior work has been conducted on investigating Inconsistency in LLMs. [14] is the first to find the Inconsistency phenomenon in question-answering and validator tasks and define GV-consistency. Besides, it leverages consistency pair for training to improve LLMs’ performance. While [17] utilizes Consistency to check for hallucination detection in LLMs, a logic consistency-basedmethod that involves logic-related questions and answers. Compared to LLMs, Inconsistency in LVLMs is more likely to occur due to the additional visual modality, which deserves further exploration.

3 ConBench

We propose a novel multimodal evaluation pipeline named ConBench to comprehensively assess LVLMs. The ConBench has a total of 4K questions on 1K images and corresponding 3K discriminative ground truths, guaranteeing evaluation quality in terms of the quantity and diversity of questions. In Sec. 3.1, we present the generation of ConBench and the construction pipeline for prompts. Sec. 3.2 introduces the hierarchical core capabilities and discusses the design philosophy. Sec. 3.3 and 3.4 describe the evaluation strategy for scoring various types of answers.

3.2 Hierarchical Core Capabilities

The ConBench comprises three core capabilities, arranged in ascending order of difficulty, namely: Sensation, Cognition, and Knowledge, with nineteen fine-grained dimensions shown in Figure 2.

[Easy Mode] Sensation: What you see is what you get. We assume that sensation is the most fundamental expertise of LVLMs, and it is the "eye" of the LLMs. While perceived questions appear simple and basic, they are nonetheless essential. Therefore, this capability accounts for $50\%$ of the ConBench. Count, color, optical character recognition (OCR) and scene categories focus on subtle details, while poster, attribute recognition and position types emphasize the overall picture.

[Medium Mode] Cognition: Go beyond the surface. The cognitive process needs the model to integrate visual and language modalities: observing the content of an image, combining it with the text of question, and retrieving knowledge from within the LLMs. It is more challenging than the single sensation task. This section constitutes $26\%$ of the ConBench, including numerical calculation, code inference, text translation, math, cross-instance reasoning and attribute reasoning categories.

See Also

Find Jeopardy Games About _OX\_[_X™[_[™ÊØXÝ_]š]_ILÐJٚYÝ\™JÌL‹ŒLXJÌÊÛيÌɜ_YÙOLMSommaire du brevet 3003006 - Base de données sur les brevets canadiensDUTY CYCLING CONTROL OF NEURAL DOSE IN A SPINAL CORD STIMULATOR SYSTEM - BOSTON SCIENT NEUROMODULATION CORPCompute's Third Book of Atari (1984) - [PDF Document]

[Hard Mode] Knowledge: Master the art of synthesis and integration. Mastering professional knowledge is an essential pathway for next-generation LVLMs to become Expert AGI, as it requires a higher level of understanding of images and the application of expert knowledge. We carefully selected knowledge from diverse and extensive fields, such as celebrities, chemistry, physics, biology, art and geography. This part takes up $24\%$ of the total, and functions as the upper limit of ConBench.

3.4 Multidimensional Evaluation Metric

Here we provide two evaluation metrics, each from the perspective of discriminative and generative domains, aiming to provide a more comprehensive understanding of LVLMs consistency. The former does not rely on AI tools and quickly produces Consistency results among discriminative responses via Sec. 3.3, primarily evaluating the knowledge. The latter employs GPT to indirectly assess the quality of captions, by judging the consistency between discriminative responses and captions.

Discriminative Domain Evaluation MetricWe define the ConScore[D] as that: when all three discriminative types of questions within the same case are answered correctly, the model gets one point. The maximum score is 1000 points. The final format is presented as a percentage (%).

Generative Domain Evaluation MetricDue to the high variability in captions, it is not possible to calculate Consistency based on character matching alone. Therefore, we rely on GPT/GPT4 for judgment. The judging process and the constructed prompts are shown in Figure 4. We formulate it as a machine reading comprehension task. We manually sample the judgment results, and GPT4 achieved an accuracy rate of $95\%$, which is reliable and trustworthy. Next, we define the ConScore[C] as the average score of Consistency between the caption and the other three discriminative responses.

4 Analysis

4.1 Evaluation Results

Method	ConScore[D]	Sensation				Cognition				Knowledge
		T	C	V	Con	T	C	V	Con	T	C	V	Con
Closed-source Vision Language Models
GPT-4V† [1]	$29.20_{6}$	$80.4$	$79.0$	$61.7$	$48.3$	$68.8$	$53.2$	$39.9$	$20.4$	$63.1$	$57.2$	$30.0$	$14.2$
GPT-4-Omni [1]	$\underline{35.70}_{2}$	$89.2$	$79.4$	$64.4$	$\underline{55.0}$	$71.8$	$62.8$	$44.9$	$27.8$	$64.7$	$61.7$	$39.7$	$\underline{23.3}$
Gemini-Pro-Vision [24]	$25.00_{10}$	$85.2$	$60.7$	$63.4$	$39.3$	$71.8$	$45.0$	$44.2$	$15.1$	$65.0$	$51.4$	$39.7$	$15.8$
Gemini-Ultra-Vision [24]	$33.10_{4}$	$78.9$	$78.6$	$66.3$	$50.3$	$68.1$	$58.5$	$47.9$	$\underline{28.5}$	$62.9$	$62.2$	$44.7$	$19.7$
Qwen-VL-Plus [3]	$28.10_{7}$	$82.7$	$74.9$	$60.4$	$45.0$	$64.2$	$41.7$	$30.8$	$16.3$	$63.6$	$54.2$	$33.3$	$15.8$
Qwen-VL-Max [3]	$\mathbf{37.00}_{1}$	$86.4$	$80.7$	$65.4$	$\mathbf{56.3}$	$72.9$	$51.4$	$51.3$	$28.1$	$68.3$	$58.6$	$38.9$	$\mathbf{24.2}$
7B Vision Language Models
LLaVA-v1.5-7B [18]	$16.60_{14}$	$79.3$	$56.8$	$44.3$	$28.3$	$51.4$	$33.5$	$15.8$	$4.7$	$61.7$	$44.4$	$16.9$	$7.8$
Qwen-VL-Chat [3]	$26.40_{9}$	$81.0$	$79.6$	$54.2$	$39.0$	$55.0$	$46.3$	$33.2$	$13.5$	$60.3$	$54.2$	$28.9$	$14.7$
$\sim$ 13B Vision Language Models
LLaVA-v1.5-13B [18]	$24.00_{11}$	$82.9$	$77.1$	$49.6$	$39.5$	$53.6$	$37.8$	$20.1$	$10.4$	$65.6$	$50.3$	$17.2$	$9.7$
MiniGemini-13B [15]	$21.80_{13}$	$81.9$	$69.7$	$52.8$	$39.3$	$51.9$	$38.2$	$21.1$	$6.9$	$52.8$	$36.7$	$17.5$	$9.2$
InternVL-v1.5-26B [7]	$31.40_{5}$	$85.6$	$84.8$	$65.0$	$54.3$	$59.7$	$58.6$	$44.4$	$19.4$	$58.1$	$55.8$	$25.3$	$12.2$
$\sim$ 34B Vision Language Models
LLaVA-NeXT-34B [19]	$27.70_{8}$	$82.4$	$81.7$	$55.6$	$43.6$	$50.7$	$47.5$	$25.6$	$9.9$	$60.4$	$56.1$	$27.8$	$12.8$
MiniGemini-34B [15]	$23.00_{12}$	$80.8$	$76.8$	$48.2$	$39.7$	$36.9$	$30.7$	$18.9$	$6.0$	$58.1$	$42.3$	$20.8$	$8.2$
InternVL-v1.2P-40B [8]	$34.70_{3}$	$83.7$	$83.2$	$66.6$	$53.4$	$74.2$	$67.6$	$57.1$	$\mathbf{34.9}$	$72.2$	$58.3$	$28.6$	$13.6$

Method	Rank Diff	ConScore[C]	Con[T]	Con[C]	Con[V]	Ordered
Closed-source Vision Language Models
GPT-4V [1]	$\mathbf{\mathcolor{ForestGreen}{\uparrow 3}}$	$55.6_{3}$	$51.20$	$56.50$	$59.10$	Y
GPT-4-Omni [1]	$\mathbf{\mathcolor{ForestGreen}{\uparrow 1}}$	$\mathbf{62.2}_{1}$	$58.00$	$62.50$	$66.10$	Y
Gemini-Pro-Vision [24]	$\mathbf{\mathcolor{ForestGreen}{\uparrow 1}}$	$48.4_{9}$	$43.30$	$45.20$	$56.80$	Y
Gemini-Ultra-Vision [24]	$\mathbf{-}$	$54.6_{4}$	$47.80$	$55.20$	$60.70$	Y
Qwen-VL-Plus [3]	$\mathbf{-}$	$50.2_{7}$	$47.10$	$49.10$	$54.30$	Y
Qwen-VL-Max [3]	$\mathbf{\mathcolor{Forestred}{\downarrow 1}}$	$\underline{58.4}_{2}$	$54.30$	$58.00$	$62.90$	Y
7B Vision Language Models
LLaVA-v1.5-7B [18]	$\mathbf{-}$	$38.4_{14}$	$39.20$	$36.60$	$39.50$	N
Qwen-VL-Chat [3]	$\mathbf{\mathcolor{Forestred}{\downarrow 2}}$	$48.0_{11}$	$42.00$	$50.80$	$51.30$	Y
$\sim$ 13B Vision Language Models
LLaVA-v1.5-13B [18]	$\mathbf{\mathcolor{Forestred}{\downarrow 1}}$	$44.4_{12}$	$41.50$	$45.80$	$46.00$	Y
MiniGemini-13B [15]	$\mathbf{-}$	$41.7_{13}$	$38.80$	$42.90$	$43.30$	Y
InternVL-v1.5-26B [7]	$\mathbf{\mathcolor{Forestred}{\downarrow 1}}$	$50.9_{6}$	$44.50$	$53.90$	$54.20$	Y
$\sim$ 34B Vision Language Models
LLaVA-NeXT-34B [19]	$\mathbf{\mathcolor{Forestred}{\downarrow 2}}$	$48.3_{10}$	$46.00$	$52.20$	$46.80$	N
MiniGemini-34B [15]	$\mathbf{\mathcolor{ForestGreen}{\uparrow 4}}$	$49.6_{8}$	$56.80$	$48.00$	$44.10$	N
InternVL-v1.2P-40B [8]	$\mathbf{\mathcolor{Forestred}{\downarrow 2}}$	$53.7_{5}$	$49.80$	$55.50$	$55.80$	Y

In this section, 6 closed-source and 8 open-source representative LVLMs with varying sizes and architectures are evaluated on our Consistency benchmark, including GPT-4V [1], GPT4-Omni [1], Gemini-Vision [24], Qwen-VL series [3], LLaVA series [18; 19], MiniGemini series [15] and InternVL series [8]. The evaluation results on ConBench are listed in Table 1 and 2.In the metric[D] 1, Qwen-VL-Max [3] secures the top position, leading the second-place GPT4-Omni [1] by a margin of $1.3\%$. The InternVL-v1.2P-40B [8] performs best in the open-sourced community, especially in cognition capability. The LLaVA series did not make it to the top ten.In the metric[C], the newest GPT4-Omni [1] leads the leaderboard, which is the only model that surpasses $60$. It has a significant advantage over the second-place model Qwen-VL-Max [3], with a gap of $3.8$. We observed that although the GPT series slightly underperforms Qwen-Max in metric[D], it significantly outperforms the Qwen series in metric[C], which aligns with our actual user experience. Actually, ConScore[C] provides an alternative quality description of captions, because higher recall and precision rates usually match better Consistency. Besides, rankings of LVLMs show a slight variation between metric[C] and metric[D]. The GPT series models claim better performance of caption generation.

4.3 Generative Domain

Next, we extend our attention to the generative domain. Based on Consistency, we first build a bridge between the discriminative and generative domains. We consolidate our findings as the below facts:

Fact 4.3.1 (Inconsistency to Generative Answers).

As the solution space of discriminative questions increases, the Consistency between their answers and generative answers increases.

See Also

3rd Grade Worksheets - Week 6 of 36 - [PDF Document]

As indicated in the last column of Table 2, “Ordered” means Con[T] < Con[C] < Con[V]. The answers of all closed-source models and most open-source models adhere to this pattern. Here is the theoretical explanation.Assume the distribution for the generative domain (Caption) is $\mathcal{S}$, and the sample space of $\mathcal{S}$ is $W$. For the discriminative domain, the sample space is limited to $W^{\prime}$, which only contains some candidates from $W$. Assume the model handles the discriminative domain by creating another distribution $\mathcal{S^{\prime}}$ according to $\mathcal{S}$ and $W^{\prime}$.Then the total variation distance (TVD) [11] between $\mathcal{S}$ and $\mathcal{S^{\prime}}$ is

$$\frac{1}{2}\left\|\mathcal{S}-\mathcal{S^{\prime}}\right\|_{1}.$$

(1)

This becomes larger when $|W\setminus W^{\prime}|$ becomes larger. For instance, if the model creates $\mathcal{S^{\prime}}$ by simply doing reject sampling³³3e.g, projecting or clustering $\mathcal{S}$ to $W^{\prime}$, then

$$\frac{1}{2}\left\|\mathcal{S}-\mathcal{S^{\prime}}\right\|_{1}=Pr[\mathcal{S}$$

(2)

It is obvious that when $W^{\prime}$ is more ”different” from $W$, the distance will be larger.

Fact 4.3.2 (Connection between Discriminative and Generative Domain).

The accuracy of the discriminative answer exhibits a strong positive correlation with its Consistency with the generative.

As shown in Figure 6 and 7, we conduct visualizations for all tested LVLMs: The vertical axis represents the accuracy of their discriminative answers, while the horizontal axis represents the consistency of the answers with caption. Figure 6 displays the distribution across different question types, while Figure 7 illustrates the distribution across different core capabilities. The green lines represent a fitted linear equation. Additionally, we utilize the Pearson coefficient $P[X,Y]$ to quantitatively analyze the degree of linear correlation, and the 6 coefficients in the above figures are all more than $0.85$.

5 Trigger-based Diagnostic Refinement

In light of the previous findings, we summarize two key insights: (1) LVLMs exhibit higher accuracy when operating within a narrower discriminative solution space; (2) Incorrect answers are usually associated with significantly lower confidence and logits. Consequently, we propose a simple but efficient method dubbed Trigger-based Diagnostic Refinement (TDR) to ameliorate the generation skill of LVLMs without any additional training. The proposed pipeline is presented in Figure 8.

MethodWe start by making the LVLM generate a caption, with each word accompanied by its corresponding probability. Next, uninformative words are dropped based on their parts of speech, and we only keep nouns, adjectives and quantifiers. When the remaining words with probabilities below a threshold $\tau$ (we set $\tau$ = 0.85 here), trigger subsequent diagnostic processes. Since low probabilities of words indicate a lack of confidence, we formulate True/False discriminative questions to force the LVLM to self-verify (e.g., Is there {cat} in the picture?). The self-diagnostic prompt and its response will be drafted into a new prompt, which is fed back into the LVLM to generate a higher-quality caption.

ResultsWe carried out experiments on the LLaVA-NeXT-34B and MiniGemini-34B and evaluated them on the metric[C] of ConBench. The experimental results are detailed in Table 4.Notably, the LLaVA-NeXT-34B sees an improvement of $9.1$ points, while the MiniGemini experiences an overall enhancement of $9.6$ points. Although our approach primarily employs True/False questions for self-verify, there is still a noticeable improvement in ConScore[C].Hence, our method effectively boosts the quality of captions by triggering the model to self-check.

In theory, we can further construct multiple discriminative questions for the caption, enabling the model to verify multiple elements within the caption. Additionally, the process can be iterated multiple rounds, leading to ongoing enhancements in the quality of the generated output. Our method is a simplified implementation of the above approaches.

6 Conclusion

In this study, we investigate the Consistency issues in large vision-language models (LVLMs). Consistency reflects the overall ability of LVLMs, as it not only requires LVLMs to provide correct answers but also demands sufficient confidence in their knowledge point, regardless of the type of question encountered. We first introduce the ConBench, a benchmark that fills the gap in assessing Consistency. It includes 1K images with 4K prompts and two evaluation metrics: ConScore[D] and ConScore[C]. Then, our findings shed light on the nature of Consistency in LVLMs according to the ConBench. We observe that as the solution space of a prompt increases, the accuracy of the answers tends to decrease. Besides, we establish a relationship between the discriminative and generative realms, highlighting the importance of Consistency between the discriminative answer and caption. Furthermore, we discover that closed-source models exhibit a bias advantage over open-source models in terms of consistency. Finally, we propose a solution by forcing LVLMs to self-think, where a discriminative prompt is constructed via uncertain words in the caption. Our method makes the quality of LVLMs’ captions an impressive achievement. We believe that our research contributes to the evaluation of LVLMs and encourages future advancements for achieving Consistency in LVLMs.

References

• [1]J.Achiam, S.Adler, S.Agarwal, L.Ahmad, I.Akkaya, F.L. Aleman, D.Almeida, J.Altenschmidt, S.Altman, S.Anadkat, etal.Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023.
• [2]J.Bai, S.Bai, Y.Chu, Z.Cui, K.Dang, X.Deng, Y.Fan, W.Ge, Y.Han, F.Huang, etal.Qwen technical report.arXiv preprint arXiv:2309.16609, 2023.
• [3]J.Bai, S.Bai, S.Yang, S.Wang, S.Tan, P.Wang, J.Lin, C.Zhou, and J.Zhou.Qwen-vl: A frontier large vision-language model with versatile abilities.arXiv preprint arXiv:2308.12966, 2023.
• [4]X.Bi, D.Chen, G.Chen, S.Chen, D.Dai, C.Deng, H.Ding, K.Dong, Q.Du, Z.Fu, etal.Deepseek llm: Scaling open-source language models with longtermism.arXiv preprint arXiv:2401.02954, 2024.
• [5]L.Chen, J.Li, X.Dong, P.Zhang, C.He, J.Wang, F.Zhao, and D.Lin.Sharegpt4v: Improving large multi-modal models with better captions.arXiv preprint arXiv:2311.12793, 2023.
• [6]L.Chen, J.Li, X.Dong, P.Zhang, Y.Zang, Z.Chen, H.Duan, J.Wang, Y.Qiao, D.Lin, etal.Are we on the right way for evaluating large vision-language models?arXiv preprint arXiv:2403.20330, 2024.
• [7]Z.Chen, W.Wang, H.Tian, S.Ye, Z.Gao, E.Cui, W.Tong, K.Hu, J.Luo, Z.Ma, etal.How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites.arXiv preprint arXiv:2404.16821, 2024.
• [8]Z.Chen, J.Wu, W.Wang, W.Su, G.Chen, S.Xing, M.Zhong, Q.Zhang, X.Zhu, L.Lu, B.Li, P.Luo, T.Lu, Y.Qiao, and J.Dai.Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks.arXiv preprint arXiv:2312.14238, 2023.
• [9]M.Cordts, M.Omran, S.Ramos, T.Rehfeld, M.Enzweiler, R.Benenson, U.Franke, S.Roth, and B.Schiele.The cityscapes dataset for semantic urban scene understanding.In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016.
• [10]C.Fu, P.Chen, Y.Shen, Y.Qin, M.Zhang, X.Lin, J.Yang, X.Zheng, K.Li, X.Sun, Y.Wu, and R.Ji.Mme: A comprehensive evaluation benchmark for multimodal large language models.arXiv preprint arXiv:2306.13394, 2023.
• [11]J.Kennedy and M.Quine.The total variation distance between the binomial and poisson distributions.The Annals of Probability, pages 396–400, 1989.
• [12]B.Li, R.Wang, G.Wang, Y.Ge, Y.Ge, and Y.Shan.Seed-bench: Benchmarking multimodal llms with generative comprehension.arXiv preprint arXiv:2307.16125, 2023.
• [13]J.Li, D.Li, S.Savarese, and S.Hoi.Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models.In International conference on machine learning, pages 19730–19742. PMLR, 2023.
• [14]X.L. Li, V.Shrivastava, S.Li, T.Hashimoto, and P.Liang.Benchmarking and improving generator-validator consistency of language models.arXiv preprint arXiv:2310.01846, 2023.
• [15]Y.Li, Y.Zhang, C.Wang, Z.Zhong, Y.Chen, R.Chu, S.Liu, and J.Jia.Mini-gemini: Mining the potential of multi-modality vision language models.arXiv preprint arXiv:2403.18814, 2024.
• [16]T.-Y. Lin, M.Maire, S.Belongie, J.Hays, P.Perona, D.Ramanan, P.Dollár, and C.L. Zitnick.Microsoft coco: Common objects in context.In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740–755. Springer, 2014.
• [17]Z.Lin, S.Trivedi, and J.Sun.Generating with confidence: Uncertainty quantification for black-box large language models.arXiv preprint arXiv:2305.19187, 2023.
• [18]H.Liu, C.Li, Y.Li, and Y.J. Lee.Improved baselines with visual instruction tuning, 2023.
• [19]H.Liu, C.Li, Y.Li, B.Li, Y.Zhang, S.Shen, and Y.J. Lee.Llava-next: Improved reasoning, ocr, and world knowledge, January 2024.
• [20]H.Liu, C.Li, Q.Wu, and Y.J. Lee.Visual instruction tuning, 2023.
• [21]Y.Liu, H.Duan, Y.Zhang, B.Li, S.Zhang, W.Zhao, Y.Yuan, J.Wang, C.He, Z.Liu, etal.Mmbench: Is your multi-modal model an all-around player?arXiv preprint arXiv:2307.06281, 2023.
• [22]M.Mathew, D.Karatzas, and C.Jawahar.Docvqa: A dataset for vqa on document images.In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200–2209, 2021.
• [23]A.Radford, J.W. Kim, C.Hallacy, A.Ramesh, G.Goh, S.Agarwal, G.Sastry, A.Askell, P.Mishkin, J.Clark, etal.Learning transferable visual models from natural language supervision.In International conference on machine learning, pages 8748–8763. PMLR, 2021.
• [24]G.Team, R.Anil, S.Borgeaud, Y.Wu, J.-B. Alayrac, J.Yu, R.Soricut, J.Schalkwyk, A.M. Dai, A.Hauth, etal.Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023.
• [25]H.Touvron, T.Lavril, G.Izacard, X.Martinet, M.-A. Lachaux, T.Lacroix, B.Rozière, N.Goyal, E.Hambro, F.Azhar, etal.Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023.
• [26]W.Wang, Q.Lv, W.Yu, W.Hong, J.Qi, Y.Wang, J.Ji, Z.Yang, L.Zhao, X.Song, etal.Cogvlm: Visual expert for pretrained language models.arXiv preprint arXiv:2311.03079, 2023.
• [27]Q.Ye, H.Xu, G.Xu, J.Ye, M.Yan, Y.Zhou, J.Wang, A.Hu, P.Shi, Y.Shi, etal.mplug-owl: Modularization empowers large language models with multimodality.arXiv preprint arXiv:2304.14178, 2023.
• [28]X.Yue, Y.Ni, K.Zhang, T.Zheng, R.Liu, G.Zhang, S.Stevens, D.Jiang, W.Ren, Y.Sun, etal.Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi.arXiv preprint arXiv:2311.16502, 2023.
• [29]R.Zhang, J.Han, C.Liu, P.Gao, A.Zhou, X.Hu, S.Yan, P.Lu, H.Li, and Y.Qiao.Llama-adapter: Efficient fine-tuning of language models with zero-init attention.arXiv preprint arXiv:2303.16199, 2023.

Appendix

Appendix A Showcases of Trigger-based Diagnostic Refinement

MGM-34B

Image. Count-000000236721.jpg

Original Caption. The image contains a variety of objects, which can be counted as follows: one tray, one cake, one bowl of sauce, one carton of milk, one banana, one glass of orange juice, one cup of coffee, and one sugar packet. The objects are arranged neatly on the tray, which is placed on a table.

Refined Caption The image contains a total of two bananas, which are positioned on the right side of the tray. The bananas are yellow, indicating that they are ripe, and they are the only fruits visible in the image. They are placed next to a carton of milk, ice cream, a cake, two cups of coffee, a glass of orange juice, which are a common accompaniment in a breakfast setting.

LLava-Next-34B

Image. Attribute Reason-1724416_2902943546.jpg

Original Caption. The image captures a vibrant live music scene. At the center of the frame, a musician is seated on a stool, engrossed in playing a red electric guitar. He is dressed casually in a white t-shirt and blue jeans, adding to the relaxed atmosphere of the performance. The stage, adorned with a colorful banner, is set against a backdrop of a brick wall, lending an urban feel to the setting. The banner prominently features the text "DOWNTOWN 5", possibly indicating the name of the band or the event. The musician is surrounded by other band members, each engrossed in their own instruments, contributing to the overall harmony of the performance. The stage is well-equipped with a microphone and a speaker, essential for amplifying the sound and reaching the audience. In the background, a crowd of people can be seen, their attention riveted on the performance.

Refined Caption The image captures a vibrant scene of a live music performance. At the center of the stage, a musician is seated on a stool, engrossed in playing an electric guitar. He is dressed in a casual white t-shirt, and his focus is evident as he strums the strings. The stage is set against a backdrop of a colorful banner that proudly proclaims "DOWNTOWN 5". The banner’s bold colors contrast with the musician’s white t-shirt, drawing attention to both the performer and the event. The stage is equipped with various musical instruments, including a drum set and a keyboard, suggesting a full band performance. A microphone stands ready, capturing the musician’s performance for the audience. The stage is situated in a space that appears to be partially open to the elements, as suggested by the presence of a roof structure in the background. This setup indicates that the concert is being held in a venue that can accommodate outdoor or semi-outdoor performances, adding a unique ambiance to the event. Overall, the image portrays a dynamic and engaging live music event, with the musician at the heart of the performance, surrounded by his instruments.

Appendix B Computation for Probabilities and Logits

For each model (MGM-13B and LLaVA-13B), we randomly selected 50 consistent and 50 inconsistent samples, respectively. The JSON samples are in support materials. To measure the LVLMs’ response confidence accurately, we manually simplified the answers and maintained the words that are the most critical to the judgment. The specific principles are as listed:

Appendix C Limitations

The introduced ConBench offers a new perspective on evaluating model performance through the consistency between multiple types of questions, providing a more comprehensive measurement and understanding of existing LVLMs. However, due to the distinct response forms of captions, assessing the consistency between captions and discriminative answers is judged by GPT, posing a risk of inaccurate evaluations. Besides, by delving deeper into our benchmark analysis, we propose trigger-based diagnostic refinement to improve the consistency and accuracy of LVLMs. This, however, introduces additional computational costs and is limited by the inherent capabilities of the LVLMs. Further improvements can be achieved by designing and training LVLMs with a focus on consistency.

Appendix D Broader Impacts

Overall, this research has broader impacts on the evaluation, performance, fairness, and future development of LVLMs, fostering progress and advancements in the field of vision-language models.

Advancing Evaluation: The introduction of ConBench, a benchmark for assessing Consistency in LVLMs, fills a crucial gap in the evaluation of these models. This benchmark provides a standardized framework for measuring the performance and reliability of LVLMs across different prompts.

Novel Insights: we are the first to reveal the tapestry and get the following findings:(1) In the discriminate realm, the larger the solution space of the prompt, the lower the accuracy of the answers.(2) Establish the relationship between the discriminative and generative realms.(3) Compared to open-source models, closed-source models exhibit a bias advantage in terms of Consistency.

Inspiring Future Research: By contributing to the evaluation and understanding of Consistency in LVLMs, this research paves the way for future advancements in the field. It encourages researchers to explore new techniques, methodologies, and approaches to achieve higher levels of Consistency in LVLMs, ultimately pushing the boundaries of language and vision understanding.

Appendix E Detailed Cases in ConBench

We have uploaded the ConBench dataset, including images and their prompts, to the Hugging Face platform. The dataset can be accessed at the following URL: https://huggingface.co/datasets/ConBench/ConBench. Here, we enumerate several representative cases from ConBench. Arrange in order from easy to difficult, respectively, based on sensation, cognition, and knowledge.