Add Batch d3ff2ff4-624e-4230-9b13-43dd56541b03

ICLR/2025/MLLM as Retriever_ Interactively Learning Multimodal Retrieval for Embodied Agents/15105407-cc64-4a31-aa06-20898f67a02d_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed47ac3a4c4b82cc79305a7d58d5b6646517aba4e046007a8cd52324e94b41fc
+size 181658

ICLR/2025/MLLM as Retriever_ Interactively Learning Multimodal Retrieval for Embodied Agents/15105407-cc64-4a31-aa06-20898f67a02d_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f38cb34c80bccee91194ed4ffbb1b0e1be45485da1bbdbb44c23270b309b151
+size 221910

ICLR/2025/MLLM as Retriever_ Interactively Learning Multimodal Retrieval for Embodied Agents/15105407-cc64-4a31-aa06-20898f67a02d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c2bcb2a7fb443100c39f544891e5d7ee85d7b61a77e0816201bc8aad6629f6a
+size 5012125

ICLR/2025/MLLM as Retriever_ Interactively Learning Multimodal Retrieval for Embodied Agents/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c7229deef536bb7feef2ca20a9bb391f46b10e4663e0d1802f70dedb3c90902
+size 1994230

ICLR/2025/MLLM as Retriever_ Interactively Learning Multimodal Retrieval for Embodied Agents/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb7924f52a61709e7fbe6a39e1ab780dee56efdd2b0ccf6eead5e330c60e21b5
+size 901699

ICLR/2025/MLLM can see_ Dynamic Correction Decoding for Hallucination Mitigation/fe126150-43f5-41e5-8e8e-709b0b20c754_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b420fde360532dcb04b9c2023c9ca1fb5be44c58963342451a541547b511f4b8
+size 144875

ICLR/2025/MLLM can see_ Dynamic Correction Decoding for Hallucination Mitigation/fe126150-43f5-41e5-8e8e-709b0b20c754_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c168a09889acf5842d437901273329dc5264c026c9f0fa9c2a8ca508ec4774b
+size 173920

ICLR/2025/MLLM can see_ Dynamic Correction Decoding for Hallucination Mitigation/fe126150-43f5-41e5-8e8e-709b0b20c754_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65218442b2639538df6c3d39170bcfdad552c62aee44bd394110059b88341c90
+size 32589675

ICLR/2025/MLLM can see_ Dynamic Correction Decoding for Hallucination Mitigation/full.md

@@ -0,0 +1,557 @@
+# MLLM CAN SEE? DYNAMIC CORRECTION DECODING FOR HALLUCINATION MITIGATION
+
+Chenxi Wang\*, Xiang Chen\*, Ningyu Zhang\*, Bozhong Tian\*, Haoming Xu\*, Shumin Deng\*, Huajun Chen\*
+
+$\spadesuit$ Zhejiang University  $\text{♥}$  Nanjing University of Aeronautics and Astronautics  
+$\diamond$  National University of Singapore, NUS-NCS Joint Lab, Singapore  
+$\clubsuit$ Zhejiang Key Laboratory of Big Data Intelligent Computing
+
+{sunnywcx,zhangningyu}@zju.edu.cn, xiang-chen@nuaa.edu.cn
+
+# ABSTRACT
+
+Multimodal Large Language Models (MLLMs) frequently exhibit hallucination phenomena, but the underlying reasons remain poorly understood. In this paper, we present an empirical analysis and find that, although MLLMs incorrectly generate the objects in the final output, they are actually able to recognize visual objects in the preceding layers. We speculate that this may be due to the strong knowledge priors of the language model suppressing the visual information, leading to hallucinations. Motivated by this, we propose a novel dynamic correction decoding method for MLLMs (DeCo), which adaptively selects the appropriate preceding layers and proportionally integrates knowledge into the final layer to adjust the output logits. Note that DeCo is model agnostic and can be seamlessly incorporated with various classic decoding strategies and applied to different MLLMs. We evaluate DeCo on widely-used benchmarks, demonstrating that it can reduce hallucination rates by a large margin compared to baselines, highlighting its potential to mitigate hallucinations<sup>1</sup>.
+
+"The first principle is that you must not fool yourself—and you are the easiest person to fool."
+
+Richard Feynman
+
+# 1 INTRODUCTION
+
+Recently, the rapid development of Multimodal Large Language Models (MLLMs) has demonstrated a potential pathway towards achieving Artificial General Intelligence (AGI) (Wang et al., 2024b; Yao et al., 2024; Lu et al., 2024a; Team, 2024; OpenAI, 2023; Liu et al., 2023b; Chern et al., 2024). However, in practice, the development of MLLMs is hindered by the phenomenon of hallucination, which typically results in the model generating statements about non-existent images while neglecting to mention certain visible objects, effectively causing it to fool itself (Bai et al., 2024; Liu et al., 2024a; Li et al., 2023b; Liu et al., 2023a; Rawte et al., 2023). This issue poses significant risks in high-stakes fields such as medical imaging (Chen et al., 2024b; Hu et al., 2023; Wang et al., 2023c), autonomous driving (Cui et al., 2024; Wang et al., 2023d), and human-computer interaction systems (Brie et al., 2023), where such errors could result in irreparable consequences.
+
+The reasons behind hallucinations in MLLMs are complex. Unlike analyses focused on unimodal LLMs (Chuang et al., 2024; Chen et al., 2024d; Orgad et al., 2024; Chen et al., 2024e; Lu et al., 2024b; Wang et al., 2024a), many current works assume that MLLM may indeed 'see' visual information. However, due to factors such as excessive model depth (Chen et al., 2024c; Zhang et al., 2024a), aggregation patterns (Huang et al., 2024), or priors knowledge inherent in the MLLMs (Leng et al., 2023; Zhang et al., 2024b), these models ultimately still experience hallucinations (The details can be found in Appendix B). Concretely, our understanding of the underlying mechanisms of hallucinations in MLLMs remains limited. It is still uncertain whether the visual information is never correctly recognized or if it is recognized but subsequently suppressed by later information streams.
+
+![](images/d11535f295ff87565f5878c8ef21b420784dee9d0ec8bc98af2946f74f02eec5.jpg)  
+(a) Object probing results.
+
+![](images/7be84555d04c875c17b53027c0b6eeef16ebeca9ba9c8ff28eed7f7640b6eb79.jpg)
+
+![](images/ba762c21fa2e15bcf301a891b034084d0f5fc7dc9b733cbfe1f99f37235c1278.jpg)  
+(b) Different resolution results.  
+Figure 1: Overall results of the probing experiment with MLLMs, indicating that they possess a certain level of awareness regarding the presence of visual objects (Figure 1(a)), with prediction accuracy being higher in the preceding layers (Figure 1(b)) but gradually decline afterward.
+
+Hallucinated MLLM can see (to some extent). Inspired by the aforementioned works, we conduct an empirical analysis and find that MLLMs are not blind; they can recognize objects in the preceding layers, but this recognition is suppressed in later layers, leading to hallucinations. Specifically, we focus on object hallucinations $^2$  and conduct experiments with MLLMs, demonstrating that they know to some extent whether an object exists (as shown in Figure 1 and Section 2.1). We further observe that the confidence of generated tokens is influenced by the knowledge priors of MLLMs (Section 2.2), leading to a reduction in the probability of ground truth tokens in the deeper layers.
+
+Dynamic correction decoding with preceding-layer knowledge. Based on those findings, we propose Dynamic Correction Decoding with preCeding-Layer Knowledge (DeCo) to mitigate hallucinations for MLLMs. Our core hypothesis is that preceding layers exhibit higher confidence for ground truth tokens, and the logits for these tokens should rank prominently at the last layer's outputs. To enhance the logits of ground truth tokens, DeCo dynamically selects preceding layer and utilizes its prior knowledge to correct the final output logits. Additionally, we introduce a dynamic soft modulation to preserve the original style of the generated responses. DeCo is training-free and can be integrated with any popular decoding strategies, such as greedy search, nucleus sampling as well as beam search, and can seamlessly incorporate into any MLLMs for hallucination mitigation.
+
+Contributions. Our primary contribution lies in exploring the internal mechanisms of hallucinations in MLLMs. We find that the confidence of generated tokens is influenced by the knowledge priors of MLLMs, leading to a reduction in the probability of ground truth tokens in the deeper layers. We further propose DeCo, a dynamic correction decoding method guided by preceding-layer knowledge. DeCo is integrated with InstructBLIP, MiniGPT-4, LLaVA, and Qwen-VL using three popular decoding strategies: greedy search, nucleus sampling, and beam search. Experimental results show that DeCo achieves an average hallucination suppression rate of  $10.8\%$  in image captioning dataset, demonstrating superior suppression effectiveness. Additionally, DeCo outperforms baselines on visual question answering datasets including POPE, and MME. Additionally, we analyze the latency and throughput, showing that DeCo introduces an approximate 1.2x increase in latency compared to the basic decoding process, much faster than previous baselines such as VCD and OPERA.
+
+# 2 WHY DO MLLMs GENERATE NON-EXIST OBJECTS?
+
+In this section, we conduct a series of empirical analysis to investigate the internal mechanisms of MLLM and elucidate the underlying reasons for its generation of non-existent objects. To strike a balance between the realism and complexity of the experiments, we primarily focus on the generation of objects in image description scenarios (image caption tasks).
+
+Preliminaries of MLLM generation. MLLMs typically concatenate visual tokens, processed by the visual encoder and projection layer, with embedded textual tokens before feeding them into an autoregressive language model. We denote the visual tokens as  $\mathbf{X}^V = \{x_{v_1}, x_{v_2}, \ldots, x_{v_P}\}$  and textual tokens as  $\mathbf{X}^C = \{x_{c_1}, x_{c_2}, \ldots, x_{c_Q}\}$ . Here  $P$  and  $Q$  are the lengths of the visual tokens
+
+![](images/89e9e593e4e4f358bd0e1027f90ea3fc6016e28ca81a12ae0633e86858b6d9b1.jpg)  
+Figure 2: Illustration of token probabilities across transformer layers, which reveals distinct trends for target hallucinated (orange) and non-hallucinated (green) tokens. In the preceding layers, non-hallucinated tokens exhibit a higher probability. In the final layers, hallucinated tokens demonstrate increased probabilities, while the probability of non-hallucinated tokens drops sharply.
+
+and textual tokens respectively. Finally, the input is  $\mathbf{X} = \mathrm{concat}\{\mathbf{X}^V,\mathbf{X}^C\}$ . Then  $\mathbf{X}$  would be passed into MLLM with  $\mathbf{N}$  stacked transformer layer. The intermediate variable generated by the i-th layer is called hidden states, denoted as  $\mathbf{h}^i = \{h_0^i,h_1^i,\dots ,h_{T - 1}^i\}$ , where  $T = P + Q$ . During the generation phase, we use the hidden state at the last position in the final layer, which is mapped to the vocabulary dimension through an affine layer  $\phi (\cdot)$ , to predict the probability of the next token. Formally, we have:
+
+$$
+p \left(x _ {T} \mid x _ {<   T}\right) = \operatorname {s o f t m a x} \left(\phi \left(h _ {T - 1} ^ {N}\right)\right) _ {x _ {T}}, x _ {T} \in \mathcal {V} \tag {1}
+$$
+
+where we use  $x_{< T}$  to simplify the sequence  $\{x_i\}_{i=0}^{T-1}$  and  $\mathcal{V}$  refers to the whole vocabulary set.
+
+# 2.1 FINDING 1:MLLMKNOWSTOSOMEEXTENTWHETHERANOBJECTEXISTS
+
+Inspired by (Ye et al., 2024), we explore how MLLMs comprehend objects in the image captioning task. For simplicity, we abstract this process into a function called isexist(obj), which determines whether an object is present in an image. To examine the application of this function within the MLLM's image captioning workflow, we conduct probing experiments at the conclusion of object descriptions in each layer of the MLLM's language model component, which consists of 32 transformer layers in a 7-billion-parameter model (Detailed setup in Appendix C.1).
+
+We employ the prompt template, "USER: <image>Describe the image. ASSISTANT: The image contains obj." Both the training and testing datasets are formatted accordingly before being input into MLLMs. We train a probe classifier at the final position of the hidden state outputs for each transformer layer, resulting in a total of 32 classifiers. (For details on the subset division, OOD and in-distribution splits, and prompt templates, please refer to Appendix C.1.) The model is evaluated using the test set, as shown in Figure 1(a) (left). Further experiments are conducted on three splits of the evaluation dataset proposed by POPE, with results reported in Figure 1(a) (right). These evaluations provide a comprehensive understanding of the model's object recognition capabilities across diverse scenarios.
+
+We select the best-performing probe classifier from the 32 classifiers to compare accuracy across all objects, existing objects, and non-existing objects. Our results show that the MLLM achieves high accuracy for correctly generated objects in image captions. Despite generating many non-existent objects, the MLLM still maintains around  $80\%$  accuracy in our probing experiments. This suggests that MLLMs possess a certain level of understanding regarding object existence in images.
+
+Additionally, our probing experiments reveal higher accuracy in the preceding layers, as illustrated in Figure 1(b), which aligns with previous findings (Zhang et al., 2024b; Leng et al., 2023). Furthermore, we show that increasing the resolution of the visual encoder (from 224px to 336px) enhances
+
+accuracy for non-existing objects, indicating that token information at the last position in the preceding layers better represents visual information. (For a detailed explanation of the different visual resolutions, please refer to Appendix C.1). These findings suggest that the utilization of the preceding-layers in MLLMs enables the model to perform self-correction.
+
+# 2.2 FINDING 2: LANGUAGE MODEL PRIORS SUPPRESS THE VISUAL INFORMATION THAT MLLM ALREADY SEE.
+
+We hypothesize that the representations in the preceding layers effectively capture (to some extent) visual information. However, the prior knowledge embedded in the MLLM reduces the probabilities of ground truth tokens in deeper layers. Figure 2 illustrates this hypothesis with running examples. We analyze the Top-4 tokens ranked by probability in the final layer's output. Non-hallucinated tokens like "people", "left", "blue", and "umbrella" exhibit high probabilities from the 18th layer. In contrast, hallucinated tokens like "bird" and "green" only show comparatively high probabilities around the 30-th layer. Interestingly, the probabilities of ground truth tokens "umbrella" and "blue" sharply decline from the 30-th layer onwards, eventually falling below the hallucinated tokens' probabilities in the final layer.
+
+To further investigate this phenomenon, we conduct an early exit experiment (Teerapittayanon et al., 2016; Elbayad et al., 2020; Schuster et al., 2022) to analyze the evolution of the MLLM's internal representations across transformer layers. We randomly select 500 images from the MSOCO dataset and use random prompts to elicit raw responses from LLaVA1.5-7b. We then extract all non-existent objects along with their corresponding preceding text and input this data into the MLLM. We observe the probabilities of the next token across the transformer layers to gain insights into the model's behavior (see Appendix C.2 for detailed experimental setup). The output of the  $i$ -th layer is denoted as  $h^i$ , and the probability
+
+![](images/b659b252b8b61bff2d5d4bd29b46b1362a0fa289a1abbe143284697969f81bf7.jpg)  
+Figure 3: Distribution of activated ground-truth tokens across layers.
+
+distribution of the next token is represented as  $p(\cdot | x_{<s})^i = \text{softmax}(\phi(h_{s-1}^i))$ . To reduce the observation tokens and simulate the real sampling process, we truncate the vocabulary, similar to Top- $p$  sampling, and obtain the candidate tokens, denoted as  $\mathcal{V}_{candidate}$  with a default threshold of 0.9. We then label the tokens in  $\mathcal{V}_{candidate}$ . Specifically, we filter out data where  $\mathcal{V}_{candidate}$  contains at least one ground truth token and observe whether an activated ground truth token exists among the candidate tokens, formally expressed as:
+
+$$
+\exists x _ {a} \in \mathcal {V} _ {\text {c a n d i d a t e}} \wedge i \in (0, N ], p \left(x _ {a} \mid x _ {<   s}\right) ^ {i} - p \left(x _ {h} \mid x _ {<   s}\right) ^ {i} \geq \text {t h r e s h o l d}, \tag {2}
+$$
+
+where  $x_{a}$  is the activated ground truth token,  $x_{h}$  is the token with the highest probability of being a hallucinated token in the probability distribution of the final layer and threshold  $\in (0,1)$ . Based on the experimental setup described above, we conducted the following investigation:
+
+What suppresses the expression of visual facts? We analyze the occurrence of  $x_{a}$  at each decoding layer, as shown in Figure 3. The results reveal that the activated ground truth tokens are primarily present between layers 20 and 28, indicating that MLLMs accurately recognize the image content in the latter layers. Notably, differences in experimental setups account for the variation in interval layers observed between Finding 1 and Finding 2. However, the activated ground truth tokens are suppressed in the final output layer. This suppression may stem from the guidance of the input image or the inherent knowledge bias of the MLLM. To investigate this, we generate candidate tokens  $\mathcal{V}_{candidate}^{\prime}$  in the absence of an input image, representing tokens based on the MLLM's inherent knowledge. We calculate that the overlap rate of  $x_{h}$  existing in  $\mathcal{V}_{candidate}^{\prime}$  reaches  $91.05\%$ , suggesting that even without expressing image information, MLLMs still tend to generate the original hallucination tokens. This finding reveals that the inherent knowledge in MLLMs may diminish the probability of the ground truth token in the deeper layers.
+
+![](images/6468f168a55d0a92a0cd7f9b92555275856a2e060ae9eb10655d37754acba7b2.jpg)  
+Figure 4: Framework of DeCo. DeCo first dynamically selects an appropriate anchor layer from the preceding layers and then correct the knowledge in the final layer with dynamic coefficient.
+
+# 3 PROPOSED APPROACH: DYNAMIC CORRECTION DECODING WITH PRECEDING-LAYER KNOWLEDGE
+
+After investigating the reasons why MLLMs generate non-existent objects, inspired by (Chuang et al., 2024), we introduce Dynamic Correction Decoding with preCeding-Layer Knowledge (DeCo), which can alleviate hallucinations during inference. The overall framework of Deco is illustrated in Figure 4, consisting of dynamic preceding layer selection (Section 3.1) and decoding correction with preceding-layer knowledge (Section 3.2).
+
+# 3.1 DYNAMIC PRECEDING-LAYER SELECTION
+
+Candidate token acquisition. Due to the vast vocabulary space, we track only the changes in the top-ranked tokens as candidate tokens across different layers for computational convenience. This is based on the hypothesis that ground tokens usually appear in the top position of the MLLM's last layer output logits. Inspired by (Li et al., 2023a), we use a truncation strategy to select the candidate tokens, with the default truncation strategy being top- $p$  truncation, formally:
+
+$$
+\mathcal {V} _ {\text {c a n d i d a t e}} \left(x _ {T} \mid x _ {<   T}\right) = \left\{x _ {T} \in \mathcal {V}: \sum_ {v \in \mathcal {V} p} P _ {\tau} \left(x _ {T} = v \mid x _ {0}, x _ {1}, \dots , x _ {T - 1}\right) \leq p \right\} \tag {3}
+$$
+
+where  $\mathcal{V}$  is the whole vocabulary, and  $p$  refers to the parameter used in top- $p$ . The selected candidate tokens are theoretically ensured to be of high quality, thereby preventing the inclusion of low quality tokens (e.g., semantically incorrect tokens) that exhibit high probabilities in preceding-layers but low probabilities in the final layer.
+
+Table 1: Hit Rate of layers across different intervals.  
+
+<table><tr><td>Layer Range</td><td>20-28</td><td>15-28</td></tr><tr><td>Hit Rate (%)</td><td>61.69</td><td>71.14</td></tr></table>
+
+**Preceding-layer selection.** Our findings in Section 2 demonstrate that activated ground truth tokens typically exhibit higher probabilities in preceding layers compared to hallucinated tokens. Based on this observation, we hypothesize that selecting the token  $x_{th}$ , where  $x_{th} \in \mathcal{V}_{\text {candidate}}$ , with the highest probability from the interval layers corresponds to the ground truth token. We compute the accuracy of  $x_{th}$  as the
+
+ground truth token and denote this metric as the hit rate, as shown in Table 1. The results indicate that within a specific range of layers (e.g., 15-28),  $x_{th}$  indeed has a high universal probability of representing the ground truth token. Intuitively, we track candidate tokens and dynamically choose the layer in which the token with the highest probability among the preceding layers resides to calibrate the final logit distribution of the MLLM. The selected preceding layer is referred to as the anchor layer, formally defined as:
+
+$$
+\mathcal {A} = \operatorname {a r g m a x} _ {i} \left\{x _ {T} \in \mathcal {V} _ {\text {c a n d i d a t e}}: \operatorname {s o f t m a x} \left(\phi \left(h _ {T - 1} ^ {i}\right)\right) _ {x _ {T}}, i \in [ a, b ] \right\}, \tag {4}
+$$
+
+where  $a \leq b, a, b \in [1, N]$ , and  $[a, b]$  represents the layer interval for MLLMs. Expanding the range of layers can improve the hit rate. To avoid increased search computation time, we assign default values of  $a = 20$  and  $b = 28$  for our subsequent experiments.
+
+# 3.2 DECODING CORRECTION WITH PRECEDING-LAYER KNOWLEDGE
+
+Dynamic soft modulation. We introduce a dynamic modulation coefficient, defaulting to the maximum probability. Formally, we have:
+
+$$
+\max  _ {-} \operatorname {p r o b} = \max  (\operatorname {s o f t m a x} (\phi (h _ {T - 1} ^ {A}))). \tag {5}
+$$
+
+This coefficient can help prevent hard changes in logits, particularly when the probability differences between candidate tokens in preceding layers are insignificant. From the example in Figure 4, we can observe that the absence of the dynamic modulation coefficient may lead to semantic incoherence or even more severe hallucinations.
+
+Preceding-layer knowledge guided decoding. Given the selected preceding layers, we integrate information from these layers into the final layer to correct the logit distribution. We utilize a hyperparameter,  $\alpha$ , to control the proportion of early-layer information incorporated. Additionally, dynamic soft modulation is employed to preserve the generative style of the original model. By utilizing the correction of preceding-layer representations, the probability of predicting the next token and the logits are updated as follows:
+
+$$
+\hat {p} \left(x _ {T} \mid x _ {<   T}\right) = \operatorname {s o f t m a x} \left(\log_ {t} \left(\log_ {t} \left(x _ {T}\right)\right) \right. \tag {6}
+$$
+
+$$
+\operatorname {l o g i t s} = \phi \left(h _ {T - 1} ^ {N}\right) + \alpha \times \max  _ {-} \operatorname {p r o b} \times \phi \left(h _ {T - 1} ^ {\mathcal {A}}\right), \tag {7}
+$$
+
+where  $N$  is the last layer of MLLM and  $\mathcal{A}$  is the selected preceding layer.
+
+# 4 EXPERIMENT
+
+# 4.1 SETUP
+
+Baselines. We integrate DeCo with various decoding methods, including greedy decoding, nucleus sampling, and beam search, and compare it against several baselines for mitigating hallucinations, as outlined below: Dola (Chuang et al., 2024) is specifically designed for alleviating hallucinations in factual tasks for LLMs by reducing shallow semantic influences to improve the factuality of the final layer's output. VCD (Leng et al., 2023) mitigates the influence of language model's priors in MLLMs by generating representations that enhance visual information through the subtraction of interfering knowledge prior during each sampling step. OPERA (Huang et al., 2024) dynamically penalizes overconfident tokens based on the emergence of aggregation patterns, while proposing a retrospective allocation strategy to avoid cases where hallucinations have already occurred. For all the baselines, we use the default hyperparameters from the source code for a fair comparison.
+
+Table 2: CHAIR hallucination evaluation results. Lower scores indicate fewer hallucinations. OPERA utilizes beam search, VCD applies nucleus sampling, and DeCo is the proposed method compatible with various decoding approaches.  
+
+<table><tr><td rowspan="2">Decoding</td><td rowspan="2">Method</td><td colspan="2">InstructBLIP</td><td colspan="2">MiniGPT-4</td><td colspan="2">LLaVA-1.5</td><td colspan="2">Qwen-VL</td></tr><tr><td>CHAIRS↓</td><td>CHAIRI↓</td><td>CHAIRS↓</td><td>CHAIRI↓</td><td>CHAIRS↓</td><td>CHAIRI↓</td><td>CHAIRS↓</td><td>CHAIRI↓</td></tr><tr><td rowspan="3">Greedy</td><td>Vanilla</td><td>58.8</td><td>23.7</td><td>31.8</td><td>9.9</td><td>45.0</td><td>14.7</td><td>46.0</td><td>12.5</td></tr><tr><td>DoLa</td><td>48.4</td><td>15.9</td><td>32.2</td><td>10.0</td><td>47.8</td><td>13.8</td><td>46.8</td><td>12.9</td></tr><tr><td>DeCo (Ours)</td><td>41.2 ↓17.6</td><td>14.4 ↓9.3</td><td>27.0 ↓4.8</td><td>8.8 ↓1.1</td><td>37.8 ↓7.2</td><td>11.1 ↓3.6</td><td>42.2 ↓3.8</td><td>10.7 ↓1.8</td></tr><tr><td rowspan="3">Beam Search</td><td>Vanilla</td><td>55.6</td><td>15.8</td><td>30.6</td><td>9.5</td><td>48.8</td><td>13.9</td><td>41.8</td><td>10.8</td></tr><tr><td>OPERA</td><td>46.4</td><td>14.2</td><td>26.2</td><td>9.5</td><td>44.6</td><td>12.8</td><td>34.6</td><td>9.5</td></tr><tr><td>DeCo (Ours)</td><td>43.8 ↓11.8</td><td>12.7 ↓3.1</td><td>24.8 ↓5.8</td><td>7.5 ↓2.0</td><td>33.0 ↓15.8</td><td>9.7 ↓4.2</td><td>32.0 ↓9.8</td><td>8.7 ↓2.1</td></tr><tr><td rowspan="3">Nucleus</td><td>Vanilla</td><td>54.6</td><td>24.8</td><td>32.6</td><td>10.7</td><td>48.8</td><td>14.2</td><td>49.2</td><td>13.1</td></tr><tr><td>VCD</td><td>58.0</td><td>17.0</td><td>33.8</td><td>11.1</td><td>54.0</td><td>16.0</td><td>46.4</td><td>11.9</td></tr><tr><td>DeCo (Ours)</td><td>43.6 ↓11.0</td><td>12.9 ↓11.9</td><td>30.8 ↓1.8</td><td>9.5 ↓1.2</td><td>42.8 ↓6.0</td><td>13.2 ↓1.0</td><td>43.8 ↓5.4</td><td>11.8 ↓1.3</td></tr></table>
+
+Table 3: POPE hallucination evaluation results. The best results are in bold.  
+
+<table><tr><td>Decoding</td><td>Method</td><td>InstructBLIP F1↑</td><td>MiniGPT-4 F1↑</td><td>LLaVA-1.5 F1↑</td><td>Qwen-VL F1↑</td></tr><tr><td rowspan="3">Greedy</td><td>Vanilla</td><td>80.0</td><td>58.5</td><td>82.2</td><td>85.2</td></tr><tr><td>DoLa</td><td>83.4</td><td>72.8</td><td>83.2</td><td>85.8</td></tr><tr><td>DeCo (Ours)</td><td>84.9 ↑4.9</td><td>77.4 ↑18.9</td><td>86.7 ↑4.5</td><td>86.3 ↑1.1</td></tr><tr><td rowspan="3">Beam Search</td><td>Vanilla</td><td>84.4</td><td>70.3</td><td>84.9</td><td>85.3</td></tr><tr><td>OPERA</td><td>84.8</td><td>73.3</td><td>85.4</td><td>86.1</td></tr><tr><td>DeCo (Ours)</td><td>84.9 ↑0.5</td><td>77.9 ↑7.6</td><td>86.7 ↑1.8</td><td>86.4 ↑1.1</td></tr><tr><td rowspan="3">Nucleus</td><td>Vanilla</td><td>79.8</td><td>52.8</td><td>83.1</td><td>84.5</td></tr><tr><td>VCD</td><td>79.9</td><td>56.0</td><td>83.1</td><td>84.7</td></tr><tr><td>DeCo (Ours)</td><td>81.8 ↑2.0</td><td>63.8 ↑11.0</td><td>85.4 ↑2.3</td><td>85.2 ↑0.7</td></tr></table>
+
+Model. We select four of the most representative MLLM models for evaluation, including Instruct-BLIP (Dai et al., 2023), MiniGPT-4 (Zhu et al., 2024), LLaVA-1.5 (Liu et al., 2023b) and Qwen-VL (Bai et al., 2023). All the MLLMs we used have a language model size of 7 billion parameters (7B).
+
+Implementation Details. To select the appropriate preceding layers for hallucination mitigation, we conduct ablation experiments, details of which can be found in the Section 4.4. For a 7B-sized, 32-layer decoder-only architecture language model, we choose layers 20-28 as candidates for the preceding layers (according to the findings in Section 2.1). For the image captioning and VQA tasks,  $\alpha$  is set within the range of 0.1 to 0.6. In all experiments, we conduct inference on a single A800 GPU. The inference of 500 image-caption pairs take approximately 40 minutes.
+
+# 4.2 BENCHMARK AND METRICS
+
+CHAIR. Caption Hallucination Assessment with Image Relevance (CHAIR) (Rohrbach et al., 2018) metric, widely used in image captioning, identifies hallucinated objects by comparing the extracted objects with ground truth labels and evaluates both at the instance level  $\left(\mathrm{CHAIR}_{\mathrm{I}}\right)$  and sentence level  $\left(\mathrm{CHAIR}_{\mathrm{S}}\right)$ , as shown in Eq. 8. Following (Huang et al., 2024), we conduct experiments using the same settings, including the consistent 500 images from the MSCOCO 2014 validation dataset and the identical prompt, "Please help me describe the image in detail."
+
+$$
+\mathrm {C H A I R} _ {\mathrm {I}} = \frac {| \{\text {h a l l u c i n a t e d o b j e c t s} \} |}{\text {a l l m e n t i o n e d o b j e c t s}}, \mathrm {C H A I R} _ {\mathrm {S}} = \frac {| \{\text {c a p t i o n s w i t h h a l l u c i n a t e d o b j e c t s} \} |}{\text {a l l c a p t i o n s}}. \tag {8}
+$$
+
+POPE. The Polling-based Object Probing Evaluation (POPE) (Li et al., 2023b) is a VQA-based metric for assessing object hallucination in MLLMs. It evaluates hallucinations by asking questions such as "Is there a <object> in the image?" where <object> is derived from three types of splits: random (randomly selected objects), popular (frequently occurring objects), and adversarial (objects closely related to those in the image). The evaluation includes 500 MSCOCO images, with six questions per image for each split. We use F1 score for performance evaluation.
+
+MME. The comprehensive MLLM Evaluation benchmark (MME) (Fu et al., 2023) assesses the perceptual and cognitive abilities of MLLMs across a total of 14 subtasks, including tasks such as OCR, visual knowledge, attribute relationships, and object recognition.
+
+GPT-4o assisted evaluation. To further assess the model's performance in image captioning, we extend beyond the CHAIR metric, which targets object hallucination. Following prior studies (Huang et al., 2024; Leng et al., 2023), an open evaluation is conducted using GPT-4o on 100 randomly sampled COCO images. GPT-4o assesses two assistants' descriptions in terms of Accuracy (A) (e.g., truthfulness), Detailedness (D) (e.g., richness) and Coherence (C). We introduce the prompt used in the experiments in Table 10 and Table 11.
+
+# 4.3 EXPERIMENTAL RESULTS
+
+Results of hallucination in image captioning. Note that we use the baseline's original decoding settings for a fair comparison and run DeCo under the same settings. From Table 2, we notice that DeCo consistently outperforms other approaches in mitigating hallucinations across four MLLMs—InstructBLIP, MiniGPT-4, LLaVA-1.5, and Qwen-VL—using three decoding strategies: greedy search, beam search, and nucleus sampling. We find that DeCo slightly outperforms OPERA, while our method demonstrates higher efficiency and simplicity in inference (see Section 4.4). Additionally, VCD does not perform as well, likely due to producing an increased number of hallucinated descriptions during the generation process. In conclusion, the proposed approach DeCo effectively reduces hallucinations in visual description tasks solely through dynamic decoding correction, achieving an average suppression rate of approximately  $10.8\%$  on image captioning datasets. Additionally, we further evaluate the performance of DeCo on the AMBER image caption dataset, as detailed in Table 7 of the Appendix.
+
+![](images/0f3cf712199779cd42261e5637d2915189839d7191eaa7bcb80ad4b0c68f535c.jpg)  
+Figure 5: DeCo generally improves the MLLM's performance.
+
+Results of hallucination in VQA. In contrast to image captioning, POPE employs a simple polling approach to assess hallucination levels in MLLMs with respect to object recognition. As shown in Table 3, DeCo demonstrates superior performance across all settings, further validating the effectiveness of the proposed approach. Additionally, Figure 5 reveals that DeCo also achieves better results on MME, which evaluates the multifaceted VQA capabilities of LLaVA-1.5. These findings suggest that the underlying mechanism we identified not only applies to object recognition but also extends to attribute-related tasks and more complex reasoning tasks.
+
+Results of GPT-4o's assistance. Following (Huang et al., 2024; Leng et al., 2023), we further use GPT-4o to evaluate our method against greedy decoding across four distinct models. From Table 4, we notice that our approach consistently outperform greedy decoding in terms of accuracy, demonstrating its efficacy in hallucination suppression. The impact of decoding intervention is evident in the level of detail produced: for some models, our method yield only marginally higher or, in certain cases, slightly lower levels of detail compared to greedy decoding. DeCo also exhibits a coherence level comparable to that of the baseline. Nonetheless, our method exhibit a clear advantage in mitigating hallucinations across all evaluated models.
+
+Table 4: GPT-4o assisted hallucination evaluation results on MSCOCO. Three aspects are verified, accuracy  $(A)$ , detailedness  $(D)$  and coherence  $(C)$ .  
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">InstructBLIP</td><td colspan="3">MiniGPT-4</td><td colspan="3">LLaVA-1.5</td><td colspan="3">Qwen-VL</td></tr><tr><td>A</td><td>D</td><td>C</td><td>A</td><td>D</td><td>C</td><td>A</td><td>D</td><td>C</td><td>A</td><td>D</td><td>C</td></tr><tr><td>Greedy Search</td><td>4.92</td><td>5.65</td><td>6.89</td><td>5.71</td><td>6.20</td><td>7.67</td><td>5.21</td><td>6.31</td><td>8.18</td><td>5.56</td><td>6.62</td><td>8.20</td></tr><tr><td>DeCo (Ours)</td><td>6.25</td><td>5.77</td><td>7.14</td><td>6.33</td><td>6.08</td><td>7.54</td><td>7.42</td><td>6.25</td><td>7.96</td><td>7.81</td><td>6.70</td><td>8.15</td></tr></table>
+
+# 4.4 ANALYSIS
+
+Latency and throughput analysis. To evaluate the efficiency of DeCo, we compare its latency and throughput with several baselines, including DoLa, OPERA, and VCD based on Greedy, Beam Search, and Nucleus Sampling, respectively. Figure 6 illustrates the results of this comparison. The findings indicate that DeCo operates within an acceptable efficiency cost, striking a balance between effectiveness and computational overhead. Compared to the basic decoding process,
+
+the latency increase introduced by our method is approximately 1.2 times. In contrast, the latency increases for VCD and OPERA are 1.8 and 5.1 times, respectively. While both VCD and OPERA demonstrate comparable efficacy in mitigating hallucinations, their computational overheads remain relatively high. This highlights the practical value of DeCo, as it can be integrated into real-world applications without significantly compromising efficiency.
+
+Perturbation in the selected preceding-layer. To evaluate the effectiveness of the dynamic layer selection method, we introduce a random perturbation strategy. Specifically, for the predetermined preceding layers, we add random values ranging from -5 to 5 to modify the selection of layers. We randomly select 200 images from the MSCOCO dataset and prompt MLLMs to generate
+
+descriptions. The results after incorporating the perturbations are presented in Table 5. Notably, the perturbed results demonstrate a significant degradation in performance, further validating the effectiveness of our proposed method.
+
+![](images/3cb22bf61659e2279557d91aa404a6fa75e7c878f258b31b79c9516cc1f7f750.jpg)  
+Figure 6: Comparison of latency and throughput across different baselines.
+
+Table 5: Comparison of results between DeCo and perturbed DeCo in image captioning tasks  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">InstructBLIP</td><td colspan="2">MiniGPT-4</td><td colspan="2">LLaVA-1.5</td><td colspan="2">Qwen-VL</td></tr><tr><td>CHAIRS↓</td><td>CHAIRI↓</td><td>CHAIRS↓</td><td>CHAIRI↓</td><td>CHAIRS↓</td><td>CHAIRI↓</td><td>CHAIRS↓</td><td>CHAIRI↓</td></tr><tr><td>DeCo</td><td>39.3</td><td>12.6</td><td>32.4</td><td>9.6</td><td>38.8</td><td>11.1</td><td>44.5</td><td>11.1</td></tr><tr><td>DeCo + ε</td><td>45.6 ↑6.3</td><td>14.3 ↑1.7</td><td>33.3 ↑0.9</td><td>10.1 ↑0.5</td><td>42.2 ↑2.4</td><td>11.3 ↑0.2</td><td>47.0 ↑2.5</td><td>12.8 ↑1.7</td></tr></table>
+
+Hyperparameter analysis. Our method incorporates two primary hyperparameters:  $\alpha$  and the selection of interval layers. In the experiments, we employ DeCo based on greedy decoding. On the one hand, the hyperparameter  $\alpha$  regulates the intensity of early information enhancement. Figure 7(a) illustrates the performance across various  $\alpha$  values. We observe that hallucination suppression is most effective when  $\alpha$  approximates 0.6. As  $\alpha$  increases, the efficacy of DeCo in mitigating hallucinations improves. However, it is crucial to note that excessively high  $\alpha$  values may lead to the generation of atypical image descriptions, characterized by repetitive word usage. Notably, we can adjust the value of alpha appropriately to balance the truthfulness and semantic coherence of the responses (e.g., by using lower alpha). Additionally, our approach and the hyperparameter for repetition penalty are orthogonal, which implies that we can introduce the repetition penalty term to mitigate repetition. On the other hand, the layer interval hyperparameter  $[a,b]$  determines the candidate layers for inclusion in the enhancement process. We conduct experiments using intervals of four layers, with results presented in Figure 7(b). Our analysis reveals that hallucination suppression for MLLM is negligible in layers 1-16, while layers 20-28 demonstrate substantial mitigation of hallucinations. Notably, layers 29-32 exhibit minimal hallucination suppression, aligning with our findings discussed in Section 2.2. For other families of MLLMs and larger scale MLLMs, the selection of interval layer should be appropriately adjusted based on empirical experimentation.
+
+![](images/1c93f413cf339c667f5a990533e731e0e5e5f53a50e85ebcc6cdba77d69926ed.jpg)  
+(a) Ablation study of  $\alpha$
+
+![](images/4dd780a14001762b1183cb529dece58d697b95a2fdebb5aa66ddf4d5a2cff0f1.jpg)  
+(b) Ablation study of interval layers.  
+Figure 7: Ablation experiment results for hyperparameter  $\alpha$  and different interval layers.
+
+Mitigating snowballing hallucinations. Snowballing hallucinations are a prevalent issue in the responses generated by MLLMs. This phenomenon occurs when an initial hallucination triggers a
+
+sequence of subsequent errors, leading to a compounding effect that significantly degrades the quality and coherence of the generated text. Figure 8 illustrates a typical example of snowballing hallucinations, where an initial misinterpretation of the visual input propagates through the decoding process, resulting in a highly inconsistent and erroneous output. Our approach can reduce the accumulation of errors and improves the overall consistency and accuracy of the generated responses. The effectiveness of DeCo is further demonstrated through additional cases based on diverse MLLMs, which can be found in Figures 9, 10, 11, and 12 in Appendix G.
+
+# 5 RELATED WORK
+
+# 5.1 MLLM HALLUCINATION MECHANISM
+
+Hallucination in MLLMs, characterized by contradictions between image input and textual output, has been a prevalent issue (Liu et al., 2024a; Chen et al., 2024g). Current research on the mechanism of hallucination in MLLMs focuses on two key aspects: the interaction between images and text at different layers, and the prior bias of the LLM during decoding. Several studies have investigated the role of image-text interaction at different layers in MLLMs. Grad-CAM (Zhang et al., 2024a) visualizations reveal that image-text interaction exists in the preceding layers (1-11) but not in the deep layers. OPERA (Huang et al., 2024) further proposes that the "Aggregation Pattern" leads to hallucination, where visual information from preceding layers is gradually aggregated to anchor tokens, and focusing solely on these tokens during prediction while ignoring visual information leads to a high probability of hallucination in the generated sequence. However, other studies have revealed that MLLMs exhibit biases towards LLM priors, even in the presence of noisy or absent visual information. VCD (Leng et al., 2023) discovers that MLLMs generate high-confidence answers even when the image is noisy or absent, indicating a bias towards LLM priors. Similarly, PAI (Liu et al., 2024b) describes this phenomenon as "Text Inertia" and posits that it stems from existing paradigms that map visual representations onto the text representations as tokens. This leads to an inference process that fails to adequately account for image tokens, resulting in hallucinations.
+
+# 5.2 HALLUCINATION MITIGATION FOR MLLMS
+
+One straightforward approach to mitigate hallucination is to reduce the knowledge gaps and data bias between vision and language during model training. Finetuning-based methods have been explored, focusing on crafting specific datasets (You et al., 2024; Gunjal et al., 2024; Chen et al., 2024f) and alignment training (Sun et al., 2023; Yu et al., 2023; Chen et al., 2023; Li et al., 2023c) to achieve better knowledge alignment between images and text. While these methods have shown promising results, they often require expensive annotated paired data and substantial computational resources.
+
+Hallucination can also be mitigated by post-processing methods, which usually involve using additional tools or self-reflection strategies to revise the response. For instance, LURE (Zhou et al., 2024) detects hallucinations using manually-crafted features and revises the generated text accordingly. Woodpecker (Yin et al., 2023) combines MLLM outputs with an expert VQA model to post-edit hallucinations. However, these methods incur additional inference costs and delays, and require task-specific procedures and prompts to be designed (Xu et al., 2024). Training-free decoding methods have been explored to mitigate hallucination. OPERA (Huang et al., 2024) identifies an abnormal attention pattern that often accompanies hallucinated descriptions and proposes the mitigation method based on this pattern. VCD (Leng et al., 2023) introduces the notion that visual uncertainty increases hallucination and proposes a contrast decoding method to alleviate the issue. VDD (Zhang et al., 2024b) proposes a "Post-Hoc debias" approach that ensures uniform scores for each answer in the absence of an image to mitigate the influence of LLM priors.
+
+# 6 CONCLUSION
+
+In this paper, we demonstrate that MLLMs exhibit an awareness of hallucinated objects, with earlier layers showing higher confidence, while tokens shaped by prior knowledge diminish the likelihood of true tokens in the final layers. Based on this insight, we introduce DeCo, dynamic correction decoding with preceding-layer knowledge to mitigate hallucinations. Extensive experiments demonstrate the efficacy of our approach, which also shows advantages in latency and throughput.
+
+# ACKNOWLEDGMENTS
+
+This work was supported by the National Natural Science Foundation of China (No. 62206246, No. NSFCU23B2055, No. NSFCU19B2027), the Fundamental Research Funds for the Central Universities (226-2023-00138), Yongjiang Talent Introduction Programme (2021A-156-G), CIPSC-SMP-Zhipu Large Model Cross-Disciplinary Fund, Ningbo Science and Technology Special Projects under Grant No. 2023Z212, Information Technology Center and State Key Lab of CAD&CG, Zhejiang University, NUS-NCS Joint Laboratory (A-0008542-00-00), the Scientific Research Starting Foundation of Nanjing University of Aeronautics and Astronautics (No.1015-YAH24096), the High Performance Computing Platform of Nanjing University of Aeronautics and Astronautics, the Ministry of Education, Singapore, under the Academic Research Fund Tier 1 (FY2023) (Grant A-8001996-00-00), NUS-NCS Joint Laboratory (A-0008542-01-00). We gratefully acknowledge the support of Zhejiang University Education Foundation Qizhen Scholar Foundation.
+
+# REPRODUCIBILITY STATEMENT
+
+We have submitted the relevant code in the supplementary materials. The names of the experimental benchmarks, the prompt templates used, and the model's hyperparameter settings can all be found in Section 4. The Appendix C.1 and C.2 provides a detailed description of the experimental setup for the mechanism experiments.
+
+# REFERENCES
+
+Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023.  
+Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. CoRR, abs/2404.18930, 2024. doi: 10.48550/ARXIV.2404.18930. URL https://doi.org/10.48550/arXiv.2404.18930.  
+Paul Brie, Nicolas Burny, Arthur Sluyters, and Jean Vanderdonckt. Evaluating a large language model on searching for GUI layouts. Proc. ACM Hum. Comput. Interact., 7(EICS):1-37, 2023. doi: 10.1145/3593230. URL https://doi.org/10.1145/3593230.  
+Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. INSIDE: llms' internal states retain the power of hallucination detection. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024a. URL https://openreview.net/forum?id=Zj12nzlQbz.  
+Junying Chen, Ruyi Ouyang, Anningzhe Gao, Shunian Chen, Guiming Hardy Chen, Xidong Wang, Ruifei Zhang, Zhenyang Cai, Ke Ji, Guangjun Yu, Xiang Wan, and Benyou Wang. Huatuogptvision, towards injecting medical visual knowledge into multimodal llms at scale. CoRR, abs/2406.19280, 2024b. doi: 10.48550/ARXIV.2406.19280. URL https://doi.org/10.48550/arXiv.2406.19280.  
+Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. CoRR, abs/2403.06764, 2024c. doi: 10.48550/ARXIV.2403.06764. URL https://doi.org/10.48550/arXiv.2403.06764.  
+Shiqi Chen, Miao Xiong, Junteng Liu, Zhengxuan Wu, Teng Xiao, Siyang Gao, and Junxian He. In-context sharpness as alerts: An inner representation perspective for hallucination mitigation. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024d. URL https://openreview.net/forum?id=s3e8poX3kb.  
+Tianxiang Chen, Zhentao Tan, Tao Gong, Yue Wu, Qi Chu, Bin Liu, Jieping Ye, and Nenghai Yu. Llama slayer 8b: Shallow layers hold the key to knowledge injection, 2024e. URL https://arxiv.org/abs/2410.02330.
+
+Xiang Chen, Duanzheng Song, Honghao Gui, Chenxi Wang, Ningyu Zhang, Yong Jiang, Fei Huang, Chengfei Lyu, Dan Zhang, and Huajun Chen. Factchd: Benchmarking fact-conflicting hallucination detection. In Kate Larson (ed.), Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, pp. 6216-6224. International Joint Conferences on Artificial Intelligence Organization, 8 2024f. doi: 10.24963/ijcai.2024/687. URL https://doi.org/10.24963/ijcai.2024/687. Main Track.  
+Xiang Chen, Chenxi Wang, Yida Xue, Ningyu Zhang, Xiaoyan Yang, Qiang Li, Yue Shen, Lei Liang, Jinjie Gu, and Huajun Chen. Unified hallucination detection for multimodal large language models. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024, pp. 3235-3252. Association for Computational Linguistics, 2024g. doi: 10.18653/V1/2024.ACL-LONG.178. URL https://doi.org/10.18653/v1/2024.acl-long.178.  
+Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. CoRR, abs/2312.14238, 2023. doi: 10.48550/ARXIV.2312.14238. URL https://doi.org/10.48550/arXiv.2312.14238.  
+Ethan Chern, Jiadi Su, Yan Ma, and Pengfei Liu. ANOLE: an open, autoregressive, native large multimodal models for interleaved image-text generation. CoRR, abs/2407.06135, 2024. doi: 10. 48550/ARXIV.2407.06135. URL https://doi.org/10.48550/arXiv.2407.06135.  
+Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James R. Glass, and Pengcheng He. Dola: Decoding by contrasting layers improves factuality in large language models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. URL https://openreview.net/forum?id=Th6NyL07na.  
+Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A survey on multimodal large language models for autonomous driving. In IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, WACVW 2024 - Workshops, Waikoloa, HI, USA, January 1-6, 2024, pp. 958-979. IEEE, 2024. doi: 10.1109/WACVW60836.2024.00106. URL https://doi.org/10.1109/WACVW60836.2024.00106.  
+Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine (eds.), Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. URL http://papers.nips.cc/paper_files/paper/2023/hash/9a6a435e75419a836fe47ab6793623e6-Abstract-Conference.html.  
+Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive transformer. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020. URL https://openreview.net/forum?id= SJg7KhVKPH.  
+Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023.  
+Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. In Michael J. Wooldridge, Jennifer G. Dy, and Siraam Natarajan (eds.), Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2014, February 20-27, 2024, Vancouver, Canada, pp. 18135-18143. AAAI Press, 2024. doi: 10.1609/AAAI.V38i16.29771. URL https://doi.org/10.1609/aaai.v38i16.29771.
+
+Mingzhe Hu, Shaoyan Pan, Yuheng Li, and Xiaofeng Yang. Advancing medical imaging with language models: A journey from n-grams to chatgpt. CoRR, abs/2304.04920, 2023. doi: 10.48550/ ARXIV.2304.04920. URL https://doi.org/10.48550/arXiv.2304.04920.  
+Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. OPERA: alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation. CVPR, abs/2311.17911, 2024.  
+Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. URL https://openreview.net/forum?id=6FXtu8clyp.  
+Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. CoRR, abs/2311.16922, 2023. doi: 10.48550/ARXIV.2311.16922. URL https://doi.org/10.48550/arXiv.2311.16922.  
+Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. Contrastive decoding: Open-ended text generation as optimization. In Anna Rogers, Jordan L. Boyd-Graber, and Naoaki Okazaki (eds.), Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pp. 12286-12312. Association for Computational Linguistics, 2023a. doi: 10.18653/V1/2023.ACL-LONG.687. URL https://doi.org/10.18653/v1/2023.acl-long.687.  
+Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pp. 292-305. Association for Computational Linguistics, 2023b. doi: 10.18653/V1/2023.EMNLP-MAIN.20. URL https://doi.org/10.18653/v1/2023.emnlp-main.20.  
+Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multimodal models. CoRR, abs/2311.06607, 2023c. doi: 10.48550/ARXIV.2311.06607. URL https://doi.org/10.48550/arXiv.2311.06607.  
+Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large multi-modal model with robust instruction tuning. CoRR, abs/2306.14565, 2023a. doi: 10.48550/ ARXIV.2306.14565. URL https://doi.org/10.48550/arXiv.2306.14565.  
+Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. CoRR, abs/2402.00253, 2024a. doi: 10.48550/ARXIV.2402.00253. URL https://doi.org/10.48550/arXiv.2402.00253.  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023b.  
+Shi Liu, Kecheng Zheng, and Wei Chen. Paying more attention to image: A training-free method for alleviating hallucination in lvlms, 2024b. URL https://arxiv.org/abs/2407.21771.  
+Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding. CoRR, abs/2403.05525, 2024a. doi: 10.48550/ARXIV.2403.05525. URL https://doi.org/10.48550/arXiv.2403.05525.  
+Taiming Lu, Muhan Gao, Kuai Yu, Adam Byerly, and Daniel Khashabi. Insights into llm long-context failures: When transformers know but don't tell, 2024b. URL https://arxiv.org/abs/2406.14673.
+
+OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023. doi: 10.48550/ARXIV.2303.08774. URL https://doi.org/10.48550/arXiv.2303.08774.  
+Hadas Orgad, Michael Toker, Zorik Gekhman, Roi Reichart, Idan Szpektor, Hadas Kotek, and Yonatan Belinkov. Llms know more than they show: On the intrinsic representation of llm hallucinations, 2024. URL https://arxiv.org/abs/2410.02707.  
+Vipula Rawte, Amit P. Sheth, and Amitava Das. A survey of hallucination in large foundation models. CoRR, abs/2309.05922, 2023. doi: 10.48550/ARXIV.2309.05922. URL https://doi.org/10.48550/arXiv.2309.05922.  
+Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Jan 2018. doi: 10.18653/v1/d18-1437. URL http://dx.doi.org/10.18653/v1/d18-1437.  
+Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Tran, Yi Tay, and Donald Metzler. Confident adaptive language modeling. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh (eds.), Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. URL http://papers.nips.cc/paper_files/paper/2022/bitstream/6fac9e316a4ae75ea244ddcef1982c71-Abstract-Conference.html.  
+Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, Kurt Keutzer, and Trevor Darrell. Aligning large multimodal models with factually augmented RLHF. CoRR, abs/2309.14525, 2023. doi: 10. 48550/ARXIV.2309.14525. URL https://doi.org/10.48550/arXiv.2309.14525.  
+Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. CoRR, abs/2405.09818, 2024. doi: 10.48550/ARXIV.2405.09818. URL https://doi.org/10.48550/arXiv.2405.09818.  
+Surat Teerapittayanon, Bradley McDanel, and H. T. Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 23rd International Conference on Pattern Recognition, ICPR 2016, Cancún, Mexico, December 4-8, 2016, pp. 2464-2469. IEEE, 2016. doi: 10.1109/ICPR.2016.7900006. URL https://doi.org/10.1109/ICPR.2016.7900006.  
+Junyang Wang, Yuhang Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Ming Yan, Ji Zhang, and Jitao Sang. An llm-free multi-dimensional benchmark for mllms hallucination evaluation. CoRR, abs/2311.07397, 2023a. doi: 10.48550/ARXIV.2311.07397. URL https://doi.org/10.48550/arXiv.2311.07397.  
+Lean Wang, Lei Li, Damai Dai, Deli Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun. Label words are anchors: An information flow perspective for understanding in-context learning. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pp. 9840-9855. Association for Computational Linguistics, 2023b. doi: 10.18653/V1/2023.EMNLP-MAIN.609. URL https://doi.org/10.18653/v1/2023.emnlp-main.609.  
+Mengru Wang, Yunzhi Yao, Ziwen Xu, Shuofei Qiao, Shumin Deng, Peng Wang, Xiang Chen, Jia-Chen Gu, Yong Jiang, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. Knowledge mechanisms in large language models: A survey and perspective. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.), Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024, pp. 7097-7135. Association for Computational Linguistics, 2024a. URL https://aclanthology.org/2024.findings-emnlp.416.  
+Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's
+
+perception of the world at any resolution, 2024b. URL https://arxiv.org/abs/2409.12191.  
+Sheng Wang, Zihao Zhao, Xi Ouyang, Qian Wang, and Dinggang Shen. Chatcad: Interactive computer-aided diagnosis on medical image using large language models. CoRR, abs/2302.07257, 2023c. doi: 10.48550/ARXIV.2302.07257. URL https://doi.org/10.48550/arXiv.2302.07257.  
+Wenhai Wang, Jiangwei Xie, Chuanyang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. *Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving.* CoRR, abs/2312.09245, 2023d. doi: 10.48550/ARXIV.2312.09245. URL https://doi.org/10.48550/arXiv.2312.09245.  
+Ziwei Xu, Sanjay Jain, and Mohan S. Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. CoRR, abs/2401.11817, 2024. doi: 10.48550/ARXIV.2401.11817. URL https://doi.org/10.48550/arXiv.2401.11817.  
+Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024.  
+Tian Ye, Zicheng Xu, Yuanzhi Li, and Zeyuan Allen-Zhu. Physics of language models: Part 2.1, grade-school math and the hidden reasoning process. CoRR, abs/2407.20311, 2024. doi: 10.48550/ARXIV.2407.20311. URL https://doi.org/10.48550/arXiv.2407.20311.  
+Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. Woodpecker: Hallucination correction for multimodal large language models. CoRR, abs/2310.16045, 2023. doi: 10.48550/ARXIV.2310.16045. URL https://doi.org/10.48550/arXiv.2310.16045.  
+Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. URL https://openreview.net/forum?id=2msbbX3ydD.  
+Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, and Tat-Seng Chua. RLHF-V: towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. CoRR, abs/2312.00849, 2023. doi: 10.48550/ARXIV.2312.00849. URL https://doi.org/10.48550/arXiv.2312.00849.  
+Xiaofeng Zhang, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multi-modal large language models. CoRR, abs/2406.06579, 2024a. doi: 10.48550/ARXIV.2406.06579. URL https://doi.org/10.48550/arXiv.2406.06579.  
+Yi-Fan Zhang, Weichen Yu, Qingsong Wen, Xue Wang, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. Debiasing multimodal large language models. CoRR, abs/2403.05262, 2024b. doi: 10. 48550/ARXIV.2403.05262. URL https://doi.org/10.48550/arXiv.2403.05262.  
+Yuji Zhang, Sha Li, Jiateng Liu, Pengfei Yu, Yi R. Fung, Jing Li, Manling Li, and Heng Ji. Knowledge overshadowing causes amalgamated hallucination in large language models. CoRR, abs/2407.08039, 2024c. doi: 10.48550/ARXIV.2407.08039. URL https://doi.org/10.48550/arXiv.2407.08039.  
+Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. Analyzing and mitigating object hallucination in large vision-language models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. URL https://openreview.net/forum?id=oZDJKTlOUe.
+
+Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. URL https://openreview.net/forum?id=1tZbq88f27.
+
+# A LIMITATIONS
+
+Lack of generalized research. Due to the GPU cost consideration, we conduct experiments solely on limited MLLMs, without exploring additional MLLMs or those with larger parameter sizes.
+
+No free lunch. The results shown in Table 4 indicate that our method has a little negative impact on the level of detailedness metric. In future work, we aim to integrate DeCo with other strategies and explore approaches that can effectively balance truthfulness and diversity.
+
+# B COMPARATIVE ANALYSIS, SUMMARY, AND FUTURE DIRECTIONS
+
+Here, we compare our work with previous works, summarize and speculate on the underlying causes of hallucinations in LLMs and MLLMs, and provide insights into future directions.
+
+Comparison of previous mechanism findings. Existing studies suggest that MLLMs may focus more on visual tokens in the early layers while paying greater attention to textual tokens in the later layers (Zhang et al., 2024a; Chen et al., 2024c). The aggregation pattern is typically positively correlated with hallucinations and tends to emerge at deeper layers (Huang et al., 2024). These conclusions align with our findings, suggesting that MLLMs exhibit a better ability to perceive visual information in the preceding-layers compared to the final layers. In the detection of hallucinations in LLMs, some studies employing probing techniques have found that the intermediate layers exhibit the best detection performance (Chen et al., 2024a; Orgad et al., 2024; Chen et al., 2024e; Lu et al., 2024b), a finding similar to our Finding 1. This suggests that the hallucination mechanisms in LLMs and MLLMs may share underlying similarities.
+
+Comparison of previous hallucination mitigation methods. Our work shares a similar assumption with OPERA (Huang et al., 2024) and VCD (Leng et al., 2023), positing that the knowledge priors inherent in MLLMs may suppress the model's ability to comprehend visual information. However, our approach is comparatively simpler than that of OPERA (Huang et al., 2024) and VCD (Leng et al., 2023). Additionally, our work differs from the assumption in unimodal LLMs, where the semantic information present in the shallow layers interferes with factual recall in the final layer (Chuang et al., 2024; Chen et al., 2024d). However, our method is actually parallel to previous approaches and can be combined to achieve better results.
+
+Summary. Combining current research, we speculate that this phenomenon observed in both LLMs and MLLMs may be due to characteristics of the Transformer architecture, specifically the anchor token effect in the attention mechanism (Huang et al., 2024; Wang et al., 2023b), which leads to information loss when processing long sequences. For instance, in MLLMs, a single token may be insufficient to summarize information from extended sequences of visual tokens. Another work suggests that the knowledge overshadowing of multiple conditions within the query leads to hallucinations in LLM (Zhang et al., 2024c). In multimodal settings, image information represents a distinct condition. When the textual modality overshadows the condition related to the image, it can result in hallucinations in visual perception. Essentially, this reflects a loss of information flow within the attention mechanism. Overall, from an architectural perspective, hallucinations in both LLMs and MLLMs arise due to the imperfect of handling such interactions within the attention patterns of the Transformer.
+
+Future directions. Overall, the present works and ours work reveals notable similarities in the internal patterns of hallucination between LLMs and MLLMs. In future research, we will adopt a unified perspective to investigate the underlying causes of hallucinations in both LLMs and MLLMs.
+
+# C DETAILED EXPERIMENTAL SETUP
+
+# C.1 DETAILED SETTINGS FOR FINDINGS 1
+
+In the probing experiment, we utilize the pipeline proposed in the POPE (Li et al., 2023b) to construct 1,200 balanced positive and negative sample pairs from the MSCOCO dataset as training data for the probe classifier, where each sample consists of an object accompanied by a label indicating its existence or non-existence. (Note: There is no overlap between the training data and the evaluation data for object hallucination proposed by the POPE). We select the AMBER dataset (Wang et al., 2023a), which has a different distribution from the MSCOCO dataset, to test whether our conclusions can generalize. The AMBER dataset contains 1,004 carefully annotated images, each labeled with existent objects as well as non-existent objects. We use the prompt "Describe the image." to generate raw responses from LLaVA-1.5 on the images and then extract all object category tokens and label them with whether they exist. Given that the training set contains only 80 object categories, we denote the object tokens in test data belonging to these 80 categories as in-distribution (in-dist), while the remaining tokens are categorized as out-of-distribution (OOD).
+
+Previous work (Karamcheti et al., 2024) has demonstrated that increasing the resolution of the vision encoder enhances the visual comprehension capabilities of MLLMs. In our study, we compare LLaVA trained with a resolution of  $224\mathrm{px}$  against the original LLaVA with a resolution of  $336\mathrm{px}$  in probing experiments. Notably, the language model's weights differ between the two MLLMs, although both initial models are based on Vicuna-1.5-7b. Our results, as illustrated in the Figure 1(b), further affirm the scaling law associated with visual resolution, while also providing indirect validation of the reliability of the probing experiments.
+
+# C.2 DETAILED SETTINGS FOR FINDINGS 2
+
+In the early exit experiment, we randomly select 500 images from MSOCO and use random prompts (shown in Table 6) to elicit raw responses from LLaVA-1.5-7b. We then extract all non-existent objects along with their corresponding preceding text. Specifically, for the sentence "Additionally, there is a car.", we extract the hallucinated object token "car" and the preceding text "Additionally, there is a". We re-input the preceding text into the MLLM and observe the changes in its internal state when pre
+
+dicting the next token. We denote that a total of  $K$  preceding texts are selected, with the j-th preceding text denoted as  $s^j$ .
+
+Table 6: Randomly prompts.
+
+Prompts
+
+Describe the image.
+
+Please describe this image in detail.
+
+Generate a caption for this image.
+
+# D EVALUATION RESULTS IN AMBER
+
+The AMBER image caption dataset consists of 1,004 images, each accompanied by meticulously annotated labels. These annotations include all objects present in the images, as well as some potential hallucinated objects. AMBER employs four evaluation metrics: CHAIR (the proportion of generated hallucinated objects among all objects), Cover (the coverage of generated objects against all ground truth objects), Hal (the proportion of hallucinations among all generated captions), and Cog (the overlap ratio with potential hallucinated objects). Lower values of CHAIR, Hal, and Cog indicate higher truthfulness for the MLLMs, while a higher Cover value signifies better diversity. We compare Deco with the baselines on the LLaVA-1.5-7b. The results are as shown in Table 7. The results reveal that Deco demonstrates a significant advantage in truthfulness, although its diversity is somewhat lacking, yet remains within an acceptable range.
+
+Table 7: Results of using DeCo on the AMBER image caption dataset with LLaVA-1.5-7b.  
+
+<table><tr><td rowspan="2">Decoding</td><td rowspan="2">Method</td><td colspan="4">LLaVA-1.5</td></tr><tr><td>CHAIR ↓</td><td>Cover ↑</td><td>Hal ↓</td><td>Cog ↓</td></tr><tr><td rowspan="3">Greedy</td><td>Vanilla</td><td>8.2</td><td>48.9</td><td>34.3</td><td>4.0</td></tr><tr><td>DoLa</td><td>8.0</td><td>50.8</td><td>37.5</td><td>4.3</td></tr><tr><td>DeCo (Ours)</td><td>6.6 ↓1.6</td><td>47.5 ↓1.4</td><td>28.1 ↓6.2</td><td>2.8 ↓1.2</td></tr><tr><td rowspan="3">Beam Search</td><td>Vanilla</td><td>7.1</td><td>50.7</td><td>32.4</td><td>3.8</td></tr><tr><td>OPERA</td><td>6.4</td><td>49.0</td><td>27.5</td><td>2.9</td></tr><tr><td>DeCo (Ours)</td><td>6.3 ↓0.8</td><td>46.8 ↓3.9</td><td>25.1 ↓7.3</td><td>2.4 ↓1.4</td></tr><tr><td rowspan="3">Nucleus</td><td>Vanilla</td><td>10.2</td><td>50.2</td><td>43.3</td><td>4.5</td></tr><tr><td>VCD</td><td>9.0</td><td>51.7</td><td>40.2</td><td>4.4</td></tr><tr><td>DeCo (Ours)</td><td>8.3 ↓1.9</td><td>48.0 ↓2.2</td><td>37.5 ↓5.8</td><td>3.4 ↓1.1</td></tr></table>
+
+# E ABLATION RESULTS OF DYNAMIC SOFT MODULATION
+
+To quantify the effect of soft modulation, we remove the "max_prob" term and use greedy decoding to describe the images. The images and prompts used in the ablation experiment are consistent with the setup in Table 2, and the ablation results are presented in the Table 8. Additionally, we provide illustrative cases that demonstrate how soft modulation helps prevent abrupt changes in logits, as shown in the Figure 13.
+
+Table 8: Ablation study of dynamic soft modulation.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">LLaVA-1.5</td><td colspan="2">Qwen-VL</td></tr><tr><td>CHAIRS↓</td><td>CHAIRI↓</td><td>CHAIRS↓</td><td>CHAIRI↓</td></tr><tr><td>DeCo</td><td>37.8</td><td>11.1</td><td>42.2</td><td>10.7</td></tr><tr><td>DeCo (wo/max_probs)</td><td>41.2 ↑3.4</td><td>11.6 ↑0.5</td><td>45.8 ↑3.6</td><td>12.3 ↑1.6</td></tr></table>
+
+# F RESULTS OF MMVET
+
+We evaluate the performance of LLava-1.5-7B and Qwen-VL-7B on a comprehensive benchmark MMVet and use their default settings based on nucleus sampling. The results are as shown in Table 9.
+
+Table 9: Results of using DeCo on MMVet with LLaVA-1.5-7b and Qwen-VL-7b.  
+
+<table><tr><td>Model</td><td>Method</td><td>Rec ↑</td><td>OCR ↑</td><td>Know ↑</td><td>Gen ↑</td><td>Spat ↑</td><td>Math ↑</td><td>Total ↑</td></tr><tr><td rowspan="2">LLaVA-1.5</td><td>Vanilla</td><td>28.8</td><td>14.1</td><td>15.5</td><td>16.4</td><td>15.6</td><td>3.5</td><td>23.6</td></tr><tr><td>DeCo (Ours)</td><td>32.1 ↑3.3</td><td>21.5 ↑7.4</td><td>18.6 ↑3.1</td><td>20.7 ↑4.3</td><td>23.7 ↑8.1</td><td>11.2 ↑7.7</td><td>27.9 ↑4.3</td></tr><tr><td rowspan="2">Qwen-VL</td><td>Vanilla</td><td>51.8</td><td>35.3</td><td>41.0</td><td>35.6</td><td>38.1</td><td>19.2</td><td>45.7</td></tr><tr><td>DeCo (Ours)</td><td>50.5 ↓1.3</td><td>38.2 ↑2.9</td><td>38.8 ↓2.2</td><td>33.8 ↓1.8</td><td>41.7 ↑3.6</td><td>26.5 ↑7.3</td><td>46.3 ↑0.6</td></tr></table>
+
+# G CASE ANALYSIS ACROSS DIVERSE MLLMS
+
+![](images/7f0c459846758c1c0edbc14c0c236b15013e61ed8508ed52832e513fdeff054f.jpg)  
+Figure 8: The case of mitigating snowballing hallucination with DeCo.
+
+![](images/0e2d91095fc93d56788e1cf7bcf13a63d0241fdb07d27ba4282ae083c47ac823.jpg)  
+Input Image
+
+Instruction: Please describe this image in detail.
+
+# InstructBlip (Greedy):
+
+The image features a large red chair with the shape of an animal, specifically in this case being represented by two horses. One horse is standing on top and another one appears to be sitting atop it from behind its back legs towards front side viewing angle. There are several people visible throughout different parts around or near these animals: three individuals can been seen walking along either sides close together; there's also someone further away who seems more distant but still within reachable distance for interaction if needed. Additionally, some cars parked nearby add movement into scene as they come closer while driving past each other during their daily activities outside.
+
+# InstructBlip (Ours):
+
+The image features a large red chair with a horse statue placed on top of it. The horse statue is positioned on the chair's seat, making it appear as if the horse is actually sitting on the chair. The chair is placed on a grassy area, likely outside, and there are several trees visible in the background. The scene is set up as a sculpture, showcasing a unique and creative way of displaying a horse statue.
+
+Instruction: Please describe this image in detail.
+
+# InstructBlip (Beam search):
+
+The image depicts a group of three motorcyclists riding down a winding road through a mountainous area. The road is surrounded by tall, rocky cliffs on both sides, creating a dramatic and scenic setting. The motorcyclists are wearing helmets and appear to be enjoying the ride, taking in the breathtaking views around them. There are two motorcycles visible in the scene, with one closer to the center of the image and the other slightly off to the right side. In addition to the motorcycles, there are several cars parked along the side of the road, providing additional transportation options for those exploring the area.
+
+# InstructBlip (Ours):
+
+The image depicts a group of motorcycles traveling down a roadway, passing through a tunnel. The motorcycles are ridden by three people who are enjoying the scenic route through the tunnel. The motorcycles are positioned in various positions along the roadway, with one motorcycle being closer to the center of the image, another motorcycle slightly behind it, and another motorcycle further away from the center of the image. The third motorcycle is positioned further away from the others, but still visible in the background. The roadway itself is surrounded by trees and rocks, adding to the scenic nature of the scene.
+
+![](images/1a2484302149e3e7c25162dd5aab3292addf2431f6d5f73e81754d3e5211c20e.jpg)  
+Input Image
+
+![](images/cbd3eb78256ce8a2d4d2b56d19aac970de123779edd4911980e953b7f3fd7f77.jpg)  
+Input Image  
+Figure 9: DeCo's performance in reducing hallucinations of InstructBlip-7B on three basic decoing methods.
+
+Instruction: Please describe this image in detail.
+
+# InstructBlip (Nucleus):
+
+The image captures a skateboarder skillfully riding a rail in a blue skate park, likely practicing his tricks. The skateboarder is wearing a helmet, providing safety for the activity. There are multiple cars present in the background, including two parked on either side of the skate park. Additionally, there are two benches scattered around the scene, likely providing seating for the skateboarder and other park visitors.
+
+# InstructBlip (Ours):
+
+The image depicts a skateboarder skillfully skating on a rail near a blue railing at a skateboarding park. The skateboarder is wearing a helmet and appears to be engaging in a trick, performing a rail grind on the rail. The skateboarding park is well-equipped with several skateboarding obstacles, including a skateboarding ramp, a skateboarding rail, and a skateboarding ledge. These obstacles are designed to challenge skateboarders and provide them with the opportunity to practice their skills and showcase their skills in various ways.
+
+Instruction: Please describe this image in detail.
+
+![](images/ca0742f77c91021a2c771cffdc0a269684a6a28617e4c7a32820b841a7907af4.jpg)  
+Input Image
+
+LLaVA-1.5 (Greedy): The image features a woman sitting on a motorcycle, which is parked in front of a white building. She is wearing a leather jacket and appears to be posing for the camera. The motorcycle is positioned in the foreground, with the woman sitting comfortably on the seat.
+
+In the background, there is a car parked further away from the motorcycle. The scene captures the woman's enthusiasm for her motorcycle and her outdoor adventures.
+
+LLaVA-1.5 (Ours): The image features a woman sitting on a motorcycle parked outside a building. She is positioned in front of the motorcycle, occupying a majority of the frame. The motorcycle itself occupies a large portion of the scene, extending from left to right.
+
+The woman appears to be dressed in motorcycle gear, possibly preparing for a ride or already having completed one. The motorcycle's positioning suggests that it might be parked outside a garage or other similar location where motorcycle enthusiasts gather.
+
+Instruction: Please describe this image in detail.
+
+LLaVA-1.5 (Beam search): The image features a delicious-looking pizza sitting in a box on a dining table. The pizza is topped with tomatoes and cheese, making it an appetizing sight. Next to the pizza, there is a bottle of beer, which complements the meal perfectly. In addition to the pizza and beer, there is a knife placed on the table, likely for cutting and serving the pizza slices. The dining table occupies most of the space in the image, emphasizing the focus on the pizza and beer.
+
+LLaVA-1.5 (Ours): The image features a delicious looking pizza sitting inside a box, ready to be eaten. The pizza is topped with tomatoes and cheese, making it look delicious and freshly prepared. Next to the pizza box, there is a bottle of beer, adding to the overall enjoyment of the meal. The scene appears to be ready for someone to enjoy a delicious combination of pizza and beer.
+
+![](images/e81ac4f3f449d01fd9ad7e49a2da7e8a1b96c7712e35c39a6867d6f4b8b421de.jpg)  
+Input Image
+
+![](images/4f5b630a176e3d8c694f33e9eb8ecd310f76dfb7d4f005a50b6fcffa6de8087c.jpg)  
+Input Image  
+Figure 10: DeCo's performance in reducing hallucinations of LLaVA-1.5-7B on three basic decoing methods.
+
+# Instruction: Please describe this image in detail.
+
+LLaVA-1.5 (Nucleus): The image features a group of four giraffes in a lush green field, surrounded by trees and a pond. The giraffes are standing close to each other, with one on the left side, two in the middle, and one on the right side of the scene. They appear to be enjoying their time in the open field, possibly in a zoo or wildlife sanctuary.
+
+In addition to the giraffes, there are two birds visible in the scene, one near the left side and the other closer to the center. The birds are likely part of the wildlife in the area. The overall atmosphere of the image is serene and peaceful, showcasing the beauty of the natural environment.
+
+LLaVA-1.5 (Ours): The image depicts a group of four giraffes in a zoo enclosure, enjoying their time together in the grassy area. They are spread out throughout the scene, with one giraffe near the left side of the enclosure, another on the right side, one close to the middle, and another giraffe further back in the background.
+
+The zoo enclosure features a pond in the middle, where the giraffes can drink water or cool off during hot days. There are also several trees within the enclosure, providing the animals with shade and a natural habitat feel.
+
+![](images/6d6bbf03825a9048ff119159b5bc16f48c2e6ed275b46c0aab6a55c43249db8c.jpg)  
+Figure 11: DeCo's performance in reducing hallucinations of Qwen-VL-7B on three basic decoding methods.
+
+![](images/02de595b431b01b4e666b4ac70532d1a2955f62c7e5d485cc26b04fb6f077a1a.jpg)  
+Input Image
+
+Instruction: Please describe this image in detail.
+
+# MiniGPT4 (Greedy):
+
+This image shows two men standing in a living room. One man is holding a white object in his right hand, while the other man is standing behind him with his arms crossed in front of his chest. The room is decorated with yellow curtains and a blue couch. The furniture is cluttered with various objects, including a lamp, a vase, and a book. The walls are adorned with yellow and white wallpaper. The overall atmosphere of the image is casual and relaxed.
+
+# MiniGPT4 (Ours):
+
+This image depicts two men standing in a room with yellow walls and curtains. One man is holding a white object in his right hand while the other man is standing behind him with his arms crossed in front of him. The furniture in the room includes a blue couch and a chair. The room appears to be well lit with sunlight coming through the window.
+
+Instruction: Please describe this image in detail.
+
+# MiniGPT4 (Beam search):
+
+The image shows a large clock tower in the middle of a city street. The clock tower is made of metal and has two large clock faces on it. The clock tower is surrounded by tall buildings on either side of it. There is a black car parked on the sidewalk next to the clock tower. The car is facing the clock tower and has its lights on. There are people walking on the sidewalk in front of the clock tower. Some of them are looking up at the clock tower while others are walking past it. The buildings on either side of the clock tower have large windows and balconies. The street is lined with trees and streetlights.
+
+# MiniGPT4 (Ours):
+
+This image depicts a large clock tower in the middle of a city street. The clock tower is made of metal and has two clock faces on it. The clock tower is surrounded by buildings and cars passing by on the street below it. There is a black car parked on the sidewalk next to the clock tower.
+
+![](images/6f084c6b4ed9365fd0302b153361af6bd8a69598fd5043b9d574f49a96ae6f69.jpg)  
+Input Image
+
+![](images/afc28513c5d9d00a0eb108d7096500f0633bbe0b4e0309b99c0549341c196eb2.jpg)  
+Input Image  
+Figure 12: DeCo's performance in reducing hallucinations of MiniGPT4-7B on three basic decoding methods.
+
+Instruction: Please describe this image in detail.
+
+# MiniGPT4 (Nucleus):
+
+The image shows a white plate with a piece of meat, a bunch of broccoli, and some cashews. There is a knife and fork on the side of the plate, and a glass of water on the other side. The plate is on a brown tablecloth and the background is a brown wall.
+
+# MiniGPT4 (Ours):
+
+The image shows a plate of food consisting of broccoli, potatoes, and meat. The plate is white with a black rim and appears to be on a table or countertop. The food is arranged in rows with broccoli on one side of the plate and potatoes on the other side. A piece of meat is in the center of the plate.
+
+![](images/9c0dee01e1d3b58069bb0fb1c887fcc6ef33b2fc56bf1f144cae84b27a8b2de1.jpg)  
+Input Image
+
+Instruction: Please describe this image in detail.
+
+# Ours:
+
+The image depicts a bedroom scene where a bed occupies a large portion of the space. The bed itself is unique, resembling a circular shape with a bedroom divider in its center. The bedroom divider separates the bed from other parts of the room.
+
+There are two pillows on the bed, one on each side, adding comfort to the bedding arrangement. The bedroom also features a wall with a few windows, allowing natural light to illuminate the room.
+
+# Ours (wo/max_probs):
+
+The image depicts a bedroom scene where a bed occupies most of the space. The bed is unique due to its shape and style; it resembles a circular bed frame suspended above the floor. It appears to be made of metal and has a bedding arrangement consisting of pillows and blankets placed on top of it.
+
+There are two pillows visible on the bed, one located near the middle of the bed frame and another one towards the bottom part of it.
+
+Additionally, two blankets can be seen on the bed frame, one occupying the middle part of it and another one slightly higher than the first one.
+
+Instruction: Please describe this image in detail.
+
+# Ours:
+
+The image depicts two giraffes standing together inside a zoo enclosure. They are both tall and majestic animals, one occupying a majority of the space in the image while the other occupies a smaller portion. Both giraffes are feeding from a feeding basket hung high above them, likely providing them with food to eat.
+
+The zoo enclosure features concrete walls and a concrete floor, creating a solid structure for the animals to stand on. The scene captures the beauty of these two majestic animals interacting with each other and feeding together.
+
+# Ours (wo/max_probs):
+
+The image depicts two giraffes standing together inside a zoo enclosure. They are both tall and majestic animals, one occupying the left side of the frame while the other occupies the right side. Both giraffes are feeding on hay placed high above them, likely hung from a feeding pole or similar device.
+
+The zoo enclosure features concrete walls and a concrete floor, creating a solid structure for the animals to stand within. Additionally, there are two clocks visible in the background, one located above the other, likely meant for zoo staff to track time during feeding or other activities within the enclosure.
+
+![](images/46de50d02d47185472eac7c9a82e9e6e86ac2d254327979fb77654b17e80bb16.jpg)  
+Input Image
+
+Instruction: Please describe this image in detail.
+
+# Ours:
+
+The image depicts a young boy standing next to two horses in a grassy field. The boy is holding the horses' reins, guiding them as they stand together. The horses are positioned close to each other, one occupying the left side of the field and the other occupying the right side.
+
+The boy is wearing jeans and a green shirt, and he appears to be enjoying his time with the horses. The scene captures a moment of bonding between the boy and the horses, showcasing the connection between humans and animals.
+
+# Ours (wo/max_probs):
+
+The image depicts a young boy standing next to two horses, one white and one black, both equipped with harnesses. The boy seems to be handling both horses simultaneously, possibly preparing them for a ride or simply admiring them.
+
+The boy is positioned between the two horses, one occupying the left side of the frame while the other occupies the right side of the frame. Both horses are equipped with harnesses, one harnessed nearer to the boy's left side while the other harnessed nearer to his right side.
+
+![](images/7cf234e2da6d8c83f16cb3cab0004110b37cb0cb11edfc7c509b590e9aae0caf.jpg)  
+Input Image  
+Figure 13: Analysis of cases of soft modulation. Soft modulation avoids the hallucination phenomena and erroneous semantics caused by abrupt changes in logits.
+
+Table 10: The prompt used for GPT-4o evaluation adopted from Leng et al. (2023); Huang et al. (2024); Liu et al. (2024b)  
+
+<table><tr><td>GPT-4o Prompt</td></tr><tr><td>You are required to score the performance of two AI assistants in describing a given image. You should pay extra attention to the hallucination, which refers to the part of descriptions that are inconsistent with the image content, such as claiming the existence of something not present in the image or describing incorrectly in terms of the counts, positions, or colors of objects in the image. Please rate the responses of the assistants on a scale of 1 to 10, where a higher score indicates better performance, according to the following criteria:1: Accuracy: whether the response is accurate with respect to the image content. Responses with fewer hallucinations should be given higher scores.2: Detailedness: whether the response is rich in necessary details. Note that hallucinated descriptions should not count as necessary details.Please output the scores for each criterion, containing only two values indicating the scores for Assistant 1 and 2, respectively. The two scores are separated by a space. Following the scores, please provide an explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment.</td></tr><tr><td>[Assistant 1]{Response of Assistant 1}[End of Assistant 1]</td></tr><tr><td>[Assistant 2]{Response of Assistant 2}[End of Assistant 2]</td></tr><tr><td>Output format:Accuracy: &lt;Scores of the two answers&gt;Reason:</td></tr><tr><td>Detailedness: &lt;Scores of the two answers&gt;Reason:</td></tr></table>
+
+Table 11: The prompt used for GPT-4o to evaluate coherence.  
+
+<table><tr><td>GPT-4o Prompt</td></tr><tr><td>You are required to score the coherence of two AI assistants in describing a given image. Please rate the responses of the assistants on a scale of 1 to 10, where a higher score indicates better coherence.</td></tr><tr><td>[Assistant 1]</td></tr><tr><td>{Response of Assistant 1}</td></tr><tr><td>[End of Assistant 1]</td></tr><tr><td>[Assistant 2]</td></tr><tr><td>{Response of Assistant 2}</td></tr><tr><td>[End of Assistant 2]</td></tr><tr><td>Output format: Coherence: &lt;Scores of the two answers &gt;Reason:</td></tr></table>

ICLR/2025/MLLM can see_ Dynamic Correction Decoding for Hallucination Mitigation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e8d8511b970af36805b35a486e8241a5bfb8b1653e4c9fad404529aae978f62
+size 1520075

ICLR/2025/MLLM can see_ Dynamic Correction Decoding for Hallucination Mitigation/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36cfa85b3b2acf74ed87a59411686150582eda5a884a2bc36341e53c9c1773d6
+size 621486

ICLR/2025/MLLMs Know Where to Look_ Training-free Perception of Small Visual Details with Multimodal LLMs/e7b06a4a-efb3-4ae7-838e-8dbd8a6db2ef_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6a3621a50c39f8ed5d2bcb2af9a54ac4dd7b322be76ec07d20094a43f6f40b7
+size 124627

ICLR/2025/MLLMs Know Where to Look_ Training-free Perception of Small Visual Details with Multimodal LLMs/e7b06a4a-efb3-4ae7-838e-8dbd8a6db2ef_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8576ed40c7bc7eb86dcec69d883ba4b28c465294832c4ae57c9c273254ec9bd0
+size 153410

ICLR/2025/MLLMs Know Where to Look_ Training-free Perception of Small Visual Details with Multimodal LLMs/e7b06a4a-efb3-4ae7-838e-8dbd8a6db2ef_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4bd18b54a5e5d1367c66a3622ba1adadb4424d513f77ab8dce03caf0d72fc24
+size 18897736

ICLR/2025/MLLMs Know Where to Look_ Training-free Perception of Small Visual Details with Multimodal LLMs/full.md

@@ -0,0 +1,512 @@
+# MLLMS KNOWWHERE TO LOOK:   
+TRAINING-FREEPERCEPTIONOF   
+SMALL VISUALDETAILSWITHMULTIMODALLLMS
+
+Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara, Filip Ilievski
+
+University of Southern California, USA  
+Vrije Universiteit Amsterdam, The Netherlands
+
+# ABSTRACT
+
+Multimodal Large Language Models (MLLMs) have experienced rapid progress in visual recognition tasks in recent years. Given their potential integration into many critical applications, it is important to understand the limitations of their visual perception. In this work, we study whether MLLMs can perceive small visual details as effectively as large ones when answering questions about images. We observe that their performance is very sensitive to the size of the visual subject of the question, and further show that this effect is in fact causal by conducting an intervention study. Next, we study the attention patterns of MLLMs when answering visual questions, and intriguingly find that they consistently know where to look, even when they provide the wrong answer. Based on these findings, we then propose training-free visual intervention methods that leverage the internal knowledge of any MLLM itself, in the form of attention and gradient maps, to enhance its perception of small visual details. We evaluate our proposed methods on two widely-used MLLMs and seven visual question answering benchmarks and show that they can significantly improve MLLMs' accuracy without requiring any training. Our results elucidate the risk of applying MLLMs to visual recognition tasks concerning small details and indicate that visual intervention using the model's internal state is a promising direction to mitigate this risk.<sup>1</sup>
+
+# 1 INTRODUCTION
+
+Multimodal large language models (MLLMs) (Hurst et al., 2024; Team et al., 2024; Anthropic, 2024; Wang et al., 2024; Li et al., 2024a; Team et al., 2025; Chen et al., 2025) have greatly progressed the state of multimodal reasoning and planning, and are rapidly being integrated into various downstream applications, ranging from robotics (Li et al., 2024b; Chen et al., 2024), biomedicine (Li et al., 2023a), autonomous driving (Xu et al., 2024b; Zhang et al., 2023a) to visual mathematical reasoning (Gao et al., 2023; Zhang et al., 2024c;b) and even food recipe generation (Chhikara et al., 2024). Given the rapidly growing application of MLLMs, especially in critical domains such as biomedicine and security, it is crucial to study the limitations of their visual perception to elucidate the potential risks that may affect their downstream applications.
+
+To motivate the limitation that will be the focus of this work, we start by presenting three revealing visual question answering examples in Fig. 1, in which we ask a popular MLLM BLIP-2 (FlanT5XL) (Li et al., 2023b) to identify an object's presence or type in each image as we vary the size of the object. In the absence of any prior evidence, we might reasonably expect the MLLM's answer to be invariant to the size of the object, because of the MLLM's large representational capacity and pretraining on a wide variety of images containing objects of various sizes. To the contrary, in Fig. 1 (left), we observe that initially the model does not recognize the existence of a small street sign and assigns a lower probability to the correct answer; however, zooming into the image (via more focused visual cropping) towards the street sign gradually increases the probability assigned to the correct answer, suggesting that the model gradually perceives more and more relevant details of the street sign.
+
+![](images/1661af008e9c2d52de273bef750b3153c150c048d6e5567be013a37e58823fc9.jpg)
+
+![](images/9a89f3bb534c51a76b8f469cfbc6b38889fecea63aec59db287c85fadda1d722.jpg)  
+Figure 1: The effect of visual cropping on the probability of answers predicted by BLIP-2  $\mathrm{FlanT5}_{\mathrm{XL}}$  zero-shot VQA model. The x-axis labels are indices to the respective cropped images displayed under each plot that the model sees at each step. The model gradually finds the correct answer.
+
+![](images/e3f2718153fd1b0fa9a9cdb1b7ac787c4d6274a64610c3e892739a3c0932837a.jpg)
+
+![](images/4b999fc3f5249e48ad1cb3a3eb698858f9d6400a64fb659da54452fff32fc333.jpg)
+
+![](images/686df0e735e2ae09c0c9998e5fa90822d2ae4968bd63fc8f85ecd685fd3449cd.jpg)
+
+![](images/3e698339abeed4365188d42d0a3965e1027f48ef3c63e6ee4cf90a3b4b240eab.jpg)
+
+In Fig. 1 (middle), we observe further evidence of this difficulty in perceiving small details: the model initially predicts white as the type of the bird, but when we zoom into the image towards the bird, without changing the question in any way, we observe that the model gradually assigns higher probability to the correct bird type of egret. This suggests that the model was not making a semantic error of misunderstanding what type means, rather it was unable to perceive sufficient details to discriminate egret from other white birds, which is mitigated by visual cropping. Similarly, in Fig. 1 (right), we observe that the model's initial answer is not entirely irrelevant ("ama" compared to the correct answer "moma"), suggesting that the model knows where to look based on the question but cannot accurately perceive the actual word, which is again mitigated by visual cropping.
+
+In this work, we will study the limitation observed in Fig. 1 extensively, elucidate its cause, and propose potential solutions to mitigate its consequences. In Sec. 3, we quantitatively show that there indeed exists a difficulty in perceiving small visual concepts across various widely-used MLLMs. Our findings are consistent with prior works on evaluating the text-image matching in vision-language joint embedding models, which have observed a reverse correlation between visual object size in images and the text-image matching score (Zhao et al., 2022), but we further establish a causal connection between visual concept size and MLLMs' perception ability through an intervention study. In Sec. 4, we study whether the MLLMs' difficulty in perceiving small visual concepts stems from a difficulty in perceiving visual details, or from a difficulty in locating the concept due to its small size. We quantitatively show that MLLMs consistently know where to look, even when they fail to answer the question correctly. In Sec. 5, we propose three automatic visual cropping methods—leveraging the attention maps and gradients of the MLLM itself—as scalable and training-free solutions to the visual perception limitation. Finally, in Sec. 6, we apply our proposed methods to two popular MLLMs and evaluate them on seven visual question answering (VQA) benchmarks, showing their efficacy in enhancing MLLMs' accuracy, especially on detail-sensitive benchmarks.
+
+# 2 RELATED WORKS
+
+Multimodal Large Language Models (MLLMs). MLLMs are foundation models capable of handling diverse language and vision tasks. These models fall into two categories: end-to-end pretrained and modular pretrained. End-to-end models process joint image-language data through architectures such as dual-encoder (Radford et al., 2021), fusion-encoder (Li et al., 2021), encoder-decoder (Cho et al., 2021), and unified transformer (Wang et al., 2022), using objectives like image-text matching, contrastive learning, and masked language modeling. Modular pretrained models, which dominate recent state-of-the-art approaches, avoid costly full pretraining by adapting existing components: BLIP2 (Li et al., 2023b) and InstructBLIP (Dai et al., 2023) train a Transformer-based connector between a frozen pretrained ViT (Dosovitskiy et al., 2021) image encoder and a frozen LLM, which transforms ViT output tokens into a fixed set of image tokens in the input space of the LLM; Qwen-VL (Bai et al., 2023), similarly uses a fixed-length token connector (a single cross-attention layer), but trains both the connector and the LLM; LLaVA (Liu et al., 2023b) and LLaVA-1.5 (Liu et al., 2023a) instead use a linear projection and a two-layer MLP as their connectors, respectively, and train both. Our work will contribute to a better understanding of the perception limitations of MLLM and improve their perception scalability and without training, offering orthogonal benefits to existing approaches.
+
+Visual Localization Methods. Dedicated visual localization methods, such as YOLO (Redmon et al., 2016), SAM (Kirillov et al., 2023), and GLIP (Li et al., 2022b), rely heavily on dense spatial annotations for identifying salient image regions. Native approaches, such as Grad-CAM (Selvaraju et al., 2017), localize regions by analyzing gradients from classifier decisions without spatial supervision. Prior work adapts Grad-CAM to BLIP (Li et al., 2022a) leveraging its dedicated image-text similarity computation neural network called the Image-Text Matching network (Tiong et al., 2022; Guo et al., 2023). In this work, we derived a more general way for localizing the attention of MLLMs to images, not relying on the specific BLIP architecture. Several recent works have explored ways to improve the visual localization capability of MLLMs for visual question answering, including chain-of-thought (Shao et al., 2024; Liu et al., 2024b), tool-using (Wu and Xie, 2023), and visual programming approaches (Suris et al., 2023; Gupta and Kembhavi, 2023). In contrast, we demonstrate that MLLMs can often effectively localize the visual subject of a question in their internal states, and propose training-free methods to leverage their internal states for improving their visual perception.
+
+Visual Perception Limitations in MLLMs. The difficulty of answering questions about small objects in images has been observed by several prior and concurrent works (Zhang et al., 2023b; 2024a; Liu et al., 2024a; Wu and Xie, 2023), which have explored mitigating solutions based on high-resolution fine-tuning (Liu et al., 2024a; Dehghani et al., 2023; Wang et al., 2024), multi-agent pipelines (Wu and Xie, 2023), and use of visual cropping (Zhang et al., 2023b). In this work, we provide more extensive evidence for this difficulty, establish its causal effect on MLLMs' performance, and show that it is rooted in a failure to observe small visual details as opposed to a failure to locate small objects. Several works have also shown that MLLMs suffer from object hallucination (Li et al., 2023c; Yu et al., 2024). Furthermore, Zhang et al. (2024a) have shown visual blind spots in MLLMs—i.e., locations on the image where the MLLMs' perception degrades—as well as their sensitivity to visual quality, presence of visual distractors in the image, and even local object location perturbations.
+
+# 3 MLLMs' SENSITIVITY TO THE SIZE OF VISUAL CONCEPTS
+
+In this section, our goal is to quantitatively study our qualitative observations in Fig. 1 that MLLMs struggle with describing small visual details in images. To that end, we consider the TextVQA dataset, in which for each question we can find the image ground-truth bounding box that contains the correct textual answer. We partition its validation set into three groups based on the relative size of the ground-truth bounding box  $S = \frac{A_{bb}}{A_{total}}$ , where  $A_{bb}$  denotes the area of the ground-truth bounding box, and  $A_{total}$  the total area of the image: 1)  $S < 0.005$  (small) consisting of 773 question-image pairs, 2)  $0.005 \leq S < 0.05$  (medium) consisting of 2411 question-image pairs, and 3)  $S \geq 0.05$  (large) consisting of 1186 question-image pairs. We chose TextVQA for this study because it contains
+
+Table 1: Sensitivity of the accuracy of MLLMs to the size of visual concepts in TextVQA. As the relative visual size of the answer decreases (right to left in each row), we observe a decline in the accuracy of the original models (no cropping) in answering questions, whereas visual cropping (human-CROP) significantly improves accuracy on smaller objects.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Method</td><td colspan="3">Answer Bbox Size (S)</td></tr><tr><td>small</td><td>medium</td><td>large</td></tr><tr><td rowspan="2">BLIP-2 (FlanT5XL)</td><td>no cropping</td><td>12.13</td><td>19.57</td><td>36.32</td></tr><tr><td>human-CROP</td><td>55.76</td><td>52.02</td><td>45.73</td></tr><tr><td rowspan="2">InstructBLIP (Vicuna-7B)</td><td>no cropping</td><td>21.79</td><td>30.58</td><td>45.30</td></tr><tr><td>human-CROP</td><td>69.60</td><td>61.56</td><td>53.39</td></tr><tr><td rowspan="2">LLaVA-1.5 (Vicuna-7B)</td><td>no cropping</td><td>39.38</td><td>47.74</td><td>50.65</td></tr><tr><td>human-CROP</td><td>69.95</td><td>65.36</td><td>56.96</td></tr><tr><td rowspan="2">Qwen-VL (Qwen-7B)</td><td>no cropping</td><td>56.42</td><td>65.09</td><td>68.60</td></tr><tr><td>human-CROP</td><td>70.35</td><td>75.49</td><td>71.05</td></tr><tr><td rowspan="2">GPT-4o</td><td>no cropping</td><td>65.76</td><td>72.81</td><td>69.17</td></tr><tr><td>human-CROP</td><td>75.63</td><td>81.32</td><td>71.72</td></tr></table>
+
+a significant number of questions about small visual concepts, and textual answers are harder for MLLMs to guess from other side information in the image (compared to object types and attributes).
+
+Sensitivity Study. If a model's perception is not sensitive to the size of visual concepts, we expect it to have similar accuracy in all three partitions. In Tab. 1, we observe that the accuracy of all MLLMs declines as the ground-truth bounding box becomes relatively smaller (right to left on the no cropping rows). BLIP-2 and InstructBLIP are not trained on TextVQA (i.e., are zero-shot models), and their accuracy declines by 24 and 23 absolute percentage points between the large and small partitions, respectively. LLaVA-1.5 and Qwen-VL are trained on the training set of TextVQA, yet, their accuracy also declines by 11 and 12 between the large and small partitions, respectively. Lastly, even the most recent commercial GPT-4o, with an unknown training set that might include all of TextVQA, is suffering from a 7 percentage point decline in accuracy between small and medium partitions. These findings suggest that MLLMs have a bias against perceiving smaller visual concepts.
+
+**Intervention Study.** The perceptual limitation we observed above might be merely correlated with size. To study whether this limitation is causally related to size, we conduct an intervention study where we provide the MLLMs with visually cropped images based on the ground-truth bounding boxes, denoted as human-CROP. More specifically, for each image-question pair and each MLLM, we crop the smallest square-shaped region containing the ground-truth bounding box from the image, and resize it to the input image resolution of the MLLM (the square-shaped cropping prevents extreme deformation of the cropped image when resizing). The cropped image is then provided to the MLLM in addition to the original image-question pair (see more details in Fig. 4). We observe in Tab. 1 that human-CROP significantly improves the accuracy of all MLLMs on the small and medium partitions, and to a lesser extent on the large partition. These findings show that the perception limitation is indeed caused by the size of the visual concepts, and that visual cropping can be a promising direction to mitigate this limitation.
+
+# 4 Do MLLMs KNOW WHERE TO LOOK?
+
+The limitation in perceiving small visual concepts can have two primary reasons: 1) they are hard to locate in the larger image, and 2) their small details are hard to perceive correctly. In Fig. 1, we observed that the MLLM's incorrect answer may contain partially correct information, hinting that it might know where to look in the image. In this section, we quantitatively study that observation to answer whether MLLMs' sensitivity to size is rooted in a perception limitation or a localization limitation. To that end, we first utilize the attention maps computed inside the Transformer layers of an MLLM to quantify its spatial attention over the image and then compare the total amount of this attention inside the ground-truth bounding box to other bounding boxes of the same size.
+
+MLLMs' Setup. The considered MLLMs process a given image-question pair  $(x, q)$  in four steps (depicted in Fig. 4): 1) the image is divided into  $N \times N$  non-overlapping patches and processed by the ViT image encoder into  $N \times N$  output tokens; 2) the ViT output tokens are transformed into the input space of the backbone LLM—by either an MLP (LLaVA-1.5) or a Transformer connector (BLIP-2, InstructBLIP and Qwen-VL)—which we refer to as image tokens; 3) the image tokens are then prepended to the question tokens and a predefined starting answer token, and fed to the LLM; 4) the LLM is sampled auto-regressively following the starting answer token (we use greedy sampling).
+
+Quantifying MLLMs' Spatial Attention over the Image. We first measure how important each image token is to the MLLM's decision (answer-to-token attention) by extracting the softmax cross-attention of the starting answer token to all image tokens in all layers of the backbone LLM, resulting in  $A_{st}(x,q) \in \mathbb{R}^{L \times H \times 1 \times T}$ , where  $L, H$  are the number of layers and heads-per-layer in the LLM, and  $T$  is the number of image tokens provided to the LLM. We then take the average over attention heads to arrive at the answer-to-token attention  $\hat{A}_{st}(x,q) = \frac{1}{H}\sum_{h=1}^{H}A_{st}(x,q)$ . Next, we measure how important each image region is to each image token (token-to-image attention). For the MLLMs that use a Transformer connector to resample ViT output tokens into a fixed number of image tokens (BLIP-2, InstructBLIP and Qwen-VL), we extract the softmax cross-attention of each image token to all ViT output tokens in all layers of the connector, resulting in  $A_{ti} \in \mathbb{R}^{L_c \times H_c \times T \times N^2}$ , where  $L_c, H_c$  are the number of layers and heads-per-layer in the connector,  $T$  the number of learnable query tokens (that become input image tokens to the LLM), and  $N^2$  the number of image patches of the ViT image encoder. We then take the average over attention heads to arrive at the token-to-image attention
+
+![](images/dee525ac38d397579db2b5492045c82e8086ec9ff049917270ace558bd622204.jpg)  
+Figure 2: Examples of MLLMs knowing where to look despite answering incorrectly. The right panel in each example displays relative attention to the image (defined in Sec. 4) of one layer in the MLLM.
+
+$\hat{A}_{ti}(x) = \frac{1}{H_c}\sum_{h = 1}^{H_c}A_{ti}(x)$ . For LLaVA-1.5 which uses an MLP to transform ViT output tokens to image tokens, we set  $\hat{A}_{ti}(x)$  to the identity tensor. Finally, we compute the answer-to-image attention by computing the tensor product of the answer-to-token and token-to-image attention, resulting in  $A_{si}(x,q)\in \mathbb{R}^{L\times L_c\times 1\times N^2}$  where  $A_{si}^{mk}(x,q) = \hat{A}_{st}^{m}(x,q)\hat{A}_{ti}^{k}(x)$  (superscripts  $m$  and  $k$  denote layer indices on the LLM and the connector, respectively).
+
+Relative Attention. One issue with using the softmax cross-attention is that not all highly attended tokens are semantically relevant to the input question. For example, recent work has observed that Transformers may use several tokens as registers to aggregate global information (Darcet et al., 2023). To emphasize semantically relevant attention, we propose to normalize the answer-to-image attention of an image-question pair  $(x,q)$  by its value on a generic instruction  $q^{\prime}$ . Specifically, we consider a fixed instruction  $q^{\prime} =$  "Write a general description of the image," and compute relative attention as  $A_{rel}(x,q) = \frac{A_{si}(x,q)}{A_{si}(x,q^{\prime})}$  under element-wise division. Fig. 2 shows examples of relative attention for LLaVA-1.5 and InstructBLIP ( $A_{rel}^{mk}$  at layers  $m = 14$ ,  $k = 0$  and  $m = 15$ ,  $k = 2$ , respectively).
+
+Do MLLMs Know Where to Look? Equipped with relative attention, we now return to our question of whether MLLMs have a localization limitation or perception limitation. To that end, we consider the validation set of TextVQA again. For each image-question pair, we first compute the relative attention. We then define attention ratio as the ratio of the total (sum) relative attention inside the answer ground-truth bounding box to its average across all bounding boxes of the same size
+
+![](images/75e19021729123719c43e91b18ce5b0c577f065f95426aae77ed8b81d245e4b8.jpg)
+
+![](images/747d93fce1ed2a1d880752c6083744d1da594e22d067ed0da6bef29065c1054a.jpg)
+
+![](images/39f4d67af96ed3adfa555d2e814c9613c248df7869a2f7f44a80cbbd1fe021c7.jpg)  
+Figure 3: MLLMs' attention ratio across all layers (average with  $95\%$  CI over TextVQA). The attention ratio measures how significantly the MLLM is attending to the ground-truth bounding box (defined in Sec. 4). We observe that it is greater than 1 in most layers, showing that the MLLMs know where to look in the image even when they fail to answer correctly.
+
+![](images/0d7b659cb82bdc593e5d63cb73201ce727d195bcc7f01ba783410fc1f6402456.jpg)
+
+as the ground-truth on the image. This ratio provides a measure of how significantly the MLLM is attending to the ground-truth bounding box (in the sense of Markov's inequality). In Fig. 3, we plot the average (with  $95\%$  confidence interval) of the attention ratio, over the validation set of TextVQA for all layers in the considered MLLMs. The horizontal axis shows the combined layer index  $l = m + k \times L$  for  $m \in \{0 \dots L - 1\}$  spanning the number of cross-attention layers in the backbone LLM, and  $k \in \{0 \dots L_c - 1\}$  spanning the number of cross-attention layers in the connector (BLIP-2:  $L = 24$ ,  $L_c = 6$ ; InstructBLIP:  $L = 32$ ,  $L_c = 6$ ; Qwen-VL:  $L = 32$ ,  $L_c = 1$ ; LLaVA-1.5:  $L = 32$ ,  $L_c = 1$ ). In all MLLMs, we observe a significantly larger than 1 attention ratio in most layers, suggesting that the models are attending significantly to the ground-truth bounding box region on the image. Intriguingly, the models show similarly strong attention to the correct region regardless of whether they can answer the question correctly or incorrectly. These observations show that the MLLMs tend to know where to look, even when they answer incorrectly.
+
+# 5 AUTOMATIC VISUAL CROPPING (VICROP)
+
+We observed in Sec. 4 that the sensitivity of MLLMs to visual concept size is primarily a perception limitation (rather than a localization limitation). Therefore, one solution to mitigate this limitation is to simply train MLLMs with a larger number of image patches while maintaining per-patch resolution (hence increasing the image resolution of MLLMs). However, increasing the input image resolution by a factor of  $\alpha$ , increases the number of ViT input patches (and output tokens) from  $N^2$  to  $\alpha^2 N^2$ , which in turn increases the softmax attention computation complexity on the order of  $\alpha^4 N^4$ . Given that this solution is not scalable for current Transformer-based MLLMs, we choose to explore an alternative solution that does not require any training and is scalable to any image resolution. We note that several concurrent works have explored the first direction of training MLLMs with higher resolution image patches (Li et al., 2024c; Sun et al., 2024; Li et al., 2024d; McKinzie et al., 2024; Xu et al., 2024a; Luo et al., 2024), and notably LLaVA-Next (Liu et al., 2024a) has achieved the VQA state-of-the-art in several VQA benchmarks at the time of writing. We believe our work is parallel to these efforts in the following sense: rather than training higher and higher resolution MLLMs to enable them to see all resolutions (which is inevitably upper bounded), we explore how to smartly adjust the input image towards what an MLLM already can see without any additional training. We provide evidence showing that our training-free method can provide orthogonal benefits to the training-based methods in Appendices D and E.
+
+Inspired by our findings that MLLMs tend to know where to look (Sec. 4) and that visual cropping can mitigate the perception limitation (Sec. 3), in this section we construct three automatic visual cropping methods in order to mitigate the perception limitation of MLLMs. These methods seek to use the internal information of an MLLM itself—in the form of attention maps and gradients—to find the approximate region of interest in images (i.e., the region containing the subject of a question), and then to zoom into that region via visual cropping. One potential drawback of visual cropping is that some questions might need to have a global view of the image. To address this issue, we utilize the fact that MLLMs typically convert the image into a series of tokens. This allows us to directly
+
+![](images/2d314d5e442ed9c06ceb32a01d5aa39ce4988ef3d9d69c131814af43a262f7ed.jpg)  
+Figure 4: Illustration of the proposed visual cropping approach applied to two MLLMs.
+
+extend the original image tokens by concatenating the visually cropped image tokens, as illustrated in Fig. 4. We use this concatenation approach when applying all our methods to MLLMs.
+
+Relative Attention ViCrop (rel-att). In this method, we directly compute the relative attention  $A_{rel}(x,q)$  defined in Sec. 4 for each image-question pair  $(x,q)$ . We then select a target layer (in LLM and connector) based on a small held-out set of samples in TextVQA and use its relative attention as the importance map for visual cropping (discussed below). We ablate on the choice of layer in Sec. 6.
+
+Gradient-Weighted Attention ViCrop (grad-att). The relative attention runs an additional generic instruction through the MLLM to normalize the answer-to-image attention and emphasize semantically relevant attention. As an alternative that does not require a second forward pass, we consider using the gradients to normalize attention, because the gradient of the model's decision with respect to an attention score shows how sensitive the decision is to changes in that attention, hence how semantically relevant the attention is for answering the question. To get a differentiable representation of the model's decision, we consider the logarithm of the maximum output probability at the starting answer token,  $v = \log \text{softmax}(z(x,q))_{t^*} \in \mathbb{R}$ , where  $z \in \mathbb{R}^D$  is the output logit of the LLM at the starting answer position,  $D$  the vocabulary size, and  $t^* = \arg \max_t z_t$ . Then, following our notation in Sec. 4, we can compute the gradient-weighted versions of answer-to-token attention  $\tilde{A}_{st}(x,q) = A_{st}(x,q) \odot \sigma(\nabla_{A_{st}}v(x,q))$  and token-to-image attention  $\tilde{A}_{ti}(x,q) = A_{ti}(x) \odot \sigma(\nabla_{A_{ti}}v(x,q))$ , where  $\odot$  is element-wise product and  $\sigma(w) = \max(0,w)$ . We remove negative gradients because they correspond to tokens that if attended to will reduce the model certainty, hence often distracting locations Selvaraju et al. (2017). Finally, we compute the gradient-weighted answer-to-image attention as the following tensor product  $\tilde{A}_{si}(x,q) = \tilde{A}_{st}(x,q) \otimes \tilde{A}_{ti}(x,q) \in \mathbb{R}^{L \times L_c \times 1 \times N^2}$ . We will select the same target layer identified in rel-att from  $\tilde{A}_{si}(x,q)$  as the importance map for visual cropping.
+
+Input Gradient ViCrop (pure-grad). In this method, we seek to find the relevant regions on the image directly using the gradient of the MLLM's decision with respect to the input image. Compared to the previous attention-based methods, pure-grad is more versatile because it does not rely on the Transformer-based architecture. Specifically, for each image-question pair  $(x,q)$ , we will compute  $G(x,q) = \| \nabla_{x}v(x,q)\|_{2}$ , where  $v(x,q)$  is the logarithm of the maximum output probability of the LLM at the starting answer token (as defined in grad-att above), and the L2-norm is taken over the image channel dimension. However, gradients sometimes show high values in entirely constant-color regions (e.g., blue skies). Given that these non-edge regions do not contain any visual details, we will explicitly diminish them in  $G$ . To that end, we first apply a  $3\times 3$  size Gaussian high-pass filter to the image, followed by a median filter of the same size to reduce salt-and-pepper noise. The resulting filtered image is then thresholded at its spatial median value to become a binary mask and is element-wise multiplied by  $G$ . Finally, the edge-emphasized  $G$  is spatially average-pooled into the  $N\times N$  patches of the MLLM to become an importance map for visual cropping.
+
+Bounding Box Selection for Visual Cropping. To convert the importance map (from each of the methods described above) to a bounding box, we use sliding windows of different sizes inspired by object detection literature Redmon et al. (2016). Specifically, for each MLLM, we define a set of windows with sizes equal to a multiple of the input image resolution of the MLLM. The multiples are in  $\{1, 1.2, \ldots, 2\}$ . We slide each window over the image with a stride of 1 and find its best position where the sum of the importance map inside the window is maximized. After selecting the best position per window, we choose the window whose internal sum has the largest difference from the average internal sum of its adjacent positions. This latter step is a heuristic to avoid choosing too small or too large windows (notice that in both cases, moving the window slightly left/right or up/down will not change its internal sum significantly). The chosen window is then cropped from the image, resized to the input image resolution of the MLLM, and provided to the MLLM in addition to the image-question pair.
+
+High-Resolution Visual Cropping. In one of the benchmarks we consider in this work,  $\mathrm{V}^{*}$  Wu and Xie (2023), the images are of very high resolution (always more than 1K) and consequently, the resized input image provided to the MLLM might completely lose the visual concept of interest for a question. To mitigate this, on this benchmark, we employ a two-stage strategy. In the first stage, we divide images into non-overlapping blocks of smaller than  $1024\times 1024$  with an aspect ratio close to 1, compute the importance map separately for each block according to the ViCrop methods, and then re-attach the blocks back together. In the second stage, we find the bounding box for visual
+
+![](images/99fe9a4c88a06d342612187dba5bb306e590c6598a077d3d2b27dc32a082dee2.jpg)  
+Q: What is the color of the clock? (A) black (B) yellow (C) green (D) red
+
+![](images/5a8f3ef46ec6c640cc2809edc32e8c82d02409873fd9a3b7ae28fbea2ed10b64.jpg)  
+LLaVA 1.5:A  
+LLaVA 1.5 (w/ViCrop): C
+
+Q: What is the last on the list the lady is pointing at?  
+![](images/3d64131558d8ce4bb0bf98e4a1a3638e2cabaef8ceb46333eecc0fcd37e2fb2f.jpg)  
+InstructBLIP: Rudolph InstructBLIP (w/ ViCrop): Holland
+
+Q: What is the name of the player?  
+Figure 5: Examples of rel-att helping MLLMs correct their mistakes (cyan-colored bounding box shows cropped region by rel-att; zoom-in insets are displayed for better readability).  
+![](images/6cfac236c20cf8a50efe059b8011622274226e8705f2ddffa464920c3656f4b7.jpg)  
+LLaVA 1.5:10 LLaVA 1.5 (w/ ViCrop): Use numbers
+
+cropping on this re-attached importance map exactly as described before and provide the original image-question pair with the resized cropped image to the MLLM.
+
+# 6 VICROP METHOD ANALYSIS
+
+In this section, we apply our proposed visual cropping methods to two open-source SOTA MLLMs, InstructBLIP (Vicuna-7B) (Dai et al., 2023) and LLaVA-1.5 (Vicuna-7B) (Liu et al., 2023a). We evaluate their effectiveness in improving the perception of smaller visual concepts on 4 detail-sensitive datasets (TextVQA $^2$  (Singh et al., 2019),  $\mathbf{V}^*$  (Wu and Xie, 2023), POPE (Li et al., 2023c), DocVQA (Mathew et al., 2021)), and their ability to maintain performance on larger visual concepts in 3 general-purpose datasets containing mostly large objects (GQA (Hudson and Manning, 2019), AOKVQA (Schwenk et al., 2022), VQAv2 (Goyal et al., 2017)). InstructBLIP uses the hyperparameters  $N = 16$ ,  $m = 15$ ,  $k = 2$  and input image resolution of  $224 \times 224$ . LLaVA-1.5 uses  $N = 24$ ,  $m = 14$  and input image resolution of  $336 \times 336$ . When reporting accuracy, we compute VQA-score $^3$  for all benchmarks except GQA. For GQA, we compute accuracy using the official code. $^4$ . See Appendices A to C for further details about implementation, datasets, and prompts.
+
+ViCrop Improves VQA Accuracy. In Fig. 5, we show a few examples of the ViCrop helping the MLLM correct itself (more examples in Appendix G), and in Tab. 2, we report the accuracy of the proposed ViCrop methods on the VQA benchmarks. We observe that all methods significantly improve the accuracy of the original MLLMs (no cropping) on detail-sensitive benchmarks, without requiring any training, while maintaining the MLLMs' performance on benchmarks with larger visual concepts. Thus, the accuracy gain on fine details (most notably in TextVQA and  $\mathbf{V}^{*}$ ) does not seem to come at the cost of accuracy on larger visual details and relations. We also observe that the accuracy gain for LLaVA-1.5 is more substantial than for InstructBLIP. This can be explained by the
+
+Table 2: Accuracy of the proposed ViCrop methods on visual question answering benchmarks.  
+
+<table><tr><td rowspan="2" colspan="2">Model</td><td colspan="4">Smaller Visual Concepts</td><td colspan="3">Larger Visual Concepts</td></tr><tr><td>TextVQA†</td><td>V*</td><td>POPE</td><td>DocVQA</td><td>AOKVQA</td><td>GQA</td><td>VQAv2</td></tr><tr><td rowspan="4">LLaVA-1.5</td><td>no cropping</td><td>47.80</td><td>42.41</td><td>85.27</td><td>15.97</td><td>59.01</td><td>60.48</td><td>75.57</td></tr><tr><td>rel-att</td><td>55.17</td><td>62.30</td><td>87.25</td><td>19.63</td><td>60.66</td><td>60.97</td><td>76.51</td></tr><tr><td>grad-att</td><td>56.06</td><td>57.07</td><td>87.03</td><td>19.84</td><td>59.94</td><td>60.98</td><td>76.06</td></tr><tr><td>pure-grad</td><td>51.67</td><td>46.07</td><td>86.06</td><td>17.70</td><td>59.92</td><td>60.54</td><td>75.94</td></tr><tr><td rowspan="4">InstructBLIP</td><td>no cropping</td><td>33.48</td><td>35.60</td><td>84.89</td><td>9.20</td><td>60.06</td><td>49.41</td><td>76.25</td></tr><tr><td>rel-att</td><td>45.44</td><td>42.41</td><td>86.64</td><td>9.95</td><td>61.28</td><td>49.75</td><td>76.84</td></tr><tr><td>grad-att</td><td>45.71</td><td>37.70</td><td>86.99</td><td>10.81</td><td>61.77</td><td>50.33</td><td>76.08</td></tr><tr><td>pure-grad</td><td>42.23</td><td>37.17</td><td>86.84</td><td>8.99</td><td>61.60</td><td>50.08</td><td>76.71</td></tr></table>
+
+Table 3: Ablation study on the choice of layer and the use of high-resolution visual cropping.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2"></td><td colspan="3">Choice of Layer</td><td colspan="3">High-Resolution ViCrop</td></tr><tr><td>Selective</td><td>Average</td><td>Δ</td><td>w/ High-Res</td><td>w/o High-Res</td><td>Δ</td></tr><tr><td rowspan="4">LLaVA-1.5</td><td>no cropping</td><td>47.80</td><td>-</td><td>-</td><td>42.41</td><td>42.41</td><td>-</td></tr><tr><td>rel-att</td><td>55.17</td><td>55.45</td><td>+0.28</td><td>62.30</td><td>47.64</td><td>-14.66</td></tr><tr><td>grad-att</td><td>56.06</td><td>56.26</td><td>+0.20</td><td>57.07</td><td>49.74</td><td>-7.33</td></tr><tr><td>pure-grad</td><td>51.67</td><td>-</td><td>-</td><td>46.07</td><td>45.03</td><td>-1.04</td></tr><tr><td rowspan="4">InstructBLIP</td><td>no cropping</td><td>33.48</td><td>-</td><td>-</td><td>35.60</td><td>35.60</td><td>-</td></tr><tr><td>rel-att</td><td>45.44</td><td>44.40</td><td>-1.04</td><td>42.41</td><td>38.74</td><td>-3.67</td></tr><tr><td>grad-att</td><td>45.71</td><td>44.98</td><td>-0.73</td><td>37.70</td><td>42.41</td><td>+4.71</td></tr><tr><td>pure-grad</td><td>42.23</td><td>-</td><td>-</td><td>37.17</td><td>42.41</td><td>+5.24</td></tr></table>
+
+fact that InstructBLIP only trains its connector and not its backbone LLM during tuning—the LLM does not adapt to use the image tokens, rather the image tokens are adapted to optimally prompt the LLM—and therefore the LLM cannot effectively use the additional (cropped) image tokens provided through visual cropping. Nonetheless, the results show that ViCrop can be effectively applied to different MLLMs, and is a promising inference-time solution for mitigating the perception limitation observed in Sec. 3.
+
+Ablation Study on the Choice of Layer. To understand the importance of the choice of an informative layer for rel-att and grad-att (as discussed in Sec. 5), in Tab. 3 we compare the accuracy of these methods when simply taking the average of all layers in  $A_{rel}$  and  $\tilde{A}_{si}$ , respectively, on TextVQA. We observe that rel-att is robust to this choice and grad-att declines about 3.5 percentage points in accuracy. Importantly, both methods still improve the MLLMs' accuracy even when using the layer average, suggesting that averaging is a suitable choice in the absence of any data for selecting a layer.
+
+Ablation Study on the High-Resolution ViCrop. In Sec. 5, we proposed a two-stage strategy for processing the very high-resolution images in the  $\mathbf{V}^{*}$  benchmark. To see how effective this strategy is, in Tab. 3 we compare the accuracy of ViCrop methods with and without this high-resolution strategy on  $\mathbf{V}^{*}$ . We observe that while this strategy is very beneficial to LLaVA-1.5, it declines the performance of grad-att and pure-grad for InstructBLIP. However, all methods, with and without this strategy, still improve the MLLMs' accuracy.
+
+ViCrop with External Tools. In addition to the internal ViCrop methods, we also considered the use of external off-the-shelf models to find the region of interest in an image for visual cropping. Specifically, we utilized SAM (Kirillov et al., 2023),YOLO (Redmon et al., 2016), and CLIP (Radford et al., 2021) to find the most relevant part of an image to a given question (details of these external ViCrop methods are provided in Appendix F). In Tab. 4, we compare the accuracy of external ViCrop methods to the internal methods on TextVQA. While external models are also effective in improving the accuracy of MLLMs, they are weaker than all the proposed internal ViCrop methods, thus we did not explore them further.
+
+Table 4: Accuracy of ViCrop using external tools compared to attention/gradient (on TextVQA); and the inference time overhead of ViCrop methods (in seconds). Original's time is per answer token.  
+
+<table><tr><td></td><td>Model</td><td>Original</td><td>SAM</td><td>YOLO</td><td>CLIP</td><td>rel-att</td><td>grad-att</td><td>pure-grad</td></tr><tr><td rowspan="2">Accuracy (TextVQA)</td><td>LLaVA-1.5</td><td>47.80</td><td>49.42</td><td>48.84</td><td>48.55</td><td>55.17</td><td>56.06</td><td>51.67</td></tr><tr><td>InstructBLIP</td><td>33.48</td><td>39.23</td><td>36.49</td><td>39.61</td><td>45.44</td><td>45.71</td><td>42.23</td></tr><tr><td rowspan="2">CPU Time</td><td>LLaVA-1.5</td><td>2.26</td><td rowspan="2">91.53</td><td rowspan="2">0.97</td><td rowspan="2">5.46</td><td>14.43</td><td>11.33</td><td>29.86</td></tr><tr><td>InstructBLIP</td><td>0.66</td><td>4.35</td><td>3.78</td><td>7.04</td></tr><tr><td rowspan="2">GPU Time</td><td>LLaVA-1.5</td><td>0.17</td><td rowspan="2">3.33</td><td rowspan="2">0.35</td><td rowspan="2">1.07</td><td>1.16</td><td>0.89</td><td>2.36</td></tr><tr><td>InstructBLIP</td><td>0.06</td><td>0.28</td><td>0.29</td><td>0.60</td></tr></table>
+
+Inference Time Overhead. In Tab. 4, we report the average inference-time overhead of the proposed visual cropping methods on GPU (NVIDIA RTX A6000) and CPU (Intel(R) Gold 5317 CPU @ 3.00GHz) and compare with the per-answer-token processing time of the MLLMs. We see that all proposed methods (except SAM) are reasonably fast (1 to 2 seconds overhead on GPU). For example, computing the visual cropping with rel-att takes the time of generating only 5 tokens by the MLLM. Note that our methods' time overhead will not scale with the number of answer tokens and is constant regardless of how long the answer is because our external methods do not need any answer token, and internal methods only need the starting answer token (see Sec. 5). In contrast, MLLMs' inference time scales approximately linearly with the number of answer tokens.
+
+# 7 CONCLUSION
+
+In this work, we qualitatively and quantitatively showed that there exists a perception bias against small visual details in widely-used MLLMs. Then we found that MLLMs often know where to look even if they fail to answer the question, indicating that the bias toward small visual details is rooted in a perception limitation rather than a localization limitation. To mitigate this limitation, we proposed multiple automatic visual localization methods as scalable and training-free solutions based on models' internal dynamics while answering the visual questions. Through evaluation of multiple multimodal benchmarks, we showed that our method can significantly improve MLLMs' accuracy without requiring any training, especially in detail-sensitive scenarios. Our findings suggest that MLLMs should be used with caution in detail-sensitive applications, and that visual cropping/localization with the model's own knowledge is a promising direction to enhance their performance.
+
+Limitations and Future Work. The proposed ViCrop methods do not enhance all types of questions equally. We have observed that questions concerning relations and counting are particularly difficult for ViCrop methods to help answer. This is expected as the proposed ViCrop can only focus on one region in the image. We leave extending ViCrop to focus on multiple regions simultaneously for future work. Another limitation of the proposed methods is their time overhead and the addition of visual tokens. While the overhead is reasonable (a few seconds), we believe it can be significantly optimized as an inference-time mechanism, for example by utilizing lower precision, and weight quantization. Furthermore, Matryoshka Query Transformer (MQT) (Hu et al., 2024) enables MLLMs to have varying visual context size during inference. In our current results, we have shown that our methods can work with two different MLLMs with distinct visual context sizes, so it seems entirely possible that our method can still work with varying visual context size under MQT, which can further reduce our computational cost through rescaling the cropped image. We leave these inference cost optimizations to future works. Lastly, we have observed that the proposed methods tend to have some complementary benefits, and therefore exploring ways to combine them (for example based on the prediction uncertainty) is also a very interesting direction for future research.
+
+# ACKNOWLEDGMENTS
+
+We thank Jinyi Hu and Joe Mathai for their very useful insights. We also express our gratitude to anonymous reviewers for their valuable feedback. This research was supported in part by the National Science Foundation under Contract No. IIS-2153546.
+
+# REFERENCES
+
+Anthropic. The claude 3 model family: Opus, Sonnet, Haiku, March 2024. URL https:// www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/ Model_Card_Claude_3.pdf.  
+Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023.  
+Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14455-14465, June 2024.  
+Liang Chen, Lei Li, Haozhe Zhao, Yifan Song, and Vinci. R1-v: Reinforcing super generalization ability in vision-language models with less than $3. https://github.com/Deep-Agent/R1-V, 2025. Accessed: 2025-02-02.  
+Prateek Chhikara, Dhiraj Chaurasia, Yifan Jiang, Omkar Masur, and Filip Ilievski. Fire: Food image to recipe generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 8184-8194, 2024.  
+Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021.  
+Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning, 2023.  
+Timothée Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023.  
+Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36:2252-2274, 2023.  
+Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021.  
+Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023.  
+Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017.  
+Jiaxian Guo, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Boyang Li, Dacheng Tao, and Steven Hoi. From images to textual prompts: Zero-shot visual question answering with frozen large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10867-10877, 2023.
+
+Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023.  
+Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models. arXiv preprint arXiv:2405.19315, 2024.  
+Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019.  
+Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024.  
+Glenn Jocher, Ayush Chaurasia, and Jing Qiu. YOLO by Ultralytics. January 2023. URL https://github.com/ultralytics/ultralytics.  
+Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023.  
+Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024a.  
+Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. arXiv preprint arXiv:2306.00890, 2023a.  
+Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021.  
+Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022a.  
+Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023b.  
+Lianian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022b.  
+Xiang Li, Cristina Mata, Jongwoo Park, Kumara Kahatapitiya, Yoo Sung Jang, Jinghuan Shang, Kanchana Ranasinghe, Ryan Burgert, Mu Cai, Yong Jae Lee, et al. Llara: Supercharging robot learning data for vision-language policy. arXiv preprint arXiv:2406.20095, 2024b.  
+Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024c.  
+Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023c.  
+Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763-26773, 2024d.
+
+Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence 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.  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023a.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023b.  
+Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, January 2024a. URL https://llava-v1.github.io/blog/2024-01-30-llava-last/.  
+Shilong Liu, Hao Cheng, Haotian Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, et al. Llava-plus: Learning to use tools for creating multimodal agents. In European Conference on Computer Vision, pages 126-142. Springer, 2024b.  
+Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of-resolution adaptation for multimodal large language models. arXiv preprint arXiv:2403.03003, 2024.  
+Minesh Mathew, Dimosthenis Karatzas, and CV 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.  
+Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal llm pre-training. arXiv preprint arXiv:2403.09611, 2024.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021.  
+Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016.  
+Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VIII, pages 146–162. Springer, 2022.  
+Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017.  
+Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models, 2024.  
+Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019.  
+Hai-Long Sun, Da-Wei Zhou, Yang Li, Shiyin Lu, Chao Yi, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, De-Chuan Zhan, et al. Parrot: Multilingual visual instruction tuning. arXiv preprint arXiv:2406.02539, 2024.  
+Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. arXiv preprint arXiv:2303.08128, 2023.
+
+Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024.  
+Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599, 2025.  
+Anthony Meng Huat Tiong, Junnan Li, Boyang Li, Silvio Savarese, and Steven CH Hoi. Plug-and-play vqa: Zero-shot vqa by conjoining large pretrained models with zero training. arXiv preprint arXiv:2210.08773, 2022.  
+Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024.  
+Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022.  
+Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. arXiv preprint arXiv:2312.14135, 2023.  
+Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024a.  
+Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024b.  
+Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13807-13816, 2024.  
+Jiarui Zhang, Filip Ilievski, Kaixin Ma, Aravinda Kollaa, Jonathan Francis, and Alessandro Oltramari. A study of situational reasoning for traffic understanding. KDD, 2023a.  
+Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara, and Filip Ilievski. Visual cropping improves zero-shot question answering of multimodal large language models. In R0-FoMo: Robustness of Few-shot and Zero-shot Learning in Large Foundation Models, 2023b.  
+Jiarui Zhang, Jinyi Hu, Mahyar Khayatkhoei, Filip Ilievski, and Maosong Sun. Exploring perceptual limitation of multimodal large language models. arXiv preprint arXiv:2402.07384, 2024a.  
+Jiarui Zhang, Ollie Liu, Tianyu Yu, Jinyi Hu, and Willie Neiswanger. Euclid: Supercharging multimodal llms with synthetic high-fidelity visual descriptions. arXiv preprint arXiv:2412.08737, 2024b.  
+Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Ziyu Guo, Shicheng Li, Yichi Zhang, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, et al. Mavis: Mathematical visual instruction tuning with an automatic data engine. arXiv preprint arXiv:2407.08739, 2024c.  
+Tiancheng Zhao, Tianqi Zhang, Mingwei Zhu, Haozhan Shen, Kyusong Lee, Xiaopeng Lu, and Jianwei Yin. Vl-checklist: Evaluating pre-trained vision-language models with objects, attributes and relations. arXiv preprint arXiv:2207.00221, 2022.
+
+# A IMPLEMENTATION DETAILS
+
+We use python 3.10.6, transformers 4.29.1 and torch 2.1.2 for all the experiments. Our environment consists of an Intel(R) Gold 5317 CPU @ 3.00GHz with 48 cores and 756 GB of RAM, and we utilize NVIDIA RTX A6000 GPUs for our experiments. We use the huggingface implementations of all studied MLLMs with the recommended hyper-parameters according to the respective papers. For GPT-4o, we use the official public API, which is available at the time of submission.
+
+Regarding the evaluation setting of the TextVQA dataset in Tab. 2, our setting is slightly different from the one used by the LLaVA-1.5 original paper Liu et al. (2023a). They report accuracy on TextVQA by using externally extracted OCR tokens to enrich its text prompt. This is a text-specific trick that essentially out-sources the perception of text to an external OCR model. This text-specific trick is not mentioned in their paper or supplementary material, but see their clarification in response to a GitHub issue here: https://github.com/haotian-liu/LLaVA/issues/515#issuecomment-1763779341. In contrast, we treat TextVQA the same as any other vision dataset in our experiments, that is, we do not provide any OCR extracted tokens to MLLMs when applying them to TextVQA (only image and question, in the evaluation prompt format specified in their respective papers). This results in a slightly lower accuracy compared to the one reported in the original paper, but instead, this number shows the true perception ability of LLaVA-1.5 on TextVQA, not confounded by the ability of an external OCR model. For completeness, we also measured TextVQA accuracy in the presence of OCR tokens, which results in 59.8 for LLaVA-1.5 without any visual cropping, and 63.95 with rel-att, showing that our proposed visual cropping can still be beneficial even when OCR tokens are provided to the MLLM.
+
+# B DATASET STATISTICS
+
+In this section, we present the details of the datasets used for evaluation in this paper. We report the average height and weight of the images in the dataset. We also report the number of images and questions in each dataset. For VQAv2, we run our experiment on a random 50K subset of the official validation set. We use the entire validation set in all other datasets.
+
+Table 5: Average width  $(\bar{W})$  and height  $(\bar{H})$  of images, number of images, and number of questions on all datasets.  
+
+<table><tr><td></td><td>V*</td><td>DocVQA</td><td>TextVQA</td><td>POPE</td><td>AOKVQA</td><td>GQA</td><td>VQAv2</td></tr><tr><td>W</td><td>2246</td><td>1776</td><td>954</td><td>584</td><td>581</td><td>578</td><td>577</td></tr><tr><td>H</td><td>1582</td><td>2084</td><td>818</td><td>478</td><td>480</td><td>482</td><td>485</td></tr><tr><td># Images</td><td>191</td><td>1286</td><td>3166</td><td>500</td><td>1122</td><td>398</td><td>14206</td></tr><tr><td># Questions</td><td>191</td><td>5349</td><td>5000</td><td>8910</td><td>1145</td><td>10781</td><td>50000</td></tr></table>
+
+For our analysis presented in Table 1 and Figure 3, we focused on TextVQA dataset, which includes bounding box annotations for OCR-detected text within images. However, this dataset does not specify which bounding boxes correspond to the regions where answers are located, necessitating a manual annotation process. The TextVQA dataset comprises 5000 questions and 3166 images. We manually annotated these question-image pairs, ensuring accurate bounding boxes over all the regions of interest where the answers could be found. This manual annotation process was essential for our analysis, allowing us to provide precise and reliable ground-truth data for the study. Given that some questions were associated with multiple bounding boxes in their corresponding images, we undertook a filtering process to isolate the question-image pairs. This effort resulted in a refined set of 4370 question-image pairs, where there is only one instance of the subject of the question in the image. For example, if the question is "what type of drink is sold here?" and there are two different cans of drinks in the image, we remove this image-question pair.
+
+# C PROMPT FORMAT FOR ZERO-SHOT INFERENCE
+
+In this section, we provide details about the prompt format used in models for zero-shot inference. We use a different prompt format for LLaVA and InstructBLIP which we adapt from the original papers, as shown below.
+
+# LLaVA-1.5
+
+<image> USER:{question} Answer the question using a single word or phrase. ASSISTANT:
+
+# InstructBLIP
+
+<image> Question:{question} Short Answer:
+
+# D ORTHOGONAL BENEFITS TO LLAVA-NEXT
+
+We apply our proposed rel-att visual cropping method to an additional newer MLLM - LLaVA-NeXT (Liu et al., 2024a) current SOTA in several VQA benchmarks - that has support for higher-resolution compared to LLaVA-1.5. In Tab. 6, we observe that our method can still boost the MLLM's performance, without requiring any training. This provides further evidence for the generalizability of our proposed visual cropping and its orthogonal benefits to training MLLMs with higher image patch resolution.
+
+Table 6: Orthogonal benefits of visual cropping when applied to LLaV-NeXT that is trained to adapt to processing high-resolution images.  
+
+<table><tr><td>Model</td><td>TextVQA</td><td>V*</td></tr><tr><td>LLaVA-NeXT (Mistral-7B)</td><td>65.17</td><td>58.11</td></tr><tr><td>LLaVA-NeXT (Mistral-7B) + rel-att</td><td>68.65</td><td>61.78</td></tr></table>
+
+# E COMPARISON WITH THE V* METHOD (SEAL)
+
+The V* method (SEAL) (Wu and Xie, 2023) proposes a multi-agent fine-tuning approach to enhance the ability of an underlying MLLM to answer questions about small visual concepts. However, SEAL requires substantial training and finetuning of several neural networks, whereas our methods are completely training-free, so a direct comparison would not be fair. Nonetheless, to provide an idea of how our method compares to SEAL in an "as-is" fashion (i.e. if a user just wants to pick one method as-is off-the-shelf), we report the accuracy of SEAL compared to LLaVA-1.5+rel-att in Tab. 7. We observe that our method outperforms SEAL except on the V* benchmark. We think this might be because SEAL is designed and tuned specifically toward high-resolution images in its V* benchmark. We also note that the inference time of SEAL is slower than our method (4.44s compared to 1.88s on average per question, tested on the same random 100 TextVQA samples with one A6000 GPU). That being said, we note that our methods and SEAL can both help enhance MLLMs, and our methods can be integrated into SEAL or other multi-agent pipelines.
+
+Table 7: Performance comparison between our rel-att applied on LLaVA-1.5 and SEAL (Wu and Xie, 2023) across multiple vision-language benchmarks.  
+
+<table><tr><td>Model</td><td>TextVQA</td><td>V*</td><td>POPE</td><td>DocVQA</td><td>AOKVQA</td><td>GQA</td><td>VQAV2</td></tr><tr><td>SEAL</td><td>36.30</td><td>75.30</td><td>82.40</td><td>5.31</td><td>55.34</td><td>50.18</td><td>65.35</td></tr><tr><td>LLaVA-1.5+rel-att</td><td>55.17</td><td>62.30</td><td>87.25</td><td>19.63</td><td>60.66</td><td>60.97</td><td>76.29</td></tr></table>
+
+# F EXTERNAL TOOLS VICROP
+
+In this section, we present three automatic question-guided localization methods based on popular off-the-shelf vision-based models, namely CLIP Radford et al. (2021), YOLO Redmon et al. (2016), and SAM Kirillov et al. (2023). These three methods utilize external vision-based knowledge for the localization process through multimodal encoding, object detection, and semantic segmentation, respectively. See Tab. 4 for their results compared to internal ViCrop methods.
+
+CLIP ViCrop. The intuition of this method is to progressively refine the image towards the region of highest relevance to a given question using CLIP Radford et al. (2021). CLIP consists of an image encoder and a text encoder, which are trained on a large dataset of image-caption pairs to map each image (caption) close to its caption (image) and far from all other captions (images). The result is an aligned shared space where various images can be directly compared with various texts. To find the region of interest, given an image-question pair, we first crop the image from the four sides (top, bottom, left, and right) at a cropping ratio of 0.9 to produce four overlapping cropped images. We then use CLIP to assess the semantic similarity between these cropped images and the question. The highest-scoring crop is chosen as the input for the next iteration. This process is repeated for 20 iterations, and the cropped image with the highest CLIP similarity to the question is selected for visual cropping.
+
+YOLO ViCrop. Instead of a progressive approach to finding the region of interest, in this method we select candidate regions based on a state-of-the-art object detection method: YOLOv8 (Jocher et al., 2023) pretrained on COCO Lin et al. (2014). Using YOLO, we filter out regions that contain no salient objects - i.e., regions for which CLIP could mistakenly assign high similarity. More concretely, for each question-image pair, we first use YOLO to collect bounding boxes for all predicted objects with confidence higher than 0.25 (the recommended default). Then, for each predicted bounding box, we crop its corresponding image and compute its similarity to the question using CLIP. Finally, the bounding box with the highest similarity score is selected as the region of interest for visual cropping.
+
+SAM ViCrop. A limitation of YOLO is that it only provides bounding boxes corresponding to a fixed number of object classes. To overcome this issue, we use the segment anything model (SAM) Kirillov et al. (2023), which has shown state-of-the-art zero-shot segmentation performance. SAM can provide an extensive set of segmentation masks for each image, thus providing a more granular set of salient candidate regions compared to YOLO. More concretely, for each image-question pair, we feed the image into SAM, which provides an extensive set of segmentation masks corresponding to all objects and object parts. Then, we translate these masks into bounding boxes by computing the smallest bounding box that covers each segmentation mask. Finally, the bounding box with the highest CLIP similarity to the question is selected as the region of interest for visual cropping.
+
+Finally, for each method, we crop the smallest covering square (so that the cropped image is not deformed when resized to the input resolution of the MLLM), and provide it to the MLLM in addition to the original image-question pair (as depicted in Fig. 4).
+
+# G ADDITIONAL EXAMPLES ON MODEL'S PREDICTIONS
+
+![](images/bc27cd4013c46bdf69800fde34631678a0959411a082fc18c16467368da3509d.jpg)  
+Figure 6: Success (first 3) and failure (last) examples of LLaVA-1.5 (rel-att) on the  $\mathbf{V}^*$  benchmark (cyan-colored bounding box shows cropped region by rel-att; zoom-in insets are displayed for better readability).
+
+![](images/9026944d5f3add9915c5b9fd0f914ed27408132297805efe85a2bac952a080ff.jpg)  
+Q: which brewery made this ale?
+
+![](images/13ae26209045786267680fce9e2217d47137b70f3ec2f635b746d0a6d0a0c5e4.jpg)  
+Q: what team is written on the baseball in the right hand corner?  
+LLaVA 1.5: Yankees  
+LLaVA 1.5 (w/ ViCrop): Riverdogs
+
+![](images/e24e98da288186c2245f06563a93c2480bb7e5d6edda76bc64cc42bf84689ec7.jpg)  
+Q: is city of ballard in europe?
+
+![](images/f5283aa2fbbc3556a8fed281e74ae89d60ca5b38882d94c397813c91bf9bec9d.jpg)  
+LLA 1.5: Perrys  
+LLA 1.5 (w/ VIcrop): C
+
+![](images/af0a2b458f2c6419f4f637fc0f327dad79eb49b9f750574edcb332cdecac38ea.jpg)  
+Q: what number is on the middle bike?  
+LLaVA 1.5: 39  
+LLaVA 1.5 (w/ ViCrop): 30
+
+![](images/760f1c32f8c03d001f250e757cc6ae791b37ed1c2e0d11ef50ab7f457d25de09.jpg)  
+LLAVA1.5: no  
+LLAVA1.5 (w/ VICrop): yes  
+Q: what is the company that made this game / character?
+
+![](images/d1b8810e681627b82571fe28e91ae05862639dce0a751bfb2e49a5ac57512eb2.jpg)  
+Q: what is the brand of the monitor?  
+LLaVA1.5: Postugo  
+LLaVA1.5 (w/ ViCrop): Positivo  
+Q: what does the sign at the crosswalk say?  
+LAVA1.5: Go LAVA1.5 (w/ VIcCrop): 10av
+
+![](images/c7186e5bf821a90a4b18243e2e841c6a073ee6c530b9692d0350da84ef2ffd57.jpg)  
+Q: what number is on the man in whites jersey?
+
+![](images/c5d30535537e9efe0768190035e6604f1cce022b30a289857d2312188cf21a2b.jpg)  
+LLAva 1.5: Crazy brick  
+LLAva 1.5 (w/ VIcrop): Steve Jackson games  
+Q: what country is miss universe from?  
+LAVA1.5 Usa LAVA1.5 (w/ VIcrop): Canada
+
+![](images/428a22de65ab69999d15a05b8a5e681fc9ff44745b7c54d55106bab2a5929bbb.jpg)  
+Q: what does the bubble text say for the woman?
+
+![](images/7669249f855ae3e6f1e07e49c4388083a1c70e82c24687d89129d67b846f6dd9.jpg)
+
+![](images/55de1b899e9471570b323bf3d943e1fc3da1b59eb8efd6852fc79ad9693427d0.jpg)  
+Q: what is styped in the search bar?
+
+![](images/a97e10368442b5a4a6ccd9ef17127401f36a111e14f74d841d50fedb881f6e4b.jpg)  
+Q: who is the picture of?
+
+![](images/93c0c362f8506c8aa5dc5cfbff5fe52451382d7c6bad528b4960f68a642c69cf.jpg)  
+LAva1.5: Hold it boys! LAva1.5 (w/ ViCrop): Hold it  
+Q: what is the number of the player in the middle?  
+LAVa1.5: 3 LAVa1.5 (w/VIcrop): 22  
+Figure 7: Success (first 9) and failure (last 6) examples of LLaVA-1.5 (rel-att) on the TextVQA benchmark (cyan-colored bounding box shows cropped region by rel-att).
+
+![](images/e1be1512d042c5e9d859220244c1f0fe5e3807bccfc71da68231d9e9954a8374.jpg)  
+LAVa1.5: Sushi LAVa1.5 (w/ ViCrop): Buscar diercion  
+Q: what is the wine brand?  
+LAVa1.5: Dog house LAVa1.5 (w/ VICrop): Masseto
+
+![](images/1271d2ef3eabc1ed7766da9e5c9c1a24fbc78321905bfaf3eb94cb4a67fc487a.jpg)  
+LLAVA 1.5: Charles grayson  
+LLAVA 1.5 (w/ VIcCrop): Man  
+Q: what kind of exchange shows on the banner?  
+LLA1V.5:Policy LLA1V.5(w/VIcrop): Sab miller
+
+![](images/6f821206ff1007628637f910d178e1473ebb21288e4e41418c408e926c126e60.jpg)  
+Q: what letter is on the mans ball cap?
+
+![](images/0b066742ee04ae68575c41ba0110fcdd913390b633ccbff704066d469f3bf873.jpg)
+
+rectBLIP:
+
+structBLIP(w/ViCrop):
+
+T
+
+![](images/ded565c27a2760386cce6f7fb366a544520cae27e979a65c2e21da1e55231523.jpg)  
+Q: what is the name of the group on the bottom right of the poster?
+
+![](images/269cbf2078645b5f10029129025f0a2df690c556af4de5a093d2d2ede24c6b52.jpg)
+
+uctBLIP: Public
+
+InstructBLIP (w/ ViCrop):
+
+![](images/10605a4de18fd93d296c80c133871b6c89459e074ce649f0ebd7373a0a9b8ca2.jpg)  
+Q: what is written on the milk carton?
+
+![](images/df9e63febe20890e9ebf06005c1919b03fc27d7b0fc78276bdae510a56450778.jpg)
+
+ctBLIP:
+
+ctBLIP (w/ ViCrop):
+
+U.s. forced europe
+
+![](images/272d644815eb1bfddce7004c23e080b835001259d81c0dd0c6f6407ddaf797eb.jpg)  
+Q: what brand of bike wheel is this?
+
+![](images/a7457dd4ecd98f0b8aab028fa4096b05955a5eec43ebd2b36376223e4c1b814c.jpg)
+
+Pace
+
+structBLIP (w/ ViCrop):
+
+Pacenti
+
+![](images/cbfcdd7a3bb1d2747fcf9fbc14262db8f8961a0b9e8b3f0d3cf5b76f60ce7783.jpg)  
+Q: what is the name in the very top right hand corner of this page?
+
+![](images/2c591e1c8f01b3f71eeae2e1e7446ab9f22e978295e514ce08b031034fb26e72.jpg)
+
+FructBLIP: Shakespeare
+
+uctBLIP (w/ ViCrop): Newcome
+
+![](images/3febaa4f29ea708146a3f24315c81b2cce7448be7309a252275923d749351d6f.jpg)  
+Q: what does the sign say above the wheelchair symbol?
+
+![](images/4b2d6562929ff0f38cc70cc2cfcdba2945a8ae384ed099797fafc7f0a387a67d.jpg)
+
+uctBLIP:Men
+
+InstructBLIP (w/ ViCrop): Restroom
+
+![](images/2b6e2e0d5bbb39120d20d8b72e49ceb197eb58ee9d79c188beea83cf3f0b74b1.jpg)  
+Q: what is the number at the bottom right corner?
+
+InstructBLIP: 100
+
+InstructBLIP (w/ ViCrop): 16
+
+![](images/1be8a8409e09e27825b99fa043fdf5125aaa0f0ca22bca89a84379c3f6424290.jpg)  
+Q: more sharing what?
+
+![](images/4efa3383f80b37276b04e8a97d04582a201a50e46c07b0925d2a552657393f53.jpg)
+
+ctBLIP: Photo
+
+ctBLIP (w/ ViCrop):
+
+Options
+
+![](images/3279da1b476dfcac33f9da773145a5780ead32e0225a408797e845237c4f528a.jpg)  
+Q: what degree angle has been drawn?
+
+InstructBLIP:
+
+InstructBLIP (w/ ViCrop)
+
+90
+
+![](images/7cac7a6e62ba0427d8d66d0aa472dfcb55fd3a801ae91107fcafa096ea0e614c.jpg)  
+Q: what brand are the guy's shorts?
+
+![](images/03e8ba6a549310e9940204a11409d360a2f1fdc863e958ccb37856efe0f83a1a.jpg)
+
+StructBLIP: Gators
+
+InstructBLIP (w/ ViCrop):
+
+Nike
+
+![](images/cef84df67d7df592d67c2fb0ef467142249e7ba491912fec97597dc7d94dc9cf.jpg)  
+Q: what is the word written in the bottom of the box?
+
+![](images/5fbc0b3d05608f4c233d21a101fb41596ca2fb1389101e219c6d71548bebcaa3.jpg)  
+Q: what is the word under the state name?
+
+![](images/80c482c7ed1254caeaf1404335154e0a241114ed3ebfda4767513fbb9d42daae.jpg)  
+Figure 8: Success (first 9) and failure (last 6) examples of InstructBLIP (rel-att) on the TextVQA benchmark (cyan-colored bounding box shows cropped region by rel-att).
+
+FructBLIP: Hardcast
+
+rectBLIP (w/ ViCrop):
+
+Flexible
+
+![](images/a603e64b688694caee4054b58c1cbb6629dca02af0fef9975cbb6fe2d350869c.jpg)  
+Q: who is the service provider?
+
+![](images/132b8f0085a30f7a0955e27eea8d744c74875c374a1d8b9a7c2ce77e6662ee95.jpg)
+
+ctBLIP: Mophie
+
+InstructBLIP (w/ ViCrop):
+
+Apple
+
+![](images/640a64d46240462cfdda6ac7ba380b6a9d5eba7ee3fe9abf472528cde9157519.jpg)  
+Q: what is the manufacturer of these bullets?
+
+![](images/af612214c08536fadbbcb226bde9a3fb09be96bd634fcb79c10e749500616af7.jpg)
+
+ctBLIP: Karazy
+
+InstructBLIP (w/ ViCrop):
+
+California
+
+![](images/c1053aabb0a6b532eedb8742631e0d91ad284897619bc681625f74c0357e2553.jpg)  
+Q: what is the middle book title?
+
+InstructBLIP:
+
+InstructBLIP (w/ ViCrop):
+
+Messily
+
+![](images/db0326318f84cef3633db178469915d0f339660663d1d6fa3839613c07f78bbe.jpg)
+
+![](images/ccddc3d8939d5080c164ec0952f92879d7e480fbc70faf491f8c0d315860fcea.jpg)
+
+ctBLIP: Sears
+
+InstructBLIP (w/ ViCrop):
+
+Remington

ICLR/2025/MLLMs Know Where to Look_ Training-free Perception of Small Visual Details with Multimodal LLMs/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:536fcc537d89a90991558d3f638e6fc9b0eaea895c838cb1736397630256f396
+size 1206056

ICLR/2025/MLLMs Know Where to Look_ Training-free Perception of Small Visual Details with Multimodal LLMs/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48344a19ecf9167eec135638c0a722df8ada7955bbde2abb40f62ef9092b2daa
+size 634741

ICLR/2025/MLPs Learn In-Context on Regression and Classification Tasks/0bddb98a-b970-4ead-9708-0d5ea7cebe38_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe1485288df4d542117c22441ac62f7905d4d9e074b1c94f70d4cbdaec4c382f
+size 194599

ICLR/2025/MLPs Learn In-Context on Regression and Classification Tasks/0bddb98a-b970-4ead-9708-0d5ea7cebe38_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ef0ba080fcbd8476931535b75f1df7cc727be40395780671fb9f3e60c4e0f45
+size 221031

ICLR/2025/MLPs Learn In-Context on Regression and Classification Tasks/0bddb98a-b970-4ead-9708-0d5ea7cebe38_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f2399f30136aaef70ce78fc0cdad5b85d88e93aabae79d0a222f1c47b2fa7239
+size 10627168

ICLR/2025/MLPs Learn In-Context on Regression and Classification Tasks/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31d8f59f8d85273d985a38cd9b4f5d19c75297739b694d4b9624984a1d3685ff
+size 1135088

ICLR/2025/MLPs Learn In-Context on Regression and Classification Tasks/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45e06726c71280a3d7ebcb20aa99f12cb23c27f360004b3b5163277a75802668
+size 1154391

ICLR/2025/MM-EMBED_ UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS/65163941-032d-440e-8e2d-7c812aac6d5b_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a421a0926c6e23e89c6879f0f09c8eb8087422bde450d712a206abb3b33d0144
+size 128546

ICLR/2025/MM-EMBED_ UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS/65163941-032d-440e-8e2d-7c812aac6d5b_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45cb9c1e5dd03b055b5f09aaeff73cb8ec9aed207664c868a0a662af21065992
+size 152325

ICLR/2025/MM-EMBED_ UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS/65163941-032d-440e-8e2d-7c812aac6d5b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05e231b3b621e8b858ddce685f8a8aead44f1aac673d268423bd8ad57ccb3fc0
+size 4064395

ICLR/2025/MM-EMBED_ UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS/full.md

@@ -0,0 +1,316 @@
+# MM-EMBED: UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS
+
+Sheng-Chieh Lin  $^{*1,2}$  Chankyu Lee 1 Mohammad Shoeybi 1 Jimmy Lin 2
+
+Bryan Catanzaro<sup>1</sup> Wei Ping*<sup>1</sup>
+
+$^{1}$  NVIDIA  $^{2}$  University of Waterloo
+
+# ABSTRACT
+
+State-of-the-art retrieval models typically address a straightforward search scenario, in which retrieval tasks are fixed (e.g., finding a passage to answer a specific question) and only a single modality is supported for both queries and retrieved results. This paper introduces techniques for advancing information retrieval with multimodal large language models (MLLMs), enabling a broader search scenario, termed universal multimodal retrieval, where multiple modalities and diverse retrieval tasks are accommodated. To this end, we first study fine-tuning an MLLM as a bi-encoder retriever on 10 datasets with 16 retrieval tasks. Our empirical results show that the fine-tuned MLLM retriever is capable of understanding challenging queries, composed of both text and image, but it underperforms compared to a smaller CLIP retriever in cross-modal retrieval tasks due to the modality bias exhibited by MLLMs. To address the issue, we propose modality-aware hard negative mining to mitigate the modality bias exhibited by MLLM retrievers. Second, we propose continuously fine-tuning the universal multimodal retriever to enhance its text retrieval capability while preserving multimodal retrieval capability. As a result, our model, MM-Embed, achieves state-of-the-art performance on the multimodal retrieval benchmark MBEIR, which spans multiple domains and tasks, while also surpassing the state-of-the-art text retrieval model, NV-Embed-v1, on the MTEB retrieval benchmark. Finally, we explore prompting the off-the-shelf MLLMs as zero-shot rerankers to refine the ranking of the candidates from the multimodal retriever. We find that, through prompt-and-eranking, MLLMs can further improve multimodal retrieval when the user queries (e.g., text-image composed queries) are more complex and challenging to understand. These findings also pave the way for advancing universal multimodal retrieval in the future. We release the model weights at: https://huggingface.co/nvidia/MM-Embed.
+
+# 1 INTRODUCTION
+
+Information retrieval is crucial for a variety of downstream tasks, such as question answering (Kwiatkowski et al., 2019), fact-checking (Thorne et al., 2018), and retrieval-augmented generation (Lewis et al., 2020). State-of-the-art retrievers often focus on narrow scenarios. For example, LLM-based retrievers (Wang et al., 2023; Lee et al., 2024; Meng et al., 2024) are limited to text-to-text retrieval tasks, in which both the query and the retrieved results are text-only. Recent work on multimodal retrieval (Zhang et al., 2024; Jiang et al., 2024) focuses on specific tasks and assumes a homogeneous document format. However, in real-world applications, documents and queries often incorporate diverse formats or modalities, such as text, images, and interleaved text-image content. To advance information retrieval and support broader search scenarios, this work explores leveraging multimodal LLMs (MLLMs; Liu et al., 2023a; 2024; Dai et al., 2024) for universal multimodal
+
+retrieval, accommodating diverse user-instructed tasks with multimodal queries and documents, as illustrated in Figure 1.
+
+We first explore fine-tuning MLLM-based bi-encoder retrievers with instructions as a guide (Asai et al., 2023) on 16 multimodal retrieval tasks from M-BIER (Wei et al., 2023). We find that MLLM-based retrievers significantly outperform CLIP-based retrievers in the challenging tasks that involve interleaved text-image queries, such as visual question answering and composed image retrieval (tasks 3 and 7 in Figure 1). However, MLLM-based retrievers underperform in cross-modal retrieval tasks due to the modality bias exhibited by MLLMs. That is, given a text-based query with the instruction to retrieve an image (e.g., task 9 in Figure 1), an MLLM-based retriever tends to retrieve a relevant text-only document rather than one containing images, especially when we enhance the retriever's text retrieval capability. To address the issue, we propose modality-aware hard negative mining (Section 4.1.1) and continuously fine-tuning for text-to-text retrieval (Section 4.1.2). Our final retriever, coined MM-Embed, is the first universal multimodal retriever to achieve state-of-the-art performance while maintaining competitive text-to-text retrieval across diverse tasks.
+
+Finally, we explore prompting MLLMs as zero-shot rerankers. Surprisingly, we find that the zero-shot MLLM-based rerankers can further boost retrieval accuracy in the tasks where user queries are interleaved text-image and particularly challenging to understand. For example, in the composed image retrieval dataset, CIRCO (Baldrati et al., 2023), the zero-shot reranker can refine the ranked lists and significantly boost the accuracy  $(\mathrm{mAP}@5)$  over 7 points from the existing state-of-the-art composed-image retriever (Zhang et al., 2024) and our universal multimodal retrievers. This finding indicates that there is still room for improvement in such challenging tasks to advance universal multimodal retrieval. Moreover, knowledge distillation from zero-shot or few-shot MLLM-based rerankers to retrievers is a promising direction.
+
+We summarize our contributions as follows: i) We present a study on applying MLLMs to universal multimodal retrieval. ii) We are the first to develop MLLM-based universal multimodal retrievers. Notably, our MM-Embed, initialized from the existing best-performing text retriever (NV-Embed-v1; Lee et al., 2024), not only achieves state-of-the-art results in the universal multimodal retrieval benchmark M-BEIR (Wei et al., 2023), but also surpasses NV-Embed-v1 in text-to-text retrieval tasks on MTEB. iii) We explore prompting MLLMs as zero-shot rerankers in various multimodal retrieval tasks. Surprisingly, we find that zero-shot MLLM-based rerankers can improve ranking accuracy over strong retrievers in challenging tasks involving interleaved text-image queries.
+
+We organize the rest of the paper as follows: We discuss related work in  $\S 2$ . We introduce the definition of universal multimodal retrieval in  $\S 3$  and present the proposed method in  $\S 4$ . We report experimental results in  $\S 5$  and conclude the paper in  $\S 6$ .
+
+# 2 RELATED WORK
+
+Instruction-Aware Dense Representation Learning. Asai et al. (2023) is the first study to identify the implicit search intent behind each retrieval task and proposes fine-tuning a retriever to learn diverse retrieval tasks with handwritten task instructions. Su et al. (2023) and existing state-of-the-art LLM-based text embedding models (Wang et al., 2023; Meng et al., 2024; Lee et al., 2024) adopt this approach to broader tasks beyond text retrieval, such as text classification and clustering. Recently, Wei et al. (2023) propose a universal multimodal retrieval dataset, M-BEIR, and find that instruction-aware dense retrieval fine-tuning is crucial to tackle universal multimodal retrieval.
+
+Vision-Language Models for Multimodal Retrieval. With the advancement of pre-trained vision-language models (Radford et al., 2021; Li et al., 2022), the research focus shifts from single-modal (Bajaj et al., 2016; Fu et al., 2023) to cross-modal (Lin et al., 2014; Han et al., 2017; Liu et al., 2021a) or more complex multimodal retrieval tasks (Liu et al., 2021b; Wu et al., 2021; Baldrati et al., 2023). However, the aforementioned tasks assume a homogeneous modality for queries and documents, thus limiting their application. Liu et al. (2023c) take one step further to tackle the retrieval scenario involving a candidate pool with heterogeneous modalities but still limit itself to a single retrieval task.
+
+Wei et al. (2023) extend the study to a more general scenario, where retrievers are required to deal with queries, a candidate pool with heterogeneous modalities and diverse retrieval tasks. However,
+
+![](images/39a339fcb726e0fe46c1bc579b314580904fac193a7d4e16e13a2189d977c4e5.jpg)  
+Figure 1: Illustration of universal multimodal retrieval in (a), where each task consists of a task-specific instruction and query. Both queries and candidate documents are in heterogeneous formats (i.e., text, image or, interleaved text-image). In this work, we explore (a) fine-tuning MLLM-based universal multimodal retrievers and (b) prompting pre-trained MLLMs for zero-shot reranking over retrieved candidates. We adopt LLaVa-Next (Liu et al., 2024) as our MLLM backbone.
+
+the study is limited to CLIP-based retrievers and ignores important text-to-text retrieval tasks, such as fact-checking (Thorne et al., 2018) and entity retrieval (Hasibi et al., 2017). While Koukounas et al. (2024) aim to fine-tune a CLIP-based retriever with both strong text-to-text and multimodal retrieval capabilities, they only consider simple multimodal retrieval tasks: image-caption retrieval (Young et al., 2014; Lin et al., 2014). Concurrent with our work, Jiang et al. (2024) propose to fine-tune MLLMs on the NLI dataset (Bowman et al., 2015) and demonstrate their transferability to multimodal retrieval. In this paper, we are the first to study how to fine-tune an MLLM-based universal multimodal retriever while maintaining strong text-to-text retrieval capability.
+
+Prompting Multimodal LLMs for Reranking. Instruction tuning has enabled large language models (LLMs) to tackle a wide range of tasks in a zero-shot setting. Building on this, prior studies have investigated prompting LLMs for text reranking (Ma et al., 2023; Sun et al., 2023; Zhuang et al., 2024b). In this work, we extend this line of research into multimodal LLMs, exploring their potential as zero-shot rerankers for multimodal tasks. Notably, Qu et al. (2024) introduce a framework using multimodal LLMs for zero-shot reranking through a generative retrieval approach (Li et al., 2024). However, their method is constrained to retrieval tasks with text-only queries. In contrast, our approach broadens the scope by prompting multimodal LLMs to handle diverse multimodal reranking tasks, accommodating queries and documents that can be in text, image, or interleaved text-image formats. This generalization enables versatile applications in multimodal ranking settings.
+
+# 3 UNIVERSAL MULTIMODAL RETRIEVAL
+
+Following the framework of Lin et al. (2021), we formulate the task of retrieval as follows: given a query  $q$ , the goal is to retrieve a ranked list of candidates  $\{c_1, c_2, \dots, c_k\} \in C$  to maximize some ranking metrics such as nDCG, where  $C$  is the collection of documents. In this work, we borrow the setting of universal multimodal retrieval from Wei et al. (2023), where user queries and candidates may consist of text, image or interleaved text-image; i.e.,  $q \in \{q^{\mathrm{txt}}, q^{\mathrm{img}}, (q^{\mathrm{txt}}, q^{\mathrm{img}})\}$ ;  $c \in \{c^{\mathrm{txt}}, c^{\mathrm{img}}, (c^{\mathrm{txt}}, c^{\mathrm{img}})\}$ . Additionally, there are multiple search intents behind a search query, which can be elaborated through task-specific instructions (Asai et al., 2023). For example, in tasks 1 and 2 of Figure 1, given the same image as a query, the search intent is to find an image caption and similar image, respectively. Thus, in universal multimodal retrieval, given a multimodal query and task instruction inst, we aim to retrieve a list of candidates from a pool of multimodal documents to
+
+maximize a specified ranking metric. Note that we only consider text and image in this work while other modalities, such as audio and video can be included, which we leave for future work.
+
+# 4 METHOD
+
+In this section, we describe our approach to universal multimodal retrieval by leveraging multimodal LLMs (MLLMs), specifically LLaVa-Next (Liu et al., 2024). In Section 4.1, we first fine-tune an MLLM-based retriever to project multimodal user queries, along with task instructions, into the same semantic space as multimodal documents, enabling  $k$ -nearest neighbor search (Johnson et al., 2021). In Section 4.2, we present our method to use MLLMs to rerank the top- $k$  candidates retrieved by the universal multimodal retriever.
+
+# 4.1 FINE-TUNING MULTIMODAL LLMS FOR UNIVERSAL MULTIMODAL RETRIEVAL
+
+We fine-tune an MLLM-based retriever parameterized by  $\theta$  (i.e.,  $\eta^{\theta}$ ) under the guidance of task-specific instructions, aiming to capture the implicit intent behind retrieval tasks. Specifically, given a user query  $q_{i}$  with the specified task instruction inst  $i$  and its corresponding relevant candidate,  $c_{i}^{+}$ , we minimize the contrastive loss (Gutmann & Hyvarinen, 2010):
+
+$$
+L = - \frac {1}{| \mathcal {B} |} \sum_ {i = 1} ^ {| \mathcal {B} |} \log \frac {\exp \left(\eta^ {\theta} (i n s t _ {i} , q _ {i}) \cdot \eta^ {\theta} \left(c _ {i} ^ {+}\right) / \tau\right)}{\sum_ {c ^ {\prime} \in \mathcal {D}} \exp \left(\eta^ {\theta} (i n s t _ {i} , q _ {i}) \cdot \eta^ {\theta} \left(c ^ {\prime}\right) / \tau\right)}, \tag {1}
+$$
+
+where  $\eta^{\theta}(\cdot)\in \mathbb{R}^{d}$  is a normalized vector and  $\tau$  is the temperature scaling factor. Ideally,  $\mathcal{D}$  includes all the candidate documents. However, including all the candidate documents is not computationally feasible; thus, mining informative negative candidates as an alternative to  $\mathcal{D}$  is crucial for successful contrastive learning. In this work, we propose modality-aware negative mining for contrastive learning in the scenario for universal multimodal retrieval.
+
+# 4.1.1 MODALITY-AWARE HARD NEGATIVE MINING
+
+Prior work (Karpukhin et al., 2020; Xiong et al., 2021; de Souza P. Moreira et al., 2024) has demonstrated that hard negative mining significantly improves representation learning for text-to-text retrieval. In the traditional retrieval setting, where the corpus consists of documents with a homogeneous modality, a document is considered a hard negative if it lacks the required information but is still retrieved by a model. However, in the scenario of universal multimodal retrieval, where the corpus contains documents involving diverse modalities, a user's desired modality as specified in task instructions (i.e., text, image or interleaved text-image) should be taken into consideration. For example, as shown in Figure 1, the first and second users issue the same query along with different instructions, requiring the documents in text and image formats, respectively. To address this, we propose modality-aware hard negative mining to guide models in retrieving candidates that meet both the users' information needs and their preferred modality.
+
+Specifically, we first fine-tune an MLLM-based retriever by using in-batch samples as random negatives; i.e.,  $\mathcal{D} = (c_1^+, \dots, c_{|\mathcal{B}|}^+)$  in Eq. (1). The candidate documents in the mini-batch, except  $c_i^+$ , are considered random negatives for  $(inst_i, q_i)$ . The fine-tuned model is denoted as  $M^{\mathrm{rand}}$ . For each query  $q_i$  and its associated instruction  $inst_i$  in the training set, we generate two types of negative samples from the top-50 candidates retrieved by  $M^{\mathrm{rand}}$ :  $i$  negatives with an incorrect modality  $(C_i^1)$ , where the candidate ranks higher than the labeled positive but has a different modality from the desired one, and  $ii$  negatives with unsatisfactory information  $(C_i^2)$ , where the candidate ranks lower than position  $k'$  but has the same desired modality.
+
+Previous studies on text retrieval (Xiong et al., 2021; de Souza P. Moreira et al., 2024) have shown that setting  $k'$  to a small number may introduce false positives while setting  $k'$  to a large number could make the negative samples too easy. In our experiment, we set  $k' = 45$  following the state-of-the-art text retrieval training in Lin et al. (2023). During training, given the query  $q_i$  with the associated instruction inst  $i$ , we generate a triplet,  $((\text{inst}_i, q_i), c_i^+, c_i^-)$ , by sampling a hard negative  $c_i^-$  from either  $C_i^1$  or  $C_i^2$  with the same probability; i.e.,  $\mathcal{D} = (c_1^+, c_1^-, \dots, c_{|\mathcal{B}|}^+, c_{|\mathcal{B}|}^-)$  in Eq. (1). Thus, in the setting of hard negative mining, the negatives mined for  $(\text{inst}_i, q_i)$  include  $i$ ) the hard
+
+negative  $c_{i}^{-}$  and  $ii)$  all positives and hard negatives from other queries, which are considered random negatives. Note that the setting of hard negative mining includes twice as many candidate documents as random negative mining under the same batch size  $|\mathcal{B}|$ . For a fair comparison, we use  $2 \cdot |\mathcal{B}|$  and  $|\mathcal{B}|$  when fine-tuning with random and hard negatives, respectively. Figure 2 in the Appendix showcases both types of negative samples. We observe that the negatives from  $C_1$  are sentences that are semantically similar to the queries but do not match the user's desired modality.
+
+# 4.1.2 CONTINUOUS TEXT-TO-TEXT RETRIEVAL FINE-TUNING
+
+Since text-to-text retrieval remains one of the most widely used retrieval tasks, we further fine-tune  $M^{\text{hard}}$  on a diverse set of public text-to-text retrieval tasks, including MS MARCO (Bajaj et al., 2016), HotpotQA (Yang et al., 2018), Natural Question (Kwiatkowski et al., 2019), PAQ (Lewis et al., 2021), StackExchange (Stack-Exchange-Community., 2023), Natural Language Inference (Bowman et al., 2015), SQuAD (Rajpurkar et al., 2016), ArguAna (Wachsmuth et al., 2018), BioASQ (Nentidis et al., 2023), FiQA (Maia et al., 2018), and FEVER (Thorne et al., 2018). As these datasets do not contain negative samples, we employ the fine-tuned LLM-based retriever (NV-Embed-v1; Lee et al., 2024) to mine hard negatives in our experiments (see de Souza P. Moreira et al. (2024) for details).
+
+During continuous fine-tuning, we uniformly sample triplets from both universal multimodal and text-to-text retrieval training data. Note that for each query  $q_{i}$  in the universal multimodal retrieval training data, we use  $M^{\mathrm{hard}}$  to mine second-type hard negatives  $C_i^2$  again. Since no first-type hard negatives (i.e.,  $C_i^1 = \emptyset$ ) are found by  $M^{\mathrm{hard}}$ , we retain the first-type hard negatives mined by  $M^{\mathrm{rand}}$ .
+
+# 4.2 PROMPTING MULTIMODAL LLMS FOR RERANKING
+
+Prior work (Sun et al., 2023; Jin et al., 2024) has demonstrated that instruction-fine-tuned LLMs can be prompted to rerank candidates in text-to-text retrieval tasks. In this work, we directly prompt pre-trained LLaVa-Next (i.e., the same MLLM backbone for retrievers but without fine-tuning) to further rerank the top 10 retrieved candidates from universal multimodal retrievers. Following the approach of Nogueira et al. (2020), we frame the reranking task as a series of True/False questions. Specifically, given a query and a retrieved candidate, we prompt LLaVa-Next to determine whether the retrieved candidate satisfies the query by answering "True" or "False". For example, in the image caption retrieval (task 1 in Figure 1), given an image query,  $q^{\mathrm{img}}$ , and a retrieved text-based candidate,  $c^{\mathrm{txt}}$ , we use the following prompt: " $<q^{\mathrm{img}}>\backslash nCaption:<c^{\mathrm{txt}}>\backslash nDoes\ the\ above\ daily-life\ image\ match\ the\ caption? \ True\ or\ False$ ". Additionally, in visual question answering retrieval (task 5 in Figure 1), given a visual question,  $<Qry\ image><Qry\ text>$ , and a retrieved text-based candidate,  $<Doc\ text>$ , we use the following prompt:  $<Qry\ image>\backslash nQuestion:<Qry\ text>\backslash nAnswer:<Doc\ text>\backslash nDoes\ the\ answer\ correctly\ answer\ the\ question? \ True\ or\ False$ . We refer readers to Table 18 in Appendix for the specific prompts used in different multimodal retrieval tasks.
+
+To compute relevance scores, we apply the Softmax operation to the logits of the "True" and "False" tokens, and use the probability of the "True" token as the relevance score for reranking. Our preliminary study in Section 5.3.3 shows that zero-shot MLLM-based rerankers mainly improve the tasks where queries are interleaved text and images, such as composed image retrieval and visual question answering, as shown in the tasks 3, 5, and 6 of Figure 1.
+
+# 5 EXPERIMENTS
+
+# 5.1 DATASETS AND MODELS
+
+Multimodal Retrieval Dataset. We evaluate models' universal multimodal retrieval capabilities using the M-BEIR dataset (Wei et al., 2023), which is constructed from 10 datasets with 16 diverse multimodal retrieval tasks across 4 domains, as listed in Appendix Table 10. We train our models on the M-BEIR 1.1M training queries and evaluate their effectiveness on the 190K test queries. Following the global evaluation setting of the M-BEIR dataset, for each query, candidates are retrieved from a unified candidate pool of 5.6M multimodal documents spanning all 10 datasets. We report the average Recall@5 (R@5) as the retrieval accuracy across all test queries in each dataset, except
+
+for Fashion200K and FashionIQ, for which we report Recall@10 (R@10). We refer readers to Wei et al. (2023) for more details on the construction of the M-BEIR dataset.
+
+Text-to-Text Retrieval Dataset. While M-BEIR contains the WebQA dataset for text-to-text retrieval evaluation, we conduct a more comprehensive text-to-text retrieval evaluation using the MTEB dataset (Muennighoff et al., 2023). Specifically, we evaluate our models on 15 diverse text retrieval datasets. Following the established procedure, we report the average nDCG@10 across the 15 text retrieval datasets. Note that, unlike in M-BEIR, where candidates are retrieved from a unified pool across all tasks, in the MTEB retrieval tasks, candidates are retrieved from separate corpora for each task.
+
+Backbone Model Choices. In this work, we utilize two representative backbones of vision-language models to build universal multimodal retrievers, CLIP (Radford et al., 2021) and LLaVa-Next (Liu et al., 2024). For CLIP, we initialize from the CLIP-Large model and employ the best-performing modeling approach from Wei et al. (2023), denoted as  $\mathrm{CLIP}_{\mathrm{SF}}$ . This method fuses the input image and text features by separately encoding each input (query or document) image and text into distinct vectors, which are then summed to create a fused vector (Liu et al., 2023c). Additionally, we report results for BLIP (Li et al., 2022), which integrates text information into the image encoder through cross-attention. We use  $\mathrm{BLIP}_{\mathrm{FF}}$  from Wei et al. (2023), which is fine-tuned on the M-BEIR dataset with random negative.
+
+LLaVa-Next (Liu et al., 2024) is a multimodal LLM (MLLM), which integrates a CLIP image encoder, LLM and a vision-language MLP projector to align image features to the input embedding space of the LLM. We use LLaVa-Next with the Mistral 7B (Jiang et al., 2023) as the backbone LLM. We experiment with three variants: (1) LLaVa-E: the token embedding is used to aggregate information from the multimodal input, a method commonly employed in prior work on text retrieval (Wang et al., 2023; Ma et al., 2024b); (2) LLaVa-P: the MLLM is prompted to summarize each multimodal query (or document) input in one word, using embedding of the last token to encode multimodal input; (3) NVEmb: The LLM from LLaVa-Next is replaced by the fine-tuned LLM-based text retrieval model NV-Embed-v1 (Lee et al., 2024) while all other components (i.e., image encoder and vision-language MLP projector) remain unchanged. Note that the backbone of NV-Embed-v1 is also Mistral 7B. The instructions for LLaVa-E (or NVEmb) and LLaVa-P are illustrated in Appendix Table 16 and 17, respectively. For reranking experiments, we also utilize LLaVa-Next with Mistral 7B, and the prompts are listed in Appendix Table 18.
+
+Retriever Training Details. For each backbone, we start by fine-tuning  $M^{\mathrm{rand}}$  with random negatives; i.e.,  $\mathcal{D} = (c_1^+, \dots, c_{|\mathcal{B}|}^+)$  in Eq. (1). The fine-tuned model is denoted  $M^{\mathrm{rand}}$ . For the CLIP backbone, following (Wei et al., 2023), we fine-tune  $\mathrm{CLIP}_{\mathrm{SF}}$  for 20 epochs with a learning rate of  $1e - 5$ . For LLaVa-Next backbone, we fine-tune the models for 2 epochs with a learning rate of  $1e - 4$ . Note that for LLaVa-Next backbone, we only fine-tune the vision-language projector and LoRA ( $r = 8, \alpha = 64$ ) added to the language model. At the stage of fine-tuning  $M^{\mathrm{hard}}$  with hard negatives, we mine the two types of hard negatives following Section 4.1.1 using each retriever. Then, we fine-tune each retriever using its own mined hard negatives with the same training procedure as the first stage; i.e.,  $\mathcal{D} = (c_1^+, c_1^-, \dots, c_{|\mathcal{B}|}^+, c_{|\mathcal{B}|}^-)$  in Eq. (1). We fine-tune the models with the batch size of  $128 \times 8$  and  $64 \times 8$  when using random and hard negatives, respectively. When GPU memory is insufficient for the designated batch size, we use gradient accumulation. Note that when fine-tuning  $M^{\mathrm{hard}}$ , we initialize the models using the pre-trained model, rather than continuously fine-tuning  $M^{\mathrm{rand}}$ . We denote the models fine-tuned with random and hard negatives as  $M^{\mathrm{rand}}(\cdot)$  and  $M^{\mathrm{hard}}(\cdot)$ , respectively. We refer readers to Appendix A.1 for more details.
+
+Table 1: Main results on retrieval. Following Wei et al. (2023), we report R@5 for all the datasets, except for Fashion200K and FashionIQ, for which we report R@10. The tasks of single-modal and multi-modal queries are tasks 1-5 and 6-8, respectively. For MTEB text retrieval, we report nDCG@10 averaged across 15 retrieval tasks (detailed in Appendix Table 15). See more comparison in Appendix Table 12 and 13.  
+
+<table><tr><td rowspan="2">Task</td><td rowspan="2">Dataset</td><td colspan="5">\( M^{\text{rand}} \)</td><td colspan="3">\( M^{\text{hard}} \)</td><td rowspan="2">MM-Embed</td></tr><tr><td>CLIPSF</td><td>BLIPFF</td><td>LLaVa-E</td><td>LLaVa-P</td><td>NVEmb</td><td>CLIPSF</td><td>LLaVa-P</td><td>NVEmb</td></tr><tr><td rowspan="3">1. \( q^{txt} \rightarrow c^{img} \)</td><td>VisualNews</td><td>43.8</td><td>23.0</td><td>33.2</td><td>34.2</td><td>32.1</td><td>42.7</td><td>39.7</td><td>41.1</td><td>41.0</td></tr><tr><td>MSCOCO</td><td>72.0</td><td>75.6</td><td>69.3</td><td>70.8</td><td>64.6</td><td>69.2</td><td>73.8</td><td>72.7</td><td>71.3</td></tr><tr><td>Fashion200K</td><td>16.4</td><td>25.4</td><td>13.5</td><td>13.3</td><td>10.4</td><td>19.7</td><td>17.4</td><td>18.6</td><td>17.1</td></tr><tr><td>2. \( q^{txt} \rightarrow c^{txt} \)</td><td>WebQA</td><td>83.2</td><td>79.5</td><td>88.6</td><td>88.8</td><td>92.1</td><td>88.2</td><td>93.6</td><td>95.6</td><td>95.9</td></tr><tr><td rowspan="2">3. \( q^{txt} \rightarrow (c^{img},c^{txt}) \)</td><td>EDIS</td><td>46.5</td><td>50.3</td><td>55.9</td><td>56.6</td><td>55.1</td><td>54.2</td><td>68.8</td><td>69.8</td><td>68.8</td></tr><tr><td>WebQA</td><td>76.0</td><td>79.7</td><td>80.3</td><td>81.6</td><td>81.3</td><td>80.1</td><td>84.9</td><td>84.8</td><td>85.0</td></tr><tr><td rowspan="3">4. \( q^{img} \rightarrow c^{txt} \)</td><td>VisualNews</td><td>39.5</td><td>21.1</td><td>32.4</td><td>33.3</td><td>30.4</td><td>40.6</td><td>39.4</td><td>41.4</td><td>41.3</td></tr><tr><td>MSCOCO</td><td>91.0</td><td>88.8</td><td>91.8</td><td>92.2</td><td>90.3</td><td>88.5</td><td>89.5</td><td>88.9</td><td>90.1</td></tr><tr><td>Fashion200K</td><td>17.2</td><td>27.6</td><td>13.9</td><td>14.7</td><td>13.2</td><td>20.0</td><td>17.5</td><td>19.9</td><td>18.4</td></tr><tr><td>5. \( q^{img} \rightarrow c^{img} \)</td><td>NIGHTS</td><td>31.6</td><td>33.0</td><td>31.8</td><td>30.7</td><td>30.4</td><td>31.9</td><td>31.8</td><td>31.1</td><td>32.4</td></tr><tr><td rowspan="2">6. \( (q^{img},q^{txt}) \rightarrow c^{txt} \)</td><td>OVEN</td><td>40.4</td><td>38.7</td><td>37.9</td><td>39.1</td><td>36.3</td><td>40.9</td><td>42.9</td><td>42.6</td><td>42.1</td></tr><tr><td>InfoSeek</td><td>26.1</td><td>19.7</td><td>31.0</td><td>32.9</td><td>33.3</td><td>27.6</td><td>37.2</td><td>35.8</td><td>42.3</td></tr><tr><td rowspan="2">7. \( (q^{img},q^{txt}) \rightarrow c^{img} \)</td><td>FashionIQ</td><td>24.2</td><td>28.5</td><td>27.4</td><td>27.0</td><td>26.0</td><td>21.7</td><td>25.8</td><td>26.6</td><td>25.7</td></tr><tr><td>CIRR</td><td>43.2</td><td>51.4</td><td>48.1</td><td>45.4</td><td>45.3</td><td>38.3</td><td>49.5</td><td>50.8</td><td>50.0</td></tr><tr><td rowspan="2">8. \( (q^{img},q^{txt}) \rightarrow (c^{img},c^{txt}) \)</td><td>OVEN</td><td>60.9</td><td>57.8</td><td>61.6</td><td>62.6</td><td>61.7</td><td>61.6</td><td>63.9</td><td>63.5</td><td>64.1</td></tr><tr><td>InfoSeek</td><td>45.9</td><td>27.7</td><td>50.3</td><td>50.0</td><td>53.4</td><td>47.1</td><td>54.4</td><td>53.5</td><td>57.7</td></tr><tr><td rowspan="3">M-BEIR Avg.</td><td>All</td><td>47.4</td><td>45.5</td><td>47.9</td><td>48.3</td><td>47.2</td><td>48.3</td><td>51.9</td><td>52.3</td><td>52.7</td></tr><tr><td>Single-modal Qry</td><td>51.7</td><td>50.4</td><td>51.0</td><td>51.6</td><td>50.0</td><td>53.5</td><td>55.6</td><td>56.4</td><td>56.1</td></tr><tr><td>Multi-modal Qry</td><td>40.1</td><td>37.3</td><td>42.7</td><td>42.8</td><td>42.7</td><td>39.5</td><td>45.6</td><td>45.5</td><td>47.0</td></tr><tr><td colspan="2">MTEB (Muennighoff et al., 2023) Text Retrieval Avg.</td><td>-</td><td>-</td><td>-</td><td>40.8</td><td>51.6</td><td>-</td><td>46.4</td><td>49.7</td><td>60.3*</td></tr></table>
+
+* ranked top-5 on MTEB retrieval task leaderboard. NV-Embed-v1 (Lee et al., 2024) scores 59.36 in MTEB retrieval task.
+
+To enhance text-to-text retrieval capability, we continuously fine-tune  $M^{\mathrm{hard}}(\mathrm{NVEmb})$  with a learning rate of  $2e - 5$  using a mixture of training data from M-BEIR and public text retrieval datasets mentioned in Section 4.1.2 for 4.5K steps. The final model is called MM-Embed.
+
+# 5.2 MAIN RESULTS
+
+Universal Multimodal Retrieval. Table 1 presents the retrieval accuracy of different retrievers. In the M-BEIR evaluation, we observe that when fine-tuning with random negatives, LLaVa-P achieves the highest overall retrieval effectiveness. This result indicates that LLaVa-P effectively aggregates multimodal input information into a single word representation. While MLLM-based retrievers outperform  $\mathrm{CLIP}_{\mathrm{SF}}$  on tasks involving multimodal queries, they still lag behind  $\mathrm{CLIP}_{\mathrm{SF}}$  on tasks with single-modal queries, especially in cross-modality retrieval (i.e., tasks 1 and 4). In addition, NVEmb achieves the best text-to-text retrieval accuracy on WebQA Task 2. It is worth noting that although  $\mathrm{BLIP}_{\mathrm{FF}}$  performs the worst overall, it demonstrates notably strong performance in the fashion domain (e.g., Fashion200K and FashionIQ) but performs worse in news domain (e.g., VisualNews), likely due to differences in the text-image pairs used for pre-training between CLIP and BLIP.
+
+Observing the models fine-tuned with hard negatives, MLLM-based retrievers show significant improvements in retrieval accuracy, particularly in tasks involving single-modal queries. On the other hand,  $\mathrm{CLIP}_{\mathrm{SF}}$  does not show a similar improvement. This could be attributed to the fact that CLIP has been well pre-trained for cross-modal retrieval, whereas MLLM-based retrievers, fine-tuned with a contrastive learning objective for only 2 epochs, may still be underfitting. Fine-tuning with hard negatives accelerates the contrastive learning process of MLLM-based retrievers.
+
+Table 2 reveals another factor contributing to the lower retrieval accuracy of MLLM-based retrievers for single-modal queries: text retrieval bias. This issue is particularly obvious for NVEmb. We compare models' retrieval accuracy on text-image and image-text retrieval (tasks 1 and 4) on MSCOCO. The compar
+
+Table 2: Retrieval analysis on MSCOCO. M.A.@1 denotes the modality accuracy of the top-1 candidate. More results of M.A.@1 are presented in Appendix Table 14.  
+
+<table><tr><td rowspan="2">Task</td><td rowspan="2">Metric</td><td colspan="4">\( M^{\text{rand}} \)</td><td colspan="3">\( M^{\text{hard}} \)</td></tr><tr><td>\( CLIP_{SF} \)</td><td>LLaVa-E</td><td>LLaVa-P</td><td>NVEmb</td><td>\( CLIP_{SF} \)</td><td>LLaVa-P</td><td>NVEmb</td></tr><tr><td rowspan="3">1.</td><td>R@1</td><td>42.6</td><td>33.9</td><td>41.7</td><td>14.1</td><td>45.8</td><td>50.7</td><td>49.8</td></tr><tr><td>R@5</td><td>72.0</td><td>69.3</td><td>70.8</td><td>64.6</td><td>69.2</td><td>73.8</td><td>72.7</td></tr><tr><td>M.A.@1</td><td>92.6</td><td>79.9</td><td>91.0</td><td>42.1</td><td>98.3</td><td>100.0</td><td>100.0</td></tr><tr><td rowspan="3">4.</td><td>R@1</td><td>72.3</td><td>73.0</td><td>73.4</td><td>69.3</td><td>63.8</td><td>72.7</td><td>72.4</td></tr><tr><td>R@5</td><td>91.0</td><td>91.8</td><td>92.2</td><td>90.3</td><td>88.5</td><td>89.5</td><td>88.9</td></tr><tr><td>M.A.@1</td><td>98.7</td><td>99.2</td><td>99.8</td><td>96.3</td><td>94.2</td><td>100.0</td><td>100.0</td></tr></table>
+
+Table 3: Results of zero-shot reranking on tasks 6-8 from the M-BEIR dataset.  
+
+<table><tr><td rowspan="2">Task</td><td rowspan="2">Dataset</td><td colspan="2">Mhard (NVEmb)</td><td colspan="2">MM-Embed</td></tr><tr><td>Retrieval</td><td>Rerank</td><td>Retrieval</td><td>Rerank</td></tr><tr><td rowspan="2">6.</td><td>OVEN</td><td>42.6</td><td>44.3</td><td>42.1</td><td>43.5</td></tr><tr><td>InfoSeek</td><td>35.8</td><td>37.1</td><td>42.3</td><td>43.1</td></tr><tr><td rowspan="2">7.</td><td>FashionIQ</td><td>26.6</td><td>20.0</td><td>25.7</td><td>19.0</td></tr><tr><td>CIRR</td><td>50.8</td><td>48.6</td><td>50.0</td><td>48.2</td></tr><tr><td rowspan="2">8.</td><td>OVEN</td><td>63.5</td><td>65.8</td><td>64.1</td><td>65.9</td></tr><tr><td>InfoSeek</td><td>53.5</td><td>54.5</td><td>57.7</td><td>57.3</td></tr></table>
+
+Table 4: Results of zero-shot reranking on composed image retrieval task, CIRCO (Baldrati et al., 2023).  
+
+<table><tr><td>Retrieval Model</td><td>Retrieval</td><td>Rerank</td></tr><tr><td>MagicLens (Zhang et al., 2024)</td><td>24.9</td><td>32.4</td></tr><tr><td>E5-V (Jiang et al., 2024)</td><td>19.1</td><td>31.0</td></tr><tr><td>Mrand(CLIPSF)</td><td>12.7</td><td>31.6</td></tr><tr><td>BLIPFF (Wei et al., 2023)</td><td>26.6</td><td>36.1</td></tr><tr><td>Mhard (LLaVa-P)</td><td>29.0</td><td>37.9</td></tr><tr><td>Mhard (NVEmb)</td><td>32.4</td><td>40.9</td></tr><tr><td>MM-Embed</td><td>32.3</td><td>39.9</td></tr></table>
+
+ison shows that  $M^{\mathrm{rand}}(\mathrm{LLaVa - E})$  and  $M^{\mathrm{rand}}(\mathrm{NVEmb})$  exhibit significantly lower modality accuracy (M.A. @ 1) than  $M^{\mathrm{rand}}(\mathrm{CLIP}_{\mathrm{SF}})$  in the text-to-image retrieval task. Most incorrectly retrieved top-1 candidates from the MLLM-based retrievers are relevant texts rather than images (see Appendix Figure 2). This result indicates that MLLM-based retrievers have a bias toward relevant text rather than images, a phenomenon not caused by random sampling, as evidenced in Appendix Table 11. This issue can be mitigated by our proposed modality-aware hard negative mining.
+
+Finally, we observe that  $M^{\mathrm{hard}}(\mathrm{NVEmb})$  performs worse in text-to-text retrieval tasks compared to  $M^{\mathrm{rand}}(\mathrm{NVEmb})$  but still outperforms  $M^{\mathrm{hard}}(\mathrm{LLaVa-P})$  (i.e., WebQA task 2 and MTEB). However, compared to the original NVEmb (Lee et al., 2024), its score on MTEB retrieval tasks drops by almost 10 points. After continuous fine-tuning (detailed in Section 4.1.2), the final model, MMEmbed, not only surpasses NVEmb in MTEB but also retains strong multimodal retrieval capability. We attribute the improvement in text-to-text retrieval to the effective hard negatives mined by NVEmbed-v1 as mentioned in Section 4.1.2. Notably, continuous fine-tuning significantly enhances multimodal retrieval performance in InfoSeek (col 8 vs 7 in Table 1), highlighting its effectiveness in improving the model's ability to handle knowledge-intensive multimodal retrieval tasks.
+
+Zero-Shot Reranking. Table 3 reports the reranked results of the top-10 retrieved candidates from  $M^{\text{hard}}(\text{NVEmb})$  and MM-Embed on the tasks involving multimodal queries. We observe accuracy improvements in visual question answering retrieval tasks (i.e., OVEN and InfoSeek), but no improvement in composed image retrieval tasks (i.e., FashionIQ and CIRR). However, as shown in Appendix Table 10, FashionIQ and CIRR, compared to OVEN and InfoSeek, only have one relevance label per query. We hypothesize that there may be additional relevant positive samples that are not labeled. We refer the reader to Appendix Figure 3 for case studies.
+
+We conduct experiments on the CIRCO (Baldrati et al., 2023) validation set, a composed image retrieval dataset with high-quality human annotations, consisting of 219 queries and  $123\mathrm{K}$  candidates in total. On average, 4.2 positive samples are labeled by humans per query. Table 4 presents  $\mathrm{mAP}@5$  for various retrievers and their reranking results. We directly use the models and code provided by the authors to obtain the results of MagicLens (Zhang et al.,  $2024)^9$  and E5-V (Jiang et al.,  $2024)^{10}$  retrievers. For our M-BEIR fine-tuned retrievers-  $M^{\mathrm{rand}}(\mathrm{CLIP}_{\mathrm{SF}})$ ,  $M^{\mathrm{hard}}(\mathrm{LLaVa - P})$ ,  $M^{\mathrm{hard}}(\mathrm{NVEmb})$  and MM-Embed—we directly use the same instructions as CIRR in M-BEIR for query encoding. First, we observe that our MLLM-based retrievers outperform MagicLens and E5-V. More importantly, reranking the top-10 retrieved candidates from different retrievers significantly improves  $\mathrm{mAP}@5$  by at least 7 points. This result demonstrates the effectiveness of prompting an MLLM as a reranker in composed image retrieval tasks.
+
+# 5.3 ABLATION STUDIES
+
+# 5.3.1 IS FINE-TUNING WITH INSTRUCTION NECESSARY?
+
+We fine-tune NVEmb with random negatives on the M-BEIR subtasks listed in Table 5 and evaluate the models' retrieval accuracy on the development queries from each subtask. Note that, for
+
+Table 5: Ablation study on fine-tuning NVEmb w/o  $(\mathcal{X})$  and w/  $(\checkmark)$  instructions.  
+
+<table><tr><td rowspan="3">Task</td><td rowspan="3">Dataset</td><td colspan="4">zero-shot</td><td colspan="2">fine-tuning</td></tr><tr><td>CLIP</td><td>LLaVa-P</td><td colspan="2">NVEmb</td><td colspan="2">NVEmb</td></tr><tr><td>X</td><td>X</td><td>X</td><td>✓</td><td>X</td><td>✓</td></tr><tr><td rowspan="3">1. qtxt → cimg</td><td>VisualNews</td><td>40.9</td><td>11.7</td><td>15.3</td><td>17.4</td><td>33.1</td><td>38.7</td></tr><tr><td>MSCOCO</td><td>55.4</td><td>58.1</td><td>64.2</td><td>59.9</td><td>76.7</td><td>82.8</td></tr><tr><td>Fashion200K</td><td>8.9</td><td>2.4</td><td>4.2</td><td>3.2</td><td>12.3</td><td>15.6</td></tr><tr><td rowspan="3">4. qimg → ctxt</td><td>VisualNews</td><td>42.0</td><td>6.3</td><td>6.5</td><td>5.9</td><td>29.3</td><td>37.2</td></tr><tr><td>MSCOCO</td><td>79.6</td><td>66.8</td><td>70.6</td><td>68.2</td><td>88.9</td><td>93.0</td></tr><tr><td>Fashion200K</td><td>7.7</td><td>2.9</td><td>4.0</td><td>3.6</td><td>12.0</td><td>16.8</td></tr><tr><td>5. qimg → cimg</td><td>NIGHTS</td><td>25.4</td><td>28.4</td><td>29.3</td><td>27.7</td><td>31.6</td><td>30.9</td></tr></table>
+
+simplicity, we encode only the dataset-specific corpus containing documents of the targeted modality. For example, when evaluating retrieval accuracy for VisualNews Task 1, we encode the 542K images from VisualNews (see Appendix Table 10) into the index rather than the entire 5.6M documents from M-BEIR. We also report the CLIP and LLaVa-P (w/o instruction) zero-shot retrieval effectiveness as a reference point.[11]
+
+From Table 5, we observe that NVEmb, as a zero-shot MLLM-based retriever, outperforms LLaVa-P and even competes with CLIP in the tasks in tasks within the Miscellaneous domain (i.e., MSCOCO and NIGHTS). This result indicates that a fine-tuned MLLM-based text retriever is capable of performing multimodal retrieval tasks (the same finding as in (Jiang et al., 2024)). Although incorporating task instructions with queries degrades the model's retrieval effectiveness (col 4 vs 3), the model fine-tuned with instructions significantly outperforms the one fine-tuned without instructions (col 6 vs 5). This indicates that task instructions can help elicit a model's task- or domain-specific knowledge for diverse multimodal retrieval tasks.
+
+# 5.3.2 EFFECTIVENESS OF CONTINUOUS TEXT-TO-TEXT RETRIEVAL FINE-TUNING
+
+In this section, we study the best strategy to enhance a model's capabilities in both multimodal and text-to-text retrieval. We begin by fine-tuning NVEmb on training data for both universal multimodal retrieval and text-to-text retrieval (detailed in Section 4.1.2) for 2K steps. As shown in Table 6, joint fine-tuning allows the model to maintain its text retrieval capability (row 3 vs 1), though it results in a drop of over 2 points in multimodal retrieval accuracy (row 3 vs 2). In contrast, explicitly fine-tuning  $M^{\mathrm{hard}}(\mathrm{NVEmb})$  for an addition 2K steps significantly boosts
+
+Table 6: Ablation study to enhance model's text-to-text retrieval capability.  
+
+<table><tr><td rowspan="2">Initialization</td><td colspan="2">Training data</td><td rowspan="2">M-BEIR*</td><td rowspan="2">BEIR*</td></tr><tr><td>Multimodal</td><td>Text-to-Text</td></tr><tr><td rowspan="3">NVEmb</td><td>-</td><td>-</td><td>-</td><td>62.9</td></tr><tr><td>✓</td><td>×</td><td>54.3</td><td>51.7</td></tr><tr><td>✓</td><td>✓</td><td>52.2</td><td>63.0</td></tr><tr><td rowspan="2">Mhard(NVEmb)</td><td>-</td><td>-</td><td>56.4</td><td>51.7</td></tr><tr><td>✓</td><td>✓</td><td>55.6</td><td>63.1</td></tr></table>
+
+* For M-BEIR, we only evaluate on the tasks with single-modality queries (i.e., tasks 1-5) while for BIER, we evaluate on 7 tasks: ArguAna, FiQA, NFCorpus, Quora, SCIDOCS, SciFact and TREC-COVID.
+
+its text-to-text retrieval capability with a slight drop of 0.8 points in multimodal retrieval (row 5 vs 4). This experiment shows that continuously fine-tuning a multimodal retriever to enhance its text-to-text retrieval is more effective than fine-tuning a retriever across all retrieval tasks simultaneously. This finding suggests that a better optimized curriculum learning strategy (Bengio et al., 2009) could further improve performance in universal multimodal retrieval, a direction we leave for future work.
+
+# 5.3.3 STUDY ON PROMPTING MLLMS FOR RERANKING
+
+In this section, we study the reranking effectiveness of MLLMs on all the tasks in the M-BEIR dataset. Specifically, for each development query, we rerank the top-10 retrieved candidates from  $M^{\mathrm{rand}}(\mathrm{CLIP}_{\mathrm{SF}})$ . As shown in Table 7, prompting LLaVa-Next for reranking further boosts the rank
+
+Table 7: Reranking study on top-10 retrieved candidates from  $M^{\mathrm{rand}}(\mathrm{CLIP}_{\mathrm{SF}})$  on the M-BEIR development query set.  
+
+<table><tr><td rowspan="2">Task</td><td rowspan="2">Dataset</td><td rowspan="2">Retrieval</td><td colspan="2">Rerank</td></tr><tr><td>7B</td><td>34B</td></tr><tr><td rowspan="3">1. qtxt → cimg</td><td>VisualNews</td><td>44.2</td><td>38.8</td><td>42.5</td></tr><tr><td>MSCOCO</td><td>72.0</td><td>68.0</td><td>69.7</td></tr><tr><td>Fashion200K</td><td>17.8</td><td>14.7</td><td>15.6</td></tr><tr><td>2. qtxt → ctxt</td><td>WebQA</td><td>78.2</td><td>79.2</td><td>82.9</td></tr><tr><td rowspan="2">3. qtxt → (cimg, ctxt)</td><td>EDIS</td><td>48.3</td><td>46.5</td><td>47.4</td></tr><tr><td>WebQA</td><td>78.2</td><td>67.7</td><td>68.3</td></tr><tr><td rowspan="3">4. qimg → ctxt</td><td>VisualNews</td><td>37.4</td><td>29.3</td><td>29.8</td></tr><tr><td>MSCOCO</td><td>91.0</td><td>87.3</td><td>89.0</td></tr><tr><td>Fashion200K</td><td>17.3</td><td>9.9</td><td>12.0</td></tr><tr><td>5. qimg → cimg</td><td>NIGHTS</td><td>32.1</td><td>29.4</td><td>32.7</td></tr><tr><td rowspan="2">6. (qimg, qtxt) → ctxt</td><td>OVEN</td><td>40.6</td><td>43.2</td><td>43.7</td></tr><tr><td>InfoSeek</td><td>25.6</td><td>28.4</td><td>29.0</td></tr><tr><td rowspan="2">7. (qimg, qtxt) → cimg</td><td>FashionIQ</td><td>32.5</td><td>21.5</td><td>23.4</td></tr><tr><td>CIRR</td><td>52.4</td><td>54.1</td><td>54.2</td></tr><tr><td rowspan="2">8. (qimg, qtxt) → (cimg, ctxt)</td><td>OVEN</td><td>60.6</td><td>63.8</td><td>63.7</td></tr><tr><td>InfoSeek</td><td>45.3</td><td>48.7</td><td>50.5</td></tr></table>
+
+ing accuracy in tasks 6-8, which involve multimodal queries (except for FashionIQ). However, the reranking degrades accuracy in tasks 1-5 which involve single-modal queries (except for WebQA task 2). This trend persists even after scaling the reranker from 7B to 34B (col 3, 2 vs 1). $^{13}$  We hypothesize that MLLM rerankers, as a more robust cross-encoder compared to a bi-encoder retriever, excel at challenging tasks involving multimodal queries, even in a zero-shot manner. However, zero-shot rerankers fail to leverage task- or domain-specific knowledge, which limits their performance on relatively simple tasks involving single-modal queries. The relevance signals between queries and documents in the News, Miscellaneous, and Fashion domains can vary significantly. Thus, optimizing prompts or instruction-tuning for MLLMs to better capture domain- or task-specific knowledge offers a promising direction for improving reranking accuracy.
+
+# 6 CONCLUSION AND FUTURE WORK
+
+In this paper, we present techniques for advancing information retrieval using multimodal large language models (MLLMs). We first study fine-tuning MLLM-based retrievers to tackle a general information retrieval scenario: universal multimodal retrieval, where models are required to deal with diverse retrieval tasks, multimodal queries, and documents. Our study shows that MLLM-based retrievers exhibit a modality bias in cross-modal retrieval tasks compared to CLIP-based retrievers. To address the issue, we propose modality-aware hard negative mining, which significantly improves our MLLM-based retrievers' accuracies by 5 points in the M-BEIR dataset, a benchmark for universal multimodal retrieval. Additionally, with our proposed continuous fine-tuning, our MLLM-based retriever, MM-Embed, is the first model to yield state-of-the-art retrieval accuracy in universal multimodal retrieval tasks while maintaining strong text-to-text retrieval capability (ranked top-5 on the MTEB retrieval tasks leaderboard). Finally, we explore prompting MLLMs as rerankers in M-BEIR tasks. We find that MLLMs can be used as zero-shot rerankers to further boost retrieval accuracy in challenging tasks that require the understanding of multimodal queries, such as visual question answering and composed image retrieval. For example, our zero-shot MLLM-based reranker improves the retrieval accuracy over the state-of-the-art retrievers by over 7 points in CIRCO.
+
+Our work also suggests two promising future directions: (1) Distilling our MLLM-based retriever, MM-Embed, into smaller multimodal retrievers, such as CLIP (Radford et al., 2021) or BLIP (Li et al., 2022), to achieve better retrieval efficiency (see the efficiency comparisons in Appendix A.2); (2) Distilling an MLLM-based reranker into a retriever to further improve its retrieval capability in tasks involving multimodal queries. Other directions, such as iterative retrieval with relevance feedback (Han et al., 2024) and generative retrieval (Qu et al., 2024), are also worth exploring for universal multimodal retrieval tasks. In addition, recent work (Ma et al., 2024a; Faysse et al., 2024) has demonstrated that MLLMs can be fine-tuned to tackle visual document retrieval tasks, which could be incorporated into universal multimodal retrieval tasks.
+
+# REFERENCES
+
+Akari Asai, Timo Schick, Patrick Lewis, Xilun Chen, Gautier Izacard, Sebastian Riedel, Hannaneh Hajishirzi, and Wen-tau Yih. Task-aware retrieval with instructions. In Findings of ACL, pp. 3650-3675, 2023.  
+Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, Mir Rosenberg, Xia Song, Alina Stoica, Saurabh Tiwary, and Tong Wang. MS MARCO: A human generated machine reading comprehension dataset. arXiv:1611.09268, 2016.  
+Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proc. ICCV, pp. 15338-15347, 2023.  
+Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proc. ICML, pp. 41-48, 2009.  
+Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. A large annotated corpus for learning natural language inference. In Proc. EMNLP, pp. 632-642, 2015.  
+Yingshan Chang, Guihong Cao, Mridu Narang, Jianfeng Gao, Hisami Suzuki, and Yonatan Bisk. WebQA: Multihop and multimodal QA. In Proc. CVPR, pp. 16474-16483, 2022.  
+Yang Chen, Hexiang Hu, Yi Luan, Haitian Sun, Soravit Changpinyo, Alan Ritter, and Ming-Wei Chang. Can pre-trained vision and language models answer visual information-seeking questions? In Proc. EMNLP, pp. 14948–14968, 2023.  
+Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuoling Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. NVLM: Open frontier-class multimodal LLMs. arXiv:2409.11402, 2024.  
+Gabriel de Souza P. Moreira, Radek Osmulski, Mengyao Xu, Ronay Ak, Benedikt Schifferer, and Even Oldridge. NV-Retriever: Improving text embedding models with effective hard-negative mining. arxiv.2407.15831, 2024.  
+Manuel Faysse, Hugues Sibille, Tony Wu, Bilel Omrani, Gautier Viaud, Céline Hudelot, and Pierre Colombo. ColPali: Efficient document retrieval with vision language models. arXiv:2407.01449, 2024.  
+Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. DreamSim: Learning new dimensions of human visual similarity using synthetic data. In Proc. NeurIPS, pp. 50742-50768, 2023.  
+Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. Tevatron: An efficient and flexible toolkit for neural retrieval. In Proc. SIGIR, pp. 3120-3124, 2023.  
+Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proc. AISTATS, pp. 297-304, 2010.  
+Donghoon Han, Eunhwan Park, Gisang Lee, Adam Lee, and Nojun Kwak. MERLIN: Multimodal embedding refinement via LLM-based iterative navigation for text-video retrieval-erank pipeline. In Proc. EMNLP, pp. 547-562, 2024.  
+Xintong Han, Zuxuan Wu, Phoenix X. Huang, Xiao Zhang, Menglong Zhu, Yuan Li, Yang Zhao, and Larry S. Davis. Automatic spatially-aware fashion concept discovery. In Proc. ICCV, pp. 1472-1480, 2017.  
+Faegheh Hasibi, Fedor Nikolaev, Chenyan Xiong, Krisztian Balog, Svein Erik Bratsberg, Alexander Kotov, and Jamie Callan. DBpedia-Entity v2: A test collection for entity search. In Proc. SIGIR, pp. 1265–1268, 2017.  
+Hexiang Hu, Yi Luan, Yang Chen, Urvashi Khandelwal, Mandar Joshi, Kenton Lee, Kristina Toutanova, and Ming-Wei Chang. Open-domain Visual Entity Recognition: Towards recognizing millions of Wikipedia entities. In Proc. ICCV, 2023.
+
+Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7B. arXiv:2310.06825, 2023.  
+Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-V: Universal embeddings with multimodal large language models. arxiv.2407.12580, 2024.  
+Can Jin, Hongwu Peng, Shiyu Zhao, Zhenting Wang, Wujiang Xu, Ligong Han, Jiahui Zhao, Kai Zhong, Sanguthevar Rajasekaran, and Dimitris N. Metaxas. APEER: Automatic prompt engineering enhances large language model reranking. arXiv:2406.14449, 2024.  
+Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, pp. 535-547, 2021.  
+Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. In Proc. EMNLP, pp. 6769-6781, 2020.  
+Andreas Koukounas, Georgios Mastrupas, Michael Gunther, Bo Wang, Scott Martens, Isabelle Mohr, Saba Sturua, Mohammad Kalim Akram, Joan Fontanals Martinez, Saahil Ognawala, Susana Guzman, Maximilian Werk, Nan Wang, and Han Xiao. Jina CLIP: Your CLIP model is also your text retriever. arXiv:2405.20204, 2024.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. Natural Questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, pp. 452-466, 2019.  
+Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. NV-Embed: Improved techniques for training LLMs as generalist embedding models. arXiv:2405.17428, 2024.  
+Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Proc. NeurIPS, pp. 9459-9474, 2020.  
+Patrick Lewis, Yuxiang Wu, Linqing Liu, Pasquale Minervini, Heinrich Kuttler, Aleksandra Piktus, Pontus Stenetorp, and Sebastian Riedel. PAQ: 65 million probably-asked questions and what you can do with them. Transactions of the Association for Computational Linguistics, pp. 1098-1115, 2021.  
+Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In Proc. ICML, pp. 12888-12900, 2022.  
+Xiaoxi Li, Jiajie Jin, Yujia Zhou, Yuyao Zhang, Peitian Zhang, Yutao Zhu, and Zhicheng Dou. From matching to generation: A survey on generative information retrieval. arXiv:2404.14851, 2024.  
+Jimmy Lin, Rodrigo Nogueira, and Andrew Yates. Pretrained Transformers for Text Ranking: BERT and Beyond. Morgan & Claypool, 2021.  
+Sheng-Chieh Lin, Akari Asai, Minghan Li, Barlas Oguz, Jimmy Lin, Yashar Mehdad, Wen-tau Yih, and Xilun Chen. How to Train your DRAGON: Diverse augmentation towards generalizable dense retrieval. In Proc. Findings of EMNLP, pp. 6385-6400, 2023.  
+Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In Proc. ECCV, pp. 740-755, Cham, 2014.
+
+Fuxiao Liu, Yinghan Wang, Tianlu Wang, and Vicente Ordonez. Visual News: Benchmark and challenges in news image captioning. In Proc. EMNLP, pp. 6761-6771, 2021a.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Proc. NeurIPS, volume 36, pp. 34892-34916, 2023a.  
+Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, January 2024. URL https://llava-v1.github.io/blog/2024-01-30-llava next/.  
+Siqi Liu, Weixi Feng, Tsu-Jui Fu, Wenhu Chen, and William Wang. EDIS: Entity-driven image search over multimodal web content. In Proc. EMNLP, pp. 4877-4894, 2023b.  
+Zhenghao Liu, Chenyan Xiong, Yuanhuiyi Lv, Zhiyuan Liu, and Ge Yu. Universal vision-language dense retrieval: Learning a unified representation space for multi-modal retrieval. In Proc. ICLR, 2023c.  
+Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. In Proc. ICCV, pp. 2105-2114, 2021b.  
+Xueguang Ma, Xinyu Zhang, Ronak Pradeep, and Jimmy Lin. Zero-shot listwise document reranking with a large language model. arXiv:2305.02156, 2023.  
+Xueguang Ma, Sheng-Chieh Lin, Minghan Li, Wenhu Chen, and Jimmy Lin. Unifying multimodal retrieval via document screenshot embedding. In Proc. EMNLP, pp. 6492-6505, 2024a.  
+Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. Fine-tuning LLaMA for multi-stage text retrieval. In Proc. SIGIR, pp. 2421-2425, 2024b.  
+Macedo Maia, Siegfried Handschuh, André Freitas, Brian Davis, Ross McDermott, Manel Zarrouk, and Alexandra Balahur. WWW'18 open challenge: Financial opinion mining and question answering. In Companion Proceedings of the The Web Conference 2018, pp. 1941-1942, 2018.  
+Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. SFR-Embedding-2: Advanced text embedding with multi-stage training, 2024. URL https://huggingface.co/Salesforce/SFR-Embedding-2_R.  
+Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. MTEB: Massive text embedding benchmark. In Proc. EACL, pp. 2014-2037, 2023.  
+Anastasios Nentidis, Georgios Katsimpras, Anastasia Krithara, Salvador Lima López, Eulália Farré-Maduell, Luis Gasco, Martin Krallinger, and Georgios Paliouras. Overview of BioASQ 2023: The eleventh BioASQ challenge on large-scale biomedical semantic indexing and question answering. In Experimental IR Meets Multilinguality, Multimodality, and Interaction, pp. 227-250, 2023.  
+Rodrigo Nogueira, Zhiying Jiang, Ronak Pradeep, and Jimmy Lin. Document ranking with a pretrained sequence-to-sequence model. In Proc. Findings of EMNLP, pp. 708-718, 2020.  
+Leigang Qu, Haochuan Li, Tan Wang, Wenjie Wang, Yongqi Li, Liqiang Nie, and Tat-Seng Chua. Unified text-to-image generation and retrieval. arXiv:2406.05814, 2024.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proc. ICML, pp. 8748-8763, 2021.  
+Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. ZeRO: memory optimizations toward training trillion parameter models. In Proc. of the International Conference for High Performance Computing, Networking, Storage and Analysis, 2020.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. SQuAD: 100,000+ questions for machine comprehension of text. In Proc. EMNLP, pp. 2383-2392, 2016.
+
+Stack-Exchange-Community. Stack Exchange data dump, Sep 2023. URL https://archive.org/details/stack-exchange-data-dump-2023-09-12.  
+Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. One embedder, any task: Instruction-finetuned text embeddings. In Proc. Findings of ACL, pp. 1102-1121, July 2023.  
+Weiwei Sun, Lingyong Yan, Xinyu Ma, Shuaiqiang Wang, Pengjie Ren, Zhumin Chen, Dawei Yin, and Zhaochun Ren. Is ChatGPT good at search? Investigating large language models as re-ranking agents. In Proc. EMNLP, pp. 14918-14937, 2023.  
+Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Proc. NeurIPS, 2021.  
+James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. FEVER: a large-scale dataset for Fact Extraction and VERIFICATION. In Proc. NAACL, pp. 809-819, June 2018.  
+Henning Wachsmuth, Shahbaz Syed, and Benno Stein. Retrieval of the best counterargument without prior topic knowledge. In Proc. ACL, pp. 241-251, 2018.  
+Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. arXiv:2401.00368, 2023.  
+Cong Wei, Yang Chen, Haonan Chen, Hexiang Hu, Ge Zhang, Jie Fu, Alan Ritter, and Wenhu Chen. UniIR: Training and benchmarking universal multimodal information retrievers. arxiv.2311.17136, 2023.  
+Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The Fashion IQ dataset: Retrieving images by combining side information and relative natural language feedback. In Proc. CVPR, 2021.  
+Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In Proc. ICLR, 2021.  
+Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proc. EMNLP, pp. 2369-2380, 2018.  
+Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, pp. 67-78, 2014.  
+Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. MagicLens: Self-supervised image retrieval with open-ended instructions. In Proc. ICML, pp. to appear, 2024.  
+Shengyao Zhuang, Xueguang Ma, Bevan Koopman, Jimmy Lin, and Guido Zuccon. PromptReps: Prompting large language models to generate dense and sparse representations for zero-shot document retrieval. In Proc. EMNLP, pp. 4375-4391, 2024a.  
+Shengyao Zhuang, Honglei Zhuang, Bevan Koopman, and Guido Zuccon. A setwise approach for effective and highly efficient zero-shot ranking with large language models. In Proc. SIGIR, pp. 38-47, 2024b.
+
+# A APPENDIX
+
+# A.1 IMPLEMENTATION DETAILS
+
+We implement our training and inference using Tevatron (Gao et al., 2023). For CLIP-based retrievers, we follow all the settings from Wei et al. (2023). For the MLLM-based retrievers, we fine-tune models with DeepSpeed Zero 2 (Rajbhandari et al., 2020) and gradient checkpointing. During fine-tuning on the M-BEIR training data, we set the maximum length for queries and documents to 128. During continuously fine-tuning on both M-BEIR and text-to-text retrieval training data, we set the maximum length for queries and documents to 128 and 512, respectively. All fine-tuning is conducted on  $8 \times 80$  GB A100 GPUs. Note that image input only occupies a single token in length after being tokenized; however, each image will be converted to multiple image tokens. Thus, the actual input length for the MLLMs is longer than the maximum length we set. To speed up fine-tuning and inference for MLLM-based retrievers, we only use the global image patches, which occupy 576  $(24 \times 24)$  image tokens.
+
+# A.2 RETRIEVAL EFFICIENCY COMPARISONS
+
+Table 8: Retrieval efficiency comparisons on M-BEIR dataset.  
+
+<table><tr><td rowspan="2">Retriever</td><td>Storage (GBs)</td><td colspan="3">Latency (ms)</td></tr><tr><td>Index</td><td>Encoding (1st/50th/99thperc.)</td><td colspan="2">Vector search</td></tr><tr><td>CLIPSF</td><td>16</td><td>26/27/39</td><td>6</td><td></td></tr><tr><td>BLIPFF</td><td>16</td><td>37/38/44</td><td>6</td><td></td></tr><tr><td>MM-Embed</td><td>86</td><td>81/194/203</td><td>33</td><td></td></tr></table>
+
+Table 8 compares the retrieval efficiency in terms of storage and latency for different retrievers adopted in the paper. We measure the index storage required for the 5.6M documents from the M-BEIR dataset. As for retrieval latency, we measure the latencies of query encoding and vector search. For query latency, we randomly sample 100 queries from each test query pool in the 16 M-BEIR tasks and measure the per query encoding and vector search latency with a batch size of 1. Since query encoding latency varies with query length, we report the latency at  $1^{\text{th}}$ ,  $50^{\text{th}}$  and  $99^{\text{th}}$  percentiles. The latency is measured using one thread on a Linux machine with a 2.2 GHz Intel Xeon Silver 4210 CPU and NVIDIA RTX A6000 GPUs, respectively. Note that we perform a brute-force search on the sharded index with two GPUs since the full index from UniEmb cannot be loaded into a single A6000 GPU.
+
+# A.3 BASELINE REPRODUCING
+
+Since we implement our fine-tuning and inference following the settings from Wei et al. (2023), our fine-tuned  $M^{\mathrm{rand}}(\mathrm{CLIP}_{\mathrm{SF}})$  should be equivalent to  $\mathrm{CLIP}_{\mathrm{SF}}$  from Wei et al. (2023). In Table 9, we compare the results of our fine-tuned  $M^{\mathrm{rand}}(\mathrm{CLIP}_{\mathrm{SF}})$  with the checkpoint provided by the authors.[14]
+
+Table 9: A comparison of  $M^{\mathrm{rand}}(\mathrm{CLIP}_{\mathrm{SF}})$  fine-tuned by us and Wei et al. (2023).  
+
+<table><tr><td rowspan="2">Task</td><td rowspan="2">Dataset</td><td colspan="2">Mrand(CLIPSF)</td></tr><tr><td>Wei et al. (2023)</td><td>Ours</td></tr><tr><td rowspan="3">M-BEIR Avg.</td><td>All</td><td>47.4</td><td>47.4</td></tr><tr><td>Single-modal Qry</td><td>52.5</td><td>51.7</td></tr><tr><td>multi-modal Qry</td><td>39.1</td><td>40.1</td></tr></table>
+
+Table 10: M-BEIR dataset statistics.  
+
+<table><tr><td rowspan="2">Task</td><td rowspan="2">Dataset</td><td rowspan="2">Domain</td><td colspan="3"># Query</td><td colspan="3"># Relevance / Query</td><td rowspan="2"># Candid.</td></tr><tr><td>Train</td><td>Dev</td><td>Test</td><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td rowspan="3">1. q\( ^{\text{xt}} \) → c\( ^{\text{img}} \)</td><td>VisualNews (Liu et al., 2021a)</td><td>News</td><td>99K</td><td>20K</td><td>20K</td><td>1.0</td><td>1.0</td><td>1.0</td><td>542K</td></tr><tr><td>MSCOCO (Lin et al., 2014)</td><td>Misc.</td><td>100K</td><td>24.8K</td><td>24.8K</td><td>1.0</td><td>1.0</td><td>1.0</td><td>5K</td></tr><tr><td>Fashion200K (Han et al., 2017)</td><td>Fashion</td><td>15K</td><td>1.7K</td><td>1.7K</td><td>3.3</td><td>3.1</td><td>2.8</td><td>201K</td></tr><tr><td>2. q\( ^{\text{xt}} \) → c\( ^{\text{xt}} \)</td><td>WebQA (Chang et al., 2022)</td><td>Wiki</td><td>16K</td><td>1.7K</td><td>2.4K</td><td>2.0</td><td>2.0</td><td>2.0</td><td>544K</td></tr><tr><td rowspan="2">3. q\( ^{\text{xt}} \) → (c\( ^{\text{img}} \), c\( ^{\text{xt}} \))</td><td>EDIS (Liu et al., 2023b)</td><td>News</td><td>26K</td><td>3.2K</td><td>3.2K</td><td>2.6</td><td>2.6</td><td>2.6</td><td>1M</td></tr><tr><td>WebQA (Chang et al., 2022)</td><td>Wiki</td><td>16K</td><td>1.7K</td><td>2.4K</td><td>1.4</td><td>1.4</td><td>1.4</td><td>544K</td></tr><tr><td rowspan="3">4. q\( ^{\text{img}} \) → c\( ^{\text{xt}} \)</td><td>VisualNews (Liu et al., 2021a)</td><td>News</td><td>100K</td><td>20K</td><td>20K</td><td>1.0</td><td>1.0</td><td>1.0</td><td>537K</td></tr><tr><td>MSCOCO (Lin et al., 2014)</td><td>Misc.</td><td>113K</td><td>5K</td><td>5K</td><td>5.0</td><td>5.0</td><td>5.0</td><td>25K</td></tr><tr><td>Fashion200K (Han et al., 2017)</td><td>Fashion</td><td>15K</td><td>4.8K</td><td>4.8K</td><td>1.0</td><td>1.0</td><td>1.0</td><td>61K</td></tr><tr><td>5. q\( ^{\text{img}} \) → c\( ^{\text{img}} \)</td><td>NIGHTS (Fu et al., 2023)</td><td>Misc.</td><td>16K</td><td>2K</td><td>2K</td><td>1.0</td><td>1.0</td><td>1.0</td><td>40K</td></tr><tr><td rowspan="2">6. (q\( ^{\text{img}} \), q\( ^{\text{xt}} \)) → c\( ^{\text{xt}} \)</td><td>OVEN (Hu et al., 2023)</td><td>Wiki</td><td>150K</td><td>50K</td><td>50K</td><td>8.5</td><td>10.0</td><td>9.9</td><td>676K</td></tr><tr><td>InfoSeek (Chen et al., 2023)</td><td>Wiki</td><td>141K</td><td>11K</td><td>11K</td><td>6.8</td><td>6.7</td><td>6.5</td><td>611K</td></tr><tr><td rowspan="2">7. (q\( ^{\text{img}} \), q\( ^{\text{xt}} \)) → c\( ^{\text{img}} \)</td><td>FashionIQ (Wu et al., 2021)</td><td>Fashion</td><td>16K</td><td>2K</td><td>6K</td><td>1.0</td><td>1.0</td><td>1.0</td><td>74K</td></tr><tr><td>CIRR (Liu et al., 2021b)</td><td>Misc.</td><td>26K</td><td>2K</td><td>4K</td><td>1.0</td><td>1.0</td><td>1.0</td><td>21K</td></tr><tr><td rowspan="2">8. (q\( ^{\text{img}} \), q\( ^{\text{xt}} \)) → (c\( ^{\text{img}} \), c\( ^{\text{xt}} \))</td><td>OVEN (Hu et al., 2023)</td><td>Wiki</td><td>157K</td><td>14.7K</td><td>14.7K</td><td>17.8</td><td>17.5</td><td>17.7</td><td>335K</td></tr><tr><td>InfoSeek (Chen et al., 2023)</td><td>Wiki</td><td>143K</td><td>17.6K</td><td>17.6K</td><td>9.1</td><td>7.5</td><td>7.5</td><td>481K</td></tr><tr><td colspan="2">M-BEIR (Wei et al., 2023)</td><td>4 domains</td><td>1.1M</td><td>182K</td><td>190K</td><td>6.5</td><td>5.9</td><td>5.7</td><td>5.6M</td></tr></table>
+
+Table 11: Document modality statistics of positive samples in M-BEIR train set. We observe that there are more text-modal samples in the positive samples of the M-BIER training set. The statistics indicate that text modality has a higher probability of being randomly sampled as negatives than other modalities (i.e., image and interleaved text-image); thus, the text modality bias of MLLM-based retrievers is not from the sampling bias of the random sampling strategy.  
+
+<table><tr><td>Document modality</td><td>Text</td><td>Image</td><td>Interleaved text-image</td></tr><tr><td>number</td><td>568,757</td><td>364,903</td><td>399,130</td></tr><tr><td>percentage</td><td>42.6%</td><td>27.4%</td><td>29.9%</td></tr></table>
+
+Table 12: A comparison with other existing state-of-the-art multimodal retrieval models on M-BEIR dataset.  
+
+<table><tr><td>Task</td><td>Dataset</td><td>E5-V (Jiang et al., 2024)</td><td>MagicLens (Zhang et al., 2024)</td><td>MM-Embed</td></tr><tr><td rowspan="3">M-BEIR Avg.</td><td>All</td><td>11.5</td><td>5.8</td><td>52.7</td></tr><tr><td>Single-modal Qry</td><td>14.6</td><td>8.1</td><td>56.1</td></tr><tr><td>Multi-modal Qry</td><td>6.3</td><td>2.0</td><td>47.0</td></tr></table>
+
+Table 13: A comparison with other existing state-of-the-art multimodal retrieval models on M-BEIR subtasks in a local evaluation setting (Wei et al., 2023), where retrieval is conducted only among the corpus of each dataset consisting of less than 1M candidate documents with a single modality.  
+
+<table><tr><td>Task</td><td>Dataset</td><td>E5-V (Jiang et al., 2024)</td><td>MagicLens (Zhang et al., 2024)</td><td>MM-Embed</td></tr><tr><td>1. q\( ^{\text{txt}} \) → c\( ^{\text{img}} \)</td><td>MSCOCO</td><td>75.8</td><td>68.5</td><td>82.7</td></tr><tr><td>2. q\( ^{\text{txt}} \) → c\( ^{\text{txt}} \)</td><td>WebQA</td><td>84.8</td><td>47.9</td><td>96.6</td></tr><tr><td>4. q\( ^{\text{img}} \) → c\( ^{\text{txt}} \)</td><td>MSCOCO</td><td>83.4</td><td>17.4</td><td>91.0</td></tr><tr><td>5. q\( ^{\text{img}} \) → c\( ^{\text{img}} \)</td><td>NIGHTS</td><td>26.7</td><td>14.1</td><td>32.7</td></tr><tr><td rowspan="2">7. (q\( ^{\text{img}} \), q\( ^{\text{txt}} \)) → c\( ^{\text{img}} \)</td><td>FashionIQ</td><td>8.4</td><td>13.8</td><td>26.0</td></tr><tr><td>CIRR</td><td>30.1</td><td>37.5</td><td>53.0</td></tr></table>
+
+Table 14: Retrieval models' top-1 modality accuracy (M.A.@1). We observe that most MLLM-based retrievers suffer from low modality accuracy on Task 1 due to modality bias, particularly in NVEmb due to its superior text retrieval capability. This issue can be resolved with our modality-aware hard negative mining. Even though MM-Embed exhibits strong text retrieval effectiveness, no modality bias is observed.
+
+<table><tr><td rowspan="2">Task</td><td rowspan="2">Dataset</td><td colspan="4">\( M^{\text{rand}} \)</td><td colspan="4">\( M^{\text{hard}} \)</td><td rowspan="2">MM-Embed</td></tr><tr><td>\( CLIP_{SF} \)</td><td>\( LLaVa-E \)</td><td>\( LLaVa-P \)</td><td>NVEmb</td><td>\( CLIP_{SF} \)</td><td>\( LLaVa-P \)</td><td>NVEmb</td><td></td></tr><tr><td rowspan="3">1. \( q^{txt} \rightarrow c^{img} \)</td><td>VisualNews</td><td>0.97</td><td>0.82</td><td>0.94</td><td>0.76</td><td>0.98</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td></tr><tr><td>MSCOCO</td><td>0.93</td><td>0.80</td><td>0.91</td><td>0.42</td><td>0.98</td><td>0.99</td><td>1.00</td><td>1.00</td><td>1.00</td></tr><tr><td>Fashion200K</td><td>0.97</td><td>1.00</td><td>1.00</td><td>0.78</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>0.99</td></tr><tr><td>2. \( q^{txt} \rightarrow c^{txt} \)</td><td>WebQA</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>0.99</td><td>1.00</td><td>1.00</td></tr><tr><td rowspan="2">3. \( q^{txt} \rightarrow (c^{img},c^{txt}) \)</td><td>EDIS</td><td>0.94</td><td>1.00</td><td>0.96</td><td>1.00</td><td>0.90</td><td>0.99</td><td>1.00</td><td>0.99</td><td>0.99</td></tr><tr><td>WebQA</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td></tr><tr><td rowspan="3">4. \( q^{img} \rightarrow c^{txt} \)</td><td>VisualNews</td><td>0.33</td><td>0.90</td><td>0.84</td><td>0.84</td><td>0.97</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td></tr><tr><td>MSCOCO</td><td>0.99</td><td>0.99</td><td>1.00</td><td>0.96</td><td>0.94</td><td>0.99</td><td>1.00</td><td>0.99</td><td>0.99</td></tr><tr><td>Fashion200K</td><td>0.99</td><td>0.99</td><td>1.00</td><td>0.99</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td></tr><tr><td>5. \( q^{img} \rightarrow c^{img} \)</td><td>NIGHTS</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>0.99</td><td>1.00</td><td>1.00</td></tr><tr><td rowspan="2">6. \( (q^{img},q^{txt}) \rightarrow c^{txt} \)</td><td>OVEN</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>0.99</td></tr><tr><td>InfoSeek</td><td>0.94</td><td>1.00</td><td>1.00</td><td>1.00</td><td>0.97</td><td>0.99</td><td>1.00</td><td>1.00</td><td>0.99</td></tr><tr><td rowspan="2">7. \( (q^{img},q^{txt}) \rightarrow c^{img} \)</td><td>FashionIQ</td><td>0.99</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td></tr><tr><td>CIRR</td><td>0.99</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td></tr><tr><td rowspan="2">8. \( (q^{img},q^{txt}) \rightarrow (c^{img},c^{txt}) \)</td><td>OVEN</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td></tr><tr><td>InfoSeek</td><td>0.98</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td><td>1.00</td></tr></table>
+
+Table 15: Detailed results on MTEB retrieval tasks.  
+
+<table><tr><td>Model</td><td>AA</td><td>CF</td><td>CQ</td><td>DB</td><td>Fe</td><td>FQ</td><td>HQ</td><td>MS</td><td>NF</td><td>NQ</td><td>Qu</td><td>SD</td><td>SF</td><td>T2</td><td>TC</td><td>Avg.</td></tr><tr><td>NVEmb (Lee et al., 2024)</td><td>68.2</td><td>34.7</td><td>50.5</td><td>48.3</td><td>87.8</td><td>63.1</td><td>79.9</td><td>46.5</td><td>38.0</td><td>71.2</td><td>89.2</td><td>20.2</td><td>78.4</td><td>28.4</td><td>85.9</td><td>59.4</td></tr><tr><td>MRand (LLaVa-P)</td><td>48.4</td><td>12.9</td><td>34.0</td><td>34.0</td><td>52.2</td><td>33.7</td><td>50.1</td><td>12.3</td><td>30.4</td><td>36.5</td><td>83.8</td><td>17.9</td><td>72.3</td><td>73.4</td><td>19.6</td><td>40.8</td></tr><tr><td>MRand (NVEmb)</td><td>51.5</td><td>23.7</td><td>43.6</td><td>44.9</td><td>78.6</td><td>46.5</td><td>70.2</td><td>32.5</td><td>38.9</td><td>54.1</td><td>87.5</td><td>20.3</td><td>74.5</td><td>83.4</td><td>23.4</td><td>51.6</td></tr><tr><td>Mhard (LLaVa-P)</td><td>38.6</td><td>20.4</td><td>38.0</td><td>36.9</td><td>78.1</td><td>36.2</td><td>61.2</td><td>23.2</td><td>35.1</td><td>45.1</td><td>86.1</td><td>19.2</td><td>72.7</td><td>27.7</td><td>77.2</td><td>46.4</td></tr><tr><td>Mhard (NVEmb)</td><td>37.2</td><td>30.8</td><td>44.0</td><td>44.3</td><td>86.4</td><td>45.5</td><td>70.6</td><td>34.2</td><td>37.4</td><td>49.7</td><td>86.9</td><td>13.9</td><td>64.1</td><td>23.5</td><td>76.7</td><td>49.7</td></tr><tr><td>MM-Embed</td><td>69.0</td><td>39.3</td><td>49.7</td><td>50.6</td><td>92.6</td><td>60.1</td><td>81.4</td><td>45.1</td><td>40.5</td><td>70.6</td><td>88.7</td><td>21.8</td><td>78.3</td><td>31.1</td><td>85.4</td><td>60.3</td></tr></table>
+
+* Dataset Legend: AA=ArguAna, CF=Climate-FEVER, CQ=CQADupStack, DB=DBPedia, Fe=FEVER, FQ=FiQA, HQ=HotpotQA, MS=MSMARCO, NF=NFCorpus, NQ=NaturalQuestions, Qu=Quora, SD=SCIDOCS, SF=SciFact, T2=Touche-2020, TC=TREC-COVID
+
+Table 16: NVEmb (and LLaVa-E) instructions for M-BEIR and MTEB, which are from Wei et al. (2023) and Lee et al. (2024), respectively. For all the candidates, we use the prompt to generate the embedding:  $< c^{img}>\backslash n < c^{txt}>$
+
+<table><tr><td>Task</td><td>Dataset</td><td>M-BEIR task instruction</td></tr><tr><td rowspan="3">1. qxt → cimg</td><td>VisualNews</td><td>Identify the news-related image in line with the described event.\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>MSCOCO</td><td>Find me an everyday image that matches the given caption.\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>Fashion200K</td><td>Based on the following fashion description, retrieve the best matching image.\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>2. qxt → cxt</td><td>WebQA</td><td>Retrieve passages from Wikipedia that provide answers to the following question.\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td rowspan="2">3. qxt → (eimg, cxt)</td><td>EDIS</td><td>Find a news image that matches the provided caption.\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>WebQA</td><td>Find a Wikipedia image that answers this question.\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td rowspan="3">4. qimg → cxt</td><td>VisualNews</td><td>Find a caption for the news in the given photo.\nQuery: &lt; qimg &gt;&lt;eos&gt;</td></tr><tr><td>MSCOCO</td><td>Find an image caption describing the following everyday image.\nQuery: &lt; qimg &gt;&lt;eos&gt;</td></tr><tr><td>Fashion200K</td><td>Find a product description for the fashion item in the image.\nQuery: &lt; qimg &gt;&lt;eos&gt;</td></tr><tr><td>5. qimg → cimg</td><td>NIGHTS</td><td>Find a day-to-day image that looks similar to the provided image.\nQuery: &lt; qimg &gt;&lt;eos&gt;</td></tr><tr><td rowspan="2">6. (qimg, qxt) → cxt</td><td>OVEN</td><td>Retrieve a Wikipedia paragraph that provides an answer to the given query about the image.\nQuery: &lt; qimg &gt;n&lt;qimg &gt;&lt;eos&gt;</td></tr><tr><td>InfoSeek</td><td>Retrieve a Wikipedia paragraph that provides an answer to the given query about the image.\nQuery: &lt; qimg &gt;n&lt;qimg &gt;&lt;eos&gt;</td></tr><tr><td rowspan="2">7. (qimg, qxt) → cimg</td><td>FashionIQ</td><td>Find a fashion image that aligns with the reference image and style note.\nQuery: &lt; qimg &gt;n&lt;qimg &gt;&lt;eos&gt;</td></tr><tr><td>CIRR</td><td>Retrieve a day-to-day image that aligns with the modification instructions of the provided image.\nQuery: &lt; qimg &gt;n&lt;qimg &gt;&lt;eos&gt;</td></tr><tr><td rowspan="2">8. (qimg, qxt) → (cimg, cxt)</td><td>OVEN</td><td>Retrieve a Wikipedia image-description pair that provides evidence for the question of this image.\nQuery: &lt; qimg &gt;n&lt;qimg &gt;&lt;eos&gt;</td></tr><tr><td>InfoSeek</td><td>Retrieve a Wikipedia image-description pair that provides evidence for the question of this image.\nQuery: &lt; qimg &gt;n&lt;qimg &gt;&lt;eos&gt;</td></tr><tr><td>Task</td><td>Dataset</td><td>MTEB task instruction</td></tr><tr><td rowspan="15">9. qxt → cxt</td><td>ArguaAna</td><td>Given a claim, find documents that refute the claim\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>Climate-FEVER</td><td>Given a claim about climate change, retrieve documents that support or refute the claim\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>CQADupStack</td><td>Given a question, retrieve detailed question descriptions from Stackexchange that are duplicates to the given question\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>DBPedia</td><td>Given a query, retrieve relevant entity descriptions from DBPedia\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>FEVER</td><td>Given a claim, retrieve documents that support or refute the claim\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>FiQA</td><td>Given a financial question, retrieve user replies that best answer the question\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>HotpotQA</td><td>Given a multi-hop question, retrieve documents that can help answer the question\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>MSMARCO</td><td>Given a web search query, retrieve relevant passages that answer the query\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>NFCorpus</td><td>Given a question, retrieve relevant documents that best answer the question\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>NaturalQuestions</td><td>Given a question, retrieve Wikipedia passages that answer the question\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>Quora</td><td>Find questions that have the same meaning as the input question\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>SICDOCS</td><td>Given a scientific paper title, retrieve paper abstracts that are cited by the given paper\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>SciFact</td><td>Given a scientific claim, retrieve documents that support or refute the claim\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>Touch e-2020</td><td>Given a question, retrieve detailed and persuasive arguments that answer the question\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr><tr><td>TREC-COVID</td><td>Given a query on COVID-19, retrieve documents that answer the query\nQuery: &lt; qxt &gt;&lt;eos&gt;</td></tr></table>
+
+Table 17: LLaVa-P instructions for M-BEIR and MTEB. [image], [text] and [image,text] are used to inform LLaVa-P the user desired modality. For all the candidates, we use the prompt to generate the embedding:  $< c^{img}>\backslash n < c^{txt}>\backslash n$  Describe the above in one word:  
+
+<table><tr><td>Task</td><td>Dataset</td><td>M-BEIR task instruction</td></tr><tr><td rowspan="3">1. qxt→cimg</td><td>VisualNews</td><td>[ image ] &lt; qxt &gt;\nDescribe the news-related caption in one word:</td></tr><tr><td>MSCOCO</td><td>[ image ] &lt; qxt &gt;\nDescribe the everyday caption in one word:</td></tr><tr><td>Fashion200K</td><td>[ image ] &lt; qxt &gt;\nDescribe the fashion description in one word:</td></tr><tr><td>2. qxt→cxt</td><td>WebQA</td><td>[ text ] &lt; qxt &gt;\nAnswer the question using Wikipedia in one word:</td></tr><tr><td rowspan="2">3. qxt→(cimg, cxt)</td><td>EDIS</td><td>[ image.text ] &lt; qxt &gt;\nDescribe the news-related caption in one word:</td></tr><tr><td>WebQA</td><td>[ image.text ] &lt; qxt &gt;\nAnswer the question using Wikipedia in one word:</td></tr><tr><td rowspan="3">4. qimg→cxt</td><td>VisualNews</td><td>[ text ] &lt; qimg &gt;\nDescribe the news-related image in one word:</td></tr><tr><td>MSCOCO</td><td>[ text ] &lt; qimg &gt;\nDescribe the everyday image in one word:</td></tr><tr><td>Fashion200K</td><td>[ text ] &lt; qimg &gt;\nDescribe the fashion image in one word:</td></tr><tr><td>5. qimg→cimg</td><td>NIGHTS</td><td>[ image ] &lt; qimg &gt;\nDescribe the everyday image in one word:</td></tr><tr><td rowspan="2">6. (qimg, qxt)→cxt</td><td>OVEN</td><td rowspan="2">[ text ] &lt; qimg &gt;\n&lt; qxt &gt;\nAnswer the question based on the image from Wikipedia in one word:</td></tr><tr><td>InfoSeek</td></tr><tr><td rowspan="2">7. (qimg, qxt)→cimg</td><td>FashionIQ</td><td>[ image ] &lt; qimg &gt;\nChange the style of this shirt/dress/toptee to &lt; qxt &gt;\nDescribe this modified shirt/dress/toptee in one word:</td></tr><tr><td>CIRR</td><td>[ image ] &lt; qimg &gt;\nModify this image with &lt; qxt &gt;\nDescribe modified image in one word:</td></tr><tr><td rowspan="2">8. (qimg, qxt)→(cimg, cxt)</td><td>OVEN</td><td rowspan="2">[ image.text ] &lt; qimg &gt;\n&lt; qxt &gt;\nAnswer the question based on the interleaved image-text passage from Wikipedia in one word:</td></tr><tr><td>InfoSeek</td></tr><tr><td>Task</td><td>Dataset</td><td>MTEB task instruction</td></tr><tr><td rowspan="15">9. qxt→cxt</td><td>ArguAna</td><td>[ text ] &lt; qxt &gt;\nGiven a claim, generate a document that refute the claim in one word:</td></tr><tr><td>Climate-FEVER</td><td>[ text ] &lt; qxt &gt;\nGiven a claim about climate change, generate a document that supports or refutes the claim in one word:</td></tr><tr><td>CQADupStack</td><td>[ text ] &lt; qxt &gt;\nDescribe the Stackexchange question in one word:</td></tr><tr><td>DBPedia</td><td>[ text ] &lt; qxt &gt;\nGiven a query, generate a relevant entity description from DBPedia in one word:</td></tr><tr><td>FEVER</td><td>[ text ] &lt; qxt &gt;\nGiven a claim, generate a document that supports or refutes the claim in one word:</td></tr><tr><td>FiQA</td><td>[ text ] &lt; qxt &gt;\nAnswer the financial question in one word:</td></tr><tr><td>HotpotQA</td><td>[ text ] &lt; qxt &gt;\nAnswer the multi-hop question in one word:</td></tr><tr><td>MSMARCO</td><td>[ text ] &lt; qxt &gt;\nAnswer the web search query in one word:</td></tr><tr><td>NFCorpus</td><td>[ text ] &lt; qxt &gt;\nAnswer the question in one word:</td></tr><tr><td>NaturalQuestions</td><td>[ text ] &lt; qxt &gt;\nAnswer the question using Wikipedia in one word:</td></tr><tr><td>Quora</td><td>[ text ] &lt; qxt &gt;\nDescribe the question in one word:</td></tr><tr><td>SICDOCS</td><td>[ text ] &lt; qxt &gt;\nGiven a scientific paper title, generate a paper abstract that is cited by the given paper in one word:</td></tr><tr><td>SciFact</td><td>[ text ] &lt; qxt &gt;\nGiven a scientific claim, generate a document that support or refute the claim in one word:</td></tr><tr><td>Touch-e-2020</td><td>[ text ] &lt; qxt &gt;\nAnswer the question with detailed and persuasive arguments in one word:</td></tr><tr><td>TREC-COVID</td><td>[ text ] &lt; qxt &gt;\nAnswer the query on COVID-19 in one word:</td></tr></table>
+
+Table 18: Prompts for reranking tasks in M-BEIR.  
+
+<table><tr><td>Task</td><td>Dataset</td><td>Prompt</td></tr><tr><td rowspan="3">1. qxt→ cimg</td><td>VisualNews</td><td>&lt; cimg &gt;\nNews::&lt; qxt &gt;\nDoes the above News image match the News story? True or False</td></tr><tr><td>MSCOCO</td><td>&lt; cimg &gt;\nCaption::&lt; qxt &gt;\nDoes the above daily-life image match the caption? True or False</td></tr><tr><td>Fashion200K</td><td>&lt; cimg &gt;\nDescription::&lt; qxt &gt;\nDoes the above image match the cloth style description? True or False</td></tr><tr><td>2. qxt→ cxt</td><td>WebQA</td><td>Question: &lt; qxt &gt;\nAnswer::&lt; cxt &gt;\nDoes the answer correctly answer the question? True or False</td></tr><tr><td rowspan="2">3. qxt→ (cimg, cxt)</td><td>EDIS</td><td>Question: &lt; qxt &gt;\nAnswer::&lt; cxt &gt;\nDoes the answer correctly answer the question? True or False</td></tr><tr><td>WebQA</td><td>Question: &lt; qxt &gt;\nAnswer::&lt; cxt &gt;\nDoes the answer correctly answer the question? True or False</td></tr><tr><td rowspan="3">4. qimg→ cxt</td><td>VisualNews</td><td>&lt; qimg &gt;\nNews::&lt; cxt &gt;\nDoes the above News image match the News story? True or False</td></tr><tr><td>MSCOCO</td><td>&lt; qimg &gt;\nCaption::&lt; cxt &gt;\nDoes the above daily-life image match the caption? True or False</td></tr><tr><td>Fashion200K</td><td>&lt; qimg &gt;\nDescription::&lt; cxt &gt;\nDoes the above image match the cloth style description? True or False</td></tr><tr><td>5. qimg→ cimg</td><td>NIGHTS</td><td>&lt; qimg &gt;\n&lt; cimg &gt;\nDoes the above two images have the same scene? True or False</td></tr><tr><td rowspan="2">6. (qimg, qxt)→ cxt</td><td>OVEN</td><td>&lt; qimg &gt;\nQuestion::&lt; qxt &gt;\nAnswer::&lt; cxt &gt;Does the answer correctly answer the question? True or False</td></tr><tr><td>InfoSeek</td><td>&lt; qimg &gt;\nQuestion::&lt; qxt &gt;\nAnswer::&lt; cxt &gt;Does the answer correctly answer the question? True or False</td></tr><tr><td rowspan="2">7. (qimg, qxt)→ cimg</td><td>FashionIQ</td><td>&lt; cimg &gt;\nCaption::&lt; qxt &gt;\nDoes the above caption describe the modification of the image? True or False</td></tr><tr><td>CIRR</td><td>&lt; qimg &gt;\nQuestion::&lt; qxt &gt;\nAnswer::&lt; cxt &gt;Does the answer correctly answer the question? True or False</td></tr><tr><td rowspan="2">8. (qimg, qxt)→ (cimg, cxt)</td><td>OVEN</td><td>&lt; qimg &gt;\nQuestion::&lt; qxt &gt;\nAnswer::&lt; cxt &gt;Does the answer correctly answer the question? True or False</td></tr><tr><td>InfoSeek</td><td>&lt; qimg &gt;\nQuestion::&lt; qxt &gt;\nAnswer::&lt; cxt &gt;Does the answer correctly answer the question? True or False</td></tr><tr><td>Instruction: Find me an everyday image that matches the given caption.
+Query: A man brushes his teeth while a woman wraps in a towel.</td><td>Instruction: Identify the news-related image in line with the described event.
+Query: The Q Street NW entrance to the Dupont Circle Metro station.</td><td>Instruction: Find me an everyday image that matches the given caption.
+Query: A large tow truck towing a double decker bus.</td></tr><tr><td colspan="3">Correct Answer</td></tr><tr><td></td><td></td><td></td></tr><tr><td colspan="3">C1: Negative samples with incorrect modality</td></tr><tr><td>A man brushes his teeth while a woman behind him wraps a towel around herself.</td><td>Dupont Circle metro station, Q Street escalator.</td><td>A tow truck towing a double decker bus.</td></tr><tr><td>A man brushing his teeth with woman in wrapping herself in a towel in the background.</td><td>Riding escalator to Q Street exit of Dupont Circle Metro.</td><td>A tow truck is in front of a double decker bus.</td></tr><tr><td colspan="3">C2: Negative samples with unsatisfactory information needs</td></tr><tr><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td></tr></table>
+
+Figure 2: Examples of modality-aware negative samples mined by  $M^{\mathrm{rand}}$  (NVEmb). We observe that negative samples with incorrect modality show similar semantic meaning to queries, while negative samples with unsatisfactory information needs provide less accurate information compared to the correct answers.
+
+<table><tr><td colspan="5">M-BEIR CIRR Task 7</td></tr><tr><td colspan="2">Query</td><td>Answer</td><td>Retrieval</td><td>Reranking</td></tr><tr><td>Human and one animal from a different specie</td><td></td><td></td><td></td><td></td></tr><tr><td>Same breed dog, focus on its head.</td><td></td><td></td><td></td><td></td></tr><tr><td>Put the fries in a white plate with white background, clean.</td><td></td><td></td><td></td><td></td></tr><tr><td colspan="5">M-BEIR FashionIQ Task 7</td></tr><tr><td colspan="2">Query</td><td>Answer</td><td>Retrieval</td><td>Reranking</td></tr><tr><td>Is shiny and silver with shorter sleeves and fit and flare.</td><td></td><td></td><td></td><td></td></tr><tr><td>Is grey with black design and is a light printed short dress.</td><td></td><td></td><td></td><td></td></tr><tr><td>Is a solid red color and shorter and tighter with more blue and white.</td><td></td><td></td><td></td><td></td></tr></table>
+
+Figure 3: Top-1 candidates for the tasks of composed image retrieval and reranking. In many cases, retrieval and reranking yield different top-1 results from labeled positives, but these results appear to be correct since each query only has a single labeled positive candidate (see Table 10).

ICLR/2025/MM-EMBED_ UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de428eabb61ed47dbd025ccd6b70f133178b575a9d8faeaf48f4cc2c1537fc37
+size 1348792

ICLR/2025/MM-EMBED_ UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f507fc48237f51699c46f168b53a6d90307b0e2a726a93719447ebe7eab5871
+size 523352

ICLR/2025/MM1.5_ Methods, Analysis & Insights from Multimodal LLM Fine-tuning/2ec2a2fe-9d33-47f6-82ee-a26d6d7c895d_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc6294121821825f51f8f9219c1b4eecb1fa14527a0bfc3a077aaad53654c81e
+size 268430

ICLR/2025/MM1.5_ Methods, Analysis & Insights from Multimodal LLM Fine-tuning/2ec2a2fe-9d33-47f6-82ee-a26d6d7c895d_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4c9586d4f4010b29c2733622105bc6eda06d1ebb83c6384d249cb698a144266
+size 334673

ICLR/2025/MM1.5_ Methods, Analysis & Insights from Multimodal LLM Fine-tuning/2ec2a2fe-9d33-47f6-82ee-a26d6d7c895d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40acd98d3ee092b65705eadb4bc2da523a57ccdcfb0af22167c1e94db1e5f933
+size 19764172

ICLR/2025/MM1.5_ Methods, Analysis & Insights from Multimodal LLM Fine-tuning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5a9247e4f5f365f54a7845d32dfc65d2d2fba411e4d041ab4551547afd1c32f
+size 3212725

ICLR/2025/MM1.5_ Methods, Analysis & Insights from Multimodal LLM Fine-tuning/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1327df763a09c5fbe12ed012d098cd1c69c0d2cefa8ca92608a085bb840f3339
+size 1251853

ICLR/2025/MMAD_ A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection/f3eebb3a-0a75-41eb-a39e-4daa10442fdb_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c5c78d505e7ff01ebb17196db87156726820726740c419d708fba66dae81a3f
+size 136795

ICLR/2025/MMAD_ A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection/f3eebb3a-0a75-41eb-a39e-4daa10442fdb_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bcea00e952cb8a859877e2fb7404a3e4320e50edd78e7441880c5305848a4d7
+size 167038

ICLR/2025/MMAD_ A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection/f3eebb3a-0a75-41eb-a39e-4daa10442fdb_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c24b31259000dcb2d0f609b8e1c42ae79673c3092e386867885cddb1e2e7314
+size 9710803

ICLR/2025/MMAD_ A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection/full.md

@@ -0,0 +1,525 @@
+# MMAD: A COMPREHENSIVE BENCHMARK FOR MULTIMODAL LARGE LANGUAGE MODELS IN INDUSTRIAL ANOMALY DETECTION
+
+Xi Jiang $^{1}$  Jian Li $^{2}$  Hanqiu Deng $^{3}$  Yong Liu $^{2}$  Bin-Bin Gao $^{2}$  Yifeng Zhou $^{2}$   
+Jialin Li $^{2}$  Chengjie Wang $^{2,4}$  Feng Zheng $^{1*}$
+
+$^{1}$ Southern University of Science and Technology  $^{2}$ Tencent YouTu Lab  
+ $^{3}$ University of Alberta  $^{4}$ Shanghai Jiao Tong University  
+jiangx2020@mail.sustech.edu.cn, hanqiul@ualberta.ca,  
+{swordli, choasliu, danylgao, joefzhou, jarenli, jasoncjwang}  
+@tencent.com, f.zheng@ieee.org
+
+# ABSTRACT
+
+In the field of industrial inspection, Multimodal Large Language Models (MLLMs) have a high potential to renew the paradigms in practical applications due to their robust language capabilities and generalization abilities. However, despite their impressive problem-solving skills in many domains, MLLMs' ability in industrial anomaly detection has not been systematically studied. To bridge this gap, we present MMAD, a full-spectrum MLLM benchmark in industrial Anomaly Detection. We defined seven key subtasks of MLLMs in industrial inspection and designed a novel pipeline to generate the MMAD dataset with 39,672 questions for 8,366 industrial images. With MMAD, we have conducted a comprehensive, quantitative evaluation of various state-of-the-art MLLMs. The commercial models performed the best, with the average accuracy of GPT-40 models reaching  $74.9\%$ . However, this result falls far short of industrial requirements. Our analysis reveals that current MLLMs still have significant room for improvement in answering questions related to industrial anomalies and defects. We further explore two training-free performance enhancement strategies to help models improve in industrial scenarios, highlighting their promising potential for future research. The code and data are available at https://github.com/jam-cc/MMAD.
+
+# 1 INTRODUCTION
+
+Automatic vision inspection is a crucial challenge in realizing an unmanned factory (Benbarrad et al., 2021). Traditional AI research for automatic vision inspection, such as industrial anomaly detection (IAD) (Jiang et al., 2022b; Ren et al., 2022; Zhang et al., 2024b), typically relies on discriminative models within the conventional deep learning paradigm. These models can only perform trained detection tasks and cannot provide detailed reports like quality inspection workers. Additionally, when production lines or requirements change, traditional methods necessitate retraining or redevelopment. The development of MLLMs (Jin et al., 2024) has the potential to alter this situation. These generative models can flexibly produce the required textual output based on input language and visual prompts, allowing us to guide the model using language similar to instructing humans.
+
+Nowadays, multimodal large language models, represented by GPT-4 (Achiam et al., 2023), can already do many human jobs, especially high-paying intellectual jobs like programmers, writers, and data analysts (Eloundou et al., 2023). In comparison, the work of quality inspectors is simple, typically not requiring a high level of education but relying heavily on work experience. Therefore, we are greatly interested in the question:
+
+How well are current MLLMs performing as industrial quality inspectors?
+
+![](images/d42791e017a052a11da8349bac9fbcc0c5b791cbbb1c360c489420f70e910dc3.jpg)  
+Figure 1: Left: Innermost layer: image components, middle layer: subtasks composition, outermost layer: object categories. MMAD covers 7 key subtasks and 38 representative categories of IAD. Right: Results of 5 representative MLLMs and Human. The left-skewness indicates that models perform well on object-related questions but poorly on questions related to defects.
+
+![](images/3b76024e52f25c8268936862bf178d3ce101a72fe24538f7ad1dd800b995ee24.jpg)
+
+Recent studies have attempted to explore this issue. Through instruction following mechanisms (Dai et al., 2023; Liu et al., 2024b), some reports evaluate IAD examples with MLLMs, demonstrating the advantages of MLLMs in generalization and flexibility (Zhang et al., 2023a; Cao et al., 2023; Xu et al., 2024). Unfortunately, they only tested a few qualitative examples with no quantitative results. Other studies, such as AnomalyGPT (Gu et al., 2024) and Myriad (Li et al., 2023), specifically trained MLLMs to understand the outputs of traditional IAD models. However, they use the traditional evaluation, which measures the ability of expert models rather than MLLMs. The lack of unified test data and output format also led to an incomplete comparison with other general models. So, a comprehensive benchmark to compare the quantitative results of MLLMs in IAD is necessary.
+
+In this paper, we introduce the first-ever benchmark for MLLMs in IAD tasks. Current MLLMs are challenging to evaluate directly on IAD tasks because existing publicly available datasets only contain visual perception annotations and category labels, lacking rich semantic annotations. To address this issue, we designed a comprehensive pipeline. First, we utilized GPT-4V to generate rich semantic annotations with visual annotations and language interaction. Based on the semantic annotations, we generated questions and options for testing, which were manually reviewed to ensure their reasonableness and accuracy. Ultimately, we collected 8,366 samples from 38 classes of industrial products across 4 public datasets, generating a total of 39,672 multiple-choice questions in 7 key subtasks, as illustrated in the left panel of Figure 1. Our benchmark encompasses representative image data and language-based evaluation methods, providing a fair and reasonable assessment of MLLMs' performance in IAD. Figure 2 provides some examples.
+
+With MMAD, we have conducted a comprehensive, quantitative evaluation of various state-of-the-art (SOTA) MLLMs, including the GPT-4 series and Gemini 1.5 series (Reid et al., 2024), as well as open-source image models like InternVL2 (Chen et al., 2023) and LLaVA-NeXT (Liu et al., 2024a), and industry anomaly detection models like AnomalyGPT (Gu et al., 2024). As shown in the right panel of Figure 1, our experiments demonstrate that GPT-4o is the best-performing model, reaching  $74.9\%$ , but as the scale of models increases, some open-source models are gradually approaching the capabilities of these commercial models. However, industrial scenarios often demand high accuracy, and the current accuracy levels are far from meeting practical standards. Additionally, both open-source and proprietary models perform significantly worse when answering questions related to anomalies and defects compared to questions about objects, which reveals an important capability weakness of MLLMs. Notably, the model specifically designed for IAD, AnomalyGPT (Gu et al., 2024), performed the worst overall, as it could only handle anomaly discrimination tasks it was trained on and failed to respond adequately to questions in other subtasks. This indicates that enhancing MLLMs' capabilities in IAD through training requires larger-scale industrial data and the provision of more comprehensive industrial knowledge.
+
+![](images/5e1ce32557230325ef7f70620cd6e8f24df77ea23e19a1aee17ca9bc3016af33.jpg)  
+Figure 2: Examples of 7 subtasks of MMAD. Each question is presented in a multiple-choice format and includes several distractor options. We present different categories of objects in various examples to demonstrate the diversity.
+
+Beyond standard testing, we also conducted experiments on different settings and scaling laws, which aid in further comparative studies. Given the current shortcomings of MLLMs in addressing anomalies and defects, we explored two performance boost schemes that do not require additional training: Retrieval-Augmented Generation (RAG) and expert agent, the previous one through text augmentation and the other through visual augmentation. These methods can improve MLLMs' performance in IAD to some extent, but they are still limited by the fundamental capabilities of the models. Overall, current MLLMs are not yet capable of effectively performing the duties of a quality inspector. They require further supplementation with IAD knowledge and enhanced capabilities for fine-grained understanding and cross-comparison of multiple images.
+
+The contributions of this paper are summarized as follows:
+
+- To the best of our knowledge, our proposed MMAD is the first evaluation benchmark for MLLMs in the task of IAD. This benchmark fills the gap in the application of MLLMs in the industrial domain and sets new challenges for their capabilities.  
+- We introduce a novel pipeline for generating semantic annotations for visual anomaly detection data, addressing the issue that MLLMs cannot be directly evaluated on IAD datasets.  
+- We comprehensively evaluate the performance of representative MLLMs on MMAD, highlight the weaknesses of current MLLMs in fine-grained industrial knowledge and multi-image understanding, and provide two general boost schemes.
+
+# 2 RELATED WORK
+
+# 2.1 MULTIMODAL BENCHMARKS
+
+In recent years, substantial efforts on benchmarks have been dedicated to exploring MLLMs (Li & Lu, 2024) from various perspectives. However, the capabilities assessed by these benchmarks differ significantly from those required for IAD. Firstly, most MLLM benchmarks primarily focus on single-image input scenarios, such as MME (Fu et al., 2023), MMBench (Liu et al., 2023b), and MMVP (Tong et al., 2024b). In contrast, industrial quality inspection necessitates the ability to process multiple images, as industrial knowledge is highly specialized and often requires additional
+
+Table 1: Statistics on the composition and quantity of MMAD data.  
+
+<table><tr><td>Image Source</td><td>Specialty</td><td>Sampled Images</td><td>Generated Questions</td><td>Object Categories</td><td>Defect Categories</td></tr><tr><td>MVTec AD (Bergmann et al., 2019)</td><td>A variety of objects and textures</td><td>1691</td><td>8338</td><td>15</td><td>73</td></tr><tr><td>MVTec LOCO AD (Bergmann et al., 2022)</td><td>Including both structural and logical anomalies</td><td>1566</td><td>7624</td><td>5</td><td>89</td></tr><tr><td>VisA (Zou et al., 2022)</td><td>Containing multiple instances and complex structure</td><td>2141</td><td>10622</td><td>12</td><td>67</td></tr><tr><td>GoodsAD (Zhang et al., 2024a)</td><td>Multiple goods in each category</td><td>2968</td><td>13088</td><td>6</td><td>15</td></tr><tr><td>SUM</td><td>-</td><td>8366</td><td>39672</td><td>38</td><td>244</td></tr></table>
+
+images for product recognition. Secondly, most benchmarks emphasize the general capabilities of MLLMs rather than the specific needs of particular domains. For instance, TextVQA (Singh et al., 2019) focuses on text recognition in images, MATH-Vision (Wang et al., 2024c) and MATH-VISTA(Lu et al., 2023) evaluates mathematical reasoning in visual contexts, and Video-MME (Fu et al., 2024) emphasizes video understanding. Most importantly, these benchmarks do not address the industrial domain. For example, MMMU (Yue et al., 2024) covers lots of fields, such as Art, Business, Science, Social Science, and Engineering, but not industry. Seed-Bench (Li et al., 2024a) and CompBench (Kil et al., 2024) test multi-image comparison capabilities, but both involve fictional tasks and cannot provide realistic references. In contrast, the medical field, which has tasks and data formats similar to the industrial domain, already has numerous benchmarks for MLLMs, including Asclepius (Wang et al., 2024d), GMAI-MMBench (Chen et al., 2024b) and PMC-VQA (Zhang et al., 2023b). Therefore, proposing the first MLLMs benchmark for the industrial domain would fill a significant gap.
+
+# 2.2 INDUSTRIAL ANOMALY DETECTION
+
+Traditional IAD research primarily aims to address a significant issue in industrial visual inspection: discriminating and localizing defects without samples of defects. Consequently, traditional IAD methods typically involve training on a large number of normal samples and then using outlier detection techniques to identify anomalies in test samples. Common approaches include memory bank-based methods (Jiang et al., 2022a; Wang et al., 2025), reconstruction-based methods (Deng & Li, 2022; Jiang et al., 2024a), and methods based on training with synthetic anomalies (Zavrtanik et al., 2021). Recent research has begun to focus on the generalization capability of IAD. Leveraging vision-language models such as CLIP, some few-shot, and even zero-shot models have emerged, such as AnomalyCLIP (Zhou et al., 2023), M3DM-NR (Wang et al., 2024a) and InCTRL (Zhu & Pang, 2024). However, these discriminative models heavily rely on predefined anomaly concepts in the CLIP model, limiting their ability to generalize to new scenarios. For instance, models trained on structural anomalies struggle to detect logical anomalies (Bergmann et al., 2022). The emergence of MLLMs may help address this challenge, as they can understand complex textual inputs and provide diverse responses in conjunction with visual components.
+
+Some studies have already demonstrated the advantages of open input-output formats of MLLMs in IAD. For example, LogiCode (Zhang et al., 2024c) uses MLLMs for logical reasoning and automatically generates code to address different types of logical anomalies. Xu et al. (2024) proposes that designing visual and language prompts can directly enhance the performance of MLLMs in IAD tasks. Additionally, some research has begun to train MLLMs directly on IAD datasets, such as AnomalyGPT (Gu et al., 2024), Myriad (Li et al., 2023), and FabGPT (Jiang et al., 2024b). However, these three models heavily depend on the capabilities of expert models and cannot freely extend their capabilities by changing inputs. Moreover, we have found that due to the small amount of IAD data (Wang et al., 2024b), MLLMs like AnomalyGPT are highly prone to overfitting. In conclusion, current IAD methods rarely meet our paradigm, so we focus our evaluation on general MLLMs.
+
+# 3 THE MMAD DATASET
+
+# 3.1 DATA COLLECTIONS
+
+An excellent industrial quality inspector should be able to adapt to the inspection tasks of different products, as the skills involved in visual inspection are similar. Therefore, our designed benchmark
+
+![](images/eedc0589df662c21e641a6e9726c8238c856a6af04e2a699370502ba63a0c842.jpg)  
+Figure 3: The VQA data generation pipeline for IAD. We utilize images from the open-source IAD dataset and leverage GPT-4V to automate the generation of question-answer texts. Initially, the model is prompted to provide detailed captions for IAD images by summarizing visual cues and textual prior knowledge. Based on these image-caption pairs, the model then generates questions across different subtasks according to predefined question definitions and examples, simultaneously creating multiple-choice questions with several distractor options. Finally, manual verification is conducted to filter out low-quality VQA pairs, resulting in high-quality VQA data for the IAD.
+
+needs to cover multiple scenarios of IAD. We achieved data diversity by collecting and sampling from four different IAD datasets with distinct focuses, resulting in over 38 product categories and 244 defect types, as detailed in Table 1. Among these, MVTec AD (Bergmann et al., 2019) is one of the most prominent datasets for IAD, encompassing multiple categories of objects and textures, where we use finer annotations from Defect Spectrum (Yang et al., 2023). MVTec LOCO AD (Bergmann et al., 2022) focuses on logical anomalies, thereby testing the model's understanding of logical-level anomalies. The VisA (Zou et al., 2022) dataset includes multiple instances and complex examples, reflecting more intricate IAD scenarios. The GoodsAD (Zhang et al., 2024a) dataset primarily consists of industrial goods, and due to the varying appearances of finished products from different brands, it significantly expands the number of object categories.
+
+# 3.2 QUESTION DEFINITION
+
+To evaluate whether MLLMs can fulfill the role of an industrial quality inspector, it is necessary to test a wide range of capabilities. On the production line, a quality inspector must not only identify defective samples but also classify and grade defects, analyze causes, and diagnose faults. Therefore, in addition to basic anomaly detection, we have designed four anomaly-related subtasks and two object-related subtasks, as detailed below:
+
+- Anomaly Discrimination/Detection: Asking whether a sample has defects. These questions are binary classification problems that test the ability to judge anomalies in samples.  
+- Defect Classification: Asking about the type of defect. It verifies the model's basis for anomaly discrimination while also testing its understanding of industrial defect categories.  
+- Defect Localization: The model needs to specify the exact location of the defect, further verifying the model's basis for anomaly detection. Since most MLLMs cannot directly output masks, we standardize the testing by using textual descriptions to indicate the location.  
+- Defect Description: Describing the appearance characteristics of the defect. It simulates the steps of collecting information on the size, color, and other attributes of defects in actual production. This information can be used to determine the defect's grade according to production standards and to infer whether the defect is related to equipment malfunction.  
+- Defect Analysis: This involves analyzing the potential impact of the defect on the product, which is used to further determine the defect's severity level.  
+- Object Classification: Categorization of industrial products. An understanding of industrial products aids in recognizing normal characteristics and identifying anomalies.
+
+![](images/a2bcc9d3cd7e1f756009149c32747d19626564784b67d1861167094b859860bd.jpg)  
+Figure 4: Illustration of two proposed boost methods of MLLMs in MMAD.
+
+- Object Analysis: This involves questioning the composition, position, appearance, and function of the object, aiming to assess the detailed understanding of the specific product.
+
+Some examples are shown in the Figure 2. To better evaluate the outputs of MLLMs, we use multiple-choice questions, a method proven effective in previous research (Li et al., 2024a; Liu et al., 2023b). To avoid biases inherent in language models, we randomize the options in our questions and use text matching to select the closest option when the model cannot correctly follow instructions and output the answer's letter.
+
+# 3.3 DATA GENERATION
+
+Due to the lack of semantic annotation in open-source IAD datasets, the currently collected data cannot be directly used to evaluate MLLMs. Therefore, we designed a novel pipeline to generate evaluation questions for each IAD image. As shown in Figure 3, our process leverages the text generation capabilities of GPT-4V (Achiam et al., 2023), combined with rule-based program outputs, language prompts, and human filtering to ensure the reliability of the generated content. First, we generate rich textual descriptions for each image. Since GPT-4V is primarily trained on natural scenes and may struggle with industrial quality inspection, we provide additional visual and textual prompts. In the visual prompts, we highlight the ground truth mask in red on the original image to make the model aware of the defects. Additionally, we retrieve the most similar normal image to serve as a comparison template, using a similarity metric combining the SSIM score (Wang et al., 2004) and the Bhattacharyya distance (Bhattacharyya, 1943) of color histograms. The language prompts include object and defect category labels and textual descriptions of defect positions within the image, utilizing the nine-grid division proposed by Gu et al. (2024). To ensure diversity, we designed instructions for caption generation with variations in the prior knowledge provided, preventing the captions from merely copying the supplied text. Once we have the captions for each image, we generate questions, options, and answers based on predefined subtasks. Unlike natural images, each IAD image corresponds to multiple questions, which are generated simultaneously and then filtered through manual verification. This verification process involved 26 individuals and took over 200 working hours.
+
+# 3.4 BOOST METHODS
+
+Retrieval-Augmented Generation (RAG). RAG is a method that combines information retrieval and generation to enhance the performance of language models, particularly for tasks requiring external knowledge (Zhao et al., 2024). Knowledge related to IAD is often specialized and rarely encountered during the training of Multimodal Language Models (MLLMs). Therefore, we propose a RAG method tailored for IAD, as illustrated in Figure 4(a). Experts, with the assistance of large language models, first summarize the existing IAD datasets. For each category, they summarize the characteristics of normal samples and the features of each possible anomaly. The domain knowledge summarized from all datasets forms a retrieval database. During testing, the query image is used to retrieve the relevant category knowledge, which is then incorporated into the text prompts.
+
+Table 2: Performance comparison of both proprietary and open-source MLLMs in MMAD with the standard 1-shot setting. Anomaly Discrimination uses the average accuracy of normal and abnormal categories. (*For the methods not supporting multi-image input, the 0-shot result is reported.)  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Scale</td><td>Anomaly</td><td colspan="4">Defect</td><td colspan="2">Object</td><td rowspan="2">Average</td></tr><tr><td>Discrimination</td><td>Classification</td><td>Localization</td><td>Description</td><td>Analysis</td><td>Classification</td><td>Analysis</td></tr><tr><td>Random Chance</td><td>-</td><td>50.00</td><td>25.00</td><td>25.00</td><td>25.00</td><td>25.00</td><td>25.00</td><td>25.00</td><td>28.57</td></tr><tr><td>Human (expert)</td><td>-</td><td>95.24</td><td>75.00</td><td>92.31</td><td>83.33</td><td>94.20</td><td>86.11</td><td>80.37</td><td>86.65</td></tr><tr><td>Human (ordinary)</td><td>-</td><td>86.90</td><td>66.25</td><td>85.58</td><td>71.25</td><td>81.52</td><td>89.58</td><td>69.72</td><td>78.69</td></tr><tr><td>Claude-3.5-sonnet</td><td>-</td><td>60.14</td><td>60.14</td><td>48.81</td><td>67.13</td><td>79.11</td><td>85.19</td><td>79.83</td><td>68.36</td></tr><tr><td>Gemini-1.5-flash</td><td>-</td><td>58.58</td><td>54.70</td><td>49.10</td><td>66.53</td><td>82.24</td><td>91.47</td><td>79.71</td><td>68.90</td></tr><tr><td>Gemini-1.5-pro</td><td>-</td><td>68.63</td><td>60.12</td><td>58.56</td><td>70.38</td><td>82.46</td><td>89.20</td><td>82.25</td><td>73.09</td></tr><tr><td>GPT-4o-mini</td><td>-</td><td>64.33</td><td>48.58</td><td>38.75</td><td>63.68</td><td>80.40</td><td>88.56</td><td>79.74</td><td>66.29</td></tr><tr><td>GPT-4o</td><td>-</td><td>68.63</td><td>65.80</td><td>55.62</td><td>73.21</td><td>83.41</td><td>94.98</td><td>82.80</td><td>74.92</td></tr><tr><td>AnomalyGPT</td><td>7B</td><td>65.57</td><td>27.49</td><td>27.97</td><td>36.86</td><td>32.11</td><td>29.84</td><td>35.82</td><td>36.52</td></tr><tr><td>Qwen-VL-Chat</td><td>7B</td><td>53.65</td><td>31.33</td><td>28.62</td><td>41.66</td><td>63.99</td><td>74.46</td><td>67.94</td><td>51.66</td></tr><tr><td>LLaVA-1.5</td><td>7B</td><td>51.33</td><td>37.04</td><td>36.62</td><td>50.60</td><td>69.79</td><td>68.29</td><td>69.53</td><td>54.74</td></tr><tr><td>Cambrian-1*</td><td>8B</td><td>55.60</td><td>32.53</td><td>35.39</td><td>43.46</td><td>49.14</td><td>78.15</td><td>67.22</td><td>51.64</td></tr><tr><td>SPHINX*</td><td>7B</td><td>53.13</td><td>33.93</td><td>52.27</td><td>50.96</td><td>71.23</td><td>85.07</td><td>73.10</td><td>59.96</td></tr><tr><td>LLaVA-NEXT-Interleave</td><td>7B</td><td>57.64</td><td>33.79</td><td>47.72</td><td>51.84</td><td>67.93</td><td>81.39</td><td>74.91</td><td>59.32</td></tr><tr><td>InternLM-XComposer2-VL</td><td>7B</td><td>55.85</td><td>41.80</td><td>48.27</td><td>57.52</td><td>76.60</td><td>74.34</td><td>77.75</td><td>61.73</td></tr><tr><td>LLaVA-OneVision</td><td>7B</td><td>51.77</td><td>46.13</td><td>41.85</td><td>62.19</td><td>69.73</td><td>90.31</td><td>80.93</td><td>63.27</td></tr><tr><td>MiniCPM-V2.6</td><td>8B</td><td>57.31</td><td>49.22</td><td>43.28</td><td>65.86</td><td>75.24</td><td>92.02</td><td>80.80</td><td>66.25</td></tr><tr><td>InternVL2</td><td>8B</td><td>59.97</td><td>43.85</td><td>47.91</td><td>57.60</td><td>78.10</td><td>74.18</td><td>80.37</td><td>63.14</td></tr><tr><td>LLaVA-1.5</td><td>13B</td><td>49.96</td><td>38.78</td><td>46.17</td><td>58.17</td><td>73.09</td><td>73.62</td><td>70.98</td><td>58.68</td></tr><tr><td>LLaVA-NeXT</td><td>34B</td><td>57.92</td><td>48.79</td><td>52.87</td><td>71.34</td><td>80.28</td><td>81.12</td><td>77.80</td><td>67.16</td></tr><tr><td>InternVL2</td><td>76B</td><td>68.25</td><td>54.22</td><td>56.66</td><td>66.30</td><td>80.47</td><td>86.40</td><td>82.92</td><td>70.75</td></tr></table>
+
+Expert Agent. Tool agent is an independent module within a large model, responsible for specific functions (Xi et al., 2023). In the evaluation of MMAD, exploring whether expert models can enhance MLLMs is an intriguing topic. We designed a simple method to investigate its feasibility. As shown in Figure 4(b), we treat the IAD model as a visual expert model. Since the output anomaly map is difficult to understand by MLLMs, we visualize it and then input it into the MLLMs. We use SOTA models in IAD, such as AnomalyCLIP and PatchCore, as experts and test the differences among 3 visualization methods. Additionally, we applied a special expert, ground truth, which directly uses the mask as the output of the expert model to verify the upper limit.
+
+# 4 EXPERIMENTS
+
+# 4.1 EVALUATION SETTINGS
+
+We default to a 1-shot setting, where, in addition to the query image, we randomly provide a normal image from the dataset and inform the model that it can use this image as a template. The template image helps the model understand the normal state of objects, which is essential for anomaly detection. We also provide 0-shot and few-shot settings for further comparison. Besides, randomly providing normal images may lead to misunderstandings in some object categories due to multiple normal states existing. We introduce a 1-shot $^{+}$  setting, which retrieves the most similar image from the normal data as a template. This setting tests the model's ability to perform pairwise comparisons of industrial images.
+
+For commercial models, we conducted tests using APIs, specifically versions claude-3-5-sonnet-20241022, gemini-1.5-flash-001, gemini-1.5-pro-001, gpt-4o-mini-2024-07-18, and gpt-4o-2024-05-13. For open-source models, we adapted and tested AnomalyGPT (Gu et al., 2024), Qwen-VL-Chat (Bai et al., 2023), LLaVA-1.5 (Liu et al., 2023a), Cambrian-1 (Tong et al., 2024a), SPHINX (Lin et al., 2023), LLaVA-NeXT (Liu et al., 2024a), LLaVA-NEXT-Interleave Li et al. (2024b), InternLM-XComposer2-VL (Dong et al., 2024), MiniCPM-V2.6 (Yao et al., 2024), and InternVL2 (Chen et al., 2023). We endeavored to maintain consistency with the default hyperparameters and prompt methods provided. For models exceeding 20B parameters, we employed appropriate quantization methods. Each image's multiple VQA tasks were tested independently, and caches were cleared after each test. Some models only support single-image input by default. For LLaVA-1.5, we modified the framework to support multi-image input, while for Cambrian-1 and SPHINX, we only tested single-image input (i.e., 0-shot) performance. We will randomize the letters and order of the options and use the accuracy of responses as a metric. If the model does not
+
+Table 3: Performance comparison with and without template images. (1-shot+ indicates using a retrieved image as the template to provide more helpful visual information.)  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Scale</td><td rowspan="2">Setting</td><td colspan="2">Anomaly</td><td colspan="4">Defect</td><td rowspan="2">Average</td></tr><tr><td>Discrimination</td><td>Classification</td><td>Localization</td><td>Description</td><td colspan="2">Analysis</td></tr><tr><td rowspan="2">Gemini-1.5-flash</td><td rowspan="2">-</td><td>0-shot</td><td>58.43</td><td>49.93</td><td>53.11</td><td>63.07</td><td>82.83</td><td></td><td>68.58</td></tr><tr><td>1-shot+</td><td>62.22 (+3.79)</td><td>52.29 (+2.37)</td><td>50.99 (-2.12)</td><td>65.52 (+2.45)</td><td>83.41 (+0.59)</td><td></td><td>69.59 (+1.01)</td></tr><tr><td rowspan="2">LLaVA-1.5</td><td rowspan="2">7B</td><td>0-shot</td><td>50.79</td><td>36.94</td><td>35.94</td><td>50.86</td><td>70.45</td><td></td><td>54.91</td></tr><tr><td>1-shot+</td><td>50.99 (+0.21)</td><td>37.14 (+0.20)</td><td>35.63 (-0.31)</td><td>50.75 (-0.11)</td><td>70.01 (-0.44)</td><td></td><td>54.66 (-0.25)</td></tr><tr><td rowspan="2">LLaVA-1.5</td><td rowspan="2">13B</td><td>0-shot</td><td>49.98</td><td>38.77</td><td>48.64</td><td>57.65</td><td>72.47</td><td></td><td>58.58</td></tr><tr><td>1-shot+</td><td>49.97 (-0.01)</td><td>39.44 (+0.67)</td><td>45.22 (-3.42)</td><td>58.64 (+0.99)</td><td>73.28 (+0.81)</td><td></td><td>58.55 (-0.02)</td></tr><tr><td rowspan="2">LLaVA-NeXT</td><td rowspan="2">34B</td><td>0-shot</td><td>60.25</td><td>51.57</td><td>55.49</td><td>71.62</td><td>80.43</td><td></td><td>68.45</td></tr><tr><td>1-shot+</td><td>56.80 (-3.45)</td><td>48.87 (-2.70)</td><td>52.80 (-2.69)</td><td>70.82 (-0.79)</td><td>80.00 (-0.43)</td><td></td><td>67.16 (-1.28)</td></tr><tr><td rowspan="2">LLaVA-NEXT-Interleave</td><td rowspan="2">7B</td><td>0-shot</td><td>58.39</td><td>36.98</td><td>48.98</td><td>51.51</td><td>66.64</td><td></td><td>60.04</td></tr><tr><td>1-shot+</td><td>57.71 (-0.68)</td><td>35.87 (-1.11)</td><td>48.11 (-0.86)</td><td>52.40 (+0.89)</td><td>67.88 (+1.24)</td><td></td><td>59.82 (-0.22)</td></tr><tr><td rowspan="2">InternLM-XComposer2-VL</td><td rowspan="2">7B</td><td>0-shot</td><td>58.33</td><td>43.10</td><td>54.56</td><td>57.84</td><td>75.30</td><td></td><td>62.78</td></tr><tr><td>1-shot+</td><td>55.87 (-2.46)</td><td>42.76 (-0.34)</td><td>49.68 (-4.88)</td><td>58.03 (+0.20)</td><td>76.52 (+1.23)</td><td></td><td>61.99 (-0.79)</td></tr><tr><td rowspan="2">InternVL2</td><td rowspan="2">40B</td><td>0-shot</td><td>66.37</td><td>48.54</td><td>53.39</td><td>64.05</td><td>79.01</td><td></td><td>69.23</td></tr><tr><td>1-shot+</td><td>67.69 (+1.32)</td><td>51.15 (+2.61)</td><td>55.81 (+2.42)</td><td>66.81 (+2.76)</td><td>79.83 (+0.82)</td><td></td><td>70.71 (+1.49)</td></tr><tr><td rowspan="2">InternVL2</td><td rowspan="2">76B</td><td>0-shot</td><td>64.30</td><td>51.19</td><td>54.20</td><td>63.46</td><td>79.92</td><td></td><td>69.41</td></tr><tr><td>1-shot+</td><td>69.23 (+4.93)</td><td>54.69 (+3.50)</td><td>57.20 (+3.00)</td><td>66.75 (+3.29)</td><td>80.46 (+0.54)</td><td></td><td>71.31 (+1.90)</td></tr></table>
+
+provide any option, we will automatically match the closest option to the output as the answer. It is worth noting that, in the anomaly discrimination subtask, due to the imbalance distribution, we will separately calculate the accuracy of normal and abnormal samples and then use their mean as the final accuracy.
+
+# 4.2 EXPERIMENTAL RESULTS
+
+We compare the performance of over a dozen models, including commercial APIs, interleaved MLLMs, industrial MLLMs, and vision-centric MLLMs, as shown in Table 2. All models outperform the random baseline. The open-source models perform the best, with the average accuracy of the GPT-4o and Gemini-1.5-pro models reaching  $74.9\%$  and  $73\%$ , respectively. However, their cost-efficient counterparts, GPT-4o-mini and Gemini-1.5-flash, only achieved  $66.3\%$  and  $68.9\%$ , respectively, falling short of the best open-source model, InternVL2-76B, which achieved  $70.8\%$ . AnomalyGPT performs poorly overall, primarily due to its training on the IAD task in a fixed question-and-answer format, leading to severe overfitting issues. It demonstrates decent performance in anomaly discrimination because we specifically adapted the question format to suit its training. Similarly, the vision-centric MLLMs, Cambrian-1 and SPHINX, do not exhibit superior performance on the fine-grained visual tasks of MMAD, likely due to their foundational language models not being advanced enough. Among the general open-source MLLMs, earlier models like Qwen-VL-Chat and LLaVA-1.5 underperform compared to newer models like LLaVA-OneVision and MiniCPM-V2.6, indicating that advancements in general capabilities benefit performance on IAD tasks. MMAD uses a default 1-shot format, providing a normal image for comparison with the test image. Thus, multi-image understanding, especially image comparison, is crucial, while LLaVA-NEXT-Interleave, trained for this, does not perform outstandingly. LLaVA-NeXT-34B and InternVL2-76B, due to their larger scales, achieve the top two performances among open-source models, highlighting the importance of model size.
+
+Human evaluation. We conduct a preliminary human evaluation using 177 examples randomly sampled from the entire benchmark. Eight evaluators were divided into two groups: 3 industrial anomaly detection researchers as experts and 5 ordinary participants. As shown in Table 2, ordinary outperform the best models by  $4\%$ , and experts by  $12\%$ . Humans excel in anomaly and defect-related tasks, highlighting MMAD's challenges and MLLMs' limitations in anomaly detection. However, in object-related VQA, GPT-4o surpasses humans due to its broader knowledge base, with examples provided in the appendix B.
+
+![](images/f9711664933005f52e81981e8d18fcc85cda406c097106c3a5bfa584ca3b5263.jpg)  
+Figure 5: Left: Scaling law of model size in MMAD evaluation. We use the InternVL2 series as examples. Right: The performance trends of InternVL2-76B with varying counts of normal samples.
+
+![](images/4ba89f829740df35287af364f6da78de37f6b9a7631585674025375837a816b8.jpg)
+
+# 4.3 FURTHER ANALYSES
+
+Can MLLMs effectively utilize template normal images? Anomaly detection often benefits from template images to understand normal patterns. To examine whether MLLMs can utilize templates, we conducted a comparative experiment. As shown in Table 3, we evaluated models in 0-shot and 1-shot $^{+}$  settings. In 0-shot, no template image is provided, while in 1-shot $^{+}$ , the closest normal image to the test image is used as a template for comparison. Results indicate that most models fail to leverage template information, often resulting in performance drops. Due to cost constraints, we tested only the cost-effective Gemini-1.5-flash among commercial models, which showed notable gains, suggesting open-source models excel in contextual image understanding. Among open-source models, only the larger-scale InternVL2 effectively utilized template information.
+
+How significant is the impact of model scale on performance? In the main experiment results, we observed that larger MLLMs generally exhibit better performance. We evaluate the scaling law using the InternVL2 series models, which utilize the same training data and are among the best-performing open-source models. As shown in the left panel of Figure 5, performance significantly improves with increasing model size, with the average accuracy difference between the largest and smallest models reaching  $23.37\%$ . Specifically, the classification performance of industrial objects improves the most with increasing size, and there are notable enhancements in several subtasks related to anomalies and defects. Although the performance improvement trend slows as the model size increases, further enlarging the model remains a promising option, albeit one that must be balanced against cost considerations.
+
+Can increasing the number of images further enhance performance? For traditional few-shot IAD models, increasing the number of normal samples significantly enhances the performance of anomaly detection and localization. Therefore, we investigate whether MLLMs can leverage additional normal images. Experiments with InternVL2-76B (Figure 5, right) show performance improves from 0- to 1-shot, but further increases yield minimal gains (only slight improvements in anomaly/defect subtasks up to 4-shot). Beyond 8-shot, the performance in some subtasks even begins to decline, possibly due to information overload caused by too many input images, leading to confusion in the MLLMs. After providing multiple normal samples, the model needs to first summarize the normal characteristics and then compare them with the samples under inspection, while such tasks rarely appear in current MLLM instruction tuning data. This highlights the need for future research focusing on industrial applications' unique requirements.
+
+# 4.4 EXPLORATION
+
+Input Domain Knowledge to MLLMs. One significant challenge identified in our evaluation is the lack of industrial knowledge in MLLMs for specific tasks. During the training, MLLMs rarely encounter knowledge related to industrial quality inspection, while different products correspond to different specific knowledge. In practical applications, senior experts typically train workers. We use RAG to provide domain knowledge guidance to MLLMs to simulate this process. As shown in Table
+
+Table 4: Performance comparison of different MLLMs with and without RAG.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">RAG</td><td colspan="2">Anomaly</td><td colspan="4">Defect</td><td colspan="2">Object</td><td>Average</td></tr><tr><td>Discrimination</td><td>Classification</td><td>Localization</td><td>Description</td><td>Analysis</td><td>Classification</td><td>Analysis</td><td></td><td></td></tr><tr><td rowspan="2">LLaVA-NeXT-34B</td><td>-</td><td>57.92</td><td>48.79</td><td>52.87</td><td>71.34</td><td>80.28</td><td>81.12</td><td>77.80</td><td>67.16</td><td></td></tr><tr><td>✓</td><td>56.72 (-1.19)</td><td>68.22 (+19.43)</td><td>57.36 (+4.49)</td><td>73.12 (+1.79)</td><td>82.24 (+1.96)</td><td>91.78 (+10.66)</td><td>81.35 (+3.55)</td><td>72.97 (+5.81)</td><td></td></tr><tr><td rowspan="2">InternVL2-26B</td><td>-</td><td>67.96</td><td>49.77</td><td>49.44</td><td>62.62</td><td>80.92</td><td>88.16</td><td>81.39</td><td>68.61</td><td></td></tr><tr><td>✓</td><td>68.64 (+0.68)</td><td>67.32 (+17.55)</td><td>53.81 (+4.37)</td><td>70.84 (+8.22)</td><td>82.18 (+1.26)</td><td>93.81 (+5.64)</td><td>83.31 (+1.92)</td><td>74.27 (+5.66)</td><td></td></tr><tr><td rowspan="2">InternVL2-40B</td><td>-</td><td>64.45</td><td>50.57</td><td>53.42</td><td>66.17</td><td>79.56</td><td>90.65</td><td>82.36</td><td>69.59</td><td></td></tr><tr><td>✓</td><td>70.01 (+5.56)</td><td>70.09 (+19.53)</td><td>56.89 (+3.47)</td><td>73.29 (+7.12)</td><td>83.26 (+3.70)</td><td>96.50 (+5.85)</td><td>84.41 (+2.05)</td><td>76.35 (+6.75)</td><td></td></tr><tr><td rowspan="2">InternVL2-76B</td><td>-</td><td>68.25</td><td>54.22</td><td>56.66</td><td>66.30</td><td>80.47</td><td>86.40</td><td>82.92</td><td>70.75</td><td></td></tr><tr><td>✓</td><td>66.68 (-1.57)</td><td>70.95 (+16.73)</td><td>60.57 (+3.91)</td><td>75.32 (+9.02)</td><td>82.71 (+2.24)</td><td>91.71 (+5.31)</td><td>85.29 (+2.36)</td><td>76.18 (+5.43)</td><td></td></tr><tr><td rowspan="2">InternVL2-76B (0-shot)</td><td>-</td><td>64.30</td><td>51.19</td><td>54.20</td><td>63.46</td><td>79.92</td><td>89.34</td><td>83.48</td><td>69.41</td><td></td></tr><tr><td>✓</td><td>63.46 (-0.84)</td><td>71.50 (+20.31)</td><td>59.11 (+4.92)</td><td>74.44 (+10.97)</td><td>82.47 (+2.55)</td><td>91.78 (+2.44)</td><td>84.83 (+1.35)</td><td>75.37 (+5.96)</td><td></td></tr></table>
+
+4, the inclusion of RAG significantly improves performance on MMAD across multiple models, with the most notable improvements in anomaly classification and object classification. This is primarily because domain knowledge contains extensive category information. The performance in defect localization also improves, indicating that models can leverage textual knowledge to enhance their perception of images. Notably, the performance in anomaly discrimination shows substantial improvement for InternVL2-40B, surpassing all other models in this metric.
+
+Model Collaboration. Given that MLLMs have not been specifically trained on IAD data, their anomaly detection capabilities are relatively weak. Detecting anomalies is not particularly challenging for existing IAD models. Therefore, we are interested in whether MLLMs can perform answering with the help of the visual outputs of expert models. As shown in Table 6, we tested three agents: using AnomalyCLIP, PatchCore, and directly using ground truth (GT) as experts, while also experimenting with various visualization methods. The two IAD models only
+
+Figure 6: Performance comparison of different agent and visualization types. Based model is InternVL2-40B.  
+
+<table><tr><td rowspan="2">Expert 
+Model</td><td rowspan="2">Visualization 
+Type</td><td>Anomaly</td><td colspan="3">Defect</td></tr><tr><td>Discrimination</td><td>Classification</td><td>Localization</td><td>Description</td></tr><tr><td>baseline</td><td>-</td><td>63.84</td><td>51.58</td><td>52.94</td><td>66.43</td></tr><tr><td>AnomalyCLIP</td><td>bbox</td><td>-5.83</td><td>-0.35</td><td>+2.66</td><td>-2.63</td></tr><tr><td>AnomalyCLIP</td><td>contour</td><td>-6.40</td><td>-1.19</td><td>+2.52</td><td>-3.86</td></tr><tr><td>AnomalyCLIP</td><td>highlight</td><td>-4.21</td><td>-6.92</td><td>+7.44</td><td>-6.61</td></tr><tr><td>PatchCore</td><td>bbox</td><td>-9.58</td><td>+0.10</td><td>+1.00</td><td>-0.51</td></tr><tr><td>PatchCore</td><td>contour</td><td>-12.62</td><td>-0.02</td><td>-0.48</td><td>-2.01</td></tr><tr><td>PatchCore</td><td>highlight</td><td>-3.87</td><td>-7.23</td><td>+5.01</td><td>-5.72</td></tr><tr><td>GT</td><td>bbox</td><td>+10.10</td><td>+7.10</td><td>+21.29</td><td>+5.40</td></tr><tr><td>GT</td><td>contour</td><td>+10.86</td><td>+6.45</td><td>+20.44</td><td>+6.16</td></tr><tr><td>GT</td><td>highlight</td><td>+9.95</td><td>+1.46</td><td>+16.59</td><td>+1.21</td></tr><tr><td>GT</td><td>mask</td><td>-1.05</td><td>+0.69</td><td>+8.34</td><td>+2.16</td></tr><tr><td>GT</td><td>text</td><td>+24.89</td><td>+1.36</td><td>+28.01</td><td>-0.22</td></tr></table>
+
+improved defect localization but resulted in declines in performance for anomaly discrimination, defect classification, and defect description. In contrast, using GT improved all performance metrics, with the enhancement in defect localization far surpassing that of the two expert models. This indicates that the outputs of the current expert models are not sufficiently accurate to be used by MLLMs. Additionally, expert models will output anomaly maps for normal images, leading to overkill in anomaly discrimination. Among the various visualization methods, a simple bounding box (bbox) to highlight the anomaly region proved to be the most effective approach overall. Directly converting the GT mask into textual descriptions of defect location and size resulted in better performance in both discrimination and localization.
+
+# 5 CONCLUSION
+
+In this work, we introduce MMAD, the first benchmark for MLLMs in the IAD field, aimed at exploring the feasibility of using MLLMs for industrial quality inspection. Our evaluation of over ten SOTA MLLMs yielded less than optimistic results, revealing their weaknesses in industrial scenarios, particularly the lack of industrial knowledge and the ability to perform fine-grained comparisons across multiple images. Models may require extensive data for targeted improvement. Moreover, our further investigations suggest that MLLMs have the potential to address some of these issues through additional enhancements. We hope that MMAD will inspire future research into improving the relevant capabilities of MLLMs and promote the development of practical applications.
+
+# REFERENCES
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.  
+Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.  
+Tajeddine Benbarrad, Marouane Salhaoui, Soukaina Bakhat Kenitar, and Mounir Arioua. Intelligent machine vision model for defective product inspection based on machine learning. Journal of Sensor and Actuator Networks, 10(1):7, 2021.  
+Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 9592-9600, 2019.  
+Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130(4):947-969, 2022.  
+Anil Bhattacharyya. On a measure of divergence between two statistical populations defined by their probability distribution. Bulletin of the Calcutta Mathematical Society, 35:99-110, 1943.  
+Yunkang Cao, Xiaohao Xu, Chen Sun, Xiaonan Huang, and Weiming Shen. Towards generic anomaly detection and understanding: Large-scale visual-linguistic model (gpt-4v) takes the lead. arXiv preprint arXiv:2311.02782, 2023.  
+Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024a. URL https://openreview.net/forum?id=evP9mxNNxJ.  
+Pengcheng Chen, Jin Ye, Guoan Wang, Yanjun Li, Zhongying Deng, Wei Li, Tianbin Li, Haodong Duan, Ziyan Huang, Yanzhou Su, et al. Gmai-mmbench: A comprehensive multimodal evaluation benchmark towards general medical ai. arXiv preprint arXiv:2408.03361, 2024b.  
+Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyuan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023.  
+Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang, Weihua Peng, Ming Liu, Bing Qin, and Ting Liu. Navigate through enigmatic labyrinth a survey of chain of thought reasoning: Advances, frontiers and future. In The 62nd Annual Meeting of the Association for Computational Linguistics: ACL 2024, Bangkok, Thailand, August 11-16, 2024. Association for Computational Linguistics, 2024. URL https://arxiv.org/abs/2309.15402.  
+Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=vvoWPYqZJA.  
+Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 9737-9746, 2022.  
+Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024.
+
+Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock. Gpts are gpts: An early look at the labor market impact potential of large language models. arXiv preprint arXiv:2303.10130, 2023.  
+Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv, abs/2306.13394, 2023. URL https://api-semanticscholar.org/CorpusID:259243928.  
+Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024.  
+Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalypt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp. 1932-1940, 2024.  
+Xi Jiang, Jianlin Liu, Jinbao Wang, Qiang Nie, Kai Wu, Yong Liu, Chengjie Wang, and Feng Zheng. Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems, 35:15433-15445, 2022a.  
+Xi Jiang, Guoyang Xie, Jinbao Wang, Yong Liu, Chengjie Wang, Feng Zheng, and Yaochu Jin. A survey of visual sensory anomaly detection. arXiv preprint arXiv:2202.07006, 2022b.  
+Xi Jiang, Ying Chen, Qiang Nie, Jianlin Liu, Yong Liu, Chengjie Wang, and Feng Zheng. Toward multi-class anomaly detection: Exploring class-aware unified model against inter-class interference. arXiv preprint arXiv:2403.14213, 2024a.  
+Yuqi Jiang, Xudong Lu, Qian Jin, Qi Sun, Hanming Wu, and Cheng Zhuo. Fabgpt: An efficient large multimodal model for complex wafer defect knowledge queries. arXiv preprint arXiv:2407.10810, 2024b.  
+Yizhang Jin, Jian Li, Yexin Liu, Tianjun Gu, Kai Wu, Zhengkai Jiang, Muyang He, Bo Zhao, Xin Tan, Zhenye Gan, et al. Efficient multimodal large language models: A survey. arXiv preprint arXiv:2405.10739, 2024.  
+Jihyung Kil, Zheda Mai, Justin Lee, Zihe Wang, Kerrie Cheng, Lemeng Wang, Ye Liu, Arpita Chowdhury, and Wei-Lun Chao. Compbench: A comparative reasoning benchmark for multimodal llms. arXiv preprint arXiv:2407.16837, 2024.  
+Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13299-13308, 2024a.  
+Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-Interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024b.  
+Jian Li and Weiheng Lu. A survey on benchmarks of multimodal large language models. arXiv preprint arXiv:2408.08632, 2024.  
+Yuanze Li, Haolin Wang, Shihao Yuan, Ming Liu, Debin Zhao, Yiwen Guo, Chen Xu, Guangming Shi, and Wangmeng Zuo. Myriad: Large multimodal model by applying vision experts for industrial anomaly detection. arXiv preprint arXiv:2310.19070, 2023.  
+Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023.  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023a.
+
+Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, January 2024a. URL https://llava-v1.github.io/blog/2024-01-30-llava-last/.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024b.  
+Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023b.  
+Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023.  
+Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024.  
+Zhonghe Ren, Fengzhou Fang, Ning Yan, and You Wu. State of the art in defect detection based on machine vision. International Journal of Precision Engineering and Manufacturing-Green Technology, 9(2):661-691, 2022.  
+Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 8317-8326, 2019.  
+Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024a.  
+Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024b.  
+Chengjie Wang, Haokun Zhu, Jinlong Peng, Yue Wang, Ran Yi, Yunsheng Wu, Lizhuang Ma, and Jiangning Zhang. M3dm-nr: Rgb-3d noisy-resistant industrial anomaly detection via multimodal denoising. arXiv preprint arXiv:2406.02263, 2024a.  
+Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 22883-22892, 2024b.  
+Chengjie Wang, Xi Jiang, Bin-Bin Gao, Zhenye Gan, Yong Liu, Feng Zheng, and Lizhuang Ma. Softpatch+: Fully unsupervised anomaly classification and segmentation. Pattern Recognition, 161:111295, 2025.  
+Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. arXiv preprint arXiv:2402.14804, 2024c.  
+Wenxuan Wang, Yihang Su, Jingyuan Huan, Jie Liu, Wenting Chen, Yudi Zhang, Cheng-Yi Li, Kao-Jung Chang, Xiaohan Xin, Linlin Shen, et al. Asclepius: A spectrum evaluation benchmark for medical multi-modal large language models. arXiv preprint arXiv:2402.11217, 2024d.  
+Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004.  
+Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023.
+
+Xiaohao Xu, Yunkang Cao, Yongqi Chen, Weiming Shen, and Xiaonan Huang. Customizing visual-language foundation models for multi-modal anomaly detection and reasoning. arXiv preprint arXiv:2403.11083, 2024.  
+Shuai Yang, Zhifei Chen, Pengguang Chen, Xi Fang, Shu Liu, and Yingcong Chen. Defect spectrum: A granular look of large-scale defect datasets with rich semantics. arXiv preprint arXiv:2310.17316, 2023.  
+Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024.  
+Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 9556-9567, 2024.  
+Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pp. 8330-8339, 2021.  
+Jian Zhang, Runwei Ding, Miaoju Ban, and Linhui Dai. Pku-goodsad: A supermarket goods dataset for unsupervised anomaly detection and segmentation. IEEE Robotics and Automation Letters, 2024a.  
+Jiangning Zhang, Xuhai Chen, Zhucun Xue, Yabiao Wang, Chengjie Wang, and Yong Liu. Exploring grounding potential of vqa-oriented gpt-4v for zero-shot anomaly detection. arXiv preprint arXiv:2311.02612, 2023a.  
+Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yuxuan Cai, Zhucun Xue, Yabiao Wang, Chengjie Wang, Lei Xie, and Yong Liu. Ader: A comprehensive benchmark for multi-class visual anomaly detection. arXiv preprint arXiv:2406.03262, 2(3), 2024b.  
+Xiaoman Zhang, Chaoyi Wu, Ziheng Zhao, Weixiong Lin, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc-vqa: Visual instruction tuning for medical visual question answering. arXiv preprint arXiv:2305.10415, 2023b.  
+Yiheng Zhang, Yunkang Cao, Xiaohao Xu, and Weiming Shen. Logic: an llm-driven framework for logical anomaly detection. arXiv preprint arXiv:2406.04687, 2024c.  
+Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu, Ling Yang, Wentao Zhang, and Bin Cui. Retrieval-augmented generation for ai-generated content: A survey. arXiv preprint arXiv:2402.19473, 2024.  
+Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961, 2023.  
+Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17826-17836, 2024.  
+Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. In European Conference on Computer Vision, pp. 392-408. Springer, 2022.
+
+# A APPENDIX
+
+# A.1 DISCUSSION OF SETTINGS
+
+![](images/1b2dfee6a013a4f01c8d882a55cfa66ffbd9260730d0294b0af9d7899add6f07.jpg)  
+Figure 7: Diagram of different settings. The query image shows an abnormal screw with a manipulated front, which is difficult to identify without a compared sample.
+
+![](images/9b8acdbcedb15f816e656516e0265200b84ecd4e2ecbb1011e40c6dd0e485002.jpg)
+
+![](images/97c1e2355a196338e29fe503438884abefeb007b4ebe4850d6ef02c1fc75b0cc.jpg)
+
+![](images/ce979f61624f30fb73109df0be6efb5094046fb9e245922055303b5b23dd7905.jpg)
+
+![](images/55848f0ecf6a02738d98ccbaa633dba960306e1d79a0ae7e886602c3efce956e.jpg)
+
+Unlike other benchmarks, MMAD features various testing settings to simulate different scenarios in industrial quality inspection for MLLMs. We use 1-shot as the default setting, as a single normal image theoretically provides significant information compared to 0-shot. This approach is also used in our data annotation process. Increasing the number of normal images in few-shot settings may pose challenges for some MLLMs, particularly those not fine-tuned for multi-image tasks. In industrial production lines, the number of available normal samples often exceeds the context length of MLLMs. In such cases, similarity retrieval can filter out irrelevant images, leading us to propose the  $1\text{-shot}^{+}$  setting, where the most informative image is selected as a template. Additionally, abnormal images can sometimes be collected in production lines. To determine whether these abnormal images aid MLLMs in making judgments, we introduce abnormal image prompts in the 1-shot setting and then query the test image, termed in-context learning, as illustrated in Figure 7.
+
+# A.2 ANALYSIS WITH IN-CONTEXT LEARNING
+
+As shown in Table 5, we tested the performance under the in-context learning setting and compared it with other settings. We chose InternVL2-40B as the base model due to its superior context-handling abilities demonstrated in previous tests. The experimental results indicate that adding abnormal images in the in-context learning setting indeed helps MLLMs in anomaly detection, improving the performance from  $63.84\%$  to  $66.47\%$ . However, adding one more normal image does not significantly change this performance. In subsequent sub-tasks related to defect identification, in-context learning did not seem to provide any benefit. We believe this may be due to MLLMs not fully leveraging the information from abnormal images. Marking defects in abnormal images and combining them with textual prompts might better assist MLLMs. The  $1\text{-shot}^{+}$  setting, on the other hand, showed comprehensive improvements across sub-tasks, indicating that the quality of prompts is far more important than their quantity.
+
+# A.3 ANALYSIS OF CHAIN OF THOUGHT
+
+Chain of Thought (CoT) is a commonly used method to enhance the logical reasoning abilities of MLLMs (Chu et al., 2024). To investigate whether current MLLMs lack reasoning when addressing MMAD problems, we introduced a straightforward CoT approach. The process involves three
+
+Table 5: Performance comparison of different settings in MMAD based on InternVL2-40B.  
+
+<table><tr><td rowspan="2">Setting</td><td>Anomaly</td><td colspan="4">Defect</td><td colspan="2">Object</td><td rowspan="2">Average</td></tr><tr><td>Discrimination</td><td>Classification</td><td>Localization</td><td>Description</td><td>Analysis</td><td>Classification</td><td>Analysis</td></tr><tr><td>1-shot</td><td>63.84</td><td>51.58</td><td>52.94</td><td>66.43</td><td>79.75</td><td>90.76</td><td>82.41</td><td>69.67</td></tr><tr><td>1-shot+</td><td>67.81</td><td>54.18</td><td>55.99</td><td>68.14</td><td>81.75</td><td>90.76</td><td>82.98</td><td>71.66</td></tr><tr><td>2-shot</td><td>63.75</td><td>50.38</td><td>54.36</td><td>67.22</td><td>79.54</td><td>90.49</td><td>81.85</td><td>69.66</td></tr><tr><td>In-Context Learning</td><td>66.47</td><td>50.56</td><td>51.78</td><td>64.87</td><td>78.55</td><td>91.33</td><td>81.65</td><td>69.32</td></tr></table>
+
+Table 6: Performance comparison of different MLLMs with and without CoT.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">CoT</td><td colspan="2">Anomaly</td><td colspan="4">Defect</td><td colspan="2">Object</td><td>Average</td></tr><tr><td>Discrimination</td><td>Classification</td><td>Localization</td><td>Description</td><td>Analysis</td><td>Classification</td><td>Analysis</td><td></td><td></td></tr><tr><td rowspan="2">InternVL2-40B</td><td>-</td><td>64.45</td><td>50.57</td><td>53.42</td><td>66.17</td><td>79.56</td><td>90.65</td><td>82.36</td><td>69.59</td><td></td></tr><tr><td>✓</td><td>59.42</td><td>46.10</td><td>51.92</td><td>57.66</td><td>79.80</td><td>80.55</td><td>85.29</td><td>65.82</td><td></td></tr><tr><td rowspan="2">InternVL2-76B</td><td>-</td><td>68.25</td><td>54.22</td><td>56.66</td><td>66.30</td><td>80.47</td><td>86.40</td><td>82.92</td><td>70.75</td><td></td></tr><tr><td>✓</td><td>68.18</td><td>54.61</td><td>58.64</td><td>68.89</td><td>79.95</td><td>90.51</td><td>85.25</td><td>72.29</td><td></td></tr></table>
+
+steps: first, the model identifies objects in the image; second, it compares the differences between the template image and the query image; and finally, it determines whether the identified difference constitutes a defect in the object. We incorporated a set of rules and adjusted the instructions accordingly, dividing the CoT responses into two stages. As shown in the Table 6, InternVL2-76B, the best-performing open-source MLLM we tested, achieved a  $1.5\%$  improvement in CoT-based performance. However, its anomaly detection accuracy, which is the most critical metric, showed no improvement. On the other hand, InternVL2-40B experienced a performance decline after introducing CoT, potentially due to insufficient stability in the language model's reasoning capabilities.
+
+# A.4 ANALYSIS OF VISION DISABLE
+
+Some studies (Tong et al., 2024a; Chen et al., 2024a) have proposed that determining whether a benchmark requires visual input to be solved has been a persistent challenge in vision-language research. To validate MMAD, we masked the visual components and compared the performance of MLLMs with and without visual input. The results, as shown in the Table 7, indicate that most subtasks in MMAD are highly dependent on the visual components. Performance significantly decreases when the visual input is removed.
+
+# A.5 DETAILS OF HUMAN SUPERVISION
+
+Our textual data is generated by MLLM, so its accuracy requires human supervision. In our designed process, human supervision is responsible for filtering the final model-generated multiple-choice questions. We first perform preliminary filtering through the program and then enable annotators to conduct an item-by-item review. A total of 26 personnel were involved in the review process, all of whom are researchers in industrial inspection or computer vision and possess a certain level of expertise. However, it should be noted that the industrial inspection field is highly specialized, with significant differences between products. Our annotators' understanding may differ from the professionals' understanding of where the original image data originates. We have developed a tool to enable annotators to filter out problematic multiple-choice questions quickly. As illustrated in Figure 8, we provide the original annotation information through visual and textual means to help annotators accurately identify objects and defects. At the same time, annotators do not need to correct every issue; they simply mark the erroneous questions for exclusion, significantly improving efficiency.
+
+# A.6 DIVERSITY ANALYSIS OF DATASET
+
+To systematically analyze the diversity of the dataset, we separately calculated the frequency of phrases appearing in questions and options, presenting them in the form of word clouds in Figure 9. In the question text, "defect" appeared most frequently, followed by "object", reflecting that our benchmark is constructed for industrial inspection scenarios, focusing on seven sub-tasks. In addition to common words such as "image", "type", and "appearance", the questions also include
+
+Table 7: Performance comparison of different MLLMs with and without vision.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Vision</td><td>Anomaly</td><td colspan="4">Defect</td><td colspan="2">Object</td><td rowspan="2">Average</td></tr><tr><td>Discrimination</td><td>Classification</td><td>Localization</td><td>Description</td><td>Analysis</td><td>Classification</td><td>Analysis</td></tr><tr><td rowspan="2">InternVL2-8B</td><td>Enable</td><td>59.97</td><td>43.85</td><td>47.91</td><td>57.60</td><td>78.10</td><td>74.18</td><td>80.37</td><td>63.14</td></tr><tr><td>Disable</td><td>49.71</td><td>32.28</td><td>41.38</td><td>52.20</td><td>75.34</td><td>41.94</td><td>57.13</td><td>50.00</td></tr><tr><td rowspan="2">InternVL2-40B</td><td>Enable</td><td>63.84</td><td>51.58</td><td>52.94</td><td>66.43</td><td>79.75</td><td>90.76</td><td>82.41</td><td>69.67</td></tr><tr><td>Disable</td><td>49.61</td><td>24.43</td><td>36.35</td><td>46.39</td><td>68.05</td><td>37.29</td><td>55.75</td><td>45.41</td></tr><tr><td rowspan="2">InternVL2-76B</td><td>Enable</td><td>68.25</td><td>54.22</td><td>56.66</td><td>66.30</td><td>80.47</td><td>86.40</td><td>82.92</td><td>70.75</td></tr><tr><td>Disable</td><td>49.93</td><td>25.92</td><td>28.54</td><td>43.88</td><td>70.41</td><td>35.89</td><td>57.53</td><td>44.59</td></tr></table>
+
+![](images/eae4ff3f1146a4a9fe97873f5b387f9c8a67f19e1bdc5e007740c230998dbd48.jpg)  
+Figure 8: Illustration of human filtering tool. The tool comprises several functional areas. We use green bounding boxes to highlight defects as vision prompts, and we provide rough text prompts for object categories and defect types.
+
+diverse expressions with lower frequencies. For example, expressions indicating position such as "relative position", "arrangement", and "defect located". In the text of the options, "Yes" and "No" are the standard answers for anomaly discrimination questions, thus appearing most frequently.
+
+![](images/27d9915260718d7d9b002560d49066cea2cd066ffe4f38184ed2eaa954a19231.jpg)  
+(a) Word Cloud of All Questions
+
+![](images/6c5cb78b8ed9d4432444ccc6e1305cf4047d85b4a42912eb1594f4d834a30ed2.jpg)  
+(b) Word Cloud of All Options  
+Figure 9: Word frequency statistics presented in the form of a word cloud, with separate statistics for questions and options.
+
+Beyond these two words, the diversity of options is more evident, including specific descriptions of various positions or expressions of different types of defects.
+
+Table 8: Performance comparison of different MLLMs in Anomaly Discrimination/Detection Tasks.  
+
+<table><tr><td>Model</td><td>Scale</td><td>Accuracy</td><td>Recall</td><td>Precision</td><td>F1</td></tr><tr><td>Human (expert)</td><td>-</td><td>95.24</td><td>94.25</td><td>98.89</td><td>96.43</td></tr><tr><td>Human (ordinary)</td><td>-</td><td>86.90</td><td>87.07</td><td>94.35</td><td>89.30</td></tr><tr><td>claude-3.5-sonnet</td><td>-</td><td>60.14</td><td>30.87</td><td>76.75</td><td>41.92</td></tr><tr><td>Gemini-1.5-flash</td><td>-</td><td>58.58</td><td>78.63</td><td>67.41</td><td>72.40</td></tr><tr><td>Gemini-1.5-pro</td><td>-</td><td>68.63</td><td>45.47</td><td>86.84</td><td>57.60</td></tr><tr><td>GPT-4o-mini</td><td>-</td><td>64.33</td><td>65.47</td><td>73.04</td><td>68.67</td></tr><tr><td>GPT-4o</td><td>-</td><td>68.63</td><td>67.37</td><td>75.68</td><td>71.04</td></tr><tr><td>AnomalyGPT</td><td>7B</td><td>65.57</td><td>82.11</td><td>74.45</td><td>76.68</td></tr><tr><td>Qwen-VL-Chat</td><td>7B</td><td>53.65</td><td>43.95</td><td>65.39</td><td>47.28</td></tr><tr><td>LLaVA-1.5</td><td>7B</td><td>51.33</td><td>94.79</td><td>62.72</td><td>75.32</td></tr><tr><td>Cambrian-1*</td><td>8B</td><td>55.60</td><td>22.28</td><td>74.10</td><td>31.85</td></tr><tr><td>SPHINX*</td><td>7B</td><td>53.13</td><td>6.42</td><td>99.74</td><td>10.61</td></tr><tr><td>LLaVA-NEXT-Interleave</td><td>7B</td><td>57.64</td><td>16.58</td><td>90.83</td><td>25.64</td></tr><tr><td>InternLM-XComposer2-VL</td><td>7B</td><td>55.85</td><td>17.94</td><td>75.87</td><td>27.16</td></tr><tr><td>LLaVA-OnVision</td><td>7B</td><td>51.77</td><td>4.90</td><td>78.19</td><td>9.10</td></tr><tr><td>MiniCPM-V2.6</td><td>8B</td><td>57.31</td><td>34.38</td><td>70.98</td><td>45.31</td></tr><tr><td>InternVL2</td><td>8B</td><td>59.97</td><td>30.25</td><td>79.22</td><td>41.23</td></tr><tr><td>LLaVA-1.5</td><td>13B</td><td>49.96</td><td>99.79</td><td>62.00</td><td>76.28</td></tr><tr><td>LLaVA-NeXT</td><td>34B</td><td>57.92</td><td>46.27</td><td>69.98</td><td>54.44</td></tr><tr><td>InternVL2</td><td>76B</td><td>68.25</td><td>55.81</td><td>83.52</td><td>64.40</td></tr></table>
+
+![](images/8c40045f3ab185b206576ec377330bbac232bc312c76024d9b67ebae32ee861c.jpg)  
+Figure 10: Bubble chart with recall and precision as axes in anomaly discrimination/detection task. The size of the bubbles represents the number of parameters in the model. For commercial models or humans where the number of parameters is unknown, the bubbles are set to the maximum size.
+
+# A.7 COMPREHENSIVE METRICS FOR ANOMALY DISCRIMINATION/DETECTION TASKS
+
+To comprehensively evaluate the performance of various models in the anomaly discrimination/detection task, we follow the traditional anomaly detection setup, treating the anomaly class as the positive class and the normal class as the negative class. We measured recall, precision, and F1-score. As shown in Table 8, by analyzing the recall and precision metrics, we can identify some reasons for poor accuracy performance in certain models. For instance, both SPHINX and LLaVA-OnVision have recall rates below  $10\%$ , indicating that these models frequently misclassify anomaly samples as normal, leading to a high rate of missed detections. On the other hand, LLaVA-1.5 has a high recall but low precision, suggesting a high rate of false positives. Humans, however, outperform MLLMs across all metrics, with human experts achieving over  $94\%$  and ordinary individuals
+
+achieving over  $87\%$ . This disparity is visualized in the bubble chart, as shown in Figure 10, where there is a significant gap between the bubbles representing humans and those representing various MLLMs, and considerable differences among the MLLMs themselves. Additionally, it can be observed that the model AnomalyGPT, which is specifically trained for anomaly detection, performs better than most models but still suffers from a significant false positive issue.
+
+# B CASE STUDY
+
+# Hit:
+
+You are an industrial inspector who checks products by images. You should judge whether there is a defect in the query image and answer the questions about it.
+
+Answer with the option's letter from the given choices directly!
+
+The first image is the normal sample, which can be used as a template to compare.
+
+The last image is the query image.
+
+![](images/95032eda24db93679a5dbc5a24c121256342807a9ecfa1b2d16984082a461a3b.jpg)
+
+![](images/ba99ec72fd4642f32b40f5f7481861f425dc3183310ef7be2dbcc1a321d7b083.jpg)
+
+# Question:
+
+Should the query picture be classified as normal or abnormal?
+
+A. Abnormal
+
+B. Normal
+
+Give a short analysis and then answer a letter.  $\leftarrow$  only included in qualitative analysis
+
+# InternVL-40B:
+
+The query picture shows three objects with copper-colored spheres on top, which are not typical for a normal product. They appear to be some kind of artificial or experimental items rather than a standard product. Therefore, the query picture should be classified as abnormal.
+
+Answer: A. Abnormal.
+
+# GPT-40:
+
+Analysis: Comparing the query image to the normal template, the green section in the query image shows a visible crack or separation in its structure, which is not present in the normal image. This defect indicates that the product does not meet the standard criteria.
+
+Answer: A
+
+Figure 11: A case of a question-and-answer result in qualitative analysis. Since MMAD is testing with multiple-choice questions, we cannot directly analyze the answers. Instead, we use the chain-of-thought approach to encourage the model to provide its own analysis. In this example, the InternVL-40B model did not notice the damage to the cable, yet it still chose the correct answer from the options. In comparison, GPT-4o can output a proper analysis and letter. Subsequent issues such as anomaly classification, localization, and description can further test whether the model's correct answers are coincidental. Therefore, a comprehensive evaluation, rather than simple anomaly detection, is essential.
+
+# Hit:
+
+You are an industrial inspector who checks products by images. You should judge whether there is a defect in the query image and answer the questions about it.
+
+Answer with the option's letter from the given choices directly!
+
+The first image is the normal sample, which can be used as a template to compare.
+
+The last image is the query image.
+
+![](images/3ed5a4b025c0cd3efed9f9a90c1464f88e8ca64ad9b8c3dcc0cb3a55d5af9df1.jpg)  
+Figure 12: A case of a question-and-answer result in qualitative analysis. A human can quickly notice the defect in the query image, while the model focuses on the number of components. This illusion tends to occur when the defect is minor, and the object's composition is complex.
+
+![](images/fd0df75e122a5057235cbea10b0387c624bcd4a8d0b379cedd012fe3c21687c4.jpg)
+
+# Question:
+
+There is a defect in the object. What is the type of the defect?
+
+A: Incorrect labeling  
+B: Structural deformation  
+C: Missing component  
+D: Discoloration
+
+Give a short analysis and then answer a letter.  $\leftarrow$  only included in qualitative analysis
+
+# GPT-40:
+
+Analysis: Comparing the query image to the normal sample, the query image is missing one washer. The normal sample has two bolts, two nuts, and two washers, whereas the query image has only one washer.
+
+Answer: C: Missing component
+
+Human:
+
+Analysis:
+
+It is easy to see that the end of the long screw in the inspected image has been stained green.
+
+Answer: D: Discoloration
+
+# Hit:
+
+You are an industrial inspector who checks products by images. You should judge whether there is a defect in the query image and answer the questions about it. Answer with the option's letter from the given choices directly!
+
+![](images/7db8aee823001b4b0320d58d590f0aae64b387534e40e97ec7f8993ce53b074b.jpg)
+
+# Question:
+
+What is the purpose or function of this electronic circuit board?
+
+A: Audio processing unit  
+B: Power supply for electronic devices  
+C: Lighting control system  
+D: Wireless communication module
+
+Give a short analysis and then answer a letter.  $\leftarrow$  only included in qualitative analysis
+
+# GPT-4o:
+
+Analysis: This is a TP4056 module, commonly used as a lithium battery charger circuit. It features a micro-USB input  $(IN+)$  and battery connection points  $(BAT+$  and  $BAT-$ . The presence of components like the TC4056A IC and the layout indicates its use in charging and regulating power supply for lithium-ion batteries.
+
+Answer: B
+
+# Human:
+
+I don't know. It looks like all the options are possible.
+
+![](images/c2ed37901382a3625132ceb1cf1e0a11e8b3c885bbf34be01994beeb69a9a1bb.jpg)
+
+![](images/6065f6a3c62852e1f61133ecbf0c0e7ccd3f2c204431a1d5766bbf665c904255.jpg)  
+Figure 13: A case of a question-and-answer result in qualitative analysis. For object-related questions, GPT-4o possesses extensive knowledge and can analyze object information based on subtle clues. However, most humans do not have such comprehensive knowledge. For instance, nonspecialists may not recognize BAT+ and BAT-, let alone the TC4056A IC chip.
+
+![](images/255ee1e3a7e9ccb758869d80f945a558c0d53bf5a5a0021f34174da34297fa1b.jpg)  
+Image Information:  
+Template Image
+
+![](images/0ff9aee820310972468358048bbbb6254bf3eb8427c6d2723f406dc367d1f221.jpg)  
+Captioned Image
+
+![](images/d9298a48ea61a06dbb760d7adf5e753c91f1d5537e07e1e07a51b25bf967e0d9.jpg)  
+RGB mask  
+Prior Knowledge
+
+object category:
+
+"cable";
+
+defect category:
+
+"ab combination of
+
+multiple defects";
+
+defect localization:
+
+"center of the image
+
+# Caption:
+
+This is a close-up image of a cross-section of a cable. The cable consists of three separate conductive wires, each insulated with a different colored sheath: green with a yellow stripe, blue, and brown. The wires are encased in a gray outer sheath. Based on a comparison with the template normal image provided:
+
+1. There appear to be multiple defects present in the object.  
+2. The green and yellow wire exhibits a defect in the form of bent strands. These are visible at the approximate top center of the image, where the copper strands stick out at an irregular angle compared to the undisturbed arrangement seen in the normal image.  
+3. Another anomaly is present in the brown wire, where there seems to be damage to the insulation. This defect is located at the bottom center of the image, appearing as a deviation from the uniformly round shape of the corresponding wire in the normal sample.
+
+These anomalies could potentially affect the performance and safety of the cable.
+
+![](images/e0d08b4ed6ddfce7489671f7fdc4eeb09d06e415bd39194a72a58442f29db86d.jpg)  
+Image Information:  
+Template Image  
+Figure 15: A bad case of image caption. Although we informed the AI assistant that a screw is too long, it failed to analyze that one screw should be long and the other short. This error is primarily due to the insufficient number of template images, and such misunderstandings will be corrected during manual filtering.
+
+![](images/8a99be87bc2683034acc8a65216f75804da3c8346c082aa4f8ce008dd4d02c65.jpg)  
+Captioned Image
+
+![](images/ada705dcafc0ea9df3aaf510943a5c7a72a05f3f410cc34a6fe56112f44ccb16.jpg)  
+Figure 14: A good case of image caption. We prompt the AI assistant that the image contains multiple defects and provide corresponding annotations. The AI assistant distinguishes different defects in the caption, describing their appearance and location individually.  
+RGB mask
+
+object category:
+
+"screw bag";
+
+defect category:
+
+"screw too long";
+
+defect localization:
+
+"center of the image
+
+Prior Knowledge
+
+# Caption:
+
+The query image depicts a sealed plastic bag containing hardware items. Inside the bag, there are two bolts, four nuts, and two washers. The bolts appear metallic with threads for fastening, and the nuts and washers accompany them as components for securing or distributing the load on the material being fastened.
+
+Upon comparing the query image with the template normal image, there is a noticeable anomaly related to the length of the bolts. In the query image, one bolt, positioned in the bottom part of the bag, appears longer than the corresponding bolt in the normal image. Specifically, the defect is observed in the central bolt, which extends from the bottom left to the bottom center region of the bag. This anomaly disrupts the uniformity expected from a standard hardware set, where bolts should be of matching lengths for consistent application in construction or assembly tasks. The appearance of the anomaly is characterized by a visibly extended thread area beyond what is observed in the normal comparison image.
+
+![](images/1ec7df17ae14a304366cbbc81ef233431e6c79c9dbd8c2feedbd04180446d77a.jpg)  
+Sample Images:
+
+![](images/e80df71f0ad3e67fe1c71d29129c5ec13412429c8b958d566403d46529692738.jpg)
+
+![](images/2f3167bc06a3c481bb08a1f2f9901b0f274f8b69dc336f77b3e44497d5a07699.jpg)
+
+# Domain Knowledge:
+
+<Normal Characteristics>
+
+Description: The ultrasonic sensor module, typically characterized by its rectangular blue printed circuit board (PCB), features two round ultrasonic transducers positioned symmetrically, which function as the transmitter and receiver. The module includes a central electronic component likely responsible for signal processing, surrounded by clear white labeling indicating each pin's function, such as VCC (power supply), Trig (trigger), Echo (echo signal), and GND (ground). Upon inspection, there are no visible defects or anomalies present; all components, including solder points, appear well-formed and intact. The noise features, which may include slight variations in texture or minor color differences, are typical in manufactured electronic components and do not indicate defects. The overall condition conveys a standard representation of a fully functional ultrasonic sensor in good working order, free from any damage, discoloration, or misalignment.
+
+![](images/d61023f172eff1755a2c909b511343093a927efcdff4db26834254d373e5d5ea.jpg)  
+Figure 16: A case of domain knowledge. Domain knowledge includes descriptions of the characteristics of normal samples. In this example, the text describes the components and appearance of this category of PCBs, as well as some common types of noise that may be present.  
+Sample Images:  
+Figure 17: A case of domain knowledge. Domain knowledge includes descriptions of defects. In this example, the text outlines four major logical anomalies that may occur in these juice bottles. The actual dataset contains some rare logical anomalies that are not described, such as the incorrect icon placement in the first image. We do not expect domain knowledge to cover all possible anomalies; its primary role is to provide MLLMs with an understanding of the types of anomalies.
+
+![](images/0928ad0979226a88e6130482408423debebbabc7ffe1dbc18ae31046a6dcd4a1.jpg)
+
+![](images/595f97c685bfd19b8220378ace1834a604a7e0a54c2f653180a512e13eb162c8.jpg)
+
+![](images/8bee2ba1b365c82f59e25819d4609a731d8ce9c688afbef790c902e205133463.jpg)
+
+# Domain Knowledge:
+
+<Logical Anomalies>
+
+Description: The observed defects primarily exhibit inconsistencies related to product labeling, fill levels, or the representation of contents within the juice bottles. Common visual features for distinguishing these logical anomalies include:
+
+- Label Misalignment: Many images showed labels where graphical elements (such as fruit icons) were misplaced or not centered, deviating from their expected positions. This often causes confusion regarding branding and product identity.
+
+- Missing Labels: Several instances indicated the absence of labels essential for branding, ingredient information, or flavor identification. Missing labels create visual gaps on the product, which can mislead consumers about the product's contents.
+
+- Fill Level Deviations: Anomalies relating to insufficient or excessive juice levels compared to standard expectations were found. These incidents presented notable visual differences where expected liquid content was overflown, underfilled, or completely absent.
+
+- Label Content Mismatch: Some products contained labels that represented a different fruit than the actual juice color, indicating a logical misrepresentation of the product's identity. This created a strong disconnect between the label imagery and the liquid contents inside the bottle.
+
+All these features are significant in determining the presence and type of logical anomalies across the evaluated juice bottle batch.

ICLR/2025/MMAD_ A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9c43fc7f6dc1523f281a107aa00f76c6bb52d05206d1ccbedec247937cc7bf3
+size 1415728

ICLR/2025/MMAD_ A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc423178b82ffd42cfd21ccf48f496df9f7760ebbbb793e625ababae8a1c6e65
+size 590147

ICLR/2025/MMDT_ Decoding the Trustworthiness and Safety of Multimodal Foundation Models/c53b7d81-3431-49b6-8e43-b9ab1280f1f7_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dadc11b481661465939aa9b4966e7b4bde00c141e0a640cef64247ee3302774a
+size 565853

ICLR/2025/MMDT_ Decoding the Trustworthiness and Safety of Multimodal Foundation Models/c53b7d81-3431-49b6-8e43-b9ab1280f1f7_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29d6672c5409179938a43eadd0f6a6a4cb49d049b98a9c4aa0ccdd7098aa9333
+size 674190

ICLR/2025/MMDT_ Decoding the Trustworthiness and Safety of Multimodal Foundation Models/c53b7d81-3431-49b6-8e43-b9ab1280f1f7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c039b1965f0b418080da1ac74553993f51f01fb4c644ea397a82123da0d1cc15
+size 19111230

ICLR/2025/MMDT_ Decoding the Trustworthiness and Safety of Multimodal Foundation Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd6490579ba476ae0690648db8f8ce98f75cc438edb3ce003d99edf1ad981ca6
+size 6918724

ICLR/2025/MMDT_ Decoding the Trustworthiness and Safety of Multimodal Foundation Models/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd040253cb1adb3d6a148dc11b94198095811b5affe5944ba5597dc8a6f9bb10
+size 2205605

ICLR/2025/MMDisCo_ Multi-Modal Discriminator-Guided Cooperative Diffusion for Joint Audio and Video Generation/5ab7ff06-31f2-4d0d-8db8-8319d784a5bb_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e96ede00e2a5429665a5f1e259c87efebbeca731411d04f393ffc8e1edc36c1
+size 151518

ICLR/2025/MMDisCo_ Multi-Modal Discriminator-Guided Cooperative Diffusion for Joint Audio and Video Generation/5ab7ff06-31f2-4d0d-8db8-8319d784a5bb_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:588b14fa88f84ebc00af8ed658d834af09f59be75a61c6d363e807ef8be6b421
+size 180515