# multiobject_hallucination_in_vision_language_models__3feae5e1.pdf Multi-Object Hallucination in Vision Language Models Xuweiyi Chen1,2 Ziqiao Ma1 Xuejun Zhang1 Sihan Xu1 Shengyi Qian1 Jianing Yang1 David F. Fouhey3 Joyce Chai1 1University of Michigan 2University of Virginia 3New York University https://multi-object-hallucination.github.io/ Large vision language models (LVLMs) often suffer from object hallucination, producing objects not present in the given images. While current benchmarks for object hallucination primarily concentrate on the presence of a single object class rather than individual entities, this work systematically investigates multi-object hallucination, examining how models misperceive (e.g., invent nonexistent objects or become distracted) when tasked with focusing on multiple objects simultaneously. We introduce Recognition-based Object Probing Evaluation (ROPE), an automated evaluation protocol that considers the distribution of object classes within a single image during testing and uses visual referring prompts to eliminate ambiguity. With comprehensive empirical studies and analysis of potential factors leading to multi-object hallucination, we found that (1) LVLMs suffer more hallucinations when focusing on multiple objects compared to a single object. (2) The tested object class distribution affects hallucination behaviors, indicating that LVLMs may follow shortcuts and spurious correlations. (3) Hallucinatory behaviors are influenced by data-specific factors, salience and frequency, and model intrinsic behaviors. We hope to enable LVLMs to recognize and reason about multiple objects that often occur in realistic visual scenes, provide insights, and quantify our progress towards mitigating the issues. 1 Introduction Recent advances in large language models (LLMs) have motivated increasing efforts in adapting them for understanding visual semantics, giving rise to a surge of large vision language models (LVLMs) [1, 37, 43]. These models, whether explicitly trained with grounding data [71] or without [31], demonstrate an impressive grounded understanding of visual entities. This motivates a new prompting paradigm based on user-provided visual cues, referred to as visual prompting [45, 60, 27, 58, 58]. Despite their promising performances on various downstream applications [36], LVLMs often suffer from object hallucination [44, 10, 26], where they produce objects not present in a given image. Although object hallucination was initially observed in image captioning describing multiple objects [44], current benchmarks for object hallucination primarily concentrate on the presence of a single object class rather than individual entities. These benchmarks either verify if an object class mentioned in the caption can ground to an object in the image [44, 19], or probe the model about the existence of an object class, sometimes with additional attributes or relations to other objects [26, 32]. There are, however, two key limitations with these setups as shown by a case study in Figure 1. First, grounding is not simply one-to-one between objects and classes, but a many-to-many mapping between objects and phrases [20, 34]. For instance, apples could potentially correspond to multiple referents in Figure 1, and the model doesn t necessarily need to recognize all of them to provide such a response. Therefore, being able to produce an object that exists in an image does not necessarily Authors contributed equally to this work, alphabetized by last names. 38th Conference on Neural Information Processing Systems (Neur IPS 2024). Is there a whisk in this image? Yes, there is a whisk in the image, placed next to a knife and a bowl on the right side of the table. Is there a whisk next to a knife? Yes, there is a whisk next to a knife on the table in the image you provided. ... There are also multiple pieces of fresh fruit, such as apples, pears, and a lemon. Additionally, there are three jars of jam and a packet of M&M's Skittles on the table. To the side, there's a bottle of water, and utensils including a whisk, a knife, and some spoons placed inside a blue mug Does the caption accurately describe the image: "A whisk is placed to the right of a knife." No, the caption does not accurately describe the image. The whisk is actually placed to the left of the knife on the table. (b) Captioning-based evaluation. (c) Polling-based object probing. (d) Object attribute/relation probing. (e) Counterfactual probing. Select one and the most appropriate class for each object located within red bounding boxes from the following list: apple, orange, banana, lemon, pear, plate, bowl, jar, bottle, soda can, knife, fork, whisk, scissors, packaged snack. Provide the class names in the format: 'obj1: , obj2: , obj3: , obj4: , obj5: ', with no additional words or punctuations. obj1: apple, obj2: knife, obj3: fork, obj4: apple, obj5: jar GPT-4V (a) Recognition-based object probing. Provide a detailed description of the given image. Figure 1: A case study that compares our Recognition-based Object Probing Evaluation (ROPE) benchmark with existing benchmarks for object hallucination in GPT-4V. ROPE offers an automated evaluation protocol with controlled output formatting and uses visual prompts to distinctly ground to objects, thus mitigating referential ambiguity. Unlike binary inquiries relying solely on textual descriptions, ROPE challenges the model to identify multiple objects concurrently. We observe that, while GPT-4V can identify the whisk to the left of a knife when prompted about it, the model hallucinates a fork when directly tasked to recognize multiple objects. indicate that the model is free of hallucinations. Second, explicitly instructing the model to recognize multiple objects poses greater challenges compared to simple yes/no inquiries that contain explicit text descriptions for individual objects. For instance, while the model can correctly identify that a whisk is positioned to the left of a knife when a whisk is deliberately prompted, as shown in Figure 1(b-d), it may hallucinate a fork when directly prompted to recognize both the whisk and the knife (i.e., Figure 1a). This could be due to the common association between knives and forks, which leads to potential hallucinations when models are tasked to focus on multiple objects at the same time. In real-world applications, multi-object querying is crucial for embodied AI tasks. For example, in a cooking scenario, an agent must recognize multiple ingredients and tools simultaneously to be effective. We also present a case study on autonomous driving (see Figure 8 in the appendix), demonstrating how common associations between cars, pedestrians and traffic lights could lead to potential hallucinations. In addition, evaluating multiple objects simultaneously, rather than querying each object individually, can significantly save both time and resources. To enable LVLMs to recognize and reason about multiple objects that often occur in realistic visual scenes and to better quantify the complex phenomena we observed, this paper investigates multi-object hallucination, examining how models may misperceive (e.g., by inventing nonexistent objects or becoming distracted) when tasked to focus on multiple objects concurrently, and which factors cause the hallucinations. We start by introducing Recognition-based Object Probing Evaluation (ROPE) for assessing multiobject hallucination with formatted output control. ROPE features an automated evaluation protocol without black-box neural models or humans as evaluators, and leverages visual prompts to uniquely refer to objects to avoid ambiguity and multiple referents caused by object class names. ROPE considers the distribution of object classes within each image at test time, dividing ROPE into 4 subsets: In-the-Wild, Homogeneous, Heterogeneous, and Adversarial. For instance, we investigate scenarios where all tested objects belong to the same class or where each tested object represents a different class. We conduct an in-depth analysis of the hallucination behaviors of LVLMs of different scales and training data (e.g., whether grounding data and conversational data are used), and provide a comprehensive analysis of potential factors that lead to multi-object hallucination. Our main findings are: (1) LVLMs suffer from more hallucinations when tasked to focus on multiple objects, compared to focusing on a single object; (2) The tested object class distribution affects the hallucination behaviors, revealing that LVLMs may be following shortcuts and spurious correlations; (3) The hallucinatory behaviors of LVLMs are affected by data-specific factors, salience and frequency, and model intrinsic behaviors. These findings provide key insights for the development and application of LVLMs, suggesting for more balanced object distributions, diverse annotations, and enhanced multi-object instructions in grounded LVLMs. We hope this work takes a step towards LVLMs that recognize and reason about multiple objects that often occur in realistic visual scenes. 2 Related Work Large Vision-Language Models. There is a growing trend to harness and adapt the powerful large language models (LLMs) for multimodal understanding beyond text [51, 1, 65, 41]. Especially, visual instruction tuning has gained prominence for its competitive performance with a comparatively moderate amount of data and computational resources, leading to a variety of Large Vision-Language Models (LVLMs) [31, 30, 9, 76, 12, 56, 62, 22]. Grounding datasets have been shown to benefit vision-language pre-training [33, 25, 34]. Researchers have developed a family of grounded LVLMs focusing on object grounding to bounding box [40, 8, 72, 3, 63, 70, 39] and segmentation masks [21, 69, 57, 42, 71]. Of the large space of LVLMs, our work is most related to visual prompting [58, 60] and object hallucination [44, 10]. The paragraphs below describe the two lines of work in detail. Visual Prompting. LVLMs demonstrate their grounded understanding of user-provided visual cues, giving rise to a practical and user-friendly prompting paradigm known as visual prompting [58, 60]. Early work on visual prompting in vision-language models can date back to tuning-based methods [2, 61]. Recent studies show that LVLMs demonstrate zero-shot understanding of user-provided visual cues (e.g., a red circle) [45, 60]. This observation allows prompting LVLMs by editing images directly in the pixel space, e.g., by adding visual marks or visual text [27]. Starting from Set-of-Marks (So M) prompting [58], several training-free methods have been introduced [24, 59, 52]. Recent work further enhances visual prompt understanding by additional visual instruction tuning with diverse visual prompts overlaid on the images [6], or explicitly represent visual pointer tokens in the models [21, 63, 71]. We leverage visual prompting to avoid potential ambiguity in textual descriptions, especially when evaluating multiple object hallucinations for objects of the same class. Benchmark Design Considerations #Test Multi. Distr. Source Ref. Eval. CCEval [68] Seen Text N 0.1k GAVIE [29] Mixed Text N 1k FAITHScore [19] Unseen Text N 2k Ha ELM [54] Unseen Text N 5k M-Hal Detect [13] Unseen Text H 0.8k MMHal-Bench [46] Unseen Text N,H 0.1k CHAIR [44] Unseen A 46k AMBER [53] Unseen Text A 1k CIEM [15] Unseen Text A 5k NOPE [32] Unseen A 3k POPE [26] Train Unseen A 0.5k ROPE (Ours) Test Seen & Vis. A 5k Unseen Table 1: An overview of object hallucination benchmarks. For design considerations, we summarize the number of tested images, and if multiple classes and object class distribution (at training and test time) are considered. The image sources include those seen or unseen during instruction tuning. To refer to an object, textual descriptions and visual cues can be adopted. For evaluation, neural models, humans and automatic pipelines are used. Object Hallucination. Despite their promising performance on benchmarks, these models frequently generate objects that do not exist in the provided images, a problem known as object hallucination [44, 10]. Several methods have been suggested to mitigate the object hallucination issue, such as integrating an external object detector [68], applying visually grounded visual instruction tuning [63, 71] or reinforcement learning [46, 13], performing iterative refinement [75], and adapting the decoding strategies [17]. To quantify progress on mitigating them, various benchmarks have been developed and have revealed the prevalence of object hallucination, even in images that are seen during instruction tuning [68, 29]. We contrast our ROPE benchmark against existing benchmarks and setups in Table 3. ROPE, which is designed for evaluating multi-object hallucination, is distinguished in several ways. First, we deliberately consider the distribution of object classes within a single image at test time. Object hallucination is observed originally in image captioning applications, where multiple objects are described [44]. While existing research has demonstrated that the object class distribution in the instruction tuning dataset can influence hallucination patterns [26, 75, 53], the impact of object class distribution within an image at test time remains under-explored. Second, current benchmarks concentrate on the presence of an object class or distinguish instances using textual descriptions like attributes, which can still result in ambiguity and multiple referents. We instead leverage the visual referring prompting setups and use visual cues (i.e., marked bounding boxes) to refer to objects. Finally, our evaluation is automated, without black-box neural models or human evaluators. 3 Recognition-based Object Probing Evaluation We introduce the Recognition Object Probing Evaluation (ROPE), an automated protocol for assessing LVLMs in multi-object recognition. ROPE specifically measures object hallucination in VLMs within a multi-object setting, examining how models may misperceive (e.g., by inventing nonexistent objects or becoming distracted) when tasked to focus on multiple objects concurrently, and which factors cause the hallucinations. 3.1 Task Setup Problem Definition. To avoid ambiguity from multiple candidate referents when using text prompts, ROPE leverages visual prompts to uniquely refer to objects. ROPE tasks LVLMs with selecting the best matching class for multiple objects, as referred to by the visual prompt, from a predefined set of object classes. Specifically, each sample in the ROPE protocol consists of a quadruple {I, L, p1, , pn , o1, , on }: (1) an image I consisting of at least n objects; (2) a natural language instruction L that specifies the recognition task, including N candidate object classes c1, , c N; (3) n visual prompts p1, , pn, each queries an object in the image; and (4) n object classes o1, , on as the answers. In this work, we construct a dataset with N = 50 and n = 5, i.e., models are tasked with recognizing 5 objects out of 50 candidate object classes. Although we use this dataset as an example, ROPE can be applied to any dataset containing multiple objects and their bounding boxes. Language Instruction Prompts. For a fair comparison that accommodates both open-weight and API-based LVLMs, ROPE explicitly instructs models to generate a formatted output of object classes, e.g., obj1:, ..., obj5: (Figure 2). This format enables automated evaluation through simple parsing. This format enables automated evaluation through simple parsing, avoiding black-box neural models or human evaluators With different analytical purposes, we designed 3 types of task prompts for Multi-Object queries, as illustrated in Figure 7 and described as follows. Default: We probe the model to recognize the 5 objects referred to by the visual prompts concurrently in a single turn of prompting. This setting tasks the model with focusing on and recognizing all 5 objects simultaneously, aiming to capture the complexity involved when the model generates language that includes multiple objects. Student-Forcing: One potential confounder in the default setting is the model s ability to generate data in the specified format. To separate out errors due to following instructions, we force the model to follow the format template and decode only the object tokens for each of the five objects. Ideally, this setting allows the model to focus solely on object recognition. Teacher-Forcing: This setting eliminates cumulative error, allowing the model to condition on the correct previous context when generating object classes, leading to upper bound performance in multi-object recognition. We similarly force the model to follow the provided template and decode only the object tokens for each of the five objects, but we replace the previously generated object tokens with the ground truth. This essentially follows the few-shot in-context learning setting. Teacher forcing helps especially when LVLMs take shortcuts by repeating the object class list as ordered in the prompt (e.g., LLa VA-7B [31] and Gemini 1.0 Pro [49] in Figure 2). For comparison, we also designed task prompts for Single-Object query. We probe the model to recognize the object referred to by the visual prompts one at a time, repeating this as 5 independent and individual prompts. Unlike Default multi-object query, the model only needs to focus on one object, which can be seen as an extension of the POPE [26] setup from yes/no polling to classification. We refer to Appendix A.1 for the prompt templates for each type of task prompt. 3.2 Dataset Construction Data Sources and Curation. Since our goal is to evaluate and analyze multi-object hallucination, the image data must contain multiple objects of diverse classes with instance-level semantic annotations. We build our dataset upon existing panoptic segmentation datasets, including MSCOCOPanoptic [28, 5] and ADE20K [74], to ensure access to all object instances and their semantic classes. We note that one can build a dataset using the ROPE protocol with any dataset containing multiple objects and their bounding boxes. We describe the data curation pipeline in Appendix A.1. Splits by Query Distributions. As shown in Figure 2and 3, our initial observations indicate that LVLMs are less likely to hallucinate objects when they are tasked with recognizing the same object obj1: apple obj2: orange obj3: banana obj4: lemon obj5: pear Gemini 1.0 Pro obj1: apple obj2: knife obj3: fork obj4: apple obj5: jar obj1: apple obj2: orange obj3: banana obj4: lemon obj5: pear obj1: fork obj2: knife obj3: whisk obj4: lemon obj5: jar Ground Truth obj1: apple obj2: lemon obj3: bottle obj4: packaged snack obj5: jar Qwen-VL-Chat obj1: packaged snack obj2: knife obj3: whisk obj4: lemon obj5: jar obj1: apple obj2: apple obj3: apple obj4: lemon obj5: pear LLa VA-34B obj1: fork obj2: knife obj3: whisk obj4: lemon obj5: jar Gemini 1.5 Pro obj1: packaged snack obj2: knife obj3: soda can obj4: lemon obj5: jar Qwen-VL-Max Select one and the most appropriate class for each object located within red bounding boxes from the following list: apple, orange, banana, lemon, pear, plate, bowl, jar, bottle, soda can, knife, fork, whisk, scissors, packaged snack. Provide the class names in the format: 'obj1: , obj2: , obj3: , obj4: , obj5: ', with no additional words or punctuations. Figure 2: A heterogeneous ROPE sample tested with Deafult multi-object query, where each of the 5 objects belongs to different object classes. We label the output class as either correct or hallucinated. obj1: apple obj2: apple obj3: apple obj4: apple obj5: apple Ground Truth obj1: apple obj2: apple obj3: apple obj4: apple obj5: orange Ground Truth obj1: apple obj2: apple obj3: apple obj4: apple obj5: apple obj1: apple obj2: apple obj3: apple obj4: apple obj5: apple obj1: apple obj2: apple obj3: apple obj4: apple obj5: apple obj1: apple obj2: apple obj3: apple obj4: apple obj5: apple Figure 3: A homogeneous ROPE sample, where the 5 objects belong to the same object class, and a corresponding adversarial ROPE sample, where the last object belongs to a different object class. class multiple times. However, they tend to make more mistakes when all tasked object classes are different or when a new object class is introduced after multiple repeated tasks. We thus consider the distribution of object classes within each image at test time, dividing ROPE into 4 subsets: Homogeneous, Heterogeneous, and Adversarial, In-the-Wild. Homogeneous: All the 5 tested objects are of the same class, e.g., AAAAA. Heterogeneous: All the 5 tested objects are of different classes, e.g., ABCDE. Adversarial: The first 4 tested objects are of the same class while the last is different, e.g., AAAAB. In-the-Wild: A subset with mixed object class distribution, where the 5 tested objects are randomly chosen and ordered given a test image. Attending to Data Contamination. While data contamination has been explicitly handled in most of the existing benchmarks, object hallucination has been observed even in images that appear in the instruction tuning dataset, such as Visual Genome [68, 29]. To evaluate whether multi-object hallucination can be observed in both seen and unseen images, and to critically determine if training on these images helps reduce hallucinations, we explicitly split our dataset into Seen and Unseen based on the original split of the datasets.2 Depending on the object query distributions (4 splits) and whether the image appears in the training split (2 splits), we divide the test into 8 folders. 4 Experiments and Results 4.1 LVLM Baselines The proposed ROPE framework, in principle, applies to all LVLMs that can follow format instructions and understand multiple visual prompts. To cover a variety of LVLMs of different scales and training data (e.g., whether grounding data and conversational data are used), we selected the following LVLMs as baselines. LVLMs with base LLMs at different scales: LLa VA v1.6 (7B/13B/34B) [31, 30] and Yi-VL (6B/34B) [64]. 2We believe this approach is the best practice, but we also acknowledge that the distinction between seen and unseen images may not be strict. Uncurated web images often overlap with public test images, and researchers have no transparent access to the datasets used to train some of these LVLMs unfortunately [11]. Default Multi-Object Student-Forcing Teacher-Forcing Single-Object Models Wild Hom. Het. Wild Hom. Het. Wild Hom. Het. Wild Hom. Het. Seen Yi-VL-6B 2.95 5.65 1.99 3.44 6.80 3.78 5.45 26.25 4.36 0.19 0.30 0.13 Yi-VL-34B 8.50 15.35 3.33 8.97 16.30 4.23 10.09 19.75 4.94 0.22 2.60 0.13 LLa VA-7B 31.29 67.50 8.00 31.28 67.25 11.22 31.49 92.15 12.37 35.32 62.35 17.37 LLa VA-13B 31.54 67.63 12.64 31.49 73.25 11.54 34.97 94.25 16.03 43.13 80.60 23.91 LLa VA-34B 39.95 85.75 18.85 52.75 85.20 33.91 56.41 95.81 25.31 55.05 86.50 18.97 Qwen VL 2.73 6.60 1.03 6.25 16.00 3.65 18.74 71.50 5.45 8.73 16.05 5.58 Qwen VL-C 8.72 16.90 6.67 5.26 8.60 4.10 12.11 47.75 8.08 25.99 43.40 13.21 Cog VLM 0.04 0.00 0.00 0.00 0.00 0.00 0.10 0.95 0.00 0.00 0.00 0.00 Cog VLM-G 0.00 0.00 0.00 9.86 13.50 6.79 22.64 75.45 0.45 11.25 22.65 7.12 Cog VLM-C 12.89 22.75 7.18 25.37 43.63 12.03 28.25 72.80 17.50 30.16 56.00 16.35 LLa VA-7B* N/A 9.16 16.40 5.51 N/A 11.68 23.55 9.36 GLa MM* N/A 27.11 53.35 13.01 N/A 63.81 81.75 53.40 Ground HOG* N/A 23.57 30.80 24.23 N/A 44.80 43.10 38.97 IDEFICS 0.00 1.45 0.13 6.25 18.70 0.64 17.37 76.15 10.06 4.62 0.00 0.32 Cog VLM-2 21.51 37.55 17.31 37.02 70.85 12.69 37.10 73.50 17.44 21.16 38.75 13.65 Mini CPM-V 34.75 59.91 17.37 31.62 62.80 13.65 32.16 68.05 16.79 27.42 55.35 16.92 GPT-4V 53.80 77.55 40.83 N/A N/A 55.89 78.25 41.03 GPT-4O 71.27 89.25 66.03 N/A N/A 60.77 73.92 54.31 LLa VA-7B 21.26 52.40 7.69 N/A N/A 30.59 60.85 12.69 +OPERA 24.07 58.65 7.35 N/A N/A 30.44 60.85 13.27 Unseen Yi-VL-6B 2.74 3.88 1.14 3.18 4.24 5.20 4.04 10.90 10.57 0.14 0.45 0.08 Yi-VL-34B 7.77 15.63 4.23 10.28 18.04 7.97 11.24 22.49 12.03 0.46 2.37 0.41 LLa VA-7B 30.56 68.12 10.33 30.55 68.16 10.24 31.89 90.33 13.25 34.88 64.41 16.18 LLa VA-13B 27.56 63.10 8.37 27.41 63.10 8.37 35.65 91.09 14.80 42.66 71.92 23.41 LLa VA-34B 29.30 79.43 17.72 29.45 91.18 14.39 37.40 95.51 17.92 51.71 77.88 30.81 Qwen VL 2.80 1.95 7.06 7.17 16.41 4.15 10.34 58.00 4.07 17.73 31.22 9.51 Qwen VL-C 18.86 30.73 8.78 16.16 27.80 7.72 21.81 58.00 11.14 34.20 57.31 15.37 Cog VLM 0.03 0.00 0.00 0.00 0.00 0.00 0.00 0.15 0.00 0.00 0.00 0.00 Cog VLM-G 0.00 0.00 0.00 8.20 1.47 5.77 23.82 81.20 1.81 10.32 10.74 9.11 Cog VLM-C 15.56 26.57 5.53 17.18 41.27 6.02 22.81 56.04 6.67 30.56 52.00 13.50 LLa VA-7B* N/A 7.59 12.12 4.88 N/A 12.71 22.49 8.46 GLa MM* N/A 29.11 54.53 14.23 N/A 68.65 77.06 52.28 Ground HOG* N/A 23.11 24.69 26.26 N/A 40.73 30.37 38.13 IDEFICS 0.39 0.37 0.33 9.03 24.45 2.68 24.80 83.02 7.64 4.62 3.67 6.50 Cog VLM-2 20.99 35.06 15.93 24.64 38.04 23.17 26.74 46.04 26.59 11.13 30.94 5.77 Mini CPM-V 32.96 59.92 16.60 31.77 58.98 14.15 31.87 60.98 16.34 25.56 47.76 14.39 GPT-4V 45.46 63.12 34.17 N/A N/A 47.34 64.94 35.45 GPT-4O 63.27 80.29 54.47 N/A N/A 63.45 79.84 53.74 LLa VA-7B 13.96 31.88 3.98 N/A N/A 26.95 54.41 11.06 +OPERA 13.20 37.14 3.82 N/A N/A 27.90 56.69 11.22 * Mechanistically grounded LVLMs take visual prompts by dedicated pointer tokens. We slightly adapt the text prompt and probe the object classes with the highest probabilities. We also apply such probabilistic probing to LLa VA-7B for comparison, as all of three models adopt Vicuna-7B v1.5 [73] as the base LLM. See Appendix A.1 for details. For GPT models, student/teacher forcing doesn t apply as they are API-only. OPERA is implemented based on LLa VA-7B v1.5. Table 2: Averaged accuracy of baselines on the In-the-Wild, Homogeneous, and Heterogeneous splits. The bold marker denotes the best-performing baseline and the underlined marker denotes the second-best-performing baseline. LVLMs with conversational/grounded instruction tuning: Qwen VL-Base/Chat (7B) [3] and Cog VLM-Base/Chat/Grounding v1.1 (19B) [55]. Mechanistically grounded LVLMs: GLa MM (7B) [42] and Ground HOG (7B) [71]. LVLMs with RL-based finetuning: Mini CPM-V [66] Other LVLMs: IDEFICS-instruct (9B) [23], Mini CPM-V v2.5 (8B) [16, 67], GPT-4V [37], and GPT-4O [38]. For mechanistically grounded LVLMs that take visual prompts through specially designed mechanisms, such as pointer tokens in Ground HOG [71], we additionally experiment with their default format and report whichever yields higher performance. For other LVLMs, we overlay the visual prompts on the images using a red bounding box with a width of 2 and visual text specifying the object index, presented with a white italic font on a black background with an alpha value of 0.75 for contrast and visibility. obj 1 obj 2 obj 3 obj 4 obj 5 Object Index 0 0.0 19.2 29.0 (a) LLa VA-7B. obj 1 obj 2 obj 3 obj 4 obj 5 Object Index 100.0 100.0 0.0 13.8 26.1 (b) LLa VA-13B. obj 1 obj 2 obj 3 obj 4 obj 5 Object Index 68.2 94.8 95.3 (c) LLa VA-34B. AAAAB SO BAAAA SO AAAAB TF BAAAA TF Figure 4: The performance of the LLa VA on the adversarial split, organized by the query sequence of AAAAB and BAAAA, reveals significant vulnerabilities as the model s accuracy dramatically declines for object 5 in AAAAB. SO stands for single-object probing and TF stands for teacher-forcing probing. 4.2 Main Results and Findings We summarize the average results across the splits in Table 2 and present the most important findings below. The full tables appear in Appendix A.2. Multi-object tasks introduce more hallucinations. Our immediate observation is that LVLMs suffer from more hallucinations when tasked with focusing on and recognizing multiple objects compared to a single object. Across most of the models and test splits, we find that the average accuracy of single-object queries (i.e., probing object classes one at a time) significantly outperforms that of all three types of multi-object queries. The first exceptions are GPT-4O, Mini CPM-V, and Cog VLM-2, the latter two leverage LLa MA-3 [35]. Another exception to this is when teacher-forcing is applied to homogeneous test splits, which demonstrates an unreasonably high accuracy. We discuss them later in this section. Heterogeneous queries introduce more hallucinations. We find that for all models and query methods, more heterogeneous queries lead to substantially more hallucinations, with performance decreasing from homogeneous to in-the-wild to heterogeneous test sets. The impact of heterogeneity applies to even start-of-the-art LVLMs like GPT-4O (Figure 2), although this performance gap is more significant in open-weight models. Language bias and shortcuts can lead to multi-object hallucinations. In the teacher-forcing setting, where there are no cumulative errors, LLa VA models score over 90% accuracy. There are three possible hypotheses for this abnormal observation: (1) LVLMs are smart enough to learn object recognition in general through few-shot in-context learning in the teacher-forcing setting; (2) LVLMs learn to recognize one specific object through few-shot in-context learning in the teacher-forcing setting; or (3) LVLMs simply exploit language biases and rule-based shortcuts (e.g., repeating previous answers). To reach a conclusion on this, we examine an Adversarial split, in which the first four tested objects are of the same class and we probe an object of a different class for the last one (e.g., AAAAB). We compare the single-object query performance with the teacher-forcing performance on the fifth object (object B). We anticipate the following outcomes: If hypothesis (1) is correct, the teacher-forcing performance should outperform the single-object query. If hypothesis (2) is correct, the teacher-forcing performance should perform on par with the single-object query. If hypothesis (3) is correct, the teacher-forcing performance should underperform compared to the single-object query. For a controlled comparison in the multi-object setting, we also reverse the order of queries (i.e., BAAAA) and repeat the experiments. We present the results of LLa VA models on the unseen split in Figure 4, with the full results available in Appendix A.2. We find that the model s predictions on class A progressively improve, scoring nearly perfectly starting from the third repetition. However, the model s performance on the last object (with the different class label B) drops to nearly zero, with almost all hallucinations labeling it as A. This is in stark contrast to 23.35% if these objects are probed individually or 19.16% when these objects are placed as the first to query in multi-object settings. Our findings suggest that hypothesis (3) is true, indicating that the LVLMs high performance on homogeneous queries could be an illusion resulting from textual shortcuts. We observe that models show lower performance in identifying object class B in single-object analysis, potentially due to the higher salience of object class A in these images. Multi-object hallucinations occur in both seen and unseen images. We finally investigate whether our observations and findings hold uniformly in both seen and unseen splits. We observe that the gap between multi-object hallucination and single-object hallucination, as well as the reliance on shortcuts, persists. Although most of the models perform slightly better on seen images, the trends remain consistent across both splits. While large-scale training is involved in developing these LVLMs, it appears they might not have fully exploited the fine-grained information in the data. Training on these images does not significantly reduce object hallucinations. 4.3 What May Help and What May Not? Comparing the tested LVLMs, we discuss our observations regarding design considerations that may or may not help reduce multi-object hallucinations. Scaling the base LLM: data and parameters. We find that using base LLMs with more parameters reduces single-object hallucinations, but may not have the same effect on multi-object hallucinations. We observe a consistent increase in performance with larger LLa VA models in the seen set and in single-object queries, but not in the unseen set with multi-object queries. One possible explanation for this finding is that LLMs with more parameters are better at memorizing seen images, as the performance gap between seen and unseen images is also more significant in larger models. We also notice that the performance gap between single-object probing and multi-object probing does not apply to Mini CPM-V and Cog VLM-2, which adopt a LLa MA-3 (8B) [35] base LLM pre-trained with 15T tokens, as they fail to follow the instruction sometimes. Compared to LLa VA models developed upon LLa MA-2 (7/13B) [50] and Yi (34B) [64] with 2T and 3T pre-training tokens, these models underperform in quantitative measures due to instruction following error but exhibit greater robustness when multiple visual prompts are presented. Visual instruction fine-tuning: chat and grounding. While it s surprising that conversational tuning reduces multi-object hallucinations, we observe that models without conversational tuning struggle to follow instructions and are prone to shortcuts, such as repeating the list of all object class candidates in order or consistently repeating the first candidate. This might also explain why grounded tuning in Cog VLM-G is of little help in reducing multi-object hallucinations thus far. These models typically lack conversational fine-tuning, and there is currently no available grounded dialogue data at scale. While mechanistically grounded LVLMs show strong performances in single-object probing, there remain a gap in multi-object probing with student forcing. This could be attributed to a significant portion of the grounded instruction tuning dataset consisting mainly of short captions or questions featuring one single or few objects.RL-based finetuning approaches, such as Mini CPM-V, demonstrate promising results across diverse settings, surpassing single-object results in both the Wild and Homogenous settings. Upon inspection, we found that this model demonstrates strong visual in-context learning capability and improves correct recognition when objects of the same classes are probed together. Decoding and inference time strategy Decoding algorithms like OPERA introduce nuanced improvements in specific multi-object settings[18]. In default multi-object tasks, OPERA shows marginal performance enhancements for LLa VA-1.5, but its effectiveness declines in tests with greater heterogeneity, to the point it can even lower performance. This suggests OPERA is beneficial in homogeneous contexts but requires further refinement in handling mixed object scenarios effectively. 5 Analysis of Hallucinatory Behaviors 5.1 Potential Hallucinatory Factors The task setup described above allows us to evaluate LVLMs in multi-object hallucinations and identify hallucinatory behaviors. Based on existing literature and our case studies, we further identify potential factors that correlate to and potentially explain these hallucinations. Data-specific Factors. We consider the following factors that are specific to the tested sample (e.g., object and token positions), and are not relevant to the frequency distribution. Input Order: we consider the order in which the object classes are presented in the input prompt containing all candidates. Query Homogeneity: We define query homogeneity as the total number of task objects of the same class, normalized by the total number of queried objects (five in this work). Object Token Position: Zhou et al. [75] has shown that more hallucinations occur in the latter part of captions. In this work, the object indices directly correspond to the object token positions. Object Homogeneity: We define object homogeneity as the number of object types in the image, calculated upon panoptic annotations. Object Centrality: Previous research has identified a center bias in datasets and models, indicating that objects are disproportionately located at the center of images in detection models [47, 48]. We define object centrality as one minus the distance d between the object s bounding box center and the image center, normalized by the diagonal distance D from the center to the corner. 0.20 0.40 0.60 0.80 1.00 (a) Query Homogeneity. (b) Object token position. 64 12.9 16.2 Non Hallucinate Hallucinate Hallucinate Mean Non Hallucinate Mean (c) Object Homogeneity. 0 1 0.56 0.52 (d) Object centrality. (e) Object salience. 0 1 0.21 0.13 (f) Semantic salience. 0 250000 52832 168583 (g) Training salience. (h) Object token entropy. 0.06 0.18 0.11 0.12 (i) Visual modality contribution. Figure 5: A comparison of the distribution of hallucinatory versus non-hallucinatory object classes in LLa VA-13B, across the unseen split under student forcing. Salience and Frequency. We consider the following factors that are related to the saliency or frequency of the visual object or the object class. Object Salience: Previous research has shown that smaller objects are harder to detect and ground to [14, 34]. We define object salience as the ratio of the number of pixels occupied by the object s instance segmentation mask to the total number of pixels in the image. Semantic Salience: We observe and hypothesize that LVLMs are less likely to hallucinate objects when they co-occur with multiple copies of the same class ( jar in Figure 2). We define semantic salience as the ratio of the total number of pixels in all instances of the same class, to the total number of pixels in the image. Training Salience: Previous research has shown that spurious co-occurring patterns in the training data can lead to object hallucinations [26, 75]. We use the log frequency of classes in MSCOCO as a proxy for training salience following previous work, and hypothesize that LVLMs tend to hallucinate more on less frequent objects in the training set. Model Behaviors. We consider the following factors relevant to the mechanistic behaviors. Object Token Entropy: Zhou et al. [75] have shown that object hallucinations are more likely when the decoded object tokens have a higher log perplexity. In our work, we define object token entropy as the entropy of the logits of the first token in the generated word. Given s as the softmax logits of the generated word s first token, we calculate the entropy using the following formula: H(s) = P i si log(si). Simply put, higher entropy indicates greater uncertainty in the model s prediction for the first token, which can lead to more frequent object hallucinations. Visual Modality Contribution: We hypothesize that LVLMs pay less attention to the visual modality during object hallucinations. Motivated by the modality importance score [7], we define Visual Modality Contribution (VMC) as the proportion of attention allocated to visual tokens compared to textual tokens. To quantify this, we analyze the attention weights of the last generated token across all heads and layers. The VMC is computed as follows: VMC = P i V αij / P i V αij + P k T αkj , where αij represents the attention weight assigned to visual token i at head j, and αkj represents the attention weight assigned to textual token k at head j. The sets V and T denote the visual and textual tokens, respectively. By examining the VMC, we can determine how much attention is given to visual inputs in comparison to textual inputs. A lower VMC may indicate a higher likelihood of object hallucinations due to insufficient attention to visual cues. 5.2 When Do LVLMs Experience Multi-Object Hallucinations? In Figure 5, we compare the distribution of these factors between hallucinatory and non-hallucinatory objects in the student forcing setting on the unseen split using LLa VA-13B. For continuous values, we use ridgeline plots, and for discrete values with fewer bins, we use bar charts. 0.2 0.4 0.6 0.8 1.0 (a) Semantic salience. 0 250000 50849 106916 (b) Training salience. 0 10 20 30 40 50 Actual Class Predicted Class Mean Actual Class Mean Predicted Class (c) Input order. Figure 6: A comparison of the distribution of actual versus predicted object classes for all hallucinatory objects in the student forcing setting on the unseen split using LLa VA-13B. Data/Task-specific Factors. We observed that specific data factors, such as query and object homogeneity, significantly influence model performance, with increased hallucination occurring when models process images featuring multiple object classes or a variety of objects. For positional factors, the position of object tokens seems to have minimal impact and the object centrality has only a slight influence as LVLMs tend to hallucinate objects more frequently when they are positioned away from the center. This tendency may stem from a reporting bias, as objects mentioned in captions are typically foreground objects that distribute toward the centers of images. Salience and Frequency. We note that semantic salience significantly affects the model s performance, as it is more prone to hallucinate an object class that is less salient within the image. Conversely, the salience of individual objects does not statistically correlate with hallucination incidents. This implies that LVLMs may rely more on the presence of co-occurring objects of the same class to predict the labels of queried objects, rather than solely on the presence or salience of the objects themselves. Additionally, training salience plays a crucial role as models are less likely to hallucinate object classes that frequently appear in training. Intrinsic Behaviors. The intrinsic behaviors of the model provide significant insights into its tendencies to hallucinate. Similar to Zhou et al. [75], we find that models are more prone to hallucination when they experience uncertainty or confusion, especially in scenarios involving multiple objects, as evidenced by higher token entropy. Furthermore, the contribution from the visual modality consistently registers below 20%, suggesting that current LVLMs may rely more heavily on linguistic contexts. There is a marginal increase in the likelihood of hallucination when models pay less attention to the visual context. 5.3 How Do LVLMs Experience Multi-Object Hallucinations? In Figure 6, we conducted a detailed comparison of the distribution of actual versus predicted object classes within the context of hallucinatory objects, examining factors such as semantic salience, training salience, and input order. Although semantic salience is a key factor in determining whether a model hallucinates, it appears to have minimal impact on the prediction of hallucinated objects. Our analysis also shows that models are more likely to hallucinate object classes that are prevalent in the training data, but the reverse is not necessarily true. Additionally, there is a notable preference for models to hallucinate objects that are listed early in the input prompt as candidate classes. Overall, our findings indicate that spurious correlations may lead to hallucinations involving multiple objects. 6 Discussions and Conclusion Hallucinations in large vision-language models (LVLMs) can occur at different scales and granularities. In this study, we study the problem of multi-object hallucination, examining how LVLMs may misperceive when tasked to focus on multiple objects concurrently, and which factors cause the hallucinations. we introduce Recognition-based Object Probing Evaluation (ROPE), an automated evaluation protocol designed to account for the distribution of object classes within a single image during testing and to use visual referring prompts to reduce ambiguity. Our research provides key insights for the development and application of LVLMs. Since models tend to experience more hallucinations with multiple objects than with single ones, it may be advantageous to probe objects individually in visual prompts to enhance performance. The likelihood of a model s hallucinatory output is linked to various data factors and model behaviors. Particularly in situations involving heterogeneous data and low certainty from the model, there is an increased risk of hallucinations, and users should be vigilant. Moreover, our analysis indicates that merely adopting (grounded) instruction tuning and scaling the base language model may not be enough to fully address the issue of object hallucination. There is a need for more balanced object distributions, annotations of objects away from image centers, and an increase in diversity. Introducing instructions that require multiple visual pointers and complex multi-object reasoning is also crucial. Acknowledgement This work was supported in part by NSF IIS-1949634, NSF SES-2128623, the DARPA Perceptual Task Guidance (PTG) Program, and the DARPA Machine Common Sense Program. Our experiments have also benefited from the Microsoft Accelerate Foundation Models Research (AFMR) program. We thank the Amazon AGI team for Ground HOG model access. The authors would like to thank Yichi Zhang and anonymous reviewers for their valuable feedback. [1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. 2022. Flamingo: a visual language model for few-shot learning. volume 35, pages 23716 23736. [2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. 2022. Exploring visual prompts for adapting large-scale models. ar Xiv preprint ar Xiv:2203.17274. [3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision-language model with versatile abilities. ar Xiv preprint ar Xiv:2308.12966. [4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. 2020. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621 11631. [5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. 2018. Coco-stuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209 1218. [6] Mu Cai, Haotian Liu, Siva Karthik Mustikovela, Gregory P Meyer, Yuning Chai, Dennis Park, and Yong Jae Lee. 2024. Vip-llava: Making large multimodal models understand arbitrary visual prompts. In IEEE Conference on Computer Vision and Pattern Recognition. [7] Jize Cao, Zhe Gan, Yu Cheng, Licheng Yu, Yen-Chun Chen, and Jingjing Liu. 2020. Behind the scene: Revealing the secrets of pre-trained vision-and-language models. In Computer Vision ECCV 2020: 16th European Conference, Glasgow, UK, August 23 28, 2020, Proceedings, Part VI 16, pages 565 580. Springer. [8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. 2023. Shikra: Unleashing multimodal llm s referential dialogue magic. ar Xiv preprint ar Xiv:2306.15195. [9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. 2023. Instructblip: Towards general-purpose vision-language models with instruction tuning. volume 36. [10] Wenliang Dai, Zihan Liu, Ziwei Ji, Dan Su, and Pascale Fung. 2023. Plausible may not be faithful: Probing object hallucination in vision-language pre-training. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2128 2140. [11] Jesse Dodge, Maarten Sap, Ana Marasovi c, William Agnew, Gabriel Ilharco, Dirk Groeneveld, Margaret Mitchell, and Matt Gardner. 2021. Documenting large webtext corpora: A case study on the colossal clean crawled corpus. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1286 1305. [12] Tao Gong, Chengqi Lyu, Shilong Zhang, Yudong Wang, Miao Zheng, Qian Zhao, Kuikun Liu, Wenwei Zhang, Ping Luo, and Kai Chen. 2023. Multimodal-gpt: A vision and language model for dialogue with humans. ar Xiv preprint ar Xiv:2305.04790. [13] Anisha Gunjal, Jihan Yin, and Erhan Bas. 2024. Detecting and preventing hallucinations in large vision language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18135 18143. [14] Derek Hoiem, Yodsawalai Chodpathumwan, and Qieyun Dai. 2012. Diagnosing error in object detectors. In European conference on computer vision, pages 340 353. Springer. [15] Hongyu Hu, Jiyuan Zhang, Minyi Zhao, and Zhenbang Sun. 2023. Ciem: Contrastive instruction evaluation method for better instruction tuning. In Neur IPS 2023 Workshop on Instruction Tuning and Instruction Following. [16] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, et al. 2024. Minicpm: Unveiling the potential of small language models with scalable training strategies. ar Xiv preprint ar Xiv:2404.06395. [17] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. 2024. Opera: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. [18] Qidong Huang, Xiao wen Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Neng H. Yu. 2023. Opera: Alleviating hallucination in multi-modal large language models via over-trust penalty and retrospection-allocation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13418 13427. [19] Liqiang Jing, Ruosen Li, Yunmo Chen, Mengzhao Jia, and Xinya Du. 2023. Faithscore: Evaluating hallucinations in large vision-language models. ar Xiv preprint ar Xiv:2311.01477. [20] Aishwarya Kamath, Mannat Singh, Yann Le Cun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. 2021. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780 1790. [21] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. 2023. Lisa: Reasoning segmentation via large language model. ar Xiv preprint ar Xiv:2308.00692. [22] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. 2023. Obelics: An open web-scale filtered dataset of interleaved image-text documents. volume 36. [23] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. 2023. Obelics: An open web-scale filtered dataset of interleaved image-text documents. Advances in Neural Information Processing Systems, 36. [24] Xuanyu Lei, Zonghan Yang, Xinrui Chen, Peng Li, and Yang Liu. 2024. Scaffolding coordinates to promote vision-language coordination in large multi-modal models. ar Xiv preprint ar Xiv:2402.12058. [25] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. 2022. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965 10975. [26] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. 2023. Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. [27] Zongjie Li, Chaozheng Wang, Chaowei Liu, Pingchuan Ma, Daoyuan Wu, Shuai Wang, and Cuiyun Gao. 2023. Vrptest: Evaluating visual referring prompting in large multimodal models. ar Xiv preprint ar Xiv:2312.04087. [28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. 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. [29] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. 2023. Mitigating hallucination in large multi-modal models via robust instruction tuning. In The Twelfth International Conference on Learning Representations. [30] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023. Improved baselines with visual instruction tuning. In Neur IPS 2023 Workshop on Instruction Tuning and Instruction Following. [31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. volume 36. [32] Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. 2023. Negative object presence evaluation (nope) to measure object hallucination in vision-language models. ar Xiv preprint ar Xiv:2310.05338. [33] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. 2022. Unified-io: A unified model for vision, language, and multi-modal tasks. In The Eleventh International Conference on Learning Representations. [34] Ziqiao Ma, Jiayi Pan, and Joyce Chai. 2023. World-to-words: Grounded open vocabulary acquisition through fast mapping in vision-language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 524 544. [35] Meta. 2024. Introducing meta llama 3: The most capable openly available llm to date. [36] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. 2024. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. ar Xiv preprint ar Xiv:2402.07872. [37] Open AI. 2023. Gpt-4v(ision) system card. [38] Open AI. 2024. Hello gpt-4o. [39] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. 2024. Grounding multimodal large language models to the world. In The Twelfth International Conference on Learning Representations. [40] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. 2023. Detgpt: Detect what you need via reasoning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14172 14189. [41] Shengyi Qian, Weifeng Chen, Min Bai, Xiong Zhou, Zhuowen Tu, and Li Erran Li. 2024. Affordancellm: Grounding affordance from vision language models. ar Xiv preprint ar Xiv:2401.06341. [42] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. 2024. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. [43] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jeanbaptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, et al. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. ar Xiv preprint ar Xiv:2403.05530. [44] Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2018. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035 4045. [45] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. 2023. What does clip know about a red circle? visual prompt engineering for vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11987 11997. [46] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. 2023. Aligning large multimodal models with factually augmented rlhf. ar Xiv preprint ar Xiv:2309.14525. [47] Gergely Szabó and András Horváth. 2022. Mitigating the bias of centered objects in common datasets. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 4786 4792. IEEE. [48] Mohammad Reza Taesiri, Giang Nguyen, Sarra Habchi, Cor-Paul Bezemer, and Anh Nguyen. 2023. Imagenet-hard: The hardest images remaining from a study of the power of zoom and spatial biases in image classification. volume 36. [49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. Gemini: a family of highly capable multimodal models. ar Xiv preprint ar Xiv:2312.11805. [50] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. ar Xiv preprint ar Xiv:2307.09288. [51] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. 2021. Multimodal few-shot learning with frozen language models. volume 34, pages 200 212. [52] David Wan, Jaemin Cho, Elias Stengel-Eskin, and Mohit Bansal. 2024. Contrastive region guidance: Improving grounding in vision-language models without training. ar Xiv preprint ar Xiv:2403.02325. [53] Junyang Wang, Yuhang Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Ming Yan, Ji Zhang, and Jitao Sang. 2023. An llm-free multi-dimensional benchmark for mllms hallucination evaluation. ar Xiv preprint ar Xiv:2311.07397. [54] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. 2023. Evaluation and analysis of hallucination in large vision-language models. ar Xiv preprint ar Xiv:2308.15126. [55] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. 2023. Cogvlm: Visual expert for pretrained language models. ar Xiv preprint ar Xiv:2311.03079. [56] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. 2023. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. volume 36. [57] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. 2024. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. [58] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. 2023. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. ar Xiv preprint ar Xiv:2310.11441. [59] Lingfeng Yang, Yueze Wang, Xiang Li, Xinlong Wang, and Jian Yang. 2024. Fine-grained visual prompting. volume 36. [60] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. 2023. The dawn of lmms: Preliminary explorations with gpt-4v (ision). ar Xiv preprint ar Xiv:2309.17421, 9(1):1. [61] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. 2024. Cpt: Colorful prompt tuning for pre-trained vision-language models. AI Open, 5:30 38. [62] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. 2023. mplug-owl: Modularization empowers large language models with multimodality. ar Xiv preprint ar Xiv:2304.14178. [63] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. 2023. Ferret: Refer and ground anything anywhere at any granularity. In The Twelfth International Conference on Learning Representations. [64] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. 2024. Yi: Open foundation models by 01. ai. ar Xiv preprint ar Xiv:2403.04652. [65] Shoubin Yu, Jaehong Yoon, and Mohit Bansal. 2024. Crema: Multimodal compositional video reasoning via efficient modular adaptation and fusion. ar Xiv preprint ar Xiv:2402.05889. [66] 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. 2023. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13807 13816. [67] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. 2024. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. ar Xiv preprint ar Xiv:2405.17220. [68] Bohan Zhai, Shijia Yang, Xiangchen Zhao, Chenfeng Xu, Sheng Shen, Dongdi Zhao, Kurt Keutzer, Manling Li, Tan Yan, and Xiangjun Fan. 2023. Halle-switch: Rethinking and controlling object existence hallucinations in large vision language models for detailed caption. ar Xiv preprint ar Xiv:2310.01779. [69] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, et al. 2023. Llava-grounding: Grounded visual chat with large multimodal models. ar Xiv preprint ar Xiv:2312.02949. [70] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. 2024. Ferret-v2: An improved baseline for referring and grounding with large language models. ar Xiv preprint ar Xiv:2404.07973. [71] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. 2024. Groundhog: Grounding large language models to holistic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. [72] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. 2023. Bubogpt: Enabling visual grounding in multi-modal llms. ar Xiv preprint ar Xiv:2307.08581. [73] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Haotong Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Ar Xiv, abs/2306.05685. [74] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. 2017. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633 641. [75] Yiyang Zhou, Chenhang Cui, Jaehong Yoon, Linjun Zhang, Zhun Deng, Chelsea Finn, Mohit Bansal, and Huaxiu Yao. 2024. Analyzing and mitigating object hallucination in large visionlanguage models. In The Twelfth International Conference on Learning Representations. [76] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations. A Additional Experiments, Results, and Discussions A.1 Reproducibility Dataset Total COCO ADE Wild 1539 / 1172 732 / 547 807 / 625 Hom. 312 / 490 168 / 289 144 / 201 Het. 400 / 246 200 / 76 200 / 170 Adv. 168 / 334 54 / 170 114 / 164 Table 3: An overview of object hallucination benchmarks. For design considerations, we summarize the number of tested images, and if multiple classes and object class distribution (at training and test time) are considered. The image sources include those seen or unseen during instruction tuning. To refer to an object, textual descriptions and visual cues can be adopted. For evaluation, neural models, humans and automatic pipelines are used. Data Curation Pipeline. Our data curation pipeline involves several essential steps designed to prepare and refine our dataset for evaluating multi-object hallucination. The pipeline begins by filtering images and candidate objects to query. We consider valid objects to be those belonging to the top 50 thing classes and exclude objects with a bounding box area less than 1% of the total image area. We discard images containing fewer than 5 valid objects, and allow an intersection-over-union between bounding boxes of no more than 0.1, which preserves data integrity while ensuring high image quality. We apply this pipeline to MSCOCOPanoptic [28, 5] and ADE20K [74].3 Language Instruction Prompt Templates. We illustrate the 4 types of task prompts for Single-Object and Multi-Object queries in Figure 7, and document the prompts below. obj1: ?, obj2: ?, obj3: ?, obj4: ?, obj5: ? obj1: ?, obj2: ?, obj3: ?, obj4: ?, obj5: ? (a) Single-object. (c) Student forcing. (d) Teacher forcing. obj1: apple, obj2: obj1: apple, obj2: apple, obj3: apple, obj4: lemon, obj5: obj1: fork, obj2: obj1: fork, obj2: knife, obj3: whisk, obj4: lemon, obj5: obj1: apple, obj2: apple, obj3: obj1: fork, obj2: knife, obj3: Ground truth Ground truth obj1: apple, obj2: apple, obj3: apple, obj4: lemon, obj5: pear obj1: ?, obj2: ?, obj3: ?, obj4: ?, obj5: ? (b) Multi-object. Predicted class Predicted class Figure 7: Different types of instruction settings of ROPE. In a single turn of prompting without format enforcement, we probe the model to recognize the 5 objects referred to by the visual prompts (a) one at a time in the single-object setting and (b) concurrently in the multi-object setting. We further enforce the model to follow the format template and decode only the object tokens for each of the five objects (c) without output manipulation in student forcing and (d) replacing all previously generated object tokens with the ground truth classes in teacher forcing. Multi-Object Default Probing, Student Forcing, and Teacher Forcing: Select one and the most appropriate class for each object located within red bounding boxes from the following list: [CLASS NAMES]. Provide the class names in the format: obj1: , obj2: , obj3: , obj4: , obj5: , with no additional words or punctuations. Multi-Object Probabilistic Probing: (Ground HOG) Describe object 1 and object 2 and object 3 and object 4 and object 5 . obj1: , obj2: , obj3: , obj4: , obj5: (GLa MM) What are the classes of [ LIST]? obj1: , obj2: , obj3: , obj4: , obj5: 3Available at https://huggingface.co/datasets/sled-umich/ROPE Select the single, most appropriate class for obj{obj_num} located within the red bounding box from the following list: car, pedestrian, bicycle, motorcycle, bus, truck, traffic light, stop sign, parking meter, bench, fire hydrant, street sign, crosswalk, traffic cone, construction barrier. Your response should consist solely of the class name that obj{index} belongs to, formatted as only the class name, without any extra characters or punctuation. obj1: pedestrian obj2: car obj3: traffic light obj4: pedestrian obj5: traffic light obj1: pedestrian obj2: car obj3: car obj4: pedestrian obj5: pedestrian obj1: car obj2: car obj3: traffic light obj4: traffic light obj5: traffic light LLa VA-13B obj1: pedestrian obj2: car obj3: traffic light obj4: pedestrian obj5: traffic light obj1: car obj2: car obj3: traffic light obj4: traffic light obj5: traffic light LLa VAPhi3Mini obj1: pedestrian obj2: car obj3: traffic light obj4: pedestrian obj5: traffic light Ground Truth obj1: pedestrian obj2: car obj3: traffic light obj4: pedestrian obj5: traffic light Select one and the most appropriate class for each object located within red bounding boxes from the following list: car, pedestrian, bicycle, motorcycle, bus, truck, traffic light, stop sign, parking meter, bench, fire hydrant, street sign, crosswalk, traffic cone, construction barrier. Provide the class names in the format: 'obj1: , obj2: , obj3: , obj4: , obj5: ', with no additional words or punctuations. obj1: pedestrian obj2: car obj3: traffic light obj4: pedestrian obj5: traffic light obj1: pedestrian obj2: car obj3: car obj4: pedestrian obj5: pedestrian obj1: car obj2: car obj3: car obj4: car obj5: car LLa VA-13B obj1: pedestrian obj2: car obj3: traffic light obj4: pedestrian obj5: pedestrian obj1: car obj2: car obj3: car obj4: car obj5: car LLa VAPhi3Mini obj1: pedestrian obj2: car obj3: traffic light obj4: pedestrian obj5: traffic light Ground Truth obj1: pedestrian obj2: car obj3: traffic light obj4: pedestrian obj5: traffic light Figure 8: Single and multi-object hallucination under the default setting in nu Scenes [4]. (LLa VA) There are five red bounding boxes in this image.For each object within the red bounding boxes, identify its class. Provide the class names in the format: obj1: , obj2: , obj3: , obj4: , obj5: , with no additional words or punctuation. Single Object Default Probing: Select the single, most appropriate class for obj located within the red bounding box from the following list: [CLASS NAMES]. Your response should consist solely of the class name that obj belongs to, formatted as only the class name, without any extra characters or punctuations. Single Object Probabilistic Probing: (Ground HOG) Describe object in a word. (GLa MM) What is the class of ? (LLa VA) Describe the object in the red bounding box labeled obj in a word. Computational Resources. Our experiments were conducted on eight A40 and four A100 GPUs slightly over a week. The computational bottleneck was not the numerical accuracy values but the collection of potential hallucinatory factors for analytical purposes, including logits and attention values for each head and layer. A.2 Additional Experiments and Results Per-object Performance. We provide the per-object performance in the following Table 5 and Table 6. Autonomous Driving Case Study. Figure 8 is a case study example from the nu Scenes dataset [4] for autonomous driving. It illustrates the single and multi-object case, where each object is identified independently. The multi-object case exhibits more hallucination errors compared to the single-object case. This finding underscores the importance of studying multi-object hallucination, especially in real-world scenarios like autonomous driving, where multiple objects need to be detected accurately at the same time. B Limitations, Licenses, and Risks B.1 Limitations ROPE represents one of the pioneering efforts to publicly address the issue of multiple object hallucination. However, we acknowledge several limitations in our work: (1) The lack of transparency in the LVLMs makes it difficult to guarantee that our unseen dataset has not been previously exposed. (2) Our evaluation benchmark uses a fixed set of semantic objects, which may introduce bias and impose unnecessary constraints on the LVLMs ability to follow instructions and reason effectively. (3) The evaluation process can be slow, as it involves performing five inferences per image for both student forcing and teacher forcing. B.2 Artifacts and licenses We report a list of licenses for all datasets and models used in our experiment in Table 4. We strictly follow all the model licenses and limit the scope of these models to academic research only. Data Sources URL License MSCOCO 2017 Link CC BY 4.0 ADE20K Link BSD-3-Clause Software Code URL License LLa VA Link Llama Community Licence Qwen-VL Link Tongyi Qianwen Licence Cog VLM Link Cog VLM Licence IDEFICS Link Llama Community Licence Yi-VL Link Yi Community Licence Mini CPM-V Link Apache License 2.0 GLAMM Link Apache License 2.0 GPT-4V/4O Link Open AI Term of Use Table 4: License information for the scientific artifacts used. B.3 Ethical concerns and risks This study does not require human annotators or participants for its interactive experiments. Instead, it utilizes publicly available datasets and content created by models for evaluation purposes. We are aware that these public data might introduce biases and sensitive elements, and it is essential for future research to address these concerns, possibly by creating datasets that incorporate fairness-based filtering and metrics. Model Multi-Object Student Forcing Teacher Forcing Single-Object Wild Hom. Het. Adv. Wild Hom. Het. Adv. Wild Hom. Het. Adv. Wild Hom. Het. Adv. object 1 Yi-VL-6B 3.17 3.67 1.22 3.29 3.51 3.88 1.63 3.59 3.51 3.88 1.63 3.59 0.09 1.02 0.00 0.30 Yi-VL-34B 10.01 16.12 5.69 10.48 10.35 18.16 6.91 11.38 10.35 18.16 6.91 11.38 0.43 2.24 0.00 0.60 LLa VA-7B 34.99 67.14 13.41 55.39 34.90 67.14 13.41 55.39 34.90 67.14 13.41 55.39 34.99 65.71 15.04 53.89 LLa VA-13B 29.77 62.86 12.60 51.20 29.52 62.86 12.60 50.60 32.34 63.09 12.60 50.60 40.96 72.24 23.17 40.00 LLa VA-34B 26.61 77.55 17.48 62.87 38.58 88.19 14.23 60.97 38.58 88.19 14.23 60.97 48.33 77.55 27.64 66.47 Qwen VL 4.01 2.44 5.92 4.19 9.56 20.41 6.91 16.47 10.49 20.82 5.28 17.96 18.39 32.04 9.76 19.76 Qwen VL-C 22.53 28.16 13.41 24.25 22.61 33.88 13.41 24.25 23.21 32.86 8.94 23.65 34.13 59.80 16.67 47.01 Cog VLM 1.20 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Cog VLM-C 0.00 0.00 0.00 0.00 8.94 1.47 9.76 7.02 10.16 16.50 4.17 10.71 11.83 13.97 8.94 7.78 Cog VLM-G 17.88 27.76 7.72 27.54 23.18 44.90 9.76 33.83 23.18 44.90 9.76 33.83 30.89 50.82 10.98 42.22 Cog VLM-2 24.55 36.33 16.26 31.44 24.89 37.14 18.70 44.61 24.89 37.14 18.70 44.61 10.85 31.43 3.25 27.25 IDEFICS 1.45 1.63 1.22 1.50 11.95 24.29 6.10 14.67 11.95 24.29 6.10 14.67 4.86 3.88 5.28 2.10 Mini CPM-V 33.86 63.21 18.62 55.69 35.59 65.31 15.04 54.79 35.59 65.31 15.04 54.79 25.15 47.76 12.60 37.13 LLa VA-7B* N/A 7.08 11.43 6.50 8.68 N/A 13.45 25.75 13.14 24.40 GLa MM* N/A 53.50 50.20 41.06 50.30 N/A 68.34 77.55 54.07 73.35 Ground HOG* N/A 15.27 20.41 10.98 19.76 N/A 40.10 28.98 31.30 32.34 GPT-4V 49.53 67.35 38.21 56.29 N/A N/A 47.05 64.49 36.59 57.19 GPT-4O 64.42 80.61 56.10 73.05 N/A N/A 63.56 81.22 53.66 73.65 object 2 Yi-VL-6B 2.65 4.08 2.44 3.29 1.88 8.37 4.47 5.69 3.68 7.96 5.69 9.58 0.17 0.61 0.00 0.00 Yi-VL-34B 8.55 17.55 3.25 15.57 10.61 19.59 8.54 12.28 11.21 22.24 10.98 14.37 0.51 2.45 0.41 0.60 LLa VA-7B 29.77 67.96 8.13 56.89 29.52 67.96 8.13 56.89 32.00 84.69 5.28 79.34 34.13 61.02 15.45 50.30 LLa VA-13B 29.17 63.06 4.07 52.40 28.92 63.06 4.07 52.10 37.97 92.37 6.10 85.63 41.64 70.82 21.54 42.58 LLa VA-34B 30.89 79.39 18.29 68.56 27.56 90.55 15.45 65.81 34.40 91.84 2.85 68.25 52.18 78.16 30.08 68.86 Qwen VL 3.84 2.44 8.98 6.59 10.24 23.06 4.07 20.36 13.05 37.35 4.07 33.53 16.94 27.35 9.35 20.66 Qwen VL-C 23.12 35.71 10.57 28.74 17.75 26.12 7.32 19.46 22.61 36.33 11.38 30.24 33.45 53.88 13.82 45.81 Cog VLM 0.00 0.00 0.00 0.30 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Cog VLM-C 0.00 0.00 0.00 0.00 8.12 1.47 4.88 7.02 25.90 93.50 2.08 94.64 8.39 10.29 6.91 6.89 Cog VLM-G 16.34 28.37 4.07 29.64 15.57 39.80 4.88 34.13 21.81 44.90 2.85 39.52 30.72 51.43 14.23 43.71 Cog VLM-2 25.75 34.08 19.51 28.14 24.72 36.53 22.76 45.21 26.43 40.61 26.83 45.21 11.32 32.45 5.69 25.45 IDEFICS 0.17 0.20 0.41 0.60 11.69 27.76 4.07 20.06 31.91 91.43 2.03 85.93 4.10 3.67 6.50 2.10 Mini CPM-V 35.79 57.73 17.00 49.10 31.99 58.37 13.01 53.29 31.48 60.41 13.41 53.29 26.09 49.80 14.23 34.13 LLa VA-7B* N/A 8.28 12.86 4.47 11.08 N/A 10.59 23.00 5.45 21.43 GLa MM* N/A 23.46 54.69 6.50 52.99 N/A 69.62 75.92 49.19 72.46 Ground HOG* N/A 31.31 27.14 32.52 27.25 N/A 41.13 29.59 40.65 33.83 GPT-4V 48.16 67.55 36.18 60.18 N/A N/A 47.48 65.51 32.52 57.78 GPT-4O 63.48 80.20 56.50 73.05 N/A N/A 62.62 80.41 54.07 72.16 object 3 Yi-VL-6B 2.91 3.88 1.22 2.69 3.68 3.88 5.69 5.69 4.02 8.37 17.89 15.87 0.17 0.00 0.41 0.30 Yi-VL-34B 8.30 15.92 6.10 14.67 10.95 17.96 9.35 11.98 11.46 22.86 12.60 15.57 0.43 2.86 0.41 0.30 LLa VA-7B 27.97 68.57 12.60 57.19 27.99 68.57 12.20 57.19 29.95 99.80 11.38 99.70 31.74 63.27 16.26 51.20 LLa VA-13B 25.32 62.86 8.54 52.40 25.17 62.86 8.54 52.10 36.52 100.00 13.01 100.00 43.77 72.65 24.39 45.16 LLa VA-34B 29.97 80.00 21.54 69.16 27.56 92.91 13.01 67.42 41.20 99.32 8.54 94.79 54.32 78.37 34.15 69.46 Qwen VL 2.30 2.03 7.55 7.19 6.48 16.53 4.47 14.37 21.73 91.02 4.47 45.51 18.31 33.47 12.20 23.95 Qwen VL-C 18.09 33.27 8.54 24.25 14.85 28.16 6.50 19.16 21.67 68.37 6.91 60.78 34.47 57.96 15.85 49.10 Cog VLM 0.00 0.00 0.00 0.60 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Cog VLM-C 0.00 0.60 0.00 0.00 7.84 1.47 4.47 7.02 26.02 97.50 0.69 97.62 10.04 6.62 11.79 7.49 Cog VLM-G 12.15 25.31 4.88 25.15 13.34 37.55 5.69 31.14 18.39 56.73 9.35 50.30 30.12 52.04 14.23 43.71 Cog VLM-2 19.16 35.31 16.26 27.84 24.98 38.98 24.80 42.81 27.46 43.47 32.93 51.50 10.50 30.00 6.91 21.56 IDEFICS 0.09 0.00 0.00 0.30 6.23 23.27 1.22 13.77 26.02 99.80 8.94 100.00 5.20 3.67 8.54 2.10 Mini CPM-V 35.12 61.06 13.77 52.10 28.83 55.71 14.63 49.70 29.43 57.14 15.04 47.90 26.52 46.53 14.23 32.93 LLa VA-7B* N/A 8.62 12.65 4.07 10.18 N/A 10.53 23.00 9.62 21.43 GLa MM* N/A 22.53 56.33 7.32 54.19 N/A 67.15 77.96 54.07 73.35 Ground HOG* N/A 24.49 25.31 30.08 26.35 N/A 40.10 30.82 40.24 35.33 GPT-4V 46.96 62.24 34.15 52.69 N/A N/A 47.73 65.51 34.55 56.59 GPT-4O 61.51 80.41 57.32 74.55 N/A N/A 63.05 79.59 51.22 71.56 object 4 Yi-VL-6B 2.48 4.29 0.41 3.29 3.17 2.24 7.72 6.29 4.45 15.10 13.41 15.27 0.09 0.20 0.00 0.30 Yi-VL-34B 6.67 15.31 4.47 11.98 10.01 18.37 7.72 13.17 11.80 24.90 15.45 17.66 0.51 2.04 0.81 0.60 LLa VA-7B 30.37 68.37 7.32 56.89 30.20 68.57 7.32 56.89 29.95 100.00 17.89 100.00 37.64 66.12 15.85 57.19 LLa VA-13B 26.26 63.47 7.32 52.69 26.19 63.47 7.32 52.40 35.41 100.00 18.29 100.00 44.45 72.86 26.42 43.87 LLa VA-34B 30.58 80.20 18.29 70.36 31.50 92.13 13.41 67.10 35.60 99.32 6.91 95.26 50.56 77.14 28.05 70.06 Qwen VL 2.65 2.03 8.78 7.19 5.20 12.45 4.07 11.08 9.04 97.35 3.25 38.32 18.14 30.41 8.94 26.65 Qwen VL-C 15.44 29.18 6.50 23.35 12.54 25.51 6.91 18.86 19.97 72.04 13.41 63.77 32.85 40.41 17.48 49.10 Cog VLM 0.00 0.00 0.00 0.30 0.00 0.00 0.00 0.00 26.60 47.55 29.67 57.49 0.00 0.00 0.00 0.00 Cog VLM-C 0.00 0.00 0.00 0.00 7.98 1.47 4.07 7.02 28.00 98.50 2.08 98.81 11.14 10.29 11.38 7.78 Cog VLM-G 15.31 25.92 5.69 26.35 17.19 42.86 5.28 36.23 22.75 60.20 0.81 60.48 30.29 55.10 12.60 43.41 Cog VLM-2 18.05 34.69 16.26 29.64 25.15 38.57 28.86 41.62 27.63 49.18 20.73 57.49 10.97 31.84 6.10 26.95 IDEFICS 0.09 0.00 0.00 0.30 8.70 23.47 1.22 13.47 26.88 99.80 8.13 100.00 4.95 3.67 8.13 2.10 Mini CPM-V 30.26 60.86 20.24 52.40 31.39 58.78 13.82 53.89 31.39 61.63 19.92 55.39 25.15 48.57 15.85 38.92 LLa VA-7B* N/A 6.66 11.63 2.44 9.88 N/A 13.06 23.00 10.90 21.43 GLa MM* N/A 22.27 56.12 8.54 54.19 N/A 68.17 76.33 52.85 74.85 Ground HOG* N/A 19.54 25.31 25.61 25.75 N/A 39.42 32.45 36.18 36.83 GPT-4V 45.42 64.08 31.71 50.00 N/A N/A 47.82 66.53 35.77 53.59 GPT-4O 63.82 79.80 57.32 74.85 N/A N/A 64.16 79.59 53.66 70.96 object 5 Yi-VL-6B 2.48 3.47 0.41 2.10 3.68 2.86 6.50 5.39 4.53 19.18 14.23 17.66 0.17 0.41 0.00 0.00 Yi-VL-34B 5.30 13.27 1.63 3.29 9.50 16.12 7.32 9.88 11.38 24.29 14.23 18.26 0.43 2.24 0.41 0.30 LLa VA-7B 30.11 68.57 10.16 13.17 30.12 68.57 10.16 13.17 32.68 100.00 18.29 0.00 36.36 65.92 18.29 23.35 LLa VA-13B 27.63 63.27 9.35 10.48 27.22 63.27 9.35 10.48 36.01 100.00 23.98 0.00 42.49 71.02 21.54 21.29 LLa VA-34B 28.44 80.00 13.01 8.38 25.20 91.34 15.85 18.71 37.20 99.32 3.66 5.21 53.81 78.16 34.15 34.73 Qwen VL 1.19 0.81 4.08 1.50 4.35 9.59 1.22 1.50 7.68 47.35 3.25 0.00 17.11 32.86 7.32 11.98 Qwen VL-C 15.10 27.35 4.88 8.38 13.05 25.31 4.47 5.09 21.59 80.41 15.04 6.29 36.09 59.80 13.01 19.16 Cog VLM 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Cog VLM-C 0.00 0.00 0.00 0.00 8.12 1.47 5.69 7.44 29.00 100.00 0.00 0.60 10.18 12.50 6.50 9.58 Cog VLM-G 16.34 25.51 5.28 6.59 16.60 41.22 4.47 5.69 27.89 73.47 10.57 3.89 30.80 53.06 15.45 17.66 Cog VLM-2 17.71 34.90 11.38 12.87 23.44 38.98 20.73 12.28 27.29 59.80 33.74 12.57 12.02 28.98 6.91 5.09 IDEFICS 0.17 0.00 0.00 0.00 6.57 23.47 0.81 1.80 27.22 99.80 13.01 0.00 4.01 3.47 4.07 4.79 Mini CPM-V 29.76 56.75 13.36 19.16 31.05 56.73 14.23 14.07 31.48 60.41 18.29 17.66 25.24 46.12 15.04 17.96 LLa VA-7B* N/A 7.34 12.04 6.91 4.79 N/A 10.79 23.00 7.69 10.12 GLa MM* N/A 23.81 55.31 7.72 6.29 N/A 69.97 77.55 51.22 53.29 Ground HOG* N/A 24.91 25.31 32.11 18.56 N/A 42.92 30.00 42.28 30.54 GPT-4V 38.41 56.94 31.30 27.54 N/A N/A 46.62 62.65 37.80 31.44 GPT-4O 63.74 80.41 54.07 53.59 N/A N/A 64.67 78.37 56.10 54.49 Table 5: Complete per-object results on the unseen split. Model Multi-Object Student Forcing Teacher Forcing Single-Object Wild Hom. Het. Adv. Wild Hom. Het. Adv. Wild Hom. Het. Adv. Wild Hom. Het. Adv. object 1 Yi-VL-6B 2.92 5.50 1.92 4.76 3.12 5.75 2.56 6.55 3.12 20.75 2.56 6.55 0.06 0.25 0.00 0.60 Yi-VL-34B 8.51 15.75 4.81 13.10 8.71 16.00 5.13 11.90 8.71 16.00 5.13 11.90 0.19 2.50 0.00 0.60 LLa VA-7B 31.51 65.75 5.00 51.79 31.45 65.75 10.90 51.79 31.45 65.75 10.90 51.79 33.01 62.75 17.95 47.62 LLa VA-13B 28.14 65.83 10.91 54.17 28.14 72.00 7.69 54.76 28.14 72.00 7.69 54.76 41.46 81.25 24.04 65.48 LLa VA-34B 34.18 83.50 8.65 67.26 48.02 83.75 29.17 72.02 48.02 83.75 29.17 72.02 54.13 85.00 17.20 77.98 Qwen VL 2.21 6.00 1.60 4.17 7.41 14.25 1.28 10.71 5.46 15.25 2.24 9.52 8.06 16.50 4.49 17.86 Qwen VL-C 5.65 14.25 2.24 10.71 5.59 9.25 2.24 8.33 6.04 8.75 2.24 7.74 25.02 43.50 14.10 30.36 Cog VLM 0.00 0.00 0.00 0.60 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Cog VLM-C 0.00 0.00 0.00 0.00 11.18 13.50 4.17 16.67 3.44 3.00 1.92 4.76 12.46 24.75 5.13 25.60 Cog VLM-G 11.70 18.25 6.09 19.64 24.44 41.67 10.30 29.19 24.61 49.75 10.90 29.87 28.85 55.25 14.74 40.48 Cog VLM-2 18.97 37.00 13.78 32.14 35.48 69.50 10.90 33.33 35.48 69.50 10.90 33.33 18.78 39.25 9.62 30.36 IDEFICS 0.00 3.50 0.64 1.20 7.41 20.25 0.96 14.29 7.41 20.25 0.96 14.29 4.86 0.00 0.00 0.00 Mini CPM-V 32.81 66.25 15.38 58.33 31.51 72.00 13.14 63.69 31.45 72.25 13.14 59.52 25.87 54.00 15.38 58.93 LLa VA-7B* N/A 9.62 16.25 6.41 11.31 N/A 13.45 25.75 13.14 24.40 GLa MM* N/A 45.61 44.50 40.38 42.26 N/A 64.33 81.25 55.77 73.81 Ground HOG* N/A 16.11 24.25 16.03 22.62 N/A 43.86 43.50 45.19 51.79 GPT-4V 56.79 79.75 41.35 71.43 N/A N/A 55.30 76.25 41.35 71.43 GPT-4O 69.98 89.50 66.03 79.76 N/A N/A 61.35 73.27 54.47 69.46 object 2 Yi-VL-6B 3.83 5.75 3.21 6.55 3.31 6.25 3.53 7.74 6.17 10.25 4.81 6.55 0.19 0.25 0.00 1.19 Yi-VL-34B 9.10 15.50 3.21 13.10 8.90 16.75 5.77 10.12 10.59 21.00 5.77 23.81 0.26 2.50 0.32 1.19 LLa VA-7B 32.36 67.50 15.00 52.98 32.36 67.25 12.50 52.98 35.15 95.00 12.18 91.07 36.32 61.75 20.19 47.62 LLa VA-13B 32.42 67.67 13.48 56.55 32.36 72.75 14.42 56.55 37.04 99.25 11.54 97.02 43.66 80.50 24.36 72.02 LLa VA-34B 41.26 85.50 25.32 70.24 53.15 85.75 31.41 73.21 57.05 98.10 21.84 94.30 54.52 85.50 18.10 77.38 Qwen VL 3.38 6.00 0.96 3.57 5.78 15.75 4.49 12.50 19.36 57.50 5.13 42.86 8.45 14.75 4.81 12.50 Qwen VL-C 9.03 16.50 9.62 11.31 5.52 8.75 6.73 6.55 16.89 45.25 9.29 44.05 26.19 40.00 15.71 28.57 Cog VLM 0.06 0.00 0.00 0.60 0.00 0.00 0.00 0.00 0.26 0.00 0.00 0.00 0.00 0.00 0.00 0.00 Cog VLM-C 0.00 0.00 0.00 0.00 12.12 13.50 8.65 16.67 29.43 94.00 0.00 94.64 12.38 22.50 7.69 18.45 Cog VLM-G 12.15 21.75 6.09 22.02 25.79 42.83 7.12 30.87 27.40 55.75 16.67 33.56 31.06 55.00 18.27 42.86 Cog VLM-2 24.95 38.00 21.79 38.10 38.60 72.00 12.50 32.14 36.13 70.25 14.10 31.55 21.90 40.00 14.42 29.76 IDEFICS 0.00 1.00 0.00 0.60 5.78 17.75 0.00 8.33 22.16 77.75 12.18 63.10 4.10 0.00 0.96 0.00 Mini CPM-V 36.19 64.75 19.23 60.71 33.07 62.75 16.67 57.14 33.79 69.00 17.63 61.31 30.06 60.75 17.63 60.12 LLa VA-7B* N/A 9.10 17.75 6.41 13.10 N/A 10.59 23.00 5.45 21.43 GLa MM* N/A 22.22 55.75 6.73 56.55 N/A 64.20 80.25 56.41 72.62 Ground HOG* N/A 31.32 33.50 34.29 30.36 N/A 43.60 44.25 41.03 51.19 GPT-4V 55.69 78.75 41.35 73.81 N/A N/A 56.27 78.50 37.50 72.62 GPT-4O 71.28 89.50 64.10 82.14 N/A N/A 60.24 74.90 54.88 68.56 object 3 Yi-VL-6B 2.99 6.25 2.56 5.95 3.12 7.00 4.17 8.93 5.26 24.75 6.09 8.33 0.13 0.25 0.32 1.19 Yi-VL-34B 9.10 16.00 1.60 14.88 9.10 16.75 2.24 10.71 11.23 21.75 4.17 36.31 0.19 2.50 0.32 1.19 LLa VA-7B 29.76 68.00 5.00 52.98 29.76 67.50 12.18 52.98 28.40 100.00 11.86 100.00 33.92 61.50 14.42 48.81 LLa VA-13B 32.75 68.00 13.18 58.33 32.68 73.50 10.90 57.74 38.27 100.00 17.63 100.00 44.90 82.00 20.19 67.86 LLa VA-34B 44.70 86.50 24.68 70.24 54.00 86.25 36.86 73.81 57.89 98.80 27.63 98.32 53.41 87.50 16.32 77.98 Qwen VL 2.86 7.50 0.96 7.14 6.30 18.00 4.49 16.67 21.25 92.50 5.13 88.10 8.58 16.25 5.77 17.26 Qwen VL-C 10.66 21.25 7.05 21.43 6.04 9.00 5.13 9.52 16.89 60.75 9.29 61.31 26.19 46.25 10.90 32.74 Cog VLM 0.06 0.00 0.00 0.60 0.00 0.00 0.00 0.00 0.13 0.75 0.00 1.19 0.00 0.00 0.00 0.00 Cog VLM-C 0.00 0.00 0.00 0.00 10.84 13.50 5.77 16.67 27.03 94.00 0.00 94.64 11.46 21.25 7.05 16.67 Cog VLM-G 12.74 25.75 7.05 24.40 26.89 45.00 14.39 34.23 27.23 81.25 19.23 57.72 29.63 55.50 16.99 42.26 Cog VLM-2 20.27 36.50 15.71 36.31 39.12 72.75 13.14 32.14 36.45 74.75 17.31 49.40 19.36 34.75 12.18 31.55 IDEFICS 0.00 1.00 0.00 0.90 6.30 18.75 0.64 11.90 15.01 92.50 8.65 86.90 5.20 0.00 0.32 0.00 Mini CPM-V 36.19 68.50 20.51 61.31 31.38 60.00 12.50 52.38 30.60 68.25 15.71 54.76 26.15 53.75 19.87 95.05 LLa VA-7B* N/A 8.77 16.50 5.45 11.90 N/A 10.53 23.00 9.62 21.43 GLa MM* N/A 22.68 55.25 5.13 54.17 N/A 62.70 83.00 47.44 77.38 Ground HOG* N/A 24.82 32.75 24.68 30.36 N/A 45.09 43.50 35.26 50.00 GPT-4V 54.45 79.25 40.06 70.24 N/A N/A 55.62 80.25 39.42 72.02 GPT-4O 71.35 90.75 65.71 83.93 N/A N/A 60.07 74.29 53.66 68.26 object 4 Yi-VL-6B 3.25 5.50 1.60 4.76 4.35 7.75 5.13 10.71 6.30 37.75 4.17 8.93 0.26 0.25 0.00 1.19 Yi-VL-34B 8.32 15.75 4.49 13.69 9.16 17.00 5.13 8.93 11.05 22.75 5.45 34.52 0.26 3.00 0.00 1.19 LLa VA-7B 32.42 68.75 10.00 59.52 32.42 68.50 10.90 54.17 30.28 100.00 15.06 100.00 36.84 63.00 17.31 49.40 LLa VA-13B 31.38 68.17 14.39 94.05 31.32 73.75 11.54 58.93 36.06 100.00 24.68 100.00 44.83 80.00 25.96 64.88 LLa VA-34B 40.68 86.75 19.23 73.21 54.06 85.25 34.62 73.81 57.96 99.10 25.26 97.65 56.14 86.75 22.12 80.95 Qwen VL 2.47 7.25 0.00 4.76 5.91 14.75 2.88 12.50 22.87 95.50 7.69 95.24 9.75 14.25 6.41 16.07 Qwen VL-C 8.58 16.25 7.05 16.07 5.46 8.00 2.88 7.74 11.96 56.50 10.26 54.17 24.82 38.50 11.22 28.57 Cog VLM 0.00 0.00 0.00 0.60 0.00 0.00 0.00 0.00 0.06 1.00 0.00 0.60 0.00 0.00 0.00 0.00 Cog VLM-C 0.00 0.00 0.00 0.00 10.24 13.50 9.94 16.67 26.25 93.75 0.32 93.45 9.83 21.25 9.29 20.83 Cog VLM-G 14.10 24.75 9.62 27.98 26.09 45.33 15.30 34.23 30.91 86.00 23.40 62.08 29.89 57.75 16.99 44.05 Cog VLM-2 22.94 37.75 19.87 36.31 36.58 66.75 14.74 33.93 38.40 75.25 20.51 61.31 24.24 37.50 15.71 26.79 IDEFICS 0.00 1.25 0.00 0.60 5.91 18.25 0.64 11.90 21.12 95.00 14.74 93.45 4.95 23.00 0.00 0.00 Mini CPM-V 37.36 64.75 18.59 60.71 32.42 59.75 13.14 56.55 34.05 67.75 19.55 55.36 29.14 56.00 16.35 62.50 LLa VA-7B* N/A 8.90 15.75 5.45 11.31 N/A 13.06 23.00 10.90 21.43 GLa MM* N/A 21.70 55.75 6.73 55.95 N/A 64.91 80.25 55.45 67.86 Ground HOG* N/A 20.86 31.50 21.15 30.95 N/A 43.99 42.00 35.58 50.00 GPT-4V 53.09 77.25 43.59 72.62 N/A N/A 56.40 78.75 44.87 74.40 GPT-4O 72.25 87.00 68.27 82.14 N/A N/A 60.67 72.65 55.69 70.06 object 5 Yi-VL-6B 1.75 5.25 0.64 1.19 3.31 7.25 3.53 8.93 6.37 37.75 4.17 0.60 0.32 0.50 0.32 0.60 Yi-VL-34B 7.47 13.75 2.56 3.57 8.97 15.00 2.88 5.36 8.86 17.25 4.17 4.17 0.19 2.50 0.00 0.60 LLa VA-7B 30.41 67.50 5.00 7.74 30.41 67.25 9.62 8.93 32.16 100.00 11.86 0.00 36.52 62.75 16.99 20.83 LLa VA-13B 33.01 68.50 11.21 21.43 32.94 74.25 13.14 7.74 35.35 100.00 18.59 0.00 40.81 79.25 25.00 23.81 LLa VA-34B 38.92 86.50 16.35 9.52 54.52 85.00 37.50 32.14 61.13 99.30 22.63 8.05 57.05 87.75 21.13 32.14 Qwen VL 2.73 6.25 1.60 1.79 5.85 17.25 5.13 3.57 24.76 96.75 7.05 0.00 8.84 18.50 6.41 4.17 Qwen VL-C 9.68 16.25 7.37 4.76 3.70 8.00 3.53 1.79 8.77 67.50 9.29 0.60 27.75 48.75 14.10 15.48 Cog VLM 0.06 0.00 0.00 0.60 0.00 0.00 0.00 0.00 0.06 3.00 1.00 0.00 0.00 0.00 0.00 0.00 Cog VLM-C 0.00 0.00 0.00 0.00 9.98 13.50 5.45 8.93 27.03 92.50 0.00 0.60 10.14 21.75 6.41 9.52 Cog VLM-G 13.78 23.25 7.05 10.12 23.64 43.33 8.94 19.13 31.12 91.25 17.31 8.72 31.38 56.50 14.74 13.10 Cog VLM-2 20.40 38.50 15.38 23.21 35.35 73.25 12.18 32.74 39.05 77.75 24.36 19.64 22.09 42.25 16.35 14.88 IDEFICS 0.00 0.50 0.00 0.60 5.85 18.50 0.96 1.79 21.18 95.25 13.78 0.60 4.01 0.00 0.32 0.00 Mini CPM-V 31.19 60.00 13.14 17.26 29.69 59.50 12.82 16.07 30.93 63.00 17.95 16.67 25.87 52.25 15.38 22.02 LLa VA-7B* N/A 9.42 15.75 3.85 4.76 N/A 10.79 23.00 7.69 10.12 GLa MM* N/A 23.33 55.50 6.09 4.76 N/A 62.90 84.00 51.92 46.43 Ground HOG* N/A 24.76 32.00 25.00 26.19 N/A 43.86 42.25 37.82 42.26 GPT-4V 48.99 72.75 37.82 43.45 N/A N/A 55.88 77.50 41.99 39.88 GPT-4O 71.47 89.50 66.35 64.29 N/A N/A 61.52 74.49 52.85 53.89 Table 6: Complete per-object results on the seen split. Neur IPS Paper Checklist Question: Do the main claims made in the abstract and introduction accurately reflect the paper s contributions and scope? Answer: [Yes] Justification: Please see Abstract and Section 1. Guidelines: The answer NA means that the abstract and introduction do not include the claims made in the paper. The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A No or NA answer to this question will not be perceived well by the reviewers. The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. 2. Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: Please see Section B.1. Guidelines: The answer NA means that the paper has no limitation while the answer No means that the paper has limitations, but those are not discussed in the paper. The authors are encouraged to create a separate "Limitations" section in their paper. The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. 3. Theory Assumptions and Proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [NA] Justification: There is no theoretical results in this paper. Guidelines: The answer NA means that the paper does not include theoretical results. All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. All assumptions should be clearly stated or referenced in the statement of any theorems. The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. Theorems and Lemmas that the proof relies upon should be properly referenced. 4. Experimental Result Reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: Please see Section 3 and Appendix A.1 Guidelines: The answer NA means that the paper does not include experiments. If the paper includes experiments, a No answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. While Neur IPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We will release both the data and code to the public soon. For code, we plan to release both the evaluation template and the code that we have used for data curation. Guidelines: The answer NA means that paper does not include experiments requiring code. Please see the Neur IPS code and data submission guidelines (https://nips.cc/public/ guides/Code Submission Policy) for more details. While we encourage the release of code and data, we understand that this might not be possible, so No is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). The instructions should contain the exact command and environment needed to run to reproduce the results. See the Neur IPS code and data submission guidelines (https://nips.cc/public/ guides/Code Submission Policy) for more details. The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. 6. Experimental Setting/Details Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: Please see Section 3 and Section 4 Guidelines: The answer NA means that the paper does not include experiments. The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. The full details can be provided either with the code, in appendix, or as supplemental material. 7. Experiment Statistical Significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] Justification: Please see Section 4 and Section 5 Guidelines: The answer NA means that the paper does not include experiments. The authors should answer "Yes" if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) The assumptions made should be given (e.g., Normally distributed errors). It should be clear whether the error bar is the standard deviation or the standard error of the mean. It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified. For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g. negative error rates). If error bars are reported in tables or plots, The authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. 8. Experiments Compute Resources Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: Please see Appendix A.1. Guidelines: The answer NA means that the paper does not include experiments. The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn t make it into the paper). 9. Code Of Ethics Question: Does the research conducted in the paper conform, in every respect, with the Neur IPS Code of Ethics https://neurips.cc/public/Ethics Guidelines? Answer: [Yes] Justification: Please see Appendix B.3. Guidelines: The answer NA means that the authors have not reviewed the Neur IPS Code of Ethics. If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics. The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). 10. Broader Impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [NA] Justification: There is no societal impact of the work performed, either positive or negative. Guidelines: The answer NA means that there is no societal impact of the work performed. If the authors answer NA or No, they should explain why their work has no societal impact or why the paper does not address societal impact. Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). 11. Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: This paper does not pose such risk. Guidelines: The answer NA means that the paper poses no such risks. Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. 12. Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: We have included licenses for existing assets in the Appendix ??. Guidelines: The answer NA means that the paper does not use existing assets. The authors should cite the original paper that produced the code package or dataset. The authors should state which version of the asset is used and, if possible, include a URL. The name of the license (e.g., CC-BY 4.0) should be included for each asset. For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. If this information is not available online, the authors are encouraged to reach out to the asset s creators. 13. New Assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: We have included data curation pipeline in the subsection of Appendix A.1: Data Curation Pipeline and we will release the dataset to the public soon. Guidelines: The answer NA means that the paper does not release new assets. Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. The paper should discuss whether and how consent was obtained from people whose asset is used. At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. 14. Crowdsourcing and Research with Human Subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [NA] Justification: This paper does not involve crowdsourcing nor research with human subjects. Guidelines: The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. According to the Neur IPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. 15. Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? Answer: [NA] Justification: This paper does not involve crowdsourcing nor research with human subjects. Guidelines: The answer NA means that the paper does not involve crowdsourcing nor research with human subjects. Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the Neur IPS Code of Ethics and the guidelines for their institution. For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.