# **Retrieval Meets Reasoning: Even High-school Textbook Knowledge Benefits Multimodal Reasoning** 

**Cheng Tan**[1] _[∗]_ **Jingxuan Wei**[2] _[∗]_ **Linzhuang Sun**[2] **Zhangyang Gao**[1] **Siyuan Li**[1] **Bihui Yu**[2] **Ruifeng Guo**[2] **Stan Z. Li**[1] _[†]_ 1Westlake University 2Shenyang Institute of Computing Technology, Chinese Academy of Sciences 

## **Abstract** 

Large language models equipped with retrieval-augmented generation (RAG) represent a burgeoning field aimed at enhancing answering capabilities by leveraging external knowledge bases. Although the application of RAG with language-only models has been extensively explored, its adaptation into multimodal vision-language models remains nascent. Going beyond mere answer generation, the primary goal of multimodal RAG is to cultivate the models’ ability to reason in response to relevant queries. To this end, we introduce a novel multimodal RAG framework named RMR ( **R** etrieval **M** eets **R** easoning). The RMR framework employs a bi-modal retrieval module to identify the most relevant question-answer pairs, which then serve as scaffolds for the multimodal reasoning process. This training-free approach not only encourages the model to engage deeply with the reasoning processes inherent in the retrieved content but also facilitates the generation of answers that are precise and richly interpretable. Surprisingly, utilizing solely the ScienceQA dataset, collected from elementary and high school science curricula, RMR significantly boosts the performance of various vision-language models across a spectrum of benchmark datasets, including A-OKVQA, MMBench, and SEED. These outcomes highlight the substantial potential of our multimodal retrieval and reasoning mechanism to improve the reasoning capabilities of vision-language models. 

## **1 Introduction** 

While deep learning and its applications have been widely explored in recent years [64, 23, 44, 41, 61, 43, 60, 42, 62], retrieval-augmented generation (RAG) has rapidly emerged as a cornerstone in the development of large language models (LLMs), enabling them to enhance their capabilities by leveraging external knowledge bases [77, 30, 11]. Integrating LLMs with RAG has found its most impactful application within language-centric models, where the dynamic interplay between retrieved content and answer generation significantly elevates the quality and relevance of responses [10, 86, 10]. While early works have demonstrated that incorporating directly retrieved information into language models can improve the quality of the generated content [29], subsequent developments have involved refinement and mitigate the potential noise associated with the raw retrieval results [71, 78, 2], thus ensuring that the content generated is not only accurate but also contextually enriched. 

Although the integration of RAG with large language models has been extensively explored, its adaptation to multimodal scenarios that encompass both visual and textual inputs remains relatively nascent [12, 76, 74]. Notably, in the domain of visual question answering (VQA), where queries comprise concise textual prompts paired with complex visual data, the requirements for integration diverge significantly from traditional text-centric approaches [74, 17]. Traditional RAG systems, 

> _∗_ Equal contribution. 

> _†_ Corresponding author. 

Preprint. Under review. 

initially designed for text-heavy applications, encounter substantial challenges when applied directly to multimodal tasks. In these contexts, textual data often provides insufficient contextual cues, failing to bridge the interpretative demands of rich visual information. This fundamental limitation is critical: models trained predominantly on textual data struggle to effectively capture the nuanced complexity of visual information, leading to significant gaps in the model’s ability to accurately interpret and reason about visual content. For instance, as illustrated in Figure 1, even when the three most pertinent pieces of information are retrieved in response to a query, the model may still fail to engage with the underlying reasoning processes embedded within the retrieved content. Instead of synthesizing insights from these inputs, the model may default to merely replicating answers, which can result in inaccuracies. This highlights a critical shortfall in current multimodal RAG systems, i.e., their inability to fully leverage the cognitive reasoning demanded by complex multimodal data. 

**==> picture [389 x 147] intentionally omitted <==**

**----- Start of picture text -----**<br>
Question  𝑋!"#$% Retrieval library<br>Which i in column 3?(A) the school Retrieved  𝑋! Retrieved  𝑋# Retrieved  𝑋"<br>(B) the park Which i in row A? Which i in column 1? Which i in row B?<br>(C) the pond (A) the pond (A) the police department (A) the theater<br>(D) the gas station (B) the school (B) the pond (B) the fire department<br>(C) the fire department (C) the school (C) the grocery store<br>(D) the gas station (D) the fire department (D) the fast-food restaurant<br>Result<br>The answer is (B). The answer is (B). The answer is (B).<br>The answer is (B).<br>**----- End of picture text -----**<br>


Figure 1: limitations of multimodal retrieval enhancement with simple question-answer pairs. 

Building on the foundational understanding that RAG significantly enhances the capabilities of large language models, we hypothesize that the ultimate purpose of multimodal RAG extends beyond merely instructing models to generate direct answers. Instead, our goal is to equip models with the ability to engage in cognitive reasoning, akin to human thought processes when confronted with complex, context-rich questions. This perspective underscores the necessity for multimodal RAG to be inherently flexible and open-ended, facilitating deep contemplation and robust reasoning whether the tasks are unimodal or multimodal, and independent of the data modality being retrieved. 

To realize this vision, we develop a comprehensive multimodal RAG framework, named RMR ( **R** etrieval **M** eets **R** easoning), which seamlessly integrates multimodal retrieval capabilities with in-context learning (ICL). This framework begins by employing a bi-modality retrieval module to fetch the most pertinent question-answer pairs, which may be unimodal or multimodal. It then integrates these elements into the model’s reasoning process, guiding it through the provided rationales associated with each retrieved item. Following this retrieval phase, the model autonomously learns coherent rationales that reflect a deep and meaningful engagement with the given problem. 

Remarkably, even when retrieval is limited to the ScienceQA dataset [49], which covers only elementary and high school science curricula, our Retrieval Meets Reasoning (RMR) framework demonstrates substantial enhancements across a variety of open-source multimodal models. Notably, the LLaVA model registers a +7.66% improvement, Qwen-VL achieves a +9.93% increase, InternLMXComposer2-VL records a +5.33% improvement, Gemini achieves a +33.67% increase. Furthermore, when evaluated against diverse datasets such as A-OKVQA [58], MMBench [47], and SEEDBench [37], RMR consistently delivers significant performance gains by leveraging the specialized knowledge embedded in the high-school curriculum from the ScienceQA dataset. 

The main contributions of this work are as summarized as follows: 

- We introduce a comprehensive multimodal RAG framework, Retrieval Meets Reasoning (RMR), designed to enhance the reasoning capabilities of multimodal models, enabling them to generate answers through cognitive processes rather than merely replicating responses. 

- We develop a bi-modality retrieval module that effectively bridges the gap between unimodal and multimodal data, ensuring robust and accurate retrieval outcomes. 

- Despite its training-free manner, RMR demonstrates significant effectiveness across a variety of multimodal models and datasets, showcasing its capability to improve multimodal reasoning tasks. 

2 

## **2 Related Work** 

**RAG in LLMs** Despite recent advancements in deep learning and large language or vision models [70, 63, 9], retrieval-augmented generation (RAG) improves language model capabilities by integrating external knowledge into the generation process [38, 22, 8, 14]. The development of RAG was initiated with the introduction of dense retrievers, which radically transformed how responses are generated by utilizing externally sourced information [36, 77, 87]. Initial efforts primarily focused on refining the interaction between the retriever and the generator [36, 24], leading to the production of contextually enriched responses. As a reliable enhancement, RAG has been foundational in allowing Large Language Models (LLMs) to exploit the vast reservoir of knowledge they encompass for various applications, including question answering, dialogue systems, and summarization [25, 71, 5]. 

Recent advancements in RAG have been directed at addressing specific challenges, such as reducing hallucination phenomena and integrating outdated and obscure long-tail knowledge. An active retrieval mechanism that adaptively selects the most relevant knowledge pieces is employed to provide up-to-date information for the generation process [30]. Building on this, ActiveRAG [73] is proposed to incorporate an active learning mechanism that not only retrieves pertinent information but also synthesizes it with existing knowledge, markedly enhancing the model’s ability to handle knowledge-intensive tasks by dynamically integrating information. Although the development in RAG has significantly enhanced the synergy between retrieval and generation, the focus has primarily been on text-based applications, with limited exploration into multimodal scenarios. 

**RAG for ICL** In-context learning (ICL) has revolutionized the functionality of LLMs, enabling them to adapt to new tasks by leveraging a few contextual examples provided directly within their input. This shift towards using retrieved demonstrations to facilitate ICL has increased the flexibility of LLMs across various applications [19, 72, 85]. The technique of demonstration retrieval, which involves selecting few-shot examples specifically tailored to the query, not only boosts task performance but also helps to mitigate biases that arise from manual or random selection of demonstrations. A key development has been the optimization of retrieval objectives, which ensures that the demonstrations are both pertinent to the query and diverse enough to offer a comprehensive context [72]. 

Expanding ICL into multimodal tasks represents a significant advancement, tackling the complex challenge of integrating textual and visual data. The extension of ICL into multimodal tasks represents a significant leap forward, addressing the inherent complexity of integrating textual and non-textual data. MM-Retrieval [45] is a concurrent work that introduces a retrieval-augmented multi-modal CoT reasoning approach, which dynamically selects demonstration examples by leveraging cross-modal and intra-modal similarities. However, it operates by employing modality-specific retrievers to gather demonstrations, which are then directly structured into a chain-of-thought format. 

## **3 Retrieval Meets Reasoning** 

## **3.1 Preliminaries** 

In vision-language tasks, the objective is to develop a mapping _F_ Θ : _X →Y_ , where _X_ represents multimodal inputs that include both textual and visual elements, and _Y_ denotes the corresponding outputs. Formally, given a dataset _D_ = _{X , Y}_ , each input _X ∈X_ can be decomposed into _X_ = ( _T, I_ ), where _T_ denotes the text component and _I_ denotes the image component. In certain cases, one of them may be absent, i.e., _T_ = _∅_ or _I_ = _∅_ , resulting in modality-incomplete inputs. 

The primary goal is to accurately predict the output _Y ∈Y_ for a given input _X_ . This task can be expressed as identifying the answer _Y_ that maximizes the conditional probability _p_ ( _Y | T, I_ ) given the text component _T_ and the image component _I_ : 

**==> picture [263 x 16] intentionally omitted <==**

Here, _p_ ( _Y[′] | T, I_ ) denotes the probability of a candidate answer _Y[′]_ given the inputs _T_ and _I_ . For the supervised learning setting, the model optimal parameters Θ _[∗]_ are those that minimize the loss function _L_ , which quantifies the discrepancy between the predicted answer _F_ Θ( _X_ ) and the ground truth answer _Y_ . This optimization is: Θ _[∗]_ = arg minΘ _L_ ( _F_ Θ( _X_ ) _, Y_ ). However, in our approach, we employ an in-context learning strategy based on pre-trained large language models (LLMs). This 

3 

method uses the retrieved content as contextual information without re-training the model, enabling it to obtain strong reasoning ability effectively. 

## **3.2 Bi-modality Retrieval Module** 

To unify the retrieval module in multimodal models for both text and visual modality, we propose a bi-modality retrieval module based on the Contrastive Language-Image Pre-training (CLIP) [54] framework, as shown in Figure 2. This module is designed to handle various cases where the input may consist of complete image-text pairs, image-only inputs, or text-only inputs. The core idea is to create a robust embedding representation that can effectively capture the relevant information across different modalities and retrieve the most pertinent examples to guide the reasoning process. 

**==> picture [397 x 186] intentionally omitted <==**

**----- Start of picture text -----**<br>
embedding library<br>Question : Birds are<br>Image  retrieve retrieve warm-blooded.blooded … Warm-<br>image Encoder … Rationale : Birds, mammals, fish,<br>reptiles, and amphibians are groups of<br>animals...<br>Answer : The answer is albatross.<br>Which property do<br>these two objects have<br>in common? Select the<br>better answer. Question : Which ocean<br>is highlighted?<br>image-text pair<br>Rationale : Oceans are huge bodies of<br>What is the source of the Text  salt water. The world has five oceans. All<br>allusionbelow? Myrain thethinkssentenceMr. Encoder ofworld ocean. This is the Atlantic Ocean…the oceans are connected, making one<br>Harper is a Luddite ScienceQA-based<br>because he doesn't own a Knowledge Library Answer : The answer is the Atlantic Ocean.<br>cell phone.<br>text …<br>**----- End of picture text -----**<br>


Figure 2: The overall architecture and the retrieval mechanism of the bi-modality retrieval module. 

**Embedding Representation** The retrieval module begins by computing embeddings for the inputs using the CLIP model, which is adept at handling both textual and visual data. The embedding strategy is adaptive, accommodating the varying availability of modalities within the input. For imagetext pairs ( _T, I_ ), we calculate the mean pooling of CLIP’s text embedding _**h** T_ = CLIP _T_ ( _T_ ) _∈_ R _[d]_ and image embedding _**h** I_ = CLIP _I_ ( _I_ ) _∈_ R _[d]_ , where _d_ denotes the embedding dimension. Here, CLIP _T_ ( _·_ ) and CLIP _I_ ( _·_ ) denote the text and image encoders of CLIP, respectively. The combined embedding _**h** X_ is then computed as the average of these text and image embeddings. For imageonly inputs, we use the CLIP image embedding _**h** I_ , and for text-only inputs, we use the CLIP text embedding _**h** T_ . Thus, the item embedding _**h** X ∈_ R _[d]_ for any input _X_ is defined as: 

**==> picture [277 x 39] intentionally omitted <==**

This approach ensures robust embeddings regardless of the presence or absence of textual and visual components, thereby providing a flexible and consistent representation of multimodal inputs. 

**High-school Knowledge Library** We construct a comprehensive knowledge embedding library _H_ = _{_ _**h**[i] X[}][N] i_ =1[using][the][ScienceQA][dataset,][which][is][derived][from][elementary][and][high][school] textbooks. This dataset is particularly valuable because it includes detailed rationales for each answer, forming question-rationale-answer triplets ( _Qi, Ri, Ai_ ) for each sample. The structured nature of this dataset provides a rich source of contextual information that is essential for training models to understand and reason about both textual and visual data comprehensively. 

**Retrieval Mechanism** The retrieval mechanism employed for identifying relevant triplets operates based on the cosine similarity between the query embedding _**h**[query] X_ and each triplet embedding _**h**[i] X_ 

4 

stored in the library. The cosine similarity is defined as: 

**==> picture [270 x 28] intentionally omitted <==**

For each query, the top- _k_ triplets with the highest similarity scores are retrieved. This process ensures that the most contextually relevant examples are selected. The retrieval process can be expressed as: 

_R_ ( _Xquery_ ) = _{_ ( _Qi, Ri, Ai_ ) _| i ∈_ Top- _k_ �sim( _**h**[query] X ,_ _**h**[i] X_[)] _[,][ ∀][i][ ∈{]_[1] _[, ..., N][}]_ � _},_ (4) where _R_ ( _Xquery_ ) denotes the set of retrieved triplets for the given query _Xquery_ , and Top- _k_ ( _·_ ) represents the function that selects the top- _k_ items based on the cosine similarity scores. 

## **3.3 Learn to Reasoning from the Retrieved Content** 

Given the retrieved question-rationale-answer (QRA) data, we organize them into a structured format to teach the model how to reason, as illustrated in Figure 3. This section details the process in leveraging the retrieved content to enhance the model’s reasoning capabilities. 

**==> picture [357 x 230] intentionally omitted <==**

**----- Start of picture text -----**<br>
Query  𝓧𝒒𝒖𝒆𝒓𝒚 Result  𝓨𝒒𝒖𝒆𝒓𝒚<br>Question The rationale is: Magnet sizes affect the magnitude of the<br>Think about the magnetic force between the magnets in each pair. magnetic force … But Magnet B is smaller in Pair 2 than in Pair 1. So the answer is  (C) The magnitude of the magnetic force is<br>Which of the following statements is true? smaller in Pair 2.<br>(A) The magnitude of the magnetic force is the same in both pairs.<br>(B) The magnitude of the magnetic force is smaller in Pair 1.<br>(C) The magnitude of the magnetic force is smaller in Pair 2.<br>Multimodal large language model<br>ScienceQA-based<br>Knowledge Library Content  𝓒(𝓧𝒒𝒖𝒆𝒓𝒚)<br>Retrieved  𝓡(𝓧𝒒𝒖𝒆𝒓𝒚)<br>For the question: ( 𝑸𝟏 )  Think about the<br>Question Question  … 𝑿𝟏 𝑿𝟐 … magnetic force between the magnets in each pair. Which of the following statements is true?  Referencing picture The rationale is: ( 𝑹𝟏 () 𝑰 Magnet sizes affect the 𝟏 )  1.<br>Think about the magnetic force between the magnets in each pair.  magnitude of the magnetic force<br>Which of the following statements is true? The answer is: ( 𝑨𝟏 )  The magnitude of the<br>magnetic force is greater in Pair 2.<br>(A) The magnitude of the magnetic force is greater in Pair 2. …<br>(B) The magnitude of the magnetic force is the same in both pairs.<br>(C) The magnitude of the magnetic force is greater in Pair 1. For the Current question:  Think about ...<br>Referencing picture  6.<br>Imitate the process of generating rationales<br>based on the previous question, providing a<br>rationale and answer with the option's letter<br>from the given choices.<br>**----- End of picture text -----**<br>


Figure 3: The reasoning process from the retrieved content. The model uses the organized context from retrieved question-rationale-answer triplets to generate answers. 

For a given input _Xquery_ = ( _Tquery, Iquery_ ), suppose we retrieve the top- _k_ relevant QRA triplets _R_ ( _Xquery_ ) = _{_ ( _Qi, Ri, Ai_ ) _}[k] i_ =1[.][These][retrieved][triplets][are][used][to][form][a][context] _[C]_[(] _[X][query]_[)] which provides a structured set of examples to guide the model’s reasoning process. The context _C_ ( _Xquery_ ) is constructed as follows: 

**==> picture [307 x 11] intentionally omitted <==**

where each example is a concatenation of the question, rationale, and answer: 

**==> picture [318 x 12] intentionally omitted <==**

where _⊕_ denotes the concatenation operation. The context _C_ ( _Xquery_ ) is then used to guide the model’s reasoning process, enabling it to learn from the retrieved content and generate accurate and contextually enriched answers. The model is prompted to reason based on the structured examples provided in the context, thereby predicting the answer _Yquery_ given the input _Xquery_ and the context _C_ ( _Xquery_ ). We define the conditional probability of generating the answer _Yquery_ as: 

**==> picture [337 x 11] intentionally omitted <==**

where _F_ Θ represents the LLMs parameterized by Θ. The final answer _Yquery_ is obtained by: 

**==> picture [293 x 17] intentionally omitted <==**

5 

## **4 Experiments** 

**Datasets** We employed four multimodal reasoning benchmarks: (i) _ScienceQA_ [49], a multimodal question dataset that includes over 21k multiple-choice questions, 3 subjects, 26 topics, 127 categories, and 379 distinct skills. (ii) _A-OKVQA_ [59], a knowledge-based multimodal dataset that includes 25k questions with extensive commonsense and world knowledge. (iii) _MMBench_ [47], a dataset comprising 2,974 multiple-choice questions covering 20 ability dimensions. (iv) _SEED-Bench_ [37], a large-scale dataset that includes 19k multiple-choice questions, and 12 evaluation dimensions, including both spatial and temporal understanding. We built the high-school knowledge library using the training data from ScienceQA, which employs over 12k question-rationale-answer triplets data. 

**Baselines** For ScienceQA, we compared our approach against strong baselines across four categories, excluding vision LLMs that specifically fine-tune or train on ScienceQA for a fair comparison: (i) heuristic and expert-guided choices, such as random choice and human evaluation [49]; (ii) small multimodal visual question answering models, which include MCAN [80], Top-Down [1], BAN [34], DFAF [21], ViLT [35], Patch-TRM [51], and VisualBERT [40]; (iii) zero-shot instruction-tuned large language models like GPT-3.5 [6] and its CoT-enhanced variants [49], in addition to ChatGPT, GPT-4 [53], and Chameleon [50]; (iv) strong vision LLMs like LLaVA-1.5 [46], Qwen-VL [3], InternLM-XComposer2-VL [20], and Gemini [56]. Regarding the A-OKVQA dataset, the baselines include state-of-the-art approaches, such as Pythia [4], ViLBERT [48], LXMERT [65], KRISP [52], GPV-2 [31], BLIP-2 [39], PICa [75], IPVR [13], PromptCap [27], Prophet [79], PaLI-3-VPD [28], PaLI-X-VPD [28], and Gemini [56]. For the MMBench [47] and SEED-Bench [37] datasets, we mainly focus on augmenting Gemini with RMR for the convenient access of the Gemini API. 

## **4.1 Results on ScienceQA** 

Table 1: The comparison on ScienceQA dataset. Question classes: NAT = natural science, SOC = social science, LAN = language science, TXT = text context, IMG = image context, NO = no context, G1-6 = grades 1-6, G7-12 = grades 7-12. 

|Model<br>Size|NAT<br>SOC<br>LAN|NAT<br>SOC<br>LAN|NAT<br>SOC<br>LAN|TXT|IMG<br>NO|IMG<br>NO|G1-6 G7-12|G1-6 G7-12|AVG|
|---|---|---|---|---|---|---|---|---|---|
|Random Choice [49]<br>-<br>Human[49]<br>-|40.28 46.13 29.25 <br>90.23 84.97 87.48|||47.75 <br> 89.60|40.08 33.66 <br> 87.50 88.10||39.35 40.67 <br> 91.59 82.42||39.83<br> 88.40|
|MCAN [80]<br>95M <br>Top-Down [1]<br>70M <br>BAN [34]<br>112M <br>DFAF [21]<br>74M <br>ViLT [35]<br>113M <br>Patch-TRM [51]<br>90M <br>VisualBERT [40]<br>111M <br>UnifedQABase [33]<br>223M <br>UnifedQABase w/ CoT[49]<br>223M|56.08 46.23 58.09 <br> 59.50 54.33 61.82 <br> 60.88 46.57 66.64 <br> 64.03 48.82 63.55 <br> 60.48 63.89 60.27 <br> 65.19 46.79 65.55 <br> 59.33 69.18 61.18 <br> 68.16 69.18 74.91 <br> 71.00 76.04 78.91|||59.43 <br> 62.90 <br> 62.61 <br> 65.88 <br> 63.20 <br> 66.96 <br> 62.71 <br> 63.78 <br> 66.42|51.17 55.40 <br> 54.88 59.79 <br> 52.60 65.51 <br> 54.49 64.11 <br> 61.38 57.00 <br> 55.28 64.95 <br> 62.17 58.54 <br> 61.38 77.84 <br> 66.53 81.81||51.65 59.72 <br> 57.27 62.16 <br> 56.83 63.94 <br> 57.12 67.17 <br> 60.72 61.90 <br> 58.04 67.50 <br> 62.96 59.92 <br> 72.98 65.00 <br> 77.06 68.82||54.54<br> 59.02<br> 59.37<br> 60.72<br> 61.14<br> 61.42<br> 61.87<br> 70.12<br> 74.11|
|GPT-3.5 [49]<br>173B <br>GPT-3.5 w/ CoT [49]<br>173B <br>ChatGPT w/ CoT [53]<br>-<br>GPT-4 w/ CoT [53]<br>-<br>Chameleon + ChatGPT [50]<br>-<br>Chameleon + GPT-4[50]<br>-|74.64 69.74 76.00 <br> 75.44 70.87 78.09 <br>78.82 70.98 83.18 <br>85.48 72.44 **90.27** <br>81.62 70.64 84.00 <br>89.83 74.13 89.82|||74.44 <br> 74.68 <br> 77.37 <br> 82.65 <br> 79.77 <br> 88.27|67.28 77.42 <br> 67.43 79.93 <br> 67.92 86.13 <br> 71.49 92.89 <br> 70.80 86.62 <br> 77.64 92.13||76.80 68.89 <br> 78.23 69.68 <br> 80.72 74.03 <br> 86.66 79.04 <br> 81.86 76.53 <br> 88.03 83.72||73.97<br> 75.17<br> 78.31<br> 83.99<br> 79.93<br> 86.54|
|LLaVA-1.5 [46]<br>13B|70.12 76.72 67.64|||70.48|71.89 68.92||76.06<br>61.5||70.86|
|LLaVA-1.5+RMR<br>13B|78.11<br>84.25<br>74.73|||79.81|78.33<br>74.36||82.05<br>72.18||78.52|
|_+Improvement_<br>-|+_7.99_<br>+_7.53_<br>+_7.09_|||+_9.33_|+_6.44_<br>+_5.44_||+_5.99_<br>+_10.68_||+_7.66_|
|Qwen-VL [3]<br>/|67.01 66.37 59.36|||68.52|67.03 57.00||69.75 56.16||64.89|
|Qwen-VL+RMR<br>/|70.07<br>86.16<br>75.36|||70.77|76.50<br>72.68||78.67<br>67.90||74.82|
|_+Improvement_<br>/|+_3.06_<br>+_19.79_<br>+_16.00_|||+_2.25_|+_9.47_<br>+_15.68_||+_8.92_<br>+_11.74_||+_9.93_|
|InternLM-XComposer2-VL [20]<br>/|88.19|93.48|78.64|88.47|89.14|80.77|88.66|83.52|86.82|
|InternLM-XComposer2-VL+RMR<br>/|**94.85**|**97.19**|82.55|**95.11**|**96.03**|85.23|**93.98**|88.86|**92.15**|
|_+Improvement_<br>/|+_6.66_|+_3.71_|+_3.91_|+_6.64_|+_6.89_|+_4.46_|+_5.32_|+_5.34_|+_5.33_|
|Gemini [56]<br>/|59.68|74.24|41.73|57.72|64.01|47.46|65.20|45.29|58.08|
|Gemini+RMR<br>/|91.79|94.26|89.64|91.40|89.69|**91.01**|92.84|**89.78**|91.75|
|_+Improvement_<br>/<br>+_32.11_<br>+_20.02_<br>+_47.91_<br>+_33.68_<br>+_25.68_<br>+_43.55_<br>+_27.64_<br>+_44.49_<br>+_33.67_||||||||||



6 

We set the number of retrieved examples _k_ to 3 by default. Table 1 illustrates the performance of RMR compared to various strong baselines across the ScienceQA dataset. Notably, our approach demonstrates significant improvements in reasoning capabilities without the need for fine-tuning or re-training on the target dataset, adhering to a zero-shot training setting. Despite the relatively low baseline performance of the vanilla Gemini model, which achieved an average accuracy of 58.08%, the integration of RMR remarkably enhances its reasoning capabilities, achieving an impressive average accuracy of 91.75%. This represents a substantial improvement of +33.67% over the baseline. Similarly, for the robust InternLM-XComposer2-VL model, our RMR framework yields a notable improvement of +5.33%, showcasing its effectiveness in boosting the reasoning capabilities of strong multimodal models. These results underscore the effectiveness of our proposed RMR framework in enhancing the performance of vision-language models across various question classes and contexts, demonstrating its potential to significantly advance the field of multimodal reasoning. 

## **4.2 Results on A-OKVQA** 

Table 2 presents the performance comparison for the direct-answer task on the A-OKVQA dataset. The table includes a variety of vision-language model combinations across different architectures. The models range from earlier architectures like Pythia and ViLBERT, to more recent and powerful systems such as BLIP-2, PaLM-CoT, and PaLI-X-VPD. Our RMR framework, when integrated with the Gemini model, demonstrates a substantial improvement in performance. The baseline Gemini model achieves a direct-answer accuracy of 44.2%, but with the incorporation of RMR, this accuracy increases significantly to 63.1%. This represents an impressive improvement of +18.9%. 

This improvement highlights the capability of the RMR framework to effectively augment the reasoning and answering performance of existing multimodal models, showcasing its potential to set new benchmarks in visual question answering. The results indicate that RMR not only competes with but also surpasses state-of-the-art methods across diverse model architectures and parameter sizes. 

Table 2: The comparison of the direct-answer task on the A-OKVQA dataset. 

|Model|Vision Model|Text Model|Parameters|Direct-answer|
|---|---|---|---|---|
|Pythia [4]|ResNet [26]|BERT [32]|70M|25.2|
|ViLBERT [48]|Faster R-CNN [57]|BERT [32]|300M|30.6|
|LXMERT [65]|Transformer [67]|Transformer [67]|220M|30.7|
|KRISP [52]|Faster R-CNN [57]|BERT [32]|200M|33.7|
|GPV-2[31]|VinVL[83]|T5-Base[55]|300M|48.6|
|BLIP-2 [39]|CLIP-VIT-LARGE [54]|FlanT5XXL [16]|11B|53.2|
|PaLM-CoT [69]|-|PaLM [15]|540B|41.5|
|PICa [75]|VinVL [83]|GPT-3 [7]|175B|42.4|
|IPVR [13]|Faster-RCNN [57]|OPT [84]|66B|46.4|
|PromptCap [27]|Ofa [68]|GPT-3 [7]|175B|56.3|
|Prophet [79]|MCAN [81]|GPT-3 [7]|175B|58.2|
|PaLI-3-VPD generalist [28]|SigLIP [82]|UL2 [66]|5B|56.5|
|PaLI-X-VPDgeneralist[28]|ViT[18]|UL2[66]|55B|62.7|
|Gemini [56]|-|-|-|44.2|
|Gemini+RMR|-|-|-|**63.1**|
|_+Improvement_|-|-|-|+_18.9_|



## **4.3 Results on MMBench and SEED-Bench** 

Figure 4 illustrates the comparative performance of the Gemini model and its enhanced version, Gemini+RMR, on the MMBench-Dev and MMBench-Test datasets across various categories. The categories evaluated include overall performance, conceptual problems (CP), fine-grained perception of cross instances (FP-C), fine-grained perception of single instances, attribute reasoning (AR), logical reasoning (LR), and relation reasoning (RR). For the MMBench-Dev dataset, Gemini+RMR shows marked improvements in all categories compared to the baseline Gemini model. The overall performance of Gemini+RMR is 70.96%, compared to 57.9% for Gemini, representing a substantial improvement. Specifically, Gemini+RMR scores 76.01% in CP, 72.70% in FP-S, 65.03% in FP-C, 75.38% in AR, 55.93% in LR, and 68.70% in RR. These results indicate that the RMR framework significantly enhances the model’s capability to handle a wide range of multimodal reasoning tasks. 

7 

Similarly, on the MMBench-Test dataset, Gemini+RMR outperforms the baseline Gemini model across all categories. The overall performance of Gemini+RMR is 67.26%, compared to 57.23% for Gemini. The specific category improvements are as follows: 69.16% in CP, 69.10% in FP-S, 61.54% in FP-C, 77.43% in AR, 53.18% in LR, and 63.98% in RR. These improvements further demonstrate the effectiveness of the RMR framework in enhancing the model’s reasoning capabilities across diverse evaluation metrics. Overall, the integration of the RMR framework with the Gemini model leads to significant performance improvements in both the MMBench-Dev and MMBench-Test datasets, underscoring the efficacy of our approach in advancing multimodal reasoning performance. 

**==> picture [396 x 136] intentionally omitted <==**

**----- Start of picture text -----**<br>
(a) Performance comparison on MMBench-Dev (b) Performance comparison on MMBench-Test<br>**----- End of picture text -----**<br>


Figure 4: Comparative performance of Gemini and Gemini+RMR on the MMBench-Dev and MMBench-Test datasets. 

Figure 5 presents the comparative performance of the Gemini model and its enhanced version, Gemini+RMR, across various evaluation dimensions on the SEED-Bench dataset. The evaluation categories include overall performance, scene understanding (SU), instance identity (IId), instance attributes (IA), instance location (IL), instance counting (IC), instance interaction (IIn), visual reasoning (VR), spatial relation (SR), and text understanding (TU). The results show significant performance improvements across all categories when integrating the RMR framework with the Gemini model. The overall performance of Gemini+RMR is 64.72%, compared to 56.53% for Gemini. 

**==> picture [199 x 122] intentionally omitted <==**

Figure 5: Performance on SEED-Bench dataset. 

## **4.4 Ablation Study** 

**Effect of Retrieval Size** We investigate the impact of varying the retrieval size on the performance of the RMR framework. The retrieval size _|R_ ( _Xquery_ ) _|_ is defined as the number of question-rationaleanswer (QRA) triplets retrieved for each query _Xquery_ . Table 3 presents the results of our ablation study on the ScienceQA dataset, showing the performance of Gemini with different retrieval sizes. 

Table 3: The impact of retrieval size on the performance of RMR on ScienceQA. 

|Model|NAT|SOC|LAN|TXT|IMG|NO|G1-6|G7-12|AVG|
|---|---|---|---|---|---|---|---|---|---|
|Gemini w/o retrieval|59.68|74.24|41.73|57.72|64.01|47.46|65.20|45.29|58.08|
|Gemini w/_|R_(_Xquery_)_|_= 1|88.41|89.43|84.82|87.34|84.53|87.11|89.72|84.05|87.69|
|Gemini w/_|R_(_Xquery_)_|_= 2|87.79|91.00|84.55|86.75|84.93|86.90|88.88|85.37|87.62|
|Gemini w/_|R_(_Xquery_)_|_= 3|91.79|94.26|89.64|91.40|89.69|91.01|92.84|89.78|91.75|
|Gemini w/_|R_(_Xquery_)_|_= 4|88.54|92.80|84.27|88.03|86.37|86.41|89.76|85.76|88.33|
|Gemini w/_|R_(_Xquery_)_|_= 5|87.83|89.09|84.45|86.46|83.59|86.69|89.24|83.59|87.22|



8 

The results indicate that the optimal retrieval size is _|R_ ( _Xquery_ ) _|_ = 3, which achieves the highest average accuracy of 91.75%. Increasing the retrieval size beyond this value leads to a slight decrease in performance, which may be attributed to (i) the inclusion of less relevant examples that could potentially introduce noise into the reasoning process and (ii) the longer context length that may hinder the model’s ability to effectively reason over the retrieved content. 

**Data modality** We compare the performance of the RMR framework using different data modalities: all available data (denoted as "-All"), only text-image pairs (denoted as "-T&I"), and only text data (denoted as "-T"). Table 4 presents the results of this ablation study on ScienceQA. RMR framework consistently enhances the performance of the models regardless of the data modality used, which can be attributed to the universality of its bi-modality retrieval module. 

Table 4: The impact of single vs. multi-modal retrieval on ScienceQA. 

|Model|NAT|SOC|LAN|TXT|IMG|NO|G1-6|G7-12|AVG|
|---|---|---|---|---|---|---|---|---|---|
|LLaVA-All|70.12|76.72|67.64|70.48|71.89|68.92|76.06|61.5|70.86|
|w/ RMR|78.11|84.25|74.73|79.81|78.33|74.36|82.05|72.18|78.52|
|LLaVA-T&I|69.56|74.74|86.36|69.69|71.89|-|78.03|56.97|71.89|
|w/ RMR|75.19|82.33|95.45|75.58|78.33|-|82.58|68.03|78.33|
|LLaVA-T|70.76|88.80|66.86|71.74|-|68.92|73.90|64.37|69.92|
|w/ RMR|81.50|96.00|73.86|86.57|-|74.36|81.47|74.81|78.69|
|Qwen-VL-All|67.01|66.37|59.36|68.52|67.03|57.00|69.75|56.16|64.89|
|w/ RMR|70.07|86.16|75.36|70.77|76.50|72.68|78.67|67.90|74.82|
|Qwen-VL-T&I|67.25|66.10|77.27|65.23|67.03|-|73.06|52.38|67.03|
|w/ RMR|70.31|85.47|90.91|68.50|76.50|-|82.44|62.07|76.50|
|Qwen-VL-T|66.73|68.00|58.62|73.76|-|57.00|66.10|58.56|62.95|
|w/ RMR|69.80|90.40|74.72|74.40|-|72.68|74.52|71.58|73.29|
|InternLM-All|88.19|93.48|78.64|88.47|89.14|80.77|88.66|83.52|86.82|
|w/ RMR|94.85|97.19|82.55|95.11|96.03|85.23|93.98|88.86|92.15|
|InternLM-T&I|85.77|93.98|97.73|86.32|89.14|-|90.83|85.03|89.14|
|w/ RMR|95.29|97.25|95.45|95.31|96.03|-|96.78|94.22|96.03|
|InternLM-T|90.99|90.40|77.84|91.89|-|80.77|86.25|82.56|84.71|
|w/ RMR|94.34|96.80|82.01|94.8|-|85.23|90.89|85.47|88.62|
|Gemini-All|59.68|74.24|41.73|57.72|64.01|47.46|65.20|45.29|58.08|
|w/ RMR|91.79|94.26|89.64|91.40|89.69|91.01|92.84|89.78|91.75|
|Gemini-T&I|55.75|76.83|68.18|54.89|64.01|-|69.91|49.66|64.01|
|w/ RMR|86.85|93.72|97.73|87.03|89.69|-|91.18|86.05|89.69|
|Gemini-T|64.24|58.40|40.62|62.23|-|47.46|60.00|42.52|52.70|
|w/ RMR|97.51|97.60|89.30|98.35|-|91.01|94.67|92.14|93.62|



## **5 Conclusion and Limitation** 

In this work, we introduce RMR, a multimodal RAG framework designed to enhance the reasoning capabilities of vision LLMs. By leveraging a bi-modality retrieval module, RMR retrieves the most relevant question-rationale-answer triplets from a high-school knowledge library constructed using the ScienceQA dataset. The retrieved triplets are then utilized to form a structured context that guides the model’s reasoning process. Extensive experiments on multiple multimodal reasoning benchmarks demonstrate that RMR significantly improves the performance of various vision LLMs. 

Despite the promising results, our work has several limitations that warrant further investigation. First, while RMR operates in a training-free manner, which offers convenience and effectiveness, developing a trainable multimodal RAG model could potentially further enhance the reasoning capabilities of vision-language models by allowing the model to adapt more precisely to specific datasets and tasks. Additionally, the high-school knowledge library constructed by ScienceQA may not be comprehensive enough to cover all scenarios, especially for domain-specific questions. 

9 

## **References** 

- [1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In _Proceedings of the IEEE conference on computer vision and pattern recognition_ , pages 6077–6086, 2018. 

- [2] Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to retrieve, generate, and critique through self-reflection. _arXiv preprint arXiv:2310.11511_ , 2023. 

- [3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. 2023. 

- [4] Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. In _International Conference on Machine Learning_ , pages 2397–2430. PMLR, 2023. 

- [5] Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. Improving language models by retrieving from trillions of tokens. In _International conference on machine learning_ , pages 2206–2240. PMLR, 2022. 

- [6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. _Advances in neural information processing systems_ , 33:1877–1901, 2020. 

- [7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. _Advances in neural information processing systems_ , 33:1877–1901, 2020. 

- [8] Deng Cai, Yan Wang, Lemao Liu, and Shuming Shi. Recent advances in retrieval-augmented text generation. In _Proceedings of the 45th international ACM SIGIR conference on research and development in information retrieval_ , pages 3417–3419, 2022. 

- [9] Hanqun Cao, Cheng Tan, Zhangyang Gao, Yilun Xu, Guangyong Chen, Pheng-Ann Heng, and Stan Z Li. A survey on generative diffusion models. _IEEE Transactions on Knowledge and Data Engineering_ , 2024. 

- [10] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. _ACM Transactions on Intelligent Systems and Technology_ , 15:1–45, 2024. 

- [11] Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. Benchmarking large language models in retrieval-augmented generation. In _Proceedings of the AAAI_ , volume 38, pages 17754–17762, 2024. 

- [12] Wenhu Chen, Hexiang Hu, Xi Chen, Pat Verga, and William Cohen. Murag: Multimodal retrieval-augmented generator for open question answering over images and text. In _Proceedings of EMNLP_ , pages 5558–5570, 2022. 

- [13] Zhenfang Chen, Qinhong Zhou, Yikang Shen, Yining Hong, Hao Zhang, and Chuang Gan. See, think, confirm: Interactive prompting between vision and language models for knowledge-based visual reasoning. _arXiv preprint arXiv:2301.05226_ , 2023. 

- [14] Xin Cheng, Di Luo, Xiuying Chen, Lemao Liu, Dongyan Zhao, and Rui Yan. Lift yourself up: Retrieval-augmented text generation with self-memory. _Advances in Neural Information Processing Systems_ , 36, 2024. 

- [15] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. _arXiv preprint arXiv:2204.02311_ , 2022. 

- [16] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. _arXiv preprint arXiv:2210.11416_ , 2022. 

10 

- [17] Wanqing Cui, Keping Bi, Jiafeng Guo, and Xueqi Cheng. More: Multi-modal retrieval augmented generative commonsense reasoning. _arXiv preprint arXiv:2402.13625_ , 2024. 

- [18] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision transformers to 22 billion parameters. In _International Conference on Machine Learning_ , pages 7480–7512. PMLR, 2023. 

- [19] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. A survey on in-context learning. _arXiv preprint arXiv:2301.00234_ , 2022. 

- [20] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. _arXiv preprint arXiv:2401.16420_ , 2024. 

- [21] Peng Gao, Zhengkai Jiang, Haoxuan You, Pan Lu, Steven CH Hoi, Xiaogang Wang, and Hongsheng Li. Dynamic fusion with intra-and inter-modality attention flow for visual question answering. In _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_ , pages 6639–6648, 2019. 

- [22] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. _arXiv preprint arXiv:2312.10997_ , 2023. 

- [23] Zhangyang Gao, Cheng Tan, Lirong Wu, and Stan Z Li. Simvp: Simpler yet better video prediction. In _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_ , pages 3170–3180, 2022. 

- [24] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. Retrieval augmented language model pre-training. In _International conference on machine learning_ , pages 3929–3938. PMLR, 2020. 

- [25] Junxian He, Graham Neubig, and Taylor Berg-Kirkpatrick. Efficient nearest neighbor language models. _arXiv preprint arXiv:2109.04212_ , 2021. 

- [26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In _Proceedings of the IEEE conference on computer vision and pattern recognition_ , pages 770–778, 2016. 

- [27] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. _arXiv preprint arXiv:2211.09699_ , 2022. 

- [28] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. _arXiv preprint arXiv:2312.03052_ , 2023. 

- [29] Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. Atlas: Few-shot learning with retrieval augmented language models. _Journal of Machine Learning Research_ , 24:1–43, 2023. 

- [30] Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. Active retrieval augmented generation. _arXiv preprint arXiv:2305.06983_ , 2023. 

- [31] Amita Kamath, Christopher Clark, Tanmay Gupta, Eric Kolve, Derek Hoiem, and Aniruddha Kembhavi. Webly supervised concept expansion for general purpose vision models. In _European Conference on Computer Vision_ , pages 662–681. Springer, 2022. 

- [32] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In _Proceedings of naacL-HLT_ , volume 1, page 2, 2019. 

- [33] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. UNIFIEDQA: Crossing format boundaries with a single QA system. In Trevor Cohn, Yulan He, and Yang Liu, editors, _Findings of the Association for Computational Linguistics: EMNLP 2020_ , pages 1896–1907, Online, November 2020. Association for Computational Linguistics. 

11 

- [34] Jin-Hwa Kim, Jaehyun Jun, and Byoung-Tak Zhang. Bilinear attention networks. _Advances in neural information processing systems_ , 31, 2018. 

- [35] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In _International Conference on Machine Learning_ , pages 5583–5594. PMLR, 2021. 

- [36] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. _Advances in Neural Information Processing Systems_ , 33:9459–9474, 2020. 

- [37] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seedbench: Benchmarking multimodal llms with generative comprehension. _arXiv preprint arXiv:2307.16125_ , 2023. 

- [38] Huayang Li, Yixuan Su, Deng Cai, Yan Wang, and Lemao Liu. A survey on retrieval-augmented text generation. _arXiv preprint arXiv:2202.01110_ , 2022. 

- [39] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping languageimage pre-training with frozen image encoders and large language models. _arXiv preprint arXiv:2301.12597_ , 2023. 

- [40] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. What does bert with vision look at? In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics_ , pages 5265–5275, 2020. 

- [41] Siyuan Li, Weiyang Jin, Zedong Wang, Fang Wu, Zicheng Liu, Cheng Tan, and Stan Z Li. Semireward: A general reward model for semi-supervised learning. _arXiv preprint arXiv:2310.03013_ , 2023. 

- [42] Siyuan Li, Zicheng Liu, Juanxi Tian, Ge Wang, Zedong Wang, Weiyang Jin, Di Wu, Cheng Tan, Tao Lin, Yang Liu, et al. Switch ema: A free lunch for better flatness and sharpness. _arXiv preprint arXiv:2402.09240_ , 2024. 

- [43] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z Li. Moganet: Multi-order gated aggregation network. In _The Twelfth International Conference on Learning Representations_ , 2023. 

- [44] Siyuan Li, Luyuan Zhang, Zedong Wang, Di Wu, Lirong Wu, Zicheng Liu, Jun Xia, Cheng Tan, Yang Liu, Baigui Sun, et al. Masked modeling for self-supervised representation learning on vision and beyond. _arXiv preprint arXiv:2401.00897_ , 2023. 

- [45] Bingshuai Liu, Chenyang Lyu, Zijun Min, Zhanyu Wang, Jinsong Su, and Longyue Wang. Retrieval-augmented multi-modal chain-of-thoughts reasoning for large language models. _arXiv preprint arXiv:2312.01714_ , 2023. 

- [46] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. _arXiv preprint arXiv:2304.08485_ , 2023. 

- [47] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? _arXiv preprint arXiv:2307.06281_ , 2023. 

- [48] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. _Advances in neural information processing systems_ , 32, 2019. 

- [49] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. _Advances in Neural Information Processing Systems_ , 35:2507–2521, 2022. 

- [50] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. _arXiv preprint arXiv:2304.09842_ , 2023. 

- [51] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. _arXiv preprint arXiv:2110.13214_ , 2021. 

12 

- [52] Kenneth Marino, Xinlei Chen, Devi Parikh, Abhinav Gupta, and Marcus Rohrbach. Krisp: Integrating implicit and symbolic knowledge for open-domain knowledge-based vqa. In _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , pages 14111–14121, 2021. 

- [53] OpenAI. Gpt-4 technical report, 2023. 

- [54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In _International conference on machine learning_ , pages 8748–8763. PMLR, 2021. 

- [55] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. _The Journal of Machine Learning Research_ , 21(1):5485–5551, 2020. 

- [56] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jeanbaptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. _arXiv preprint arXiv:2403.05530_ , 2024. 

- [57] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. _Advances in neural information processing systems_ , 28, 2015. 

- [58] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In _ECCV_ , pages 146–162. Springer, 2022. 

- [59] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In _European Conference on Computer Vision_ , pages 146–162. Springer, 2022. 

- [60] Cheng Tan, Zhangyang Gao, Siyuan Li, and Stan Z Li. Simvp: Towards simple yet powerful spatiotemporal predictive learning. _arXiv preprint arXiv:2211.12509_ , 2022. 

- [61] Cheng Tan, Zhangyang Gao, Lirong Wu, Yongjie Xu, Jun Xia, Siyuan Li, and Stan Z Li. Temporal attention unit: Towards efficient spatiotemporal predictive learning. In _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , pages 18770–18782, 2023. 

- [62] Cheng Tan, Siyuan Li, Zhangyang Gao, Wenfei Guan, Zedong Wang, Zicheng Liu, Lirong Wu, and Stan Z Li. Openstl: A comprehensive benchmark of spatio-temporal predictive learning. _Advances in Neural Information Processing Systems_ , 36:69819–69831, 2023. 

- [63] Cheng Tan, Jingxuan Wei, Zhangyang Gao, Linzhuang Sun, Siyuan Li, Xihong Yang, and Stan Z Li. Boosting the power of small multimodal reasoning models to match larger models with self-consistency training. _arXiv preprint arXiv:2311.14109_ , 2023. 

- [64] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In _Proceedings of the 29th ACM International Conference on Multimedia_ , pages 1405–1413, 2021. 

- [65] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing_ , 2019. 

- [66] Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Jason Wei, Xuezhi Wang, Hyung Won Chung, Siamak Shakeri, Dara Bahri, Tal Schuster, et al. Ul2: Unifying language learning paradigms. _arXiv preprint arXiv:2205.05131_ , 2022. 

- [67] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. _Advances in neural information processing systems_ , 30, 2017. 

- [68] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In _International Conference on Machine Learning_ , pages 23318–23340. PMLR, 2022. 

13 

- [69] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. _Advances in Neural Information Processing Systems_ , 35:24824–24837, 2022. 

- [70] Jingxuan Wei, Cheng Tan, Zhangyang Gao, Linzhuang Sun, Siyuan Li, Bihui Yu, Ruifeng Guo, and Stan Z Li. Enhancing human-like multi-modal reasoning: A new challenging dataset and comprehensive framework. _arXiv preprint arXiv:2307.12626_ , 2023. 

- [71] Fangyuan Xu, Weijia Shi, and Eunsol Choi. Recomp: Improving retrieval-augmented lms with compression and selective augmentation. _arXiv preprint arXiv:2310.04408_ , 2023. 

- [72] Xin Xu, Yue Liu, Panupong Pasupat, Mehran Kazemi, et al. In-context learning with retrieved demonstrations for language models: A survey. _arXiv preprint arXiv:2401.11624_ , 2024. 

- [73] Zhipeng Xu, Zhenghao Liu, Yibin Liu, Chenyan Xiong, Yukun Yan, Shuo Wang, Shi Yu, Zhiyuan Liu, and Ge Yu. Activerag: Revealing the treasures of knowledge via active learning. _arXiv preprint arXiv:2402.13547_ , 2024. 

- [74] Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, and Zhen-Hua Ling. Corrective retrieval augmented generation. _arXiv preprint arXiv:2401.15884_ , 2024. 

- [75] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Yumao Lu, Zicheng Liu, and Lijuan Wang. An empirical study of gpt-3 for few-shot knowledge-based vqa. In _Proceedings of the AAAI Conference on Artificial Intelligence_ , volume 36, pages 3081–3089, 2022. 

- [76] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. In _Proceedings of ICML_ , pages 39755–39769, 2023. 

- [77] Wenhao Yu. Retrieval-augmented generation across heterogeneous knowledge. In _Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop_ , pages 52–58, 2022. 

- [78] Wenhao Yu, Hongming Zhang, Xiaoman Pan, Kaixin Ma, Hongwei Wang, and Dong Yu. Chain-of-note: Enhancing robustness in retrieval-augmented language models. _arXiv preprint arXiv:2311.09210_ , 2023. 

- [79] Zhou Yu, Xuecheng Ouyang, Zhenwei Shao, Meng Wang, and Jun Yu. Prophet: Prompting large language models with complementary answer heuristics for knowledge-based visual question answering. _arXiv preprint arXiv:2303.01903_ , 2023. 

- [80] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_ , pages 6281–6290, 2019. 

- [81] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_ , pages 6281–6290, 2019. 

- [82] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In _Proceedings of the IEEE/CVF International Conference on Computer Vision_ , pages 11975–11986, 2023. 

- [83] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In _Proceedings of the IEEE/CVF conference on computer vision and pattern recognition_ , pages 5579–5588, 2021. 

- [84] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. _arXiv preprint arXiv:2205.01068_ , 2022. 

- [85] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? _Advances in Neural Information Processing Systems_ , 36, 2024. 

- [86] Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. Explainability for large language models: A survey. _ACM Transactions on Intelligent Systems and Technology_ , 15:1–38, 2024. 

- [87] Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu, Ling Yang, Wentao Zhang, and Bin Cui. Retrieval-augmented generation for ai-generated content: A survey. _arXiv preprint arXiv:2402.19473_ , 2024. 

14 

## **A Qualitative Analysis** 

_**Rationale**_ : An object has different properties. A property of an object can tell you how it looks, feels, tastes, or smells. Properties can also tell you how an object will behave when something happens to it. Different objects can have properties in common. Look at each object. For each object, decide if it has that property. A sticky object can attach or stick to other things. The toothpaste is sticky, but the soccer shorts and the water slide are not. This color is blue. All three objects are blue. A hard object does not change shape when pressed or squeezed. The water slide is hard, but the toothpaste and the soccer shorts are not. The property that all three objects have in common is blue. 

**==> picture [347 x 269] intentionally omitted <==**

**----- Start of picture text -----**<br>
Query  𝓧𝒒𝒖𝒆𝒓𝒚 looks, feels, tastes, or smells. Properties can also tell you how an object will behave<br>something happens to it. Different objects can have properties in common. Look at<br>Question : Which property do these three objects object. For each object, decide if it has that property. A sticky object can attach or stick<br>have in common? other things. The toothpaste is sticky, but the soccer shorts and the water slide are not. This<br>Options :  color is blue. All three objects are blue. A hard object does not change shape when pressed<br>or squeezed. The water slide is hard, but the toothpaste and the soccer shorts are not.<br>(A) Blue.            property that all three objects have in common is blue.<br>(B) Hard.           Answer : The answer is (A) Blue.<br>(C) Sticky.<br>Multimodal large language model<br>𝑿𝟏 𝑿𝟐 𝑿𝟑<br>Question : Which property do these three objects  Question : Which property matches this  Question : Which property do these two<br>have in common? object? objects have in common?<br>Options :  Options : Options :<br>(A) Fragile. (A) Blue. (A) Bouncy.<br>(B) Blue. (B) Hard. (B) Soft.<br>(C) Sticky Rationale : An object has different<br>Rationale : An object has different properties. A Rationale : Look at the object. Think properties. A property of an object can<br>property of an object can tell you how it looks, about each property. A hard object tell you how it looks, feels, tastes, or<br>feels, tastes, or smells. Properties can also tell keeps its shape when you squeeze it. smells. Different objects can have the<br>you how an object will behave when something The jeans are not hard. Blue is a color. same properties. Look at each object.<br>happens to it. Different objects can have This color is blue. The jeans are blue. For each object, decide if it has that<br>properties in common. Look at each object. For Answer : The answer is (A) Blue. property. A soft object changes shape<br>each object, decide if it has that property. A when you squeeze it. Both objects are<br>sticky object can attach or stick to other things. soft. A bouncy object will bounce back<br>The toothpaste is sticky, but the jeans and the from the floor if you drop it. Neither of<br>track suit are not. Blue is a color. This color is the objects are bouncy. The property<br>blue. All three objects are blue. A fragile object that both objects have in common is soft.<br>will break into pieces if you drop it. The jeans<br>andthat all threethe trackobjectssuit arehave innot fragile.commonTheis blue.property Answer : The answer is (B) Soft.<br>Answer : The answer is (B) Blue.<br>**----- End of picture text -----**<br>


Figure 6: The retrieved data of an image-text pair example. 

_**Rationale**_ : Guide words appear on each page of a dictionary. They tell you the first word and Query 𝓧𝒒𝒖𝒆𝒓𝒚 last word on the page. The other words on the page come between the guide words in alphabetical order. To put words in alphabetical order, put them in order by their first letters. If _**Question**_ : Which word would you find on a the first letters are the same, look at the second letters. If the second letters are the same, look dictionary page with the following guide at the third letters, and so on. If one word is shorter, and there are no more letters to compare, words? imitate - iron then the shorter word comes first in alphabetical order. For example, be comes before bed. Put _**Options**_ : the words in alphabetical order. Since inherit is between the guide words imitate - iron, it would be found on that page. (A) Itch. _**Answer**_ : The answer is (B) Inherit. (B) Inherit. Multimodal large language model 𝑿𝟏 𝑿𝟐 𝑿𝟑 _**Question**_ : Which word would you find on a _**Question**_ : Would you find the word instruct _**Question**_ : Which word would you find on dictionary page with the following guide on a dictionary page with the following guide a dictionary page with the following guide words? idea - inspire words? imp - item words? impose - issue _**Options**_ : _**Options**_ : _**Options**_ : (A) Issue. (A) No. (A) Ill. (B) Immense. (B) Yes. (B) Increase. _**Rationale**_ : Guide words appear on each _**Rationale**_ : Guide words appear on each page _**Rationale**_ : Guide words appear on each page of a dictionary. They tell you the first of a dictionary. They tell you the first word page of a dictionary. They tell you the first word and last word on the page. The other and last word on the page. The other words word and last word on the page. The other words on the page come between the guide on the page come between the guide words in words on the page come between the guide words in alphabetical order. To put words alphabetical order. To put words in words in alphabetical order. To put words in alphabetical order, put them in order by alphabetical order, put them in order by their in alphabetical order, put them in order by their first letters. If the first letters are the first letters. If the first letters are the same, their first letters. If the first letters are the same, look at the second letters. If the look at the second letters. If the second letters same, look at the second letters. If the second letters are the same, look at the third are the same, look at the third letters, and so second letters are the same, look at the letters, and so on. If one word is shorter, on. If one word is shorter, and there are no third letters, and so on. If one word is and there are no more letters to compare, more letters to compare, then the shorter shorter, and there are no more letters to then the shorter word comes first in word comes first in alphabetical order. For compare, then the shorter word comes first alphabetical order. For example, be comes example, be comes before bed. Put the words in alphabetical order. For example, be before bed. Put the words in alphabetical in alphabetical order. Since instruct is comes before bed. Put the words in order. Since immense is between the guide between the guide words imp - item, it would alphabetical order. Since increase is words idea - inspire, it would be found on be found on that page. between the guide words impose - issue, it that page. _**Answer**_ : The answer is (B) Yes. would be found on that page. _**Answer**_ : The answer is (B) immense. _**Answer**_ : The answer is (B) Increase. 

Figure 7: The retrieved data of a text-only example. 

15 

**==> picture [133 x 58] intentionally omitted <==**

**----- Start of picture text -----**<br>
Query  𝓧𝒒𝒖𝒆𝒓𝒚<br>Question : Is marble a mineral or a rock?<br>Options :<br>(A) Rock.<br>(B) Mineral.<br>**----- End of picture text -----**<br>


_**Rationale**_ : Minerals are the building blocks of rocks. A rock can be made of one or more minerals. Minerals and rocks have the following properties: Property | Mineral | Rock It is a solid. | Yes | Yes It is formed in nature. | Yes | Yes It is not made by organisms. | Yes | Yes It is a pure substance. | Yes... Different types of minerals have different crystal structures, but all minerals have a fixed crystal structure. This means that the atoms and molecules in different pieces of the same type of mineral are always arranged the same way.\nHowever, rocks do not have a fixed crystal structure. So, the arrangement of atoms or molecules in different pieces of the same type of rock may be different! The properties of marble match the properties of a rock. So, marble is a rock. _**Answer**_ : The answer is (A) Blue. 

Multimodal large language model 𝑿𝟏 𝑿𝟐 𝑿𝟑 _**Question**_ : Is celestine a mineral or a rock? _**Question**_ : What type of rock is marble? _**Question**_ : Is soapstone a mineral or _**Options**_ : _**Options**_ : a rock? _**Options**_ : (A) Mineral. (A) Sedimentary. (B) Rock. (B) Metamorphic. (A) Rock. (B) Mineral. _**Rationale**_ : Minerals are the building blocks of (C) Igneous. rocks. A rock can be made of one or more _**Rationale**_ : Igneous rock is formed when _**Rationale**_ : Minerals are the building minerals. Minerals and rocks have the following melted rock cools and hardens into solid blocks of rocks. A rock can be made of properties: Property | Mineral | Rock. It is a solid. rock. This type of change can occur at one or more minerals. Minerals and rocks | Yes | Yes. It is formed in nature. | Yes | Yes. It Earth‘s surface or below it. Sedimentary have the following properties: Property | is not made by organisms. | Yes | Yes. It is a pure rock is formed when layers of sediment Mineral | Rock It is a solid. | Yes | Yes It substance. | Yes | No It has a fixed crystal are pressed together, or compacted, to is formed in nature. | Yes | Yes It is not structure. | Yes | No. You can use these properties make rock... Like other metamorphic made by organisms. | Yes | Yes It is a to tell whether a substance is a mineral, a rock, or rocks, it forms when a rock is changed by pure substance. | Yes | No It has a fixed neither. Look closely at the last three properties: high temperature and pressure. Heat and crystal structure. | Yes | No …So, the Minerals and rocks are not made by organisms. pressure can change the type and arrangement of atoms or molecules in Organisms make their own body parts. For arrangement of minerals in a rock. This different pieces of the same type of rock example, snails and clams make their shells. change forms a new rock with different may be different! The properties of Because they are made by organisms, body parts properties. Marble can form when soapstone match the properties of a rock. cannot be minerals or rocks.. Celestine has all the sedimentary rocks such as limestone are So, soapstone is a rock. properties of a mineral. So, celestine is a mineral. changed by heat and pressure. _**Answer**_ : The answer is (B) Soft. _**Answer**_ : The answer is (A) Mineral. _**Answer**_ : The answer is (A) Mineral. 

Figure 8: The retrieved data of an image-text pair example. 

**==> picture [365 x 316] intentionally omitted <==**

**----- Start of picture text -----**<br>
Query  𝓧𝒒𝒖𝒆𝒓𝒚 Rationale first base. A compound: A simple sentencesentenceis ais two simplesentence withsentences joined byonly one subject a commaand predicate.and a conjunction such as and, but,The pitcher threw the ball to<br>or, or so. The pitcher threw the ball, and the batter hit it. Some simple sentences have a compound subject or a<br>Question : Which is a compound sentence?  compound predicate, but they are not compound sentences… The introductory phrase is part of the predicate. In<br>Options :  the winter, Farmer Ben wears his heavy coat. This is a simple sentence. There is one subject, Farmer Ben, and<br>one predicate, wears his heavy coat in the winter. The first sentence is the compound sentence. It is made up of<br>(A) Dillon liked the sea otters, but the jellyfish  two simple sentences joined by a comma and the conjunction but. Dillon liked the sea otters, but the jellyfish<br>were his favorite. were his favorite.<br>Answer : The answer is (A) Dillon liked the sea otters, but the jellyfish were his favorite.<br>(B) The artist prepared a canvas for a new oil<br>painting. Multimodal large language model<br>𝑿𝟏 𝑿𝟐 𝑿𝟑<br>Question : Which is a compound sentence? Question : Which is a compound sentence? Question : Which is a compound sentence?<br>Options : Options : Options :<br>(A) The artist prepared a canvas for a new oil (A) The camera isn't working, so the battery must (A) The tailor measures the length of the pant leg.<br>painting. be dead. (B) Desmond liked the sea otters, but the<br>(B) Devon shot the arrow, but she missed the (B) The artist prepared a canvas for a new oil jellyfish were his favorite.<br>target. painting.<br>Rationale : A simple sentence is a sentence Rationale : A simple sentence is a sentence with Rationale : A simple sentence is a sentence with<br>with only one subject and predicate. The only one subject and predicate. The pitcher threw only one subject and predicate. The pitcher threw<br>pitcher threw the ball to first base. A the ball to first base. A compound sentence is the ball to first base. A compound sentence is two<br>compound sentence is two simple sentences two simple sentences joined by a comma and a simple sentences joined by a comma and a<br>joined by a comma and a conjunction such as conjunction such as and, but, or, or so. The conjunction such as and, but, or, or so. The pitcher<br>and, but, or, or so. The pitcher threw the ball, pitcher threw the ball, and the batter hit it. Some threw the ball, and the batter hit it. Some simple<br>and the batter hit it. Some simple sentences simple sentences have a compound subject or a sentences have a compound subject or a<br>have a compound subject or a compound compound predicate, but they are not compound compound predicate, but they are not compound<br>predicate, but they are not compound sentences…Some simple sentences have sentences…Some simple sentences have<br>sentences…Some simple sentences have introductory phrases, but they are not compound introductory phrases, but they are not compound<br>introductory phrases, but they are not sentences. The introductory phrase is part of the sentences. The introductory phrase is part of the<br>compound sentences. The introductory phrase predicate. In the winter, Farmer Ben wears his predicate. In the winter, Farmer Ben wears his<br>is part of the predicate. In the winter, Farmer heavy coat. This is a simple sentence. There is heavy coat. This is a simple sentence. There is one<br>Ben wears his heavy coat. This is a simple one subject, Farmer Ben, and one predicate, subject, Farmer Ben, and one predicate, wears his<br>sentence. There is one subject, Farmer Ben, wears his heavy coat in the winter. The first heavy coat in the winter. The second sentence is<br>and one predicate, wears his heavy coat in the sentence is the compound sentence. It is made up the compound sentence. It is made up of two<br>winter. The second sentence is the compound of two simple sentences joined by a comma and simple sentences joined by a comma and the<br>sentence. It is made up of two simple sentences the conjunction so. The camera isn't working, so conjunction but. Desmond liked the sea otters, but<br>joined by a comma and the conjunction but. the battery must be dead. the jellyfish were his favorite.<br>Devon shot the arrow, but she missed the Answer : The answer is (A) The camera isn't Answer : The answer is (B) Desmond liked the sea<br>target.<br>Answer : The answer is (B) Devon shot the working, so the battery must be dead. otters, but the jellyfish were his favorite.<br>arrow, but she missed the target.<br>**----- End of picture text -----**<br>


Figure 9: The retrieved data of a text-only example. 

16 

We present qualitative examples to illustrate the effectiveness of the RMR framework. Figure 6 showcases an example where the model is asked, "Which property do these three objects have in common?" The retrieved data plays a crucial role in guiding the model’s reasoning process. _X_ 1 provides a similar question involving different objects. The similarity in the question structure helps the model understand the type of reasoning required, offering a blueprint for approaching the problem. By seeing how the retrieved example addresses the comparison of different objects, the model can apply a similar strategy to the current query. _X_ 2 includes a related object, enhancing the model’s comprehension of various object properties. The additional context provided by _X_ 2 helps the model draw connections between the current objects and previously encountered ones, broadening its understanding and improving its ability to identify common properties. Similar to _X_ 1, _X_ 3 also presents a question about the common properties of different objects, albeit with a wider variety. The diversity in examples reinforces the model’s reasoning skills, helping it generalize its reasoning ability across different contexts and object sets. 

Figure 7 depicts a text-only example where the model is asked, "Which word would you find on a dictionary page with the following guide words?". Three similar questions are retrieved with different words, providing the model with a diverse set of examples to learn from. The retrieved examples help the model understand the structure of the question and the type of reasoning required to answer it. By observing the different words in the retrieved examples, the model can learn to identify the commonalities between the guide words and the target words, enhancing its reasoning capabilities. 

Figure 8 presents an image and text example where the model is asked, "Is marble a mineral or a rock?". The retrieved data provides relevant context that helps the model formulate a more accurate and informed response. The retrieved data _X_ 1 contains the question, "Is celestine a mineral or a rock?" This question is analogous to the query about marble, offering the model a direct parallel that aids in understanding how to distinguish between minerals and rocks. The similar structure helps the model apply the reasoning used for celestine to marble, reinforcing the process of categorizing geological substances. _X_ 2 poses the more specific question, "What type of rock is marble?" It provides the model with additional information about the classification of marble within the broader category of rocks. This specific context helps the model not only affirm that marble is a rock but also understand its specific type, thereby enriching the model’s geological knowledge base. _X_ 3 includes the question, "Is soapstone a mineral or a rock?" Similar to _X_ 1, it presents another instance of the mineral vs. rock distinction. The inclusion of different substances like soapstone reinforces the model’s ability to generalize the reasoning process across various materials, ensuring a robust understanding of the mineral and rock classification. 

Figure 9 presents a text-only example where the model is asked, "Which is a compound sentence?" The retrieved examples all ask the same question but with different options, which provide the model with varied contexts and sentence structures to learn from. The consistent retrieval of questions on compound sentences helps the model understand the syntactic characteristics that define compound sentences. By comparing and analyzing the different options presented in the retrieved examples, the model can improve its ability to identify compound sentences accurately. 

## **B Broader Impact** 

Here, we outline several broader impacts of our work: 

**Hallucination** Similar to other large language models, RMR-integrated models like LLaVA might generate outputs that are not grounded in factual information or the input data provided. This phenomenon, known as hallucination, may raise concerns. Ensuring the reliability and accuracy of outputs in such sensitive domains is crucial, and further research is needed to mitigate these risks. 

**Biases** The RMR framework inherits biases from its base models, including the vision encoder (CLIP) and the vision LLMs (LLaVA, Qwen-VL, InternLM, and Gemini). These biases may be reflected in the retrieved examples and the model’s reasoning process. It is essential to address these biases to ensure fair and unbiased reasoning capabilities. 

**Energy Consumption** Although our approach is training-free, running inference still incurs energy consumption. The computational resources required for processing large volumes of data. Efforts should be made to optimize inference efficiency and explore sustainable computing practices to minimize the energy footprint of using RMR-enhanced models. 

17