Reasoning capabilities in large language models (LLMs) have substantially advanced through methods such as chain-of-thought and explicit step-by-step explanations. However, these improvements have not yet fully transitioned to multimodal contexts, where models must proactively decide which sensory modalities such as vision, audio, or spatial perception to engage when interacting with complex real-world environments. In this paper, we introduce the Chain of Questions (CoQ) framework, a curiosity-driven reasoning approach that encourages multimodal language models to dynamically generate targeted questions regarding their surroundings. These generated questions guide the model to selectively activate relevant modalities, thereby gathering critical information necessary for accurate reasoning and response generation. We evaluate our framework on a novel multimodal benchmark dataset, assembled by integrating WebGPT, ScienceQA, AVSD, and ScanQA datasets. Experimental results demonstrate that our CoQ method improves a foundation model's ability to effectively identify and integrate pertinent sensory information. This leads to improved accuracy, interpretability, and alignment of the reasoning process with diverse multimodal tasks.

Faculty of Computing and Information Technology & Center of Research Excellence in AI and Data Science King Abdulaziz University Jeddah, Saudi Arabia Abstract --Multimodal large language models (MLLMs) have shown success in vision-language tasks, but their ability to reason over complex educational materials remains largely untested. This work presents the first evaluation of state-of-the-art MLLMs, including LLaV A-1.5 and LLaMA 3.2-Vision, on the textbook question answering (TQA) task using the CK12-QA dataset. We introduce a multimodal retrieval-augmented generation (RAG) pipeline to simulate real-world learning by providing relevant lesson paragraphs and diagrams as context. Our zero-shot experiments reveal a critical trade-off; while retrieved context improves LLaV A's performance on text-based questions, it significantly degrades the accuracy of the more powerful LLaMA 3.2-Vision on diagram-based tasks, dropping its validation accuracy from 74.07% to 25.93%. Furthermore, fine-tuning highlights architectural differences; LLaMA 3.2-Vision's performance substantially improves to 71.16% on the test set, demonstrating its capacity to learn multimodal integration, whereas LLaV A's performance declines, indicating challenges with generalization. Our results underscore the challenges MLLMs face in modality prioritization and context integration, providing a benchmark and pointing to key directions for developing more robust AI-driven educational tools. Personal use of this material is permitted. This work has been accepted to the 2nd International Generative AI and Computational Language Modelling Conference (GACLM 2025) for publication in the proceedings. Answering curriculum-related questions in multimodal educational materials is a central challenge in AI for education, requiring systems to reason across complex multimodal contexts such as lengthy lessons, diagrams, and videos.

We introduce SciVer, the first benchmark specifically designed to evaluate the ability of foundation models to verify claims within a multimodal scientific context. SciVer consists of 3,000 expert-annotated examples over 1,113 scientific papers, covering four subsets, each representing a common reasoning type in multimodal scientific claim verification. To enable fine-grained evaluation, each example includes expert-annotated supporting evidence. We assess the performance of 21 state-of-the-art multimodal foundation models, including o4-mini, Gemini-2.5-Flash, Llama-3.2-Vision, and Qwen2.5-VL. Our experiment reveals a substantial performance gap between these models and human experts on SciVer. Through an in-depth analysis of retrieval-augmented generation (RAG), and human-conducted error evaluations, we identify critical limitations in current open-source models, offering key insights to advance models' comprehension and reasoning in multimodal scientific literature tasks.

As the key component in multimodal large language models (MLLMs), the ability of the visual encoder greatly affects MLLM's understanding on diverse image content. Although some large-scale pretrained vision encoders such as vision encoders in CLIP and DINOv2 have brought promising performance, we found that there is still no single vision encoder that can dominate various image content understanding, e.g., the CLIP vision encoder leads to outstanding results on general image understanding but poor performance on document or chart content. To alleviate the bias of CLIP vision encoder, we first delve into the inherent behavior of different pre-trained vision encoders and then propose the MoVA, a powerful and novel MLLM, adaptively routing and fusing task-specific vision experts with a coarse-to-fine mechanism. In the coarse-grained stage, we design a context-aware expert routing strategy to dynamically select the most suitable vision experts according to the user instruction, input image, and expertise of vision experts. In the fine-grained stage, we elaborately conduct the mixture-of-vision-expert adapter (MoV-Adapter) to extract and fuse task-specific knowledge from various experts.

Multimodal knowledge graph completion (MMKGC) aims to predict missing links in multimodal knowledge graphs (MMKGs) by leveraging information from various modalities alongside structural data. Existing MMKGC approaches primarily extend traditional knowledge graph embedding (KGE) models, which often require creating an embedding for every entity. This results in large model sizes and inefficiencies in integrating multimodal information, particularly for real-world graphs. Meanwhile, Transformer-based models have demonstrated competitive performance in knowledge graph completion (KGC). However, their focus on single-modal knowledge limits their capacity to utilize cross-modal information. Recently, Large vision-language models (VLMs) have shown potential in cross-modal tasks but are constrained by the high cost of training. In this work, we propose a novel approach that integrates Transformer-based KGE models with cross-modal context generated by pre-trained VLMs, thereby extending their applicability to MMKGC. Specifically, we employ a pre-trained VLM to transform relevant visual information from entities and their neighbors into textual sequences. We then frame KGC as a sequence-to-sequence task, fine-tuning the model with the generated cross-modal context. This simple yet effective method significantly reduces model size compared to traditional KGE approaches while achieving competitive performance across multiple large-scale datasets with minimal hyperparameter tuning.

Automatic Video Dubbing (AVD) generates speech aligned with lip motion and facial emotion from scripts. Recent research focuses on modeling multimodal context to enhance prosody expressiveness but overlooks two key issues: 1) Multiscale prosody expression attributes in the context influence the current sentence's prosody. 2) Prosody cues in context interact with the current sentence, impacting the final prosody expressiveness. To tackle these challenges, we propose M2CI-Dubber, a Multiscale Multimodal Context Interaction scheme for AVD. This scheme includes two shared M2CI encoders to model the multiscale multimodal context and facilitate its deep interaction with the current sentence. By extracting global and local features for each modality in the context, utilizing attention-based mechanisms for aggregation and interaction, and employing an interaction-based graph attention network for fusion, the proposed approach enhances the prosody expressiveness of synthesized speech for the current sentence. Experiments on the Chem dataset show our model outperforms baselines in dubbing expressiveness. The code and demos are available at \textcolor[rgb]{0.93,0.0,0.47}{https://github.com/AI-S2-Lab/M2CI-Dubber}.

Multimodal Large Language Models (MLLMs) utilize multimodal contexts consisting of text, images, or videos to solve various multimodal tasks. However, we find that changing the order of multimodal input can cause the model's performance to fluctuate between advanced performance and random guessing. This phenomenon exists in both single-modality (text-only or image-only) and mixed-modality (image-text-pair) contexts. Furthermore, we demonstrate that popular MLLMs pay special attention to certain multimodal context positions, particularly the beginning and end. Leveraging this special attention, we place key video frames and important image/text content in special positions within the context and submit them to the MLLM for inference. This method results in average performance gains of 14.7% for video-caption matching and 17.8% for visual question answering tasks. Additionally, we propose a new metric, Position-Invariant Accuracy (PIA), to address order bias in MLLM evaluation. Our research findings contribute to a better understanding of Multi-Modal In-Context Learning (MMICL) and provide practical strategies for enhancing MLLM performance without increasing computational costs.

Banking Transaction Flow (BTF) is a sequential data found in a number of banking activities such as marketing, credit risk or banking fraud. It is a multimodal data composed of three modalities: a date, a numerical value and a wording. We propose in this work an application of self-attention mechanism to the processing of BTFs. We trained two general models on a large amount of BTFs in a self-supervised way: one RNN-based model and one Transformer-based model. We proposed a specific tokenization in order to be able to process BTFs. The performance of these two models was evaluated on two banking downstream tasks: a transaction categorization task and a credit risk task. The results show that fine-tuning these two pre-trained models allowed to perform better than the state-of-the-art approaches for both tasks.

With the bloom of Large Language Models (LLMs), Multimodal Large Language Models (MLLMs) that incorporate LLMs with pre-trained vision models have recently demonstrated impressive performance across diverse vision-language tasks. However, they fall short to comprehend context involving multiple images. A primary reason for this shortcoming is that the visual features for each images are encoded individually by frozen encoders before feeding into the LLM backbone, lacking awareness of other images and the multimodal instructions. We term this issue as prior-LLM modality isolation and propose a two phase paradigm, browse-and-concentrate, to enable in-depth multimodal context fusion prior to feeding the features into LLMs. This paradigm initially "browses" through the inputs for essential insights, and then revisits the inputs to "concentrate" on crucial details, guided by these insights, to achieve a more comprehensive understanding of the multimodal inputs. Additionally, we develop training strategies specifically to enhance the understanding of multi-image inputs. Our method markedly boosts the performance on 7 multi-image scenarios, contributing to increments on average accuracy by 2.13% and 7.60% against strong MLLMs baselines with 3B and 11B LLMs, respectively.

The rapid progress in Multimodal Large Language Models (MLLMs) has significantly advanced their ability to process and understand complex visual and textual information. However, the integration of multiple images and extensive textual contexts remains a challenge due to the inherent limitation of the models' capacity to handle long input sequences efficiently. In this paper, we introduce SEEKER, a multimodal large language model designed to tackle this issue. SEEKER aims to optimize the compact encoding of long text by compressing the text sequence into the visual pixel space via images, enabling the model to handle long text within a fixed token-length budget efficiently. Our empirical experiments on six long-context multimodal tasks demonstrate that SEEKER can leverage fewer image tokens to convey the same amount of textual information compared with the OCR-based approach, and is more efficient in understanding long-form multimodal input and generating long-form textual output, outperforming all existing proprietary and open-source MLLMs by large margins.