Vision Language Models (VLMs) have demonstrated strong reasoning capabilities in Visual Question Answering (VQA) tasks; However, their ability to perform Theory of Mind (ToM) tasks such as accurately inferring human intentions, beliefs, and other mental states remains underexplored. In this work, we propose an open-ended question framework to comprehensively evaluate VLMs' performance across diverse categories of ToM tasks. We curated and annotated a benchmark dataset composed of 30 images. We then assessed the performance of four VLMs of varying sizes on this dataset. Our experimental results show that the GPT-4 model outperformed all others, with only one smaller model, GPT-4o-mini, achieving comparable performance. Additionally, we observed that VLMs often struggle to accurately infer intentions in complex scenarios such as bullying or cheating. Moreover, our findings also reveal that smaller models can sometimes infer correct intentions despite relying on incorrect visual cues.
We introduce JEEM, a benchmark designed to evaluate Vision-Language Models (VLMs) on visual understanding across four Arabic-speaking countries: Jordan, The Emirates, Egypt, and Morocco. JEEM includes the tasks of image captioning and visual question answering, and features culturally rich and regionally diverse content. This dataset aims to assess the ability of VLMs to generalize across dialects and accurately interpret cultural elements in visual contexts. In an evaluation of five prominent open-source Arabic VLMs and GPT-4V, we find that the Arabic VLMs consistently underperform, struggling with both visual understanding and dialect-specific generation. While GPT-4V ranks best in this comparison, the model's linguistic competence varies across dialects, and its visual understanding capabilities lag behind. This underscores the need for more inclusive models and the value of culturally-diverse evaluation paradigms.
Multimodal large language models (MLLMs) have demonstrated remarkable capabilities in various tasks. However, effectively evaluating these MLLMs on face perception remains largely unexplored. To address this gap, we introduce FaceBench, a dataset featuring hierarchical multi-view and multi-level attributes specifically designed to assess the comprehensive face perception abilities of MLLMs. Initially, we construct a hierarchical facial attribute structure, which encompasses five views with up to three levels of attributes, totaling over 210 attributes and 700 attribute values. Based on the structure, the proposed FaceBench consists of 49,919 visual question-answering (VQA) pairs for evaluation and 23,841 pairs for fine-tuning. Moreover, we further develop a robust face perception MLLM baseline, Face-LLaVA, by training with our proposed face VQA data. Extensive experiments on various mainstream MLLMs and Face-LLaVA are conducted to test their face perception ability, with results also compared against human performance. The results reveal that, the existing MLLMs are far from satisfactory in understanding the fine-grained facial attributes, while our Face-LLaVA significantly outperforms existing open-source models with a small amount of training data and is comparable to commercial ones like GPT-4o and Gemini. The dataset will be released at https://github.com/CVI-SZU/FaceBench.
Document Visual Question Answering (VQA) models have evolved at an impressive rate over the past few years, coming close to or matching human performance on some benchmarks. We argue that common evaluation metrics used by popular benchmarks do not account for the semantic and multimodal groundedness of a model's outputs. As a result, hallucinations and major semantic errors are treated the same way as well-grounded outputs, and the evaluation scores do not reflect the reasoning capabilities of the model. In response, we propose a new evaluation methodology that accounts for the groundedness of predictions with regard to the semantic characteristics of the output as well as the multimodal placement of the output within the input document. Our proposed methodology is parameterized in such a way that users can configure the score according to their preferences. We validate our scoring methodology using human judgment and show its potential impact on existing popular leaderboards. Through extensive analyses, we demonstrate that our proposed method produces scores that are a better indicator of a model's robustness and tends to give higher rewards to better-calibrated answers.
Object-centric representation learning aims to decompose visual scenes into fixed-size vectors called "slots" or "object files", where each slot captures a distinct object. Current state-of-the-art object-centric models have shown remarkable success in object discovery in diverse domains, including complex real-world scenes. However, these models suffer from a key limitation: they lack controllability. Specifically, current object-centric models learn representations based on their preconceived understanding of objects, without allowing user input to guide which objects are represented. Introducing controllability into object-centric models could unlock a range of useful capabilities, such as the ability to extract instance-specific representations from a scene. In this work, we propose a novel approach for user-directed control over slot representations by conditioning slots on language descriptions. The proposed ConTRoLlable Object-centric representation learning approach, which we term CTRL-O, achieves targeted object-language binding in complex real-world scenes without requiring mask supervision. Next, we apply these controllable slot representations on two downstream vision language tasks: text-to-image generation and visual question answering. The proposed approach enables instance-specific text-to-image generation and also achieves strong performance on visual question answering.
Existing benchmarks for Vision-Language Model (VLM) on autonomous driving (AD) primarily assess interpretability through open-form visual question answering (QA) within coarse-grained tasks, which remain insufficient to assess capabilities in complex driving scenarios. To this end, we introduce $\textbf{VLADBench}$, a challenging and fine-grained dataset featuring close-form QAs that progress from static foundational knowledge and elements to advanced reasoning for dynamic on-road situations. The elaborate $\textbf{VLADBench}$ spans 5 key domains: Traffic Knowledge Understanding, General Element Recognition, Traffic Graph Generation, Target Attribute Comprehension, and Ego Decision-Making and Planning. These domains are further broken down into 11 secondary aspects and 29 tertiary tasks for a granular evaluation. A thorough assessment of general and domain-specific (DS) VLMs on this benchmark reveals both their strengths and critical limitations in AD contexts. To further exploit the cognitive and reasoning interactions among the 5 domains for AD understanding, we start from a small-scale VLM and train the DS models on individual domain datasets (collected from 1.4M DS QAs across public sources). The experimental results demonstrate that the proposed benchmark provides a crucial step toward a more comprehensive assessment of VLMs in AD, paving the way for the development of more cognitively sophisticated and reasoning-capable AD systems.
End-to-end (E2E) autonomous driving methods still struggle to make correct decisions in interactive closed-loop evaluation due to limited causal reasoning capability. Current methods attempt to leverage the powerful understanding and reasoning abilities of Vision-Language Models (VLMs) to resolve this dilemma. However, the problem is still open that few VLMs for E2E methods perform well in the closed-loop evaluation due to the gap between the semantic reasoning space and the purely numerical trajectory output in the action space. To tackle this issue, we propose ORION, a holistic E2E autonomous driving framework by vision-language instructed action generation. ORION uniquely combines a QT-Former to aggregate long-term history context, a Large Language Model (LLM) for driving scenario reasoning, and a generative planner for precision trajectory prediction. ORION further aligns the reasoning space and the action space to implement a unified E2E optimization for both visual question-answering (VQA) and planning tasks. Our method achieves an impressive closed-loop performance of 77.74 Driving Score (DS) and 54.62% Success Rate (SR) on the challenge Bench2Drive datasets, which outperforms state-of-the-art (SOTA) methods by a large margin of 14.28 DS and 19.61% SR.
Vision-Language Models (VLMs) have recently emerged as powerful tools, excelling in tasks that integrate visual and textual comprehension, such as image captioning, visual question answering, and image-text retrieval. However, existing benchmarks for VLMs include spatial components, which often fail to isolate spatial reasoning from related tasks such as object detection or semantic comprehension. In this paper, we address these deficiencies with a multi-faceted approach towards understanding spatial reasoning. Informed by the diverse and multi-dimensional nature of human spatial reasoning abilities, we present a detailed analysis that first delineates the core elements of spatial reasoning: spatial relations, orientation and navigation, mental rotation, and spatial visualization, and then assesses the performance of these models in both synthetic and real-world images, bridging controlled and naturalistic contexts. We analyze 13 state-of-the-art Vision-Language Models, uncovering pivotal insights into their spatial reasoning performance. Our results reveal profound shortcomings in current VLMs, with average accuracy across the 13 models approximating random chance, highlighting spatial reasoning as a persistent obstacle. This work not only exposes the pressing need to advance spatial reasoning within VLMs but also establishes a solid platform for future exploration. Code available on GitHub (https://github.com/stogiannidis/srbench) and dataset available on HuggingFace (https://huggingface.co/datasets/stogiannidis/srbench).
Recent advancements in vision-language models have achieved remarkable results in making language models understand vision inputs. However, a unified approach to align these models across diverse tasks such as image captioning and visual question answering remains a challenge. Existing methods either require very big language models or very big datasets which is not efficient in utilizing existing models. This paper addresses this gap and devises a training strategy of auto-regressive vision-language models, to unify vision-language tasks like image-captioning and visual question answering. We propose four training stages for aligning the vision model with the language model, in other words, the language model is given an ability to process visual inputs. We also devise different attention masks for training transformer-based language models that improve the quality of visual features. Further, we introduce some findings, 1) the attention mask should not be applied on visual inputs, 2) the Language model converges faster on AI- generated data, 3) More work should be done in the alignment stage during the pre-training of the model, 4) the model can easily adapt to any downstream tasks like visual question answering on healthcare datasets like PathVQA. After training the model for one epoch for all the stages, it outperforms large models like VILA-13 billion models on common benchmarks like CIDEr scores on COCO and Flickr30k datasets and achieves very close scores to GIT-2 on the same dataset despite being a much smaller model trained on a much smaller dataset. All of the training is done using best practices available like multi- GPU parallel training, lower-precision training with 16-bit float numbers, faster attention (SDPA), and gradient accumulation, and completed the training within 12 hours.
Multimodal learning combining pathology images and genomic sequences enhances cancer survival analysis but faces clinical implementation barriers due to limited access to genomic sequencing in under-resourced regions. To enable survival prediction using only whole-slide images (WSI), we propose the Visual-Genomic Answering-Guided Transformer (VGAT), a framework integrating Visual Question Answering (VQA) techniques for genomic modality reconstruction. By adapting VQA's text feature extraction approach, we derive stable genomic representations that circumvent dimensionality challenges in raw genomic data. Simultaneously, a cluster-based visual prompt module selectively enhances discriminative WSI patches, addressing noise from unfiltered image regions. Evaluated across five TCGA datasets, VGAT outperforms existing WSI-only methods, demonstrating the viability of genomic-informed inference without sequencing. This approach bridges multimodal research and clinical feasibility in resource-constrained settings. The code link is https://github.com/CZZZZZZZZZZZZZZZZZ/VGAT.
Recent advancements in 2D and multimodal models have achieved remarkable success by leveraging large-scale training on extensive datasets. However, extending these achievements to enable free-form interactions and high-level semantic operations with complex 3D/4D scenes remains challenging. This difficulty stems from the limited availability of large-scale, annotated 3D/4D or multi-view datasets, which are crucial for generalizable vision and language tasks such as open-vocabulary and prompt-based segmentation, language-guided editing, and visual question answering (VQA). In this paper, we introduce Feature4X, a universal framework designed to extend any functionality from 2D vision foundation model into the 4D realm, using only monocular video input, which is widely available from user-generated content. The "X" in Feature4X represents its versatility, enabling any task through adaptable, model-conditioned 4D feature field distillation. At the core of our framework is a dynamic optimization strategy that unifies multiple model capabilities into a single representation. Additionally, to the best of our knowledge, Feature4X is the first method to distill and lift the features of video foundation models (e.g., SAM2, InternVideo2) into an explicit 4D feature field using Gaussian Splatting. Our experiments showcase novel view segment anything, geometric and appearance scene editing, and free-form VQA across all time steps, empowered by LLMs in feedback loops. These advancements broaden the scope of agentic AI applications by providing a foundation for scalable, contextually and spatiotemporally aware systems capable of immersive dynamic 4D scene interaction.
The rapid development of multimodal large language models has resulted in remarkable advancements in visual perception and understanding, consolidating several tasks into a single visual question-answering framework. However, these models are prone to hallucinations, which limit their reliability as artificial intelligence systems. While this issue is extensively researched in natural language processing and image captioning, there remains a lack of investigation of hallucinations in Low-level Visual Perception and Understanding (HLPU), especially in the context of image quality assessment tasks. We consider that these hallucinations arise from an absence of clear self-awareness within the models. To address this issue, we first introduce the HLPU instruction database, the first instruction database specifically focused on hallucinations in low-level vision tasks. This database contains approximately 200K question-answer pairs and comprises four subsets, each covering different types of instructions. Subsequently, we propose the Self-Awareness Failure Elimination (SAFEQA) model, which utilizes image features, salient region features and quality features to improve the perception and comprehension abilities of the model in low-level vision tasks. Furthermore, we propose the Enhancing Self-Awareness Preference Optimization (ESA-PO) framework to increase the model's awareness of knowledge boundaries, thereby mitigating the incidence of hallucination. Finally, we conduct comprehensive experiments on low-level vision tasks, with the results demonstrating that our proposed method significantly enhances self-awareness of the model in these tasks and reduces hallucinations. Notably, our proposed method improves both accuracy and self-awareness of the proposed model and outperforms close-source models in terms of various evaluation metrics.
Multimodal large language models (MLLMs) have demonstrated significant potential in medical Visual Question Answering (VQA). Yet, they remain prone to hallucinations-incorrect responses that contradict input images, posing substantial risks in clinical decision-making. Detecting these hallucinations is essential for establishing trust in MLLMs among clinicians and patients, thereby enabling their real-world adoption. Current hallucination detection methods, especially semantic entropy (SE), have demonstrated promising hallucination detection capacity for LLMs. However, adapting SE to medical MLLMs by incorporating visual perturbations presents a dilemma. Weak perturbations preserve image content and ensure clinical validity, but may be overlooked by medical MLLMs, which tend to over rely on language priors. In contrast, strong perturbations can distort essential diagnostic features, compromising clinical interpretation. To address this issue, we propose Vision Amplified Semantic Entropy (VASE), which incorporates weak image transformations and amplifies the impact of visual input, to improve hallucination detection in medical VQA. We first estimate the semantic predictive distribution under weak visual transformations to preserve clinical validity, and then amplify visual influence by contrasting this distribution with that derived from a distorted image. The entropy of the resulting distribution is estimated as VASE. Experiments on two medical open-ended VQA datasets demonstrate that VASE consistently outperforms existing hallucination detection methods.
Preference alignment has emerged as an effective strategy to enhance the performance of Multimodal Large Language Models (MLLMs) following supervised fine-tuning. While existing preference alignment methods predominantly target hallucination factors, they overlook the factors essential for multi-modal comprehension capabilities, often narrowing their improvements on hallucination mitigation. To bridge this gap, we propose Instruction-oriented Preference Alignment (IPA), a scalable framework designed to automatically construct alignment preferences grounded in instruction fulfillment efficacy. Our method involves an automated preference construction coupled with a dedicated verification process that identifies instruction-oriented factors, avoiding significant variability in response representations. Additionally, IPA incorporates a progressive preference collection pipeline, further recalling challenging samples through model self-evolution and reference-guided refinement. Experiments conducted on Qwen2VL-7B demonstrate IPA's effectiveness across multiple benchmarks, including hallucination evaluation, visual question answering, and text understanding tasks, highlighting its capability to enhance general comprehension.
Multi-step spatial reasoning entails understanding and reasoning about spatial relationships across multiple sequential steps, which is crucial for tackling complex real-world applications, such as robotic manipulation, autonomous navigation, and automated assembly. To assess how well current Multimodal Large Language Models (MLLMs) have acquired this fundamental capability, we introduce \textbf{LEGO-Puzzles}, a scalable benchmark designed to evaluate both \textbf{spatial understanding} and \textbf{sequential reasoning} in MLLMs through LEGO-based tasks. LEGO-Puzzles consists of 1,100 carefully curated visual question-answering (VQA) samples spanning 11 distinct tasks, ranging from basic spatial understanding to complex multi-step reasoning. Based on LEGO-Puzzles, we conduct a comprehensive evaluation of state-of-the-art MLLMs and uncover significant limitations in their spatial reasoning capabilities: even the most powerful MLLMs can answer only about half of the test cases, whereas human participants achieve over 90\% accuracy. In addition to VQA tasks, we evaluate MLLMs' abilities to generate LEGO images following assembly illustrations. Our experiments show that only Gemini-2.0-Flash and GPT-4o exhibit a limited ability to follow these instructions, while other MLLMs either replicate the input image or generate completely irrelevant outputs. Overall, LEGO-Puzzles exposes critical deficiencies in existing MLLMs' spatial understanding and sequential reasoning capabilities, and underscores the need for further advancements in multimodal spatial reasoning.
Pre-trained video large language models (Video LLMs) exhibit remarkable reasoning capabilities, yet adapting these models to new tasks involving additional modalities or data types (e.g., audio or 3D information) remains challenging. In this paper, we present PAVE, a flexible framework for adapting pre-trained Video LLMs to downstream tasks with side-channel signals, such as audio, 3D cues, or multi-view videos. PAVE introduces lightweight adapters, referred to as "patches," which add a small number of parameters and operations to a base model without modifying its architecture or pre-trained weights. In doing so, PAVE can effectively adapt the pre-trained base model to support diverse downstream tasks, including audio-visual question answering, 3D reasoning, multi-view video recognition, and high frame rate video understanding. Across these tasks, PAVE significantly enhances the performance of the base model, surpassing state-of-the-art task-specific models while incurring a minor cost of ~0.1% additional FLOPs and parameters. Further, PAVE supports multi-task learning and generalizes well across different Video LLMs. Our code is available at https://github.com/dragonlzm/PAVE.
Vision-language models (VLMs) have shown promise in 2D medical image analysis, but extending them to 3D remains challenging due to the high computational demands of volumetric data and the difficulty of aligning 3D spatial features with clinical text. We present Med3DVLM, a 3D VLM designed to address these challenges through three key innovations: (1) DCFormer, an efficient encoder that uses decomposed 3D convolutions to capture fine-grained spatial features at scale; (2) SigLIP, a contrastive learning strategy with pairwise sigmoid loss that improves image-text alignment without relying on large negative batches; and (3) a dual-stream MLP-Mixer projector that fuses low- and high-level image features with text embeddings for richer multi-modal representations. We evaluate our model on the M3D dataset, which includes radiology reports and VQA data for 120,084 3D medical images. Results show that Med3DVLM achieves superior performance across multiple benchmarks. For image-text retrieval, it reaches 61.00% R@1 on 2,000 samples, significantly outperforming the current state-of-the-art M3D model (19.10%). For report generation, it achieves a METEOR score of 36.42% (vs. 14.38%). In open-ended visual question answering (VQA), it scores 36.76% METEOR (vs. 33.58%), and in closed-ended VQA, it achieves 79.95% accuracy (vs. 75.78%). These results highlight Med3DVLM's ability to bridge the gap between 3D imaging and language, enabling scalable, multi-task reasoning across clinical applications. Our code is publicly available at https://github.com/mirthAI/Med3DVLM.
The increasing use of vision-language models (VLMs) in healthcare applications presents great challenges related to hallucinations, in which the models may generate seemingly plausible results that are in fact incorrect. Such hallucinations can jeopardize clinical decision making, potentially harming the diagnosis and treatments. In this work, we propose MedHallTune, a large-scale benchmark designed specifically to evaluate and mitigate hallucinations in medical VLMs. Comprising over 100,000 images and 1,000,000 instruction pairs, MedHallTune includes both hallucination and non-hallucination samples, each with ground-truth annotations. We conduct a comprehensive evaluation of current medical and general VLMs using MedHallTune, assessing their performance across key metrics, including clinical accuracy, relevance, detail level, and risk level. The experimental results show that fine-tuning with MedHallTune successfully improves the ability of several existing models to manage hallucinations and boost their zero-shot performance on downstream visual-question-answering (VQA) tasks, making them more reliable for practical medical applications. Our work contributes to the development of more trustworthy VLMs. Codes and dataset will be available at \href{https://github.com/russellyq/MedHallTune}{MedHallTune}.
Visual Question Answering (VQA) requires reasoning across visual and textual modalities, yet Large Vision-Language Models (LVLMs) often lack integrated commonsense knowledge, limiting their robustness in real-world scenarios. To address this, we introduce MAGIC-VQA, a novel framework that enhances VQA by systematically integrating commonsense knowledge with LVLMs. MAGIC-VQA employs a three-stage process: (1) Explicit Knowledge Integration from external sources, (2) By-Type Post-Processing for contextual refinement, and (3) Implicit Knowledge Augmentation using a Graph Neural Network (GNN) for structured reasoning. While GNNs bring greater depth to structured inference, they enable superior relational inference beyond LVLMs. MAGIC-VQA bridges a key gap by unifying commonsensse knowledge with LVLM-driven reasoning, eliminating the need for extensive pre-training or complex prompt tuning. Our framework achieves state-of-the-art performance on benchmark datasets, significantly improving commonsense reasoning in VQA.
Current vision-language models (VLMs) show exceptional abilities across diverse tasks, such as visual question answering. To enhance user experience, recent studies investigate VLM personalization to understand user-provided concepts. However, they mainly focus on single-concept personalization, neglecting the existence and interplay of multiple concepts, which limits real-world applicability. This paper proposes the first multi-concept personalization paradigm, MC-LLaVA. Specifically, MC-LLaVA employs a multi-concept instruction tuning strategy, effectively integrating multiple concepts in a single training step. To reduce the costs related to joint training, we propose a personalized textual prompt that uses visual token information to initialize concept tokens. Additionally, we introduce a personalized visual prompt during inference, aggregating location confidence maps for enhanced recognition and grounding capabilities. To advance multi-concept personalization research, we further contribute a high-quality instruction tuning dataset. We carefully collect images with multiple characters and objects from movies and manually generate question-answer samples for multi-concept scenarios, featuring superior diversity. Comprehensive qualitative and quantitative experiments demonstrate that MC-LLaVA can achieve impressive multi-concept personalized responses, paving the way for VLMs to become better user-specific assistants. The code and dataset will be publicly available at $\href{https://github.com/arctanxarc/MC-LLaVA}{https://github.com/arctanxarc/MC-LLaVA}$.