Multimodal Large Language Models (MLLMs) have demonstrated remarkable advancements in multimodal understanding; however, their fundamental visual cognitive abilities remain largely underexplored. To bridge this gap, we introduce VisFactor, a novel benchmark derived from the Factor-Referenced Cognitive Test (FRCT), a well-established psychometric assessment of human cognition. VisFactor digitalizes vision-related FRCT subtests to systematically evaluate MLLMs across essential visual cognitive tasks including spatial reasoning, perceptual speed, and pattern recognition. We present a comprehensive evaluation of state-of-the-art MLLMs, such as GPT-4o, Gemini-Pro, and Qwen-VL, using VisFactor under diverse prompting strategies like Chain-of-Thought and Multi-Agent Debate. Our findings reveal a concerning deficiency in current MLLMs' fundamental visual cognition, with performance frequently approaching random guessing and showing only marginal improvements even with advanced prompting techniques. These results underscore the critical need for focused research to enhance the core visual reasoning capabilities of MLLMs. To foster further investigation in this area, we release our VisFactor benchmark at https://github.com/CUHK-ARISE/VisFactor.
Traditional evaluations of multimodal large language models (LLMs) have been limited by their focus on single-image reasoning, failing to assess crucial aspects like contextual understanding, reasoning stability, and uncertainty calibration. This study addresses these limitations by introducing a novel benchmark that integrates multi-image reasoning tasks with rejection-based evaluation and positional bias detection. To evaluate these dimensions, we further introduce entropy as a novel metric for quantifying reasoning consistency across reordered answer variants. We applied this benchmark to assess Grok 3, ChatGPT-4o, ChatGPT-o1, Gemini 2.0 Flash Experimental, DeepSeek Janus models, Qwen2.5-VL-72B-Instruct, QVQ-72B-Preview, and Pixtral 12B across eight visual reasoning tasks, including difference spotting and diagram interpretation. Our findings reveal ChatGPT-o1 leading in overall accuracy (82.5\%) and rejection accuracy (70.0\%), closely followed by Gemini 2.0 Flash Experimental (70.8\%). QVQ-72B-Preview demonstrated superior rejection accuracy (85.5\%). Notably, Pixtral 12B (51.7\%) showed promise in specific domains, while Janus models exhibited challenges in bias and uncertainty calibration, reflected in low rejection accuracies and high entropy scores. High entropy scores in Janus models (Janus 7B: 0.8392, Janus 1B: 0.787) underscore their susceptibility to positional bias and unstable reasoning, contrasting with the low entropy and robust reasoning of ChatGPT models. The study further demonstrates that model size is not the sole determinant of performance, as evidenced by Grok 3 underperformance despite its substantial parameter count. By employing multi-image contexts, rejection mechanisms, and entropy-based consistency metrics, this benchmark sets a new standard for evaluating multimodal LLMs, enabling a more robust and reliable assessment of next-generation AI systems.
Despite strong performance on vision-language tasks, Multimodal Large Language Models (MLLMs) struggle with mathematical problem-solving, with both open-source and state-of-the-art models falling short of human performance on visual-math benchmarks. To systematically examine visual-mathematical reasoning in MLLMs, we (1) evaluate their understanding of geometric primitives, (2) test multi-step reasoning, and (3) explore a potential solution to improve visual reasoning capabilities. Our findings reveal fundamental shortcomings in shape recognition, with top models achieving under 50% accuracy in identifying regular polygons. We analyze these failures through the lens of dual-process theory and show that MLLMs rely on System 1 (intuitive, memorized associations) rather than System 2 (deliberate reasoning). Consequently, MLLMs fail to count the sides of both familiar and novel shapes, suggesting they have neither learned the concept of sides nor effectively process visual inputs. Finally, we propose Visually Cued Chain-of-Thought (VC-CoT) prompting, which enhances multi-step mathematical reasoning by explicitly referencing visual annotations in diagrams, boosting GPT-4o's accuracy on an irregular polygon side-counting task from 7% to 93%. Our findings suggest that System 2 reasoning in MLLMs remains an open problem, and visually-guided prompting is essential for successfully engaging visual reasoning. Code available at: https://github.com/rsinghlab/Shape-Blind.
A scene graph is a key image representation in visual reasoning. The generalisability of Scene Graph Generation (SGG) methods is crucial for reliable reasoning and real-world applicability. However, imbalanced training datasets limit this, underrepresenting meaningful visual relationships. Current SGG methods using external knowledge sources face limitations due to these imbalances or restricted relationship coverage, impacting their reasoning and generalisation capabilities. We propose a novel neurosymbolic approach that integrates data-driven object detection with heterogeneous knowledge graph-based object refinement and zero-shot relationship retrieval, highlighting the loosely coupled synergy between neural and symbolic components. This combination addresses the limitations of imbalanced training datasets in scene graph generation and enables effective prediction of unseen visual relationships. Objects are detected using a region-based deep neural network and refined based on their positional and structural similarity, followed by retrieval of pairwise visual relationships using a heterogeneous knowledge graph. The redundant and irrelevant visual relationships are discarded based on the similarity of relationship labels and node embeddings. Finally, the visual relationships are interlinked to generate the scene graph. The employed heterogeneous knowledge graph combines diverse knowledge sources, offering rich common sense knowledge about objects and their interactions in the world. Our method, evaluated using the benchmark Visual Genome dataset and zero-shot recall (zR@K) metric, shows a 59.96% improvement over existing state-of-the-art methods, highlighting its effectiveness in generalised SGG. The object refinement step effectively improved the object detection performance by 57.1%. Additional evaluation using the GQA dataset confirms the cross-dataset generalisability of our method. We also compared various knowledge sources and embedding models to determine an optimal combination for zero-shot SGG. The source code is available at https://github.com/jaleedkhan/zsrr-sgg.
Large multimodal models (LMMs) have shown impressive capabilities in a wide range of visual tasks. However, they often struggle with fine-grained visual reasoning, failing to identify domain-specific objectives and provide justifiable explanations for their predictions. To address this, we propose a novel visual rejection sampling framework to improve the cognition and explainability of LMMs using self-synthesized data. Specifically, visual fine-tuning requires images, queries, and target answers. Our approach begins by synthesizing interpretable answers that include human-verifiable visual features. These features are based on expert-defined concepts, carefully selected based on their alignment with the image content. After each round of fine-tuning, we apply a reward model-free filtering mechanism to select the highest-quality interpretable answers for the next round of tuning. This iterative process of data synthesis and fine-tuning progressively improves the model's ability to generate accurate and reasonable explanations. Experimental results demonstrate the effectiveness of our method in improving both the accuracy and explainability of specialized visual classification tasks.
Embodied Question Answering (EQA) has primarily focused on indoor environments, leaving the complexities of urban settings - spanning environment, action, and perception - largely unexplored. To bridge this gap, we introduce CityEQA, a new task where an embodied agent answers open-vocabulary questions through active exploration in dynamic city spaces. To support this task, we present CityEQA-EC, the first benchmark dataset featuring 1,412 human-annotated tasks across six categories, grounded in a realistic 3D urban simulator. Moreover, we propose Planner-Manager-Actor (PMA), a novel agent tailored for CityEQA. PMA enables long-horizon planning and hierarchical task execution: the Planner breaks down the question answering into sub-tasks, the Manager maintains an object-centric cognitive map for spatial reasoning during the process control, and the specialized Actors handle navigation, exploration, and collection sub-tasks. Experiments demonstrate that PMA achieves 60.7% of human-level answering accuracy, significantly outperforming frontier-based baselines. While promising, the performance gap compared to humans highlights the need for enhanced visual reasoning in CityEQA. This work paves the way for future advancements in urban spatial intelligence. Dataset and code are available at https://github.com/BiluYong/CityEQA.git.
Large Language Models (LLMs) have demonstrated impressive capabilities in language processing, yet they often struggle with tasks requiring genuine visual spatial reasoning. In this paper, we introduce a novel two-stage training framework designed to equip standard LLMs with visual reasoning abilities for maze navigation. First, we leverage Supervised Fine Tuning (SFT) on a curated dataset of tokenized maze representations to teach the model to predict step-by-step movement commands. Next, we apply Group Relative Policy Optimization (GRPO)-a technique used in DeepSeekR1-with a carefully crafted reward function to refine the model's sequential decision-making and encourage emergent chain-of-thought behaviors. Experimental results on synthetically generated mazes show that while a baseline model fails to navigate the maze, the SFT-trained model achieves 86% accuracy, and further GRPO fine-tuning boosts accuracy to 93%. Qualitative analyses reveal that GRPO fosters more robust and self-corrective reasoning, highlighting the potential of our approach to bridge the gap between language models and visual spatial tasks. These findings offer promising implications for applications in robotics, autonomous navigation, and other domains that require integrated visual and sequential reasoning.
Test time scaling is currently one of the most active research areas that shows promise after training time scaling has reached its limits. Deep-thinking (DT) models are a class of recurrent models that can perform easy-to-hard generalization by assigning more compute to harder test samples. However, due to their inability to determine the complexity of a test sample, DT models have to use a large amount of computation for both easy and hard test samples. Excessive test time computation is wasteful and can cause the ``overthinking'' problem where more test time computation leads to worse results. In this paper, we introduce a test time training method for determining the optimal amount of computation needed for each sample during test time. We also propose Conv-LiGRU, a novel recurrent architecture for efficient and robust visual reasoning. Extensive experiments demonstrate that Conv-LiGRU is more stable than DT, effectively mitigates the ``overthinking'' phenomenon, and achieves superior accuracy.
Visual instruction tuning has become the predominant technology in eliciting the multimodal task-solving capabilities of large vision-language models (LVLMs). Despite the success, as visual instructions require images as the input, it would leave the gap in inheriting the task-solving capabilities from the backbone LLMs, and make it costly to collect a large-scale dataset. To address it, we propose ViFT, a visual instruction-free fine-tuning framework for LVLMs. In ViFT, we only require the text-only instructions and image caption data during training, to separately learn the task-solving and visual perception abilities. During inference, we extract and combine the representations of the text and image inputs, for fusing the two abilities to fulfill multimodal tasks. Experimental results demonstrate that ViFT can achieve state-of-the-art performance on several visual reasoning and visual instruction following benchmarks, with rather less training data. Our code and data will be publicly released.
Entity tracking is a fundamental challenge in natural language understanding, requiring models to maintain coherent representations of entities. Previous work has benchmarked entity tracking performance in purely text-based tasks. We introduce MET-Bench, a multimodal entity tracking benchmark designed to evaluate the ability of vision-language models to track entity states across modalities. Using two structured domains, Chess and the Shell Game, we assess how effectively current models integrate textual and image-based state updates. Our findings reveal a significant performance gap between text-based and image-based tracking and that this performance gap stems from deficits in visual reasoning rather than perception. We further show that explicit text-based reasoning strategies improve performance, yet substantial limitations remain, especially in long-horizon multimodal scenarios. Our results highlight the need for improved multimodal representations and reasoning techniques to bridge the gap between textual and visual entity tracking.
Visual reasoning -- the ability to interpret the visual world -- is crucial for embodied agents that operate within three-dimensional scenes. Progress in AI has led to vision and language models capable of answering questions from images. However, their performance declines when tasked with 3D spatial reasoning. To tackle the complexity of such reasoning problems, we introduce an agentic program synthesis approach where LLM agents collaboratively generate a Pythonic API with new functions to solve common subproblems. Our method overcomes limitations of prior approaches that rely on a static, human-defined API, allowing it to handle a wider range of queries. To assess AI capabilities for 3D understanding, we introduce a new benchmark of queries involving multiple steps of grounding and inference. We show that our method outperforms prior zero-shot models for visual reasoning in 3D and empirically validate the effectiveness of our agentic framework for 3D spatial reasoning tasks. Project website: https://glab-caltech.github.io/vadar/
Multimodal large language models (MLLMs) exhibit impressive capabilities but still face challenges in complex visual reasoning. While recent efforts attempt to enhance MLLMs' reasoning by incorporating OpenAI o1-like structured thinking through explicit search structures or teacher-guided distillation, they often struggle to balance performance and efficiency. A critical limitation is their heavy reliance on extensive data and search spaces, resulting in low-efficiency implicit insight extraction and data utilization. To address this, we propose AStar, an Automated Structured thinking paradigm for multimodal reasoning via Monte Carlo Tree Search (MCTS). AStar automatically derives high-level cognitive reasoning patterns from limited data using MCTS-powered hierarchical structures. Building on these explicit patterns, we design a unified reasoning framework that seamlessly integrates models' internal reasoning capabilities and external reasoning guidelines, enabling efficient inference with minimal tree iterations. This novel paradigm strikes a compelling balance between performance and efficiency. Extensive experiments demonstrate AStar's effectiveness, achieving superior accuracy (54.0$\%$) on the MathVerse benchmark with a 7B backbone, surpassing GPT-4o (50.2$\%$) while maintaining substantial data and computational efficiency.
The recent advancements in visual reasoning capabilities of large multimodal models (LMMs) and the semantic enrichment of 3D feature fields have expanded the horizons of robotic capabilities. These developments hold significant potential for bridging the gap between high-level reasoning from LMMs and low-level control policies utilizing 3D feature fields. In this work, we introduce LMM-3DP, a framework that can integrate LMM planners and 3D skill Policies. Our approach consists of three key perspectives: high-level planning, low-level control, and effective integration. For high-level planning, LMM-3DP supports dynamic scene understanding for environment disturbances, a critic agent with self-feedback, history policy memorization, and reattempts after failures. For low-level control, LMM-3DP utilizes a semantic-aware 3D feature field for accurate manipulation. In aligning high-level and low-level control for robot actions, language embeddings representing the high-level policy are jointly attended with the 3D feature field in the 3D transformer for seamless integration. We extensively evaluate our approach across multiple skills and long-horizon tasks in a real-world kitchen environment. Our results show a significant 1.45x success rate increase in low-level control and an approximate 1.5x improvement in high-level planning accuracy compared to LLM-based baselines. Demo videos and an overview of LMM-3DP are available at https://lmm-3dp-release.github.io.
Large Vision-Language Models (VLMs) have achieved remarkable performance across a wide range of tasks. However, their deployment in safety-critical domains poses significant challenges. Existing safety fine-tuning methods, which focus on textual or multimodal content, fall short in addressing challenging cases or disrupt the balance between helpfulness and harmlessness. Our evaluation highlights a safety reasoning gap: these methods lack safety visual reasoning ability, leading to such bottlenecks. To address this limitation and enhance both visual perception and reasoning in safety-critical contexts, we propose a novel dataset that integrates multi-image inputs with safety Chain-of-Thought (CoT) labels as fine-grained reasoning logic to improve model performance. Specifically, we introduce the Multi-Image Safety (MIS) dataset, an instruction-following dataset tailored for multi-image safety scenarios, consisting of training and test splits. Our experiments demonstrate that fine-tuning InternVL2.5-8B with MIS significantly outperforms both powerful open-source models and API-based models in challenging multi-image tasks requiring safety-related visual reasoning. This approach not only delivers exceptional safety performance but also preserves general capabilities without any trade-offs. Specifically, fine-tuning with MIS increases average accuracy by 0.83% across five general benchmarks and reduces the Attack Success Rate (ASR) on multiple safety benchmarks by a large margin.
3D vision and spatial reasoning have long been recognized as preferable for accurately perceiving our three-dimensional world, especially when compared with traditional visual reasoning based on 2D images. Due to the difficulties in collecting high-quality 3D data, research in this area has only recently gained momentum. With the advent of powerful large language models (LLMs), multi-modal LLMs for 3D vision have been developed over the past few years. However, most of these models focus primarily on the vision encoder for 3D data. In this paper, we propose converting existing densely activated LLMs into mixture-of-experts (MoE) models, which have proven effective for multi-modal data processing. In addition to leveraging these models' instruction-following capabilities, we further enable embodied task planning by attaching a diffusion head, Pose-DiT, that employs a novel rectified flow diffusion scheduler. Experimental results on 3D question answering and task-planning tasks demonstrate that our 3D-MoE framework achieves improved performance with fewer activated parameters.
A fundamental challenge in artificial intelligence involves understanding the cognitive mechanisms underlying visual reasoning in sophisticated models like Vision-Language Models (VLMs). How do these models integrate visual perception with abstract thought, especially when reasoning across multiple images or requiring fine-grained compositional understanding? Drawing inspiration from cognitive science, this paper introduces a structured evaluation framework using diverse visual reasoning tasks-Bongard Problems (BPs) and Winoground-to dissect the perception-reasoning interface in VLMs. We propose three distinct evaluation paradigms, mirroring human problem-solving strategies: Direct Visual Rule Learning (DVRL; holistic processing), Deductive Rule Learning (DRL; rule extraction and application), and Componential Analysis (CA; analytical decomposition via task-agnostic textual descriptions). These paradigms systematically vary cognitive load and probe processing stages. Notably, CA enables multi-image reasoning evaluation even for single-image architectures and isolates reasoning from perception by operating on textual descriptions. Applying this framework, we demonstrate that CA, leveraging powerful language models for reasoning over rich, independently generated descriptions, achieves new state-of-the-art (SOTA) performance on challenging benchmarks including Bongard-OpenWorld, Bongard-HOI, and Winoground. Ablation studies confirm reasoning improves significantly when perceptual challenges are mitigated, revealing a critical perception bottleneck. Our framework provides a valuable diagnostic tool and suggests that decoupling perception (via rich, task-agnostic description) from reasoning is a promising direction for robust and general visual intelligence.
Pre-trained on extensive text and image corpora, current Multi-Modal Large Language Models (MLLM) have shown strong capabilities in general visual reasoning tasks. However, their performance is still lacking in physical domains that require understanding diagrams with complex physical structures and quantitative analysis based on multi-modal information. To address this, we develop a new framework, named Multi-Modal Scientific Reasoning with Physics Perception and Simulation (MAPS) based on an MLLM. MAPS decomposes expert-level multi-modal reasoning task into physical diagram understanding via a Physical Perception Model (PPM) and reasoning with physical knowledge via a simulator. The PPM module is obtained by fine-tuning a visual language model using carefully designed synthetic data with paired physical diagrams and corresponding simulation language descriptions. At the inference stage, MAPS integrates the simulation language description of the input diagram provided by PPM and results obtained through a Chain-of-Simulation process with MLLM to derive the underlying rationale and the final answer. Validated using our collected college-level circuit analysis problems, MAPS significantly improves reasoning accuracy of MLLM and outperforms all existing models. The results confirm MAPS offers a promising direction for enhancing multi-modal scientific reasoning ability of MLLMs. We will release our code, model and dataset used for our experiments upon publishing of this paper.
In abstract visual reasoning, monolithic deep learning models suffer from limited interpretability and generalization, while existing neuro-symbolic approaches fall short in capturing the diversity and systematicity of attributes and relation representations. To address these challenges, we propose a Systematic Abductive Reasoning model with diverse relation representations (Rel-SAR) in Vector-symbolic Architecture (VSA) to solve Raven's Progressive Matrices (RPM). To derive attribute representations with symbolic reasoning potential, we introduce not only various types of atomic vectors that represent numeric, periodic and logical semantics, but also the structured high-dimentional representation (SHDR) for the overall Grid component. For systematic reasoning, we propose novel numerical and logical relation functions and perform rule abduction and execution in a unified framework that integrates these relation representations. Experimental results demonstrate that Rel-SAR achieves significant improvement on RPM tasks and exhibits robust out-of-distribution generalization. Rel-SAR leverages the synergy between HD attribute representations and symbolic reasoning to achieve systematic abductive reasoning with both interpretable and computable semantics.
Localizing textual descriptions within large-scale 3D scenes presents inherent ambiguities, such as identifying all traffic lights in a city. Addressing this, we introduce a method to generate distributions of camera poses conditioned on textual descriptions, facilitating robust reasoning for broadly defined concepts. Our approach employs a diffusion-based architecture to refine noisy 6DoF camera poses towards plausible locations, with conditional signals derived from pre-trained text encoders. Integration with the pretrained Vision-Language Model, CLIP, establishes a strong linkage between text descriptions and pose distributions. Enhancement of localization accuracy is achieved by rendering candidate poses using 3D Gaussian splatting, which corrects misaligned samples through visual reasoning. We validate our method's superiority by comparing it against standard distribution estimation methods across five large-scale datasets, demonstrating consistent outperformance. Code, datasets and more information will be publicly available at our project page.
Reasoning is a fundamental capability for solving complex multi-step problems, particularly in visual contexts where sequential step-wise understanding is essential. Existing approaches lack a comprehensive framework for evaluating visual reasoning and do not emphasize step-wise problem-solving. To this end, we propose a comprehensive framework for advancing step-by-step visual reasoning in large language models (LMMs) through three key contributions. First, we introduce a visual reasoning benchmark specifically designed to evaluate multi-step reasoning tasks. The benchmark presents a diverse set of challenges with eight different categories ranging from complex visual perception to scientific reasoning with over 4k reasoning steps in total, enabling robust evaluation of LLMs' abilities to perform accurate and interpretable visual reasoning across multiple steps. Second, we propose a novel metric that assesses visual reasoning quality at the granularity of individual steps, emphasizing both correctness and logical coherence. The proposed metric offers deeper insights into reasoning performance compared to traditional end-task accuracy metrics. Third, we present a new multimodal visual reasoning model, named LlamaV-o1, trained using a multi-step curriculum learning approach, where tasks are progressively organized to facilitate incremental skill acquisition and problem-solving. The proposed LlamaV-o1 is designed for multi-step reasoning and learns step-by-step through a structured training paradigm. Extensive experiments show that our LlamaV-o1 outperforms existing open-source models and performs favorably against close-source proprietary models. Compared to the recent Llava-CoT, our LlamaV-o1 achieves an average score of 67.3 with an absolute gain of 3.8\% across six benchmarks while being 5 times faster during inference scaling. Our benchmark, model, and code are publicly available.