Recent advances in large language models (LLMs) have opened new avenues for multimodal reasoning. Yet, most existing methods still rely on pretrained vision-language models (VLMs) to encode image-text pairs in isolation, ignoring the relational structure that real-world multimodal data naturally form. This motivates reasoning on multimodal graphs (MMGs), where each node has textual and visual attributes and edges provide structural cues. Enabling LLM-based reasoning on such heterogeneous multimodal signals while preserving graph topology introduces two key challenges: resolving weak cross-modal consistency and handling heterogeneous modality preference. To address this, we propose Mario, a unified framework that simultaneously resolves the two above challenges and enables effective LLM-based reasoning over MMGs. Mario consists of two innovative stages. Firstly, a graph-conditioned VLM design that jointly refines textual and visual features through fine-grained cross-modal contrastive learning guided by graph topology. Secondly, a modality-adaptive graph instruction tuning mechanism that organizes aligned multimodal features into graph-aware instruction views and employs a learnable router to surface, for each node and its neighborhood, the most informative modality configuration to the LLM. Extensive experiments across diverse MMG benchmarks demonstrate that Mario consistently outperforms state-of-the-art graph models in both supervised and zero-shot scenarios for node classification and link prediction. The code will be made available at https://github.com/sunyuanfu/Mario.
论文检索
输入标题、作者或关键词，从 3,042 篇学术成果中精准定位
Agentic vision-language models are increasingly trained to "think with images" by calling image operations. However, we show that high final-answer accuracy often hides unfaithful visual reasoning: models may invoke tools on irrelevant regions or ignore tool outputs entirely, yet still guess the correct answer. In this work, we first propose a faithfulness evaluation protocol that measures whether intermediate visual tool outputs (e.g., crops) actually contain the queried evidence. This reveals that recent visual agents achieve high final-answer accuracy but exhibit low rates of faithful tool-use on visual search benchmarks. We then introduce CodeV, a code-based visual agent trained with Tool-Aware Policy Optimization (TAPO). TAPO is a process-level RL framework that augments GRPO with dense rewards defined directly on visual tool inputs and outputs, rather than on chain-of-thought tokens, making supervision easier to verify and less susceptible to reward hacking. CodeV represents visual tools as executable Python code, and TAPO assigns step-wise rewards based solely on the question and tool output, encouraging both necessary and evidence-consistent tool use. In a two-stage SFT+RL pipeline, CodeV achieves competitive or superior accuracy while substantially increasing faithful tool-use rates on related visual search benchmarks. Beyond visual search, CodeV attains strong performance on a range of multimodal reasoning and math benchmarks, suggesting that explicitly supervising intermediate tool behavior is crucial for building trustworthy, agentic visual reasoning systems.
Socratic-Geo: Synthetic Data Generation and Cross-Modal Geometric Reasoning via Multi-Agent Interaction
PDF ↗Multimodal Large Language Models (MLLMs) have significantly advanced vision-language understanding. However, even state-of-the-art models struggle with geometric reasoning, revealing a critical bottleneck: the extreme scarcity of high-quality image-text pairs. Human annotation is prohibitively expensive, while automated methods fail to ensure both fidelity and training effectiveness. Existing approaches either passively adapt to available images or employ inefficient random exploration with post-hoc filtering, decoupling generation from learning needs. We propose Socratic-Geo, a fully autonomous framework that dynamically couples data synthesis with model learning through multi-agent interaction. The Teacher agent generates parameterized Python scripts with reflective feedback (Reflect for solvability, RePI for visual validity), ensuring image-text pair purity. The Solver agent optimizes reasoning through preference learning, with failure paths guiding Teacher's targeted augmentation. Independently, the Generator learns image generation capabilities on accumulated "image-code-instruction" triplets, distilling programmatic drawing intelligence into visual generation. Starting from only 108 seed problems, Socratic-Solver achieves 49.11 on six geometric benchmarks using one-quarter of baseline data, surpassing strong baselines by 2.43 points. Socratic-Generator achieves 42.4% on GenExam, establishing new state-of-the-art for open-source models, surpassing Seedream-4.0 (39.8%) and approaching Gemini-2.5-Flash-Image (43.1%).
What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models
PDF ↗Multimodal large language models (MLLMs) project visual tokens into the embedding space of language models, yet the internal structuring and processing of visual semantics remain poorly understood. In this work, we introduce a two-fold analytical framework featuring a novel probing tool, EmbedLens, to conduct a fine-grained analysis. We uncover a pronounced semantic sparsity at the input level: visual tokens consistently partition into sink, dead, and alive categories. Remarkably, only the alive tokens, comprising ~60% of the total input, carry image-specific meaning. Furthermore, using a targeted patch-compression benchmark, we demonstrate that these alive tokens already encode rich, fine-grained cues (e.g., objects, colors, and OCR) prior to entering the LLM. Internal visual computations (such as visual attention and feed-forward networks) are redundant for most standard tasks. For the small subset of highly vision-centric tasks that actually benefit from internal processing, we reveal that alive tokens naturally align with intermediate LLM layers rather than the initial embedding space, indicating that shallow-layer processing is unnecessary and that direct mid-layer injection is both sufficient. Ultimately, our findings provide a unified mechanistic view of visual token processing, paving the way for more efficient and interpretable MLLM architectures through selective token pruning, minimized visual computation, and mid-layer injection.
SemanticVLA: Towards Semantic Reasoning over Action Memorization via Synergistic Explicit Trace and Latent Action Planning
PDF ↗Vision-Language-Action (VLA) models have emerged as a promising paradigm where pretrained Vision-Language Models (VLMs) serve as System 2 for high-level reasoning, connected to action experts as System 1 for low-level motor control.However, current works fail to genuinely leverage VLM capabilities: VLMs produce latent embeddings that lack semantic interpretability, providing ambiguous and unstable guidance to downstream policies, while solely action supervision further causes VLMs to degenerate into parameter-heavy fusion encoders that memorize action patterns rather than perform generalized reasoning.To bridge this gap, we introduce SemanticVLA, which leverages VLM reasoning through synergistic dual-path design. Explicit trace reasoning generates interpretable spatial waypoints as textual coordinate sequences through the VLM's native language interface, directly reusing its pretrained spatial grounding to provide a thinking process for task planning. Latent action tokens complement trace reasoning by learning compact visuomotor primitives grounded in visual observations, providing more fine-grained action representations beyond pure coordinate prediction. This synergy enables trace reasoning to leverage VLM's multimodal understanding for refining latent token prediction, while latent tokens provide stable and grounded guidance that compensates for trace's numerical sensitivity.SemanticVLA achieves 97.0% average success rate on LIBERO and 65.1% on SimplerEnv WidowX, substantially outperforming strong baselines. More importantly, SemanticVLA maintains significantly more stable performance under instruction rephrasing in both simulation suites, and demonstrates strong advantages on real-world long-horizon and reasoning-intensive tasks.By bridging VLM reasoning and action expert through semantically explicit trace and visually grounded latent action tokens, our approach enables genuine reasoning rather than action memorization.
Humans perceive and reason about their surroundings in four dimensions - three spatial and one temporal axis - by building persistent, structured internal representations that encode semantic meaning, spatial layout, and temporal dynamics. These multimodal memories enable them to recall past events, infer unobserved states, and integrate new information into context-dependent reasoning. Inspired by this capability, we introduce R4, a training-free framework for retrieval-augmented reasoning in 4D spatio-temporal space that equips vision-language models (VLMs) with structured, lifelong memory. R4 continuously constructs a 4D knowledge database by anchoring object-level semantic descriptions in metric space and time, yielding a persistent world model that can be shared across agents. At inference, natural language queries are decomposed into semantic, spatial, and temporal keys to retrieve relevant observations, which are integrated into the VLM's reasoning through an iterative retrieval-reasoning loop. Unlike classical retrieval-augmented generation methods, retrieval in R4 operates directly in structured 4D space, enabling episodic and collaborative reasoning without training. Experiments on embodied question answering and navigation benchmarks demonstrate that R4 substantially improves retrieval and reasoning over spatio-temporal information compared to baselines, advancing a new paradigm for embodied 4D reasoning in dynamic environments.
This paper provides a simplification on OpenVision's architecture and loss design for enhancing its training efficiency. Following the prior vision-language pretraining works CapPa and AIMv2, as well as modern multimodal designs like LLaVA, our changes are straightforward: we remove the text encoder (and therefore the contrastive loss), retaining only the captioning loss as a purely generative training signal. We name this new version OpenVision 2. The initial results are promising: despite this simplification, OpenVision 2 competitively matches the original model's performance on a broad set of multimodal benchmarks while substantially cutting both training time and memory consumption. For example, with ViT-L/14, it reduces training time by about 1.5x (from 83h to 57h), and memory usage by about 1.8x (from 24.5GB to 13.8GB, equivalently allowing the maximum batch size to grow from 2k to 8k). This superior training efficiency also allows us to scale far beyond the largest vision encoder used in OpenVision, reaching more than 1 billion parameters. We hold a strong belief that this lightweight, generative-only paradigm is compelling for future vision encoder development in multimodal foundation models.
DPL: Decoupled Prototype Learning for Enhancing Robustness of Vision-Language Transformers to Missing Modalities
PDF ↗The performance of Vision-Language Transformers drops sharply when an input modality (e.g., image) is missing, because the model is forced to make predictions using incomplete information. Existing missing-aware prompt methods help reduce this degradation, but they still rely on conventional prediction heads (e.g., a Fully-Connected layer) that compute class scores in the same way regardless of which modality is present or absent. We introduce Decoupled Prototype Learning (DPL), a new prediction head architecture that explicitly adjusts its decision process to the observed input modalities. For each class, DPL selects a set of prototypes specific to the current missing-modality cases (image-missing, text-missing, or mixed-missing). Each prototype is then decomposed into image-specific and text-specific components, enabling the head to make decisions that depend on the information actually present. This adaptive design allows DPL to handle inputs with missing modalities more effectively while remaining fully compatible with existing prompt-based frameworks. Extensive experiments on MM-IMDb, UPMC Food-101, and Hateful Memes demonstrate that DPL outperforms state-of-the-art approaches across all widely used multimodal image-text datasets and various missing cases.
We introduce two new benchmarks REST and REST+ (Render-Equivalence Stress Tests) to enable systematic evaluation of cross-modal inconsistency in multimodal large language models (MLLMs). MLLMs are trained to represent vision and language in the same embedding space, yet they cannot perform the same tasks in both modalities. Our benchmarks contain samples with the same semantic information in three modalities (image, text, mixed) and we show that state-of-the-art MLLMs cannot consistently reason over these different modalities. We evaluate 15 MLLMs and find that the degree of modality inconsistency varies substantially, even when accounting for problems with text recognition (OCR). Neither rendering text as image nor rendering an image as text solves the inconsistency. Even if OCR is correct, we find that visual characteristics (text colour and resolution, but not font) and the number of vision tokens have an impact on model performance. Finally, we find that our consistency score correlates with the modality gap between text and images, highlighting a mechanistic interpretation of cross-modal inconsistent MLLMs.
ZOO-Prune: Training-Free Token Pruning via Zeroth-Order Gradient Estimation in Vision-Language Models
PDF ↗Large Vision-Language Models (VLMs) enable strong multimodal reasoning but incur heavy inference costs from redundant visual tokens. Token pruning alleviates this issue, yet existing approaches face limitations. Attention-based methods rely on raw attention scores, which are often unstable across layers and heads and can lead to redundant selections. Diversity-based methods improve robustness by selecting tokens far apart in feature space, but risk dropping regions needed for accurate prediction. We propose ZOO-Prune, a training-free framework built on the intuition that highly sensitive tokens have a stronger influence on the model's output and capture complementary visual cues rather than redundant ones. To achieve this, we estimate token sensitivity using zeroth-order perturbations at the lightweight projection layer. This measures how small random perturbations affect the projected features and enables efficient approximation of each token's influence without backpropagation. Extensive experiments across multiple VLMs and benchmarks show that ZOO-Prune consistently outperforms prior methods while pruning up to 94.4% of tokens without sacrificing accuracy. Our method also improves efficiency, reaching up to 2.30x faster end-to-end inference compared to the baseline. Code is available at https://aim-skku.github.io/ZOO-Prune.
UniEdit-I: Training-free Image Editing for Unified VLM via Iterative Understanding, Editing and Verifying
PDF ↗Recent advances in diffusion models and vision-language models (VLMs) have significantly enhanced the controllability of image editing. Methods like FlowEdit enable step-by-step editing along a visible, noise-free trajectory, where each intermediate result is a clear image, eliminating the need for full noise inversion. However, these approaches still operate in pixel space or VAE latent space, where intermediate outputs often suffer from visual artifacts, distortions, or unrealistic details--making reliable semantic evaluation difficult. Furthermore, they remain open-loop systems, applying static edits without feedback to guide or correct the editing process adaptively. We propose UniEdit-I, the first training-free, closed-loop image editing framework that operates entirely within the semantic latent space of a unified VLM by introducing an Understanding-Editing-Verifying (UEV) loop:(1) Understanding: parses the source image and editing instruction into a structured source prompt and a minimal target specification;(2) Editing: applies dynamic semantic offsets, with a configurable feedback weighting mechanism that adaptively modulates editing intensity based on real-time alignment feedback;(3) Verifying: leverages the VLM's own multimodal reasoning capability to evaluate the intermediate output along multiple semantic dimensions and trigger early stopping or refinement. By transforming the VLM from a post-hoc evaluator into an in-process conductor, UniEdit-I establishes the first semantics-driven, self-correcting closed-loop image editing pipeline. Evaluated on GEdit-Bench, UniEdit-I achieves state-of-the-art performance without any fine-tuning or architectural modifications, and even surpasses several large-scale pre-trained editors.
Vision-Language-Action (VLA) models have emerged as a powerful framework that unifies perception, language, and control, enabling robots to perform diverse tasks through multimodal understanding. However, current VLA models typically contain massive parameters and rely heavily on large-scale robot data pretraining, leading to high computational costs during training, as well as limited deployability for real-time inference.Moreover, most training paradigms often degrade the perceptual representations of the Vision-Language backbone, resulting in overfitting and poor generalization to downstream tasks.In this work, we present Evo-1, a lightweight VLA model that reduces computation and improves deployment efficiency, while maintaining strong performance without pretraining on robot data. Evo-1 builds on a native multimodal Vision-Language model (VLM), incorporating a novel cross-modulated diffusion transformer along with an optimized integration module, together forming an effective architecture.We further introduce a two-stage training paradigm that progressively aligns action with perception, preserving the representations of the VLM.Notably, with only 0.77 billion parameters, Evo-1 achieves state-of-the-art results on the Meta-World and RoboTwin suite, surpassing the previous best models by 12.4% and 6.9%, respectively, and also attains a competitive result of 94.8% on LIBERO.In real-world evaluations, Evo-1 attains a 78% success rate with high inference frequency and low memory overhead, outperforming all baseline methods.We release code, data, and model weights to facilitate future research on lightweight and efficient VLA models.
Open-vocabulary object detection (OVOD) aims to detect known and unknown objects in the open world by leveraging text prompts. Benefiting from the emergence of large-scale vision--language pre-trained models, OVOD has demonstrated strong zero-shot generalization capabilities. However, when dealing with camouflaged objects, the detector often fails to distinguish and localize objects because the visual features of the objects and the background are highly similar. To bridge this gap, we construct a benchmark named OVCOD-D by augmenting carefully selected camouflaged object images with fine-grained textual descriptions. Due to the limited scale of available camouflaged object datasets, we adopt detectors pre-trained on large-scale object detection datasets as our baseline methods, as they possess stronger zero-shot generalization ability. In the specificity-aware sub-descriptions generated by multimodal large models, there still exist confusing and overly decorative modifiers. To mitigate such interference, we design a sub-description principal component contrastive fusion strategy that reduces noisy textual components. Furthermore, to address the challenge that the visual features of camouflaged objects are highly similar to those of their surrounding environment, we propose a specificity-guided regional weak alignment and dynamic focusing method, which aims to strengthen the detector's ability to discriminate camouflaged objects from background. Under the open-set evaluation setting, the proposed method achieves an AP of 56.4 on the OVCOD-D benchmark. Project code and benchmark will be released at https://github.com/Zh1fen/SDDF.
Diagnosing and Repairing Unsafe Channels in Vision-Language Models via Causal Discovery and Dual-Modal Safety Subspace Projection
PDF ↗Large Vision-Language Models (LVLMs) have achieved impressive performance across multimodal understanding and reasoning tasks, yet their internal safety mechanisms remain opaque and poorly controlled. In this work, we present a comprehensive framework for diagnosing and repairing unsafe channels within LVLMs. We first perform causal mediation analysis to identify neurons and layers that are causally responsible for unsafe behaviors. Based on these findings, we introduce a dual-modal safety subspace projection method that learns generalized safety subspaces for both visual and textual modalities through generalized eigen-decomposition between benign and malicious activations. During inference, activations are dynamically projected toward these safety subspaces via a hybrid fusion mechanism that adaptively balances visual and textual corrections, effectively suppressing unsafe features while preserving semantic fidelity.Extensive experiments on multiple LVLM safety benchmarks demonstrate that our causal-subspace repair framework significantly enhances safety robustness without degrading general multimodal capabilities, outperforming prior activation steering and alignment-based baselines. Moreover, our approach demonstrates robust transferability, effectively defending against unseen and adaptive attacks.
We introduce OneCAT, a unified multimodal model that seamlessly integrates understanding, generation, and editing within a single decoder-only transformer architecture. OneCAT uniquely eliminates the need for external components such as Vision Transformers (ViT) or vision tokenizer during inference, leading to significant efficiency gains, especially for high-resolution image inputs and outputs. This is achieved through a modality-specific Mixture-of-Experts (MoE) design trained with a unified autoregressive (AR) objective, which also natively supports dynamic resolutions. Furthermore, we pioneer to achieve multi-scale visual autoregressive mechanism within the Large Language Model (LLM) with proposed scale-aware adapter (SAA) that drastically reduces decoding latency compared to diffusion-based methods while maintaining state-of-the-art performance. Our findings demonstrate the powerful potential of pure autoregressive modeling as an elegant foundation for unified multimodal intelligence. As a result, OneCAT outperforms existing unified models across benchmarks for multimodal understanding, generation, and editing.
Reinforcement learning (RL) provides a principled framework for improving vision-language models (VLMs) on complex reasoning tasks. However, existing RL approaches often depend on human-annotated labels or task-specific heuristics to define verifiable rewards--both costly and limited in scalability. We introduce VisPlay, a self-evolving RL framework that enables VLMs to autonomously improve their reasoning capabilities from massive unlabeled image data. Starting from a single base VLM, VisPlay assigns the model into two interacting roles: an Image-Conditioned Questioner that formulates challenging yet answerable visual questions, and a Multimodal Reasoner that generates silver responses. These roles are jointly trained using Group Relative Policy Optimization (GRPO), which uses diversity and difficulty rewards to balance the difficulty of generated questions with the quality of silver answers. VisPlay scales efficiently across two model families. Trained on Qwen2.5-VL and MiMo-VL, VisPlay achieves consistent improvements in visual reasoning, compositional generalization, and hallucination reduction across eight benchmarks including MM-Vet and MMMU, and establishes a scalable path toward self-evolving multimodal intelligence.
DiG: Differential Grounding for Enhancing Fine-Grained Perception in Multimodal Large Language Models
PDF ↗Multimodal Large Language Models have achieved impressive performance on a variety of vision-language tasks, yet their fine-grained visual perception and precise spatial reasoning remain limited. In this work, we introduce DiG (Differential Grounding), a novel proxy task framework where MLLMs learn fine-grained perception by identifying and localizing all differences between similar image pairs without prior knowledge of their number. To support scalable training, we develop an automated 3D rendering-based data generation pipeline that produces high-quality paired images with fully controllable discrepancies. To address the sparsity of difference signals, we further employ curriculum learning that progressively increases complexity from single to multiple differences, enabling stable optimization. Extensive experiments demonstrate that DiG significantly improves model performance across a variety of visual perception benchmarks and that the learned fine-grained perception skills transfer effectively to standard downstream tasks, including RefCOCO, RefCOCO+, RefCOCOg, and general multimodal perception benchmarks. Our results highlight differential grounding as a scalable and robust approach for advancing fine-grained visual reasoning in MLLMs.
OVOD-Agent: A Markov-Bandit Framework for Proactive Visual Reasoning and Self-Evolving Detection
PDF ↗Open-Vocabulary Object Detection (OVOD) aims to enable detectors to generalize across categories by leveraging semantic information. Although existing methods are pretrained on large vision-language datasets, their inference is still limited to fixed category names, creating a gap between multimodal training and unimodal inference. Previous work has shown that improving textual representation can significantly enhance OVOD performance, indicating that the textual space is still underexplored. To this end, we propose OVOD-Agent, which transforms passive category matching into proactive visual reasoning and self-evolving detection. Inspired by the Chain-of-Thought (CoT) paradigm, OVOD-Agent extends the textual optimization process into an interpretable Visual Chain of Thought with explicit actions. OVOD's lightweight nature makes LLM-based management unsuitable; instead, we model visual context transitions as a Weakly Markovian Decision Process (w-MDP) over eight state spaces, which naturally represents the agent's state, memory, and interaction dynamics. A Bandit module generates exploration signals under limited supervision, helping the agent focus on uncertain regions and adapt its detection policy. We further integrate Markov transition matrices with Bandit trajectories for self-supervised Reward Model (RM) optimization, forming a closed loop from Bandit exploration to RM learning. Experiments on COCO and LVIS show that OVOD-Agent improves existing OVOD baselines and outperforms prior methods on novel categories, demonstrating strong generalization and scalability.
Accurate segmentation of marine organisms is vital for biodiversity monitoring and ecological assessment, yet existing datasets and models remain largely limited to terrestrial scenes. To bridge this gap, we introduce **AquaOV255**, the first large-scale and fine-grained underwater segmentation dataset containing 255 categories and over 20K images, covering diverse marine organisms and man-made objects for open-vocabulary evaluation. Furthermore, we establish the first underwater open-vocabulary segmentation benchmark, **UOVSBench**, by integrating AquaOV255 with five additional underwater datasets to enable comprehensive cross-domain evaluation. Alongside, we present **Earth2Ocean**, a training-free open-vocabulary segmentation framework that transfers terrestrial vision-language models (VLMs) to underwater domains without any additional underwater training. Earth2Ocean consists of two core components: a Geometric-guided Visual Mask Generator (**GMG**) that refines visual features via self-similarity geometric priors for local structure perception, and a Category-visual Semantic Alignment(**CSA**) module that enhances text embeddings through multimodal large language model reasoning and scene-aware template construction. Extensive experiments on the UOVSBench benchmark demonstrate that Earth2Ocean achieves over 6+ mIoU improvement on average while maintaining efficient inference.
Multimodal large language models (MLLMs) have expanded from vision-language systems to include audio, unlocking new capabilities in cross-modal reasoning and interaction. To address the limitation that existing benchmarks focus mainly on perception tasks and lack a unified cognitive evaluation framework, we propose Hierarchical Audio-Visual Evaluation Benchmark (HAVE-Bench). It systematically evaluates the audio-related capabilities of MLLMs along a three-level cognitive hierarchy: Perception, Reasoning, and Interaction, utilizing 2,451 curated samples and manually annotated multi-turn interaction-level tasks. Experiments using this unified framework reveal significant gaps in existing models at the reasoning and interaction levels, with speech-driven visual question answering (VQA) performance significantly lagging behind the text-image setting. These findings underscore the urgency of enhancing models' handling of long and complex audio and facilitating the transfer of reasoning capabilities from the vision-text to the audio-visual domain.