Recently, Retrieval-Augmented Generation (RAG) has been proposed to expand internal knowledge of Multimodal Large Language Models (MLLMs) by incorporating external knowledge databases into the generation process, which is widely used for knowledge-based Visual Question Answering (VQA) tasks. Despite impressive advancements, vanilla RAG-based VQA methods that rely on unstructured documents and overlook the structural relationships among knowledge elements frequently introduce irrelevant or misleading content, reducing answer accuracy and reliability. To overcome these challenges, a promising solution is to integrate multimodal knowledge graphs (KGs) into RAG-based VQA frameworks to enhance the generation by introducing structured multimodal knowledge. Therefore, in this paper, we propose a novel multimodal knowledge-augmented generation framework (mKG-RAG) based on multimodal KGs for knowledge-intensive VQA tasks. Specifically, our approach leverages MLLM-powered keyword extraction and vision-text matching to distill semantically consistent and modality-aligned entities/relationships from multimodal documents, constructing high-quality multimodal KGs as structured knowledge representations. In addition, a dual-stage retrieval strategy equipped with a question-aware multimodal retriever is introduced to improve retrieval efficiency while refining precision. Comprehensive experiments demonstrate that our approach significantly outperforms existing methods, setting a new state-of-the-art for knowledge-based VQA.

Vision-language tracking has received increasing attention in recent years, as textual information can effectively address the inflexibility and inaccuracy associated with specifying the target object to be tracked. Existing works either directly fuse the fixed language with vision features or simply modify using attention, however, their performance is still limited. Recently, some researchers have explored using text generation to adapt to the variations in the target during tracking, however, these works fail to provide insights into the model's reasoning process and do not fully leverage the advantages of large models, which further limits their overall performance. To address the aforementioned issues, this paper proposes a novel reasoning-based vision-language tracking framework, named ReasoningTrack, based on a pre-trained vision-language model Qwen2.5-VL. Both SFT (Supervised Fine-Tuning) and reinforcement learning GRPO are used for the optimization of reasoning and language generation. We embed the updated language descriptions and feed them into a unified tracking backbone network together with vision features. Then, we adopt a tracking head to predict the specific location of the target object. In addition, we propose a large-scale long-term vision-language tracking benchmark dataset, termed TNLLT, which contains 200 video sequences. 20 baseline visual trackers are re-trained and evaluated on this dataset, which builds a solid foundation for the vision-language visual tracking task. Extensive experiments on multiple vision-language tracking benchmark datasets fully validated the effectiveness of our proposed reasoning-based natural language generation strategy. The source code of this paper will be released on https://github.com/Event-AHU/Open_VLTrack

Web navigation represents a critical and challenging domain for evaluating artificial general intelligence (AGI), demanding complex decision-making within high-entropy, dynamic environments with combinatorially explosive action spaces. Current approaches to building autonomous web agents either focus on offline imitation learning or online exploration, but rarely integrate both paradigms effectively. Inspired by the dual-process theory of human cognition, we derive a principled decomposition into fast System 1 and slow System 2 cognitive processes. This decomposition provides a unifying perspective on existing web agent methodologies, bridging the gap between offline learning of intuitive reactive behaviors and online acquisition of deliberative planning capabilities. We implement this framework in CogniWeb, a modular agent architecture that adaptively toggles between fast intuitive processing and deliberate reasoning based on task complexity. Our evaluation on WebArena demonstrates that CogniWeb achieves competitive performance (43.96% success rate) while maintaining significantly higher efficiency (75% reduction in token usage).

Multimodal Large Language Models (MLLMs) demonstrate remarkable capabilities but often struggle with complex, multi-step mathematical reasoning, where minor errors in visual perception or logical deduction can lead to complete failure. While Process Reward Models (PRMs) offer step-by-step supervision, existing multimodal PRMs are limited to being binary verifiers that can identify but not correct errors, offering little explanatory power. To address these deficiencies, we introduce the Generative Multimodal Process Reward Model (GM-PRM), a novel paradigm that transforms the PRM from a passive judge into an active reasoning collaborator. Instead of a simple scalar score, GM-PRM provides a fine-grained, interpretable analysis of each reasoning step, evaluating its step intent, visual alignment, and logical soundness. More critically, GM-PRM is trained to generate a corrected version of the first erroneous step it identifies. This unique corrective capability enables our new test-time inference strategy, Refined Best-of-N (Refined-BoN). This framework actively enhances solution quality by using the PRM's generated correction to guide the policy model toward a more promising reasoning trajectory, thereby improving the diversity and correctness of the solution pool. We demonstrate that GM-PRM achieves state-of-the-art results on multiple multimodal math benchmarks, significantly boosting policy model performance with remarkable data efficiency, requiring only a 20K-sample training dataset. Our code will be released upon acceptance.

Large reasoning models achieve remarkable performance through extensive chain-of-thought generation, yet they suffer from a critical inefficiency: applying uniformly extensive reasoning regardless of problem complexity. We present Hierarchical Budget Policy Optimization (HBPO), a reinforcement learning framework that enables models to learn problem-specific reasoning depths without sacrificing capability. Unlike existing approaches that impose rigid constraints or rely on discrete mode selection, HBPO partitions the exploration space into budget-constrained hierarchies (512-2560 tokens), each with differentiated reward structures that preserve both efficiency incentives and reasoning capabilities. This design addresses a fundamental challenge in efficient reasoning training: traditional length penalties systematically bias models away from necessary long reasoning paths, causing exploration space collapse. Through hierarchical sampling and budget-aware rewards, HBPO maintains exploration diversity while teaching models to recognize when extended deliberation is warranted. Extensive experiments demonstrate that HBPO reduces average token usage by up to 60.6% while improving accuracy by 3.14% across four reasoning benchmarks. Most notably, HBPO exhibits emergent adaptive behavior where models automatically adjust reasoning depth based on problem complexity. Our results suggest that reasoning efficiency and capability are not inherently conflicting, and can be simultaneously optimized through appropriately structured hierarchical training that preserves exploration diversity. Advances in large reasoning models have led to impressive performance on complex reasoning tasks through chain-of-thought methodologies (OpenAI, 2024; DeepSeek-AI, 2025).

Vision language models (VLMs) have shown remarkable capabilities in integrating linguistic and visual reasoning but remain fundamentally limited in understanding dynamic spatiotemporal interactions. Humans effortlessly track and reason about object movements, rotations, and perspective shifts-abilities essential for robust dynamic real-world understanding yet notably lacking in current VLMs. In this paper, we introduce VLM4D, the first benchmark specifically designed to evaluate the spatiotemporal reasoning capabilities of VLMs. Our benchmark comprises diverse real-world and synthetic videos accompanied by carefully curated question-answer pairs emphasizing translational and rotational motions, perspective awareness, and motion continuity. Through comprehensive evaluations of state-of-the-art open and closed-source VLMs, we identify significant performance gaps compared to human baselines, highlighting fundamental deficiencies in existing models. Extensive analysis reveals that VLMs struggle particularly with integrating multiple visual cues and maintaining temporal coherence. We further explore promising directions, such as leveraging 4D feature field reconstruction and targeted spatiotemporal supervised fine-tuning, demonstrating their effectiveness in enhancing spatiotemporal comprehension. Our work aims to encourage deeper exploration into improving VLMs' spatial and temporal grounding, paving the way towards more capable and reliable visual intelligence for dynamic environments.

Teaching robots dexterous skills from human videos remains challenging due to the reliance on low-level trajectory imitation, which fails to generalize across object types, spatial layouts, and manipulator configurations. We propose Graph-Fused Vision-Language-Action (GF-VLA), a framework that enables dual-arm robotic systems to perform task-level reasoning and execution directly from RGB and Depth human demonstrations. GF-VLA first extracts Shannon-information-based cues to identify hands and objects with the highest task relevance, then encodes these cues into temporally ordered scene graphs that capture both hand-object and object-object interactions. These graphs are fused with a language-conditioned transformer that generates hierarchical behavior trees and interpretable Cartesian motion commands. To improve execution efficiency in bimanual settings, we further introduce a cross-hand selection policy that infers optimal gripper assignment without explicit geometric reasoning. We evaluate GF-VLA on four structured dual-arm block assembly tasks involving symbolic shape construction and spatial generalization. Experimental results show that the information-theoretic scene representation achieves over 95 percent graph accuracy and 93 percent subtask segmentation, supporting the LLM planner in generating reliable and human-readable task policies. When executed by the dual-arm robot, these policies yield 94 percent grasp success, 89 percent placement accuracy, and 90 percent overall task success across stacking, letter-building, and geometric reconfiguration scenarios, demonstrating strong generalization and robustness across diverse spatial and semantic variations.

Semi-supervised continual learning (SSCL) seeks to leverage both labeled and unlabeled data in a sequential learning setup, aiming to reduce annotation costs while managing continual data arrival. SSCL introduces complex challenges, including ensuring effective unlabeled learning (UL), while balancing memory stability (MS) and learning plasticity (LP). Previous SSCL efforts have typically focused on isolated aspects of the three, while this work presents USP, a divide-and-conquer framework designed to synergistically enhance these three aspects: (1) Feature Space Reservation (FSR) strategy for LP, which constructs reserved feature locations for future classes by shaping old classes into an equiangular tight frame; (2) Divide-and-Conquer Pseudo-labeling (DCP) approach for UL, which assigns reliable pseudo-labels across both high- and low-confidence unlabeled data; and (3) Class-mean-anchored Unlabeled Distillation (CUD) for MS, which reuses DCP's outputs to anchor unlabeled data to stable class means for distillation to prevent forgetting. Comprehensive evaluations show USP outperforms prior SSCL methods, with gains up to 5.94% in the last accuracy, validating its effectiveness. The code is available at https://github.com/NJUyued/USP4SSCL.

Recent LLMs like DeepSeek-R1 have demonstrated state-of-the-art performance by integrating deep thinking and complex reasoning during generation. However, the internal mechanisms behind these reasoning processes remain unexplored. We observe reasoning LLMs consistently use vocabulary associated with human reasoning processes. We hypothesize these words correspond to specific reasoning moments within the models' internal mechanisms. To test this hypothesis, we employ Sparse Autoencoders (SAEs), a technique for sparse decomposition of neural network activations into human-interpretable features. We introduce ReasonScore, an automatic metric to identify active SAE features during these reasoning moments. We perform manual and automatic interpretation of the features detected by our metric, and find those with activation patterns matching uncertainty, exploratory thinking, and reflection. Through steering experiments, we demonstrate that amplifying these features increases performance on reasoning-intensive benchmarks (+2.2%) while producing longer reasoning traces (+20.5%). Using the model diffing technique, we provide evidence that these features are present only in models with reasoning capabilities. Our work provides the first step towards a mechanistic understanding of reasoning in LLMs. Code available at https://github.com/AIRI-Institute/SAE-Reasoning

Depression is a widespread mental disorder that affects millions worldwide. While automated depression assessment shows promise, most studies rely on limited or non-clinically validated data, and often prioritize complex model design over real-world effectiveness. In this paper, we aim to unveil the landscape of clinical depression assessment. We introduce C-MIND, a clinical neuropsychiatric multimodal diagnosis dataset collected over two years from real hospital visits. Each participant completes three structured psychiatric tasks and receives a final diagnosis from expert clinicians, with informative audio, video, transcript, and functional near-infrared spectroscopy (fNIRS) signals recorded. Using C-MIND, we first analyze behavioral signatures relevant to diagnosis. We train a range of classical models to quantify how different tasks and modalities contribute to diagnostic performance, and dissect the effectiveness of their combinations. We then explore whether LLMs can perform psychiatric reasoning like clinicians and identify their clear limitations in realistic clinical settings. In response, we propose to guide the reasoning process with clinical expertise and consistently improves LLM diagnostic performance by up to 10% in Macro-F1 score. We aim to build an infrastructure for clinical depression assessment from both data and algorithmic perspectives, enabling C-MIND to facilitate grounded and reliable research for mental healthcare.