论文检索

Le Yang, Hongping Gan

In the field of Compressive Sensing (CS), deep unrolling networks (DUNs) have demonstrated exceptional performance and interpretability by integrating traditional optimization solvers with deep networks. However, existing DUNs suffer from homogenization in cross-stage feature extraction and insufficient integration of gradient-guided information. Additionally, the feature extraction module struggles to balance the global receptive field and computational efficiency, which limits improvements in image reconstruction details. To address these challenges, we propose a multi-scale gradient-guided unrolling architecture with adaptive Mamba for CS, named MambaCS. Specifically, we utilize our customized Adaptive State-Space Block (A-SSB) to unroll the well-known Proximal Gradient Descent (PGD) algorithm across multiple feature levels to extract comprehensive image features while maintaining computational efficiency. Moreover, we design a High-Dimensional Gradient Fusion (HDGF) that ensures the persistent and stable injection of gradient-guided information across various scales and dimensions, while effectively eliminating information bottlenecks. Finally, we develop a Feature-Adaptive Proximal Operator (FAPO), using A-SSB as an extension of the sparse basis associated with the PGD proximal operator, which enhances sensitivity to multi-scale features and improves detail reconstruction. Extensive experiments demonstrate the significant advantages of our proposed MambaCS over the current SOTA methods. Our code is available at https://github.com/nikou-arch/MambaCS.

Rohit Kundu, Vishal Mohanty, Hao Xiong, Shan Jia, Athula Balachandran, Amit K. Roy-Chowdhury

The proliferation of generative AI has led to hyper-realistic synthetic videos, escalating misuse risks and outstripping binary real/fake detectors. We introduce \texttt SAGA (\underline S ource \underline A ttribution of \underline G enerative \underline A I videos), the first comprehensive framework to address the urgent need for AI-generated video source attribution at a large scale. Unlike traditional detection, \texttt SAGA identifies the specific generative model used. It uniquely provides multi-granular attribution across five levels: authenticity, generation task (e.g., T2V/I2V), model version, development team, and the precise generator, offering far richer forensic insights. Our novel video transformer architecture, leveraging features from a robust vision foundation model, effectively captures spatio-temporal artifacts. Critically, we introduce a data-efficient pretrain-and-attribute strategy, enabling \texttt SAGA to achieve state-of-the-art attribution using only 0.5% of source-labeled data per class, matching fully supervised performance. Furthermore, we propose Temporal Attention Signatures (\texttt T-Sig ), a novel interpretability method that visualizes learned temporal differences, offering the first explanation for why different video generators are distinguishable. Extensive experiments on public datasets, including cross-domain scenarios, demonstrate that \texttt SAGA sets a new benchmark for synthetic video provenance, providing crucial, interpretable insights for forensic and regulatory applications.

Jingjing Zhang, Lei Zhang, Zheren Fu, Bo Hu, Zhendong Mao

Zero-Shot Composed Image Retrieval (ZS-CIR) aims to retrieve target images using a composed query of a reference image and a textual modification, without relying on triplet-based supervision. As the two inputs describe related but semantically unaligned information, the key challenge lies in interpreting their cross-modal discrepancy to infer the user's intended semantic modification. Existing ZS-CIR methods mainly adopt a consistency-driven paradigm, training on semantically aligned image-text pairs with alignment or reconstruction objectives. This paradigm enforces cross-modal agreement but overlooks the semantic discrepancies between modalities that naturally arise during inference. To address this issue, we propose DiffComp (Differentiate-then-Compose), a difference-driven self-supervised framework that actively induces and exploits cross-modal discrepancies during training. It stimulates the model to perceive and reconcile semantic differences across visual and textual modalities, thereby improving consistency between training and inference. The framework consists of three components: Contextual Semantic Super-patches that provide localized and coherent visual representations for downstream perception and composition; Phrase-guided Masking that selectively removes text-aligned visual cues to induce controlled cross-modal discrepancies; and Difference-aware Composition that adaptively integrates visual and textual features according to their degree of semantic difference. Extensive experiments on four ZS-CIR benchmarks show that DiffComp achieves state-of-the-art performance and strong generalization. Our code will be released at https://github.com/Orange1999/DiffComp.

Alexandros Stergiou

How do video understanding models acquire their answers? Although current Vision Language Models (VLMs) reason over complex scenes with diverse objects, action performances, and scene dynamics, understanding and controlling their internal processes remains an open challenge. Motivated by recent advancements in text-to-video (T2V) generative models, this paper introduces a logits-to-video (L2V) task alongside a model-independent approach, TRANSPORTER, to generate videos that capture the underlying rules behind VLMs' predictions. Given the high-visual-fidelity produced by T2V models, TRANSPORTER learns an optimal transport coupling to VLM's high-semantic embedding spaces. In turn, logit scores define embedding directions for conditional video generation. TRANSPORTER generates videos that reflect caption changes over diverse object attributes, action adverbs, and scene context. Quantitative and qualitative evaluations across VLMs demonstrate that L2V can provide a fidelity-rich, novel direction for model interpretability that has not been previously explored.

Shubham Kumar, Narendra Ahuja

Deep vision models perform input-output computations that are hard to interpret. Concept-based explanation methods (CBEMs) increase interpretability by re-expressing parts of the model with human-understandable semantic units, or concepts. Checking if the derived explanations are faithful--that is, they represent the model's internal computation--requires a surrogate that combines concepts to compute the output. Simplifications made for interpretability inevitably reduce faithfulness, resulting in a tradeoff between the two. State-of-the-art unsupervised CBEMs (U-CBEMs) are seemingly more interpretable, while also being more faithful to the model. However, we observe that the reported improvement in faithfulness artificially results from either (1) using overly complex surrogates, which introduces an unmeasured cost to the explanation's interpretability, or (2) relying on deletion-based approaches that, as we demonstrate, do not properly measure faithfulness. We propose Surrogate Faithfulness (SURF), which (1) replaces prior complex surrogates with a simple, linear surrogate that measures faithfulness without changing the explanation's interpretability and (2) introduces well-motivated metrics that assess loss across all output classes, not just the predicted class. We validate SURF with a measure-over-measure study by proposing a simple sanity check--explanations with random concepts should be less faithful-which prior surrogates fail. SURF enables the first reliable faithfulness benchmark of U-CBEMs, revealing that many visually compelling U-CBEMs are not faithful. Code is released.

Yuting Zhang, Kaishen Yuan, Hao Lu, Yutao Yue, Jintai Chen, Kaishun Wu

Accurate and interpretable multi-disease diagnosis remains a critical challenge in medical research, particularly when leveraging heterogeneous multimodal medical data. Current approaches often rely on single-modal data, limiting their ability to comprehensively understand complex diseases. To address this, we propose MedTVT-R1, a novel Multimodal Large Language Model (MLLM) framework designed to integrate clinical multimodal data for reasoning and diagnosing multiple diseases. We construct MedTVT-QA, a curated instruction dataset that provides question-answer pairs for physiological-level interpretations and disease-level diagnoses with a Chain of Evidence approach. MedTVT-R1 incorporates a modality perception layer to capture inter-modal dependencies and adaptively weight modality contributions. Additionally, we employ Group Relative Policy Optimization (GRPO)-based Reinforcement Fine-Tuning with a Jaccard Reward function to enhance diagnostic reasoning. Experimental results demonstrate MedTVT-R1's superiority in multimodal feature utilization and multi-disease diagnosis, offering significant potential for clinical applications such as diagnostic report generation and comorbidity reasoning. The dataset and code will be available on GitHub.

Mijeong Kim, Gunhee Kim, Jungyoon Choi, Wonjae Roh, Bohyung Han

We introduce PhysGaia, a novel physics-aware benchmark for Dynamic Novel View Synthesis (DyNVS) that encompasses both structured objects and unstructured physical phenomena. While existing datasets primarily focus on photorealistic appearance, PhysGaia is specifically designed to support physics-consistent dynamic reconstruction. Our benchmark features complex scenarios with rich multi-body interactions, where objects realistically collide and exchange forces. Furthermore, it incorporates a diverse range of materials, including liquid, gas, textile, and rheological substance, moving beyond the rigid-body assumptions prevalent in prior work. To ensure physical fidelity, all scenes in PhysGaia are generated using material-specific physics solvers that strictly adhere to fundamental physical laws. We provide comprehensive ground-truth information, including 3D particle trajectories and physical parameters (e.g., viscosity), enabling the quantitative evaluation of physical modeling. To facilitate research adoption, we also provide integration pipelines for recent 4D Gaussian Splatting models along with our dataset and their results. By addressing the critical shortage of physics-aware benchmarks, PhysGaia can significantly advance research in dynamic view synthesis, physics-based scene understanding, and the integration of deep learning with physical simulation, ultimately enabling more faithful reconstruction and interpretation of complex dynamic scenes.

Yura Choi, Roy Miles, Rolandos Alexandros Potamias, Ismail Elezi, Jiankang Deng, Stefanos Zafeiriou

Understanding and answering questions based on a user's pointing gesture is essential for next-generation egocentric AI assistants. However, current Multimodal Large Language Models (MLLMs) struggle with such tasks due to the lack of gesture-rich data and their limited ability to infer fine-grained pointing intent from egocentric video.To address this, we introduce EgoPointVQA, a dataset and benchmark for gesture-grounded egocentric question answering, comprising 4000 synthetic and 400 real-world videos across multiple deictic reasoning tasks.Built upon it, we further propose Hand Intent Tokens (HINT), which encode tokens derived from 3D hand keypoints using an off-the-shelf reconstruction model and interleaves them with the model input to provide explicit spatial and temporal context for interpreting pointing intent.We show that our model outperforms others in different backbones and model sizes.In particular, HINT-14Bachieves 68.1% accuracy, on average over 6 tasks, surpassing the state-of-the-art, InternVL3-14B, by 6.6%.To further facilitate the open research, we will release the code, model, and dataset.

Camile Lendering, Erkut Akdag, Egor Bondarau

Detecting visual anomalies in industrial inspection often requires training with only a few normal images per category. Recent few-shot methods achieve strong results employing foundation-model features, but typically rely on memory banks, auxiliary datasets, or multi-modal tuning of vision-language models. We therefore question whether such complexity is necessary given the feature representations of vision foundation models. To answer this question, we introduce SubspaceAD, a training-free method, that operates in two simple stages. First, patch-level features are extracted from a small set of normal images by a frozen DINOv2 backbone. Second, a Principal Component Analysis (PCA) model is fit to these features to estimate the low-dimensional subspace of normal variations. At inference, anomalies are detected via the reconstruction residual with respect to this subspace, producing interpretable and statistically grounded anomaly scores. Despite its simplicity, SubspaceAD achieves state-of-the-art performance across one-shot and few-shot settings without training, prompt tuning, or memory banks. In the one-shot anomaly detection setting, SubspaceAD achieves image-level and pixel-level AUROC of 98.0% and 97.6% on the MVTec-AD dataset, and 93.3% and 98.3% on the VisA dataset, respectively, surpassing prior state-of-the-art results.

Jialei Zhan, Li Liu, Jiehua Zhang, Yuhang Xie, Yongxiang Liu, Jiangming Chen, Ming-Ming Cheng

Recent advancements in remote sensing object detection have predominantly focused on oriented bounding box design and small object feature enhancement, while often overlooking the intrinsic geometric properties of remote sensing images, such as rotation invariance and structural symmetry. Many aerial objects appear in multiple orientations and exhibit clear symmetrical patterns, which, if not explicitly modeled, can lead to detection failures and inaccurate localization under geometric variation or partial occlusion. To address this, we propose the Rotation Invariant and Symmetry Aware Pixel Difference Network (RIS-PiDiNet), which introduces a novel convolutional operator called Rotation Invariant and Symmetry Aware Pixel Difference Convolution (RIS-PDC). This operator replaces traditional convolution with a mathematically grounded formulation that encodes rotation group priors and symmetrical constraints. RIS-PDC utilizes pixel differences and symmetry-guided aggregation in the polar harmonic space, enabling the network to infer partially visible structures and deduce occluded symmetrical parts. Besides improving detection accuracy, RIS-PDC enhances model interpretability by embedding geometric principles into the network design. Feature visualizations demonstrate rotation-consistent activations and symmetry-complete responses, revealing how the network captures underlying object structure even under partial visibility or orientation changes. This yields geometrically interpretable detection decisions. To our knowledge, RIS-PiDiNet is the first remote sensing object detection framework that jointly incorporates rotation invariance and symmetry modeling within a unified architecture. Extensive evaluations on standard benchmarks validate its effectiveness, achieving state-of-the-art performance on DOTA-v1.0 (78.53% mAP single-scale, 81.81% multi-scale), HRSC2016 (98.60% mAP), and DIOR-R (67.28% mAP), all with acceptable computational overhead and no increase in parameter count.

Xihang Qiu, Yuhao Fang, Qing Zhou, Bin Zhai, Jialong Hong, Wanpeng Zhang, Yao Lu, Ye Zhang, Chun Li

Multimodal Emotion Recognition in Conversations (MERC) aims to understand emotions expressed in each utterance by effectively integrating audio, text, and visual modalities. However, in real-world scenarios, unavoidable missing modalities often degrade multimodal interpretation performance. To address this, we propose Hypergraph Diffusion and Evidence Fusion based Emotion Recognition (HyperEF), a novel framework designed to mitigate challenges arising from incomplete modalities in MERC. Specifically, to mitigate performance degradation caused by modality absence, we propose Masked Hypergraph Attention (MHGAT) conditioned diffusion model to recover latent features of missing modalities in the latent space. To ensure semantic consistency between recovered and available modalities within the same utterance, we introduce MHGAT that captures high-order semantic information from available modalities to guide the diffusion model's denoising process. Furthermore, to disentangle and model the complex uncertainties inherent in MERC, we propose Dual Channel Evidence Fusion (DCEF), which estimates uncertainty at both feature source level and discriminative level, thereby achieving adaptive evidence fusion. Extensive comparative experiments and interpretability demonstrate the superior performance of our model in emotion recognition, as well as the contribution of each module within the model.

Haoyu He, Yue Zhuo, Yu Zheng, Qi R. Wang

Vision-language models (VLMs) achieve strong multimodal performance, yet how computation is organized across populations of neurons remains poorly understood. In this work, we study VLMs through the lens of neural topology, representing each layer as a within-layer correlation graph derived from neuron-neuron co-activations. This view allows us to ask whether population-level structure is behaviorally meaningful, how it changes across modalities and depth, and whether it identifies causally influential internal components under intervention. We show that correlation topology carries recoverable behavioral signal; moreover, cross-modal structure progressively consolidates with depth around a compact set of recurrent hub neurons, whose targeted perturbation materially alters model output. Neural topology thus emerges as a meaningful intermediate scale for VLM interpretability: richer than local attribution, more tractable than full circuit recovery, and empirically tied to multimodal behavior. Code is publicly available at https://github.com/he-h/vlm-graph-probing.

Francesco Gentile, Nicola Dall'Asen, Francesco Tonini, Massimiliano Mancini, Lorenzo Vaquero, Elisa Ricci

As vision-language models are deployed at scale, understanding their internal mechanisms becomes increasingly critical. Existing interpretability methods predominantly rely on activations, making them dataset-dependent, vulnerable to data bias, and often restricted to coarse head-level explanations. We introduce SITH (Semantic Inspection of Transformer Heads), a fully data-free, training-free framework that directly analyzes CLIP's vision transformer in weight space. For each attention head, we decompose its value-output matrix into singular vectors and interpret each one via COMP (Coherent Orthogonal Matching Pursuit), a new algorithm that explains them as sparse, semantically coherent combinations of human-interpretable concepts. We show that SITH yields coherent, faithful intra-head explanations, validated through reconstruction fidelity and interpretability experiments. This allows us to use SITH for precise, interpretable weight-space model edits that amplify or suppress specific concepts, improving downstream performance without retraining. Furthermore, we use SITH to study model adaptation, showing how fine-tuning primarily reweights a stable semantic basis rather than learning entirely new features.

Tao Yang, Qing Zhou, Yanliang Li, Qi Wang

Reasoning segmentation increasingly employs reinforcement learning to generate explanatory reasoning chains that guide Multimodal Large Language Models. While these geometric rewards are primarily confined to guiding the final localization, they are incapable of discriminating whether the reasoning process remains anchored on the referred region or strays into irrelevant context. Lacking this discriminative guidance, the model's reasoning often devolves into unfocused and verbose chains that ultimately fail to disambiguate and perceive the target in complex scenes. This suggests a need to complement the RL objective with Discriminative Perception, an ability to actively distinguish a target from its context. To realize this, we propose DPAD to compel the model to generate a descriptive caption of the referred object, which is then used to explicitly discriminate by contrasting the caption's semantic relevance to the referred object against the wider context. By optimizing for this discriminative capability, the model is forced to focus on the unique attributes of the target, leading to a more converged and efficient reasoning chain. The descriptive caption also serves as an interpretability rationale that aligns with the segmentation. Experiments on the benchmarks confirm the validity of our approach, delivering substantial performance gains, with the cIoU on ReasonSeg increasing by 3.09% and the reasoning chain length decreasing by approximately 42%. Code is available at https://github.com/mrazhou/DPAD.

Jovana Kondic, Pengyuan Li, Dhiraj Joshi, Isaac Sanchez, Ben Wiesel, Shafiq Abedin, Amit Alfassy, Eli Schwartz, Daniel Caraballo, Yagmur Gizem Cinar 等

Understanding charts requires models to jointly reason over geometric visual patterns, structured numerical data, and natural language -- a capability where current vision-language models (VLMs) remain limited. We introduce ChartNet, a high-quality, million-scale multimodal dataset designed to advance chart interpretation and reasoning. ChartNet leverages a novel code-guided synthesis pipeline to generate 1.5 million diverse chart samples spanning 24 chart types and 6 plotting libraries. Each sample consists of five aligned components: plotting code, rendered chart image, data table, natural language summary, and question-answering with reasoning, providing fine-grained cross-modal alignment. To capture the full spectrum of chart comprehension, ChartNet additionally includes specialized subsets encompassing human annotated data, real-world data, safety, and grounding. Moreover, a rigorous quality-filtering pipeline ensures visual fidelity, semantic accuracy, and diversity across chart representations. Fine-tuning on ChartNet consistently improves results across benchmarks, demonstrating its utility as large-scale supervision for multimodal models. As the largest open-source dataset of its kind, ChartNet aims to support the development of foundation models with robust and generalizable capabilities for data visualization understanding. The dataset is publicly available at https://huggingface.co/datasets/ibm-granite/ChartNet

Abdul Rehman, Iqra Rasool, Ayisha Imran, Mohsen Ali, Waqas Sultani

Digital hematopathology requires cell-level analysis across diverse disease categories, including malignant disorders (e.g., leukemia), infectious conditions (e.g., malaria), and non-malignant red blood cell disorders (e.g., sickle cell disease). Whether single-task, vision-language, WSI- optimized, or single-cell hematology models, these approaches share a key limitation: they cannot provide unified, multi-task, multi-modal reasoning across the complexities of digital hematopathology. To overcome these limitations, we propose Uni-Hema, a multi-task, unified model for digital hematopathology integrating detection, classification, segmentation, morphology prediction, and reasoning across multiple diseases. Uni-Hema leverages 46 publicly available datasets, encompassing over 700K images and 21K question-answer pairs, and is built upon Hema-Former, a multimodal module that bridges visual and linguistic representations at the hierarchy level for the different tasks (detection, classification, segmentation, morphology, mask language modeling, and visual question answering) at different granularities. Extensive experiments demonstrate that Uni-Hema achieves comparable or superior performance compared to training on a single task and single-dataset models, across diverse hematological tasks, while providing interpretable, morphologically relevant insights at the single-cell level. Our framework establishes a new standard for multi-task and multi-modal digital hematopathology. The code is available at https://github.com/intelligentMachines-ITU/Uni-Hema

Xingyu Qiu, Mengying Yang, Xinghua Ma, Dong Liang, Fanding Li, Gongning Luo, Wei Wang, Kuanquan Wang, Shuo Li

Although EDM aims to unify the design space of diffusion models, its reliance on fixed Gaussian noise prevents it from explaining emerging flow-based methods that diffuse arbitrary noise. Moreover, our study reveals that EDM's forcible injection of Gaussian noise has adverse effects on image restoration task, as it corrupts the degraded images, overextends the restoration distance, and increases the task's complexity. To interpret diverse methods for handling distinct noise patterns within a unified theoretical framework and to minimize the restoration distance, we propose EDA, which Elucidates the Design space of Arbitrary-noise diffusion models. Theoretically, EDA expands noise pattern flexibility while preserving EDM's modularity, with rigorous proof that increased noise complexity introduces no additional computational overhead during restoration. EDA is validated on three representative medical image denoising and natural image restoration tasks: MRI bias field correction (global smooth noise), CT metal artifact removal (global sharp noise) and natural image shadow removal (local boundary-aware noise). With only 5 sampling steps, competitive results against specialized methods across medical and natural tasks demonstrate EDA's strong generalization capability for image restoration. Code is available at: https://github.com/PerceptionComputingLab/EDA.

Rujiao Long, Yang Li, Xingyao Zhang, Weixun Wang, Tianqianjin Lin, Xi Zhao, Yuchi Xu, Wenbo Su, Junchi Yan, Bo Zheng

Exploration capacity shapes both inference-time performance and reinforcement learning (RL) training for large (vision-) language models, as stochastic sampling often yields redundant reasoning paths with little high-level diversity. This paper proposes Reasoning Palette, a novel latent-modulation framework that endows the model with a stochastic latent variable for strategic contextualization, guiding its internal planning prior to token generation. This latent context is inferred from the mean-pooled embedding of a question-answer pair via a variational autoencoder (VAE), where each sampled latent potentially encodes a distinct reasoning context. During inference, a sampled latent is decoded into learnable token prefixes and prepended to the input prompt, modulating the model's internal reasoning trajectory. In this way, the model performs internal sampling over reasoning strategies prior to output generation, which shapes the style and structure of the entire response sequence. A brief supervised fine-tuning (SFT) warm-up phase allows the model to adapt to this latent conditioning. Within RL optimization, Reasoning Palette facilitates structured exploration by enabling on-demand injection for diverse reasoning modes, significantly enhancing exploration efficiency and sustained learning capability. Experiments across multiple reasoning benchmarks demonstrate that our method enables interpretable and controllable control over the (vision-) language model's strategic behavior, thereby achieving consistent performance gains over standard RL methods.

Zhiyu Xu, Weilong Yan, Yufei Shi, Xin Meng, Tao He, Huiping Zhuang, Ming Li, Hehe Fan

Recent advancements in multimodal large language models (MLLMs) and video agent systems have significantly improved general video understanding. However, when applied to scientific video understanding and educating--a domain that demands external professional knowledge integration and rigorous step-wise reasoning--existing approaches often struggle. To bridge this gap, we propose SciEducator, an iterative self-evolving multi-agent system for scientific video comprehension and education. Rooted in the classical Deming Cycle from management science, our design reformulates its Plan-Do-Study-Act philosophy into a self-evolving reasoning and feedback mechanism, which facilitates the interpretation of intricate scientific activities in videos. Moreover, SciEducator can produce multimodal educational content tailored to specific scientific processes, including textual instructions, visual guides, audio narrations, and interactive references. To support evaluation, we construct SciVBench, a benchmark consisting of 500 expert-verified and literature-grounded science QA pairs across five categories, covering physical, chemical, and everyday phenomena. Extensive experiments demonstrate that SciEducator substantially outperforms leading closed-source MLLMs (e.g., Gemini, GPT-4o) and state-of-the-art video agents on the benchmark, establishing a new paradigm for the community.

Yuzhi Huang, Kairun Wen, Rongxin Gao, Dongxuan Liu, Yibin Lou, Jie Wu, Jing Xu, Jian Zhang, Zheng Yang, Yunlong Lin 等

Humans inhabit a physical 4D world, where spatial geometry and semantic content evolve over time, forming a dynamic reality. While current Multimodal Large Language Models (MLLMs) demonstrate strong capabilities in understanding static visual inputs, it remains unclear whether they can effectively "think in dynamics," i.e., perceive, track, and reason about spatio-temporal evolution in complex scenes.To systematically evaluate these abilities, we introduce \texttt Dyn-Bench , a large-scale benchmark designed to assess spatio-temporal reasoning and localized dynamics perception. Constructed through multi-stage filtering over massive 2D and 4D data sources, \texttt Dyn-Bench provides a high-quality collection of diverse dynamic scenes, consisting of 1k videos, 7k visual question answering (VQA) pairs, and 3k dynamic object grounding samples.We comprehensively study general-purpose, spatial-aware, and region-level MLLMs to understand how they "think in dynamics" from both linguistic and visual perspectives. Our results reveal that existing models struggle to jointly excel in both spatio-temporal reasoning and dynamic object grounding, often producing inconsistent interpretations of motion and interaction. Conventional prompting strategies i.e., chain-of-thought or caption-based hints) provide only limited improvements.In contrast, structured integration approaches, including Mask-Guided Fusion and the Spatio-Temporal Textual Cognitive Map (ST-TCM), substantially enhance MLLMs' dynamic perception and spatio-temporal reasoning in an evolving 4D world. These findings underscore the importance of explicit spatio-temporal structural cues to bridge the gap between static perception and dynamic reasoning in MLLMs.