Conventional Euclidean fusion paradigms in multimodal sentiment analysis (MSA) fail to capture non-Euclidean correlations and probabilistic uncertainty across modalities, often leading to misaligned affective representations and limited robustness under real-world variability. To overcome these challenges, we introduce the novel Quantum-Inspired Geometric and Energy-Based Fusion (QEGF) framework that unifies geometric sentiment representation with probabilistic uncertainty modeling for more reliable affective perception. First, the Quantum-Inspired Geometric Fusion (QGF) operator represents each modality as a trace-normalized Symmetric Positive-Definite (SPD) density descriptor, enabling second-order correlation modeling and adaptive gating via a spectral entropy statistic from each trace-normalized SPD descriptor. This design decouples structural encoding from semantic refinement, significantly mitigating performance degradation under noisy and corrupted inputs, and achieving a 4.55% Acc-2 gain over Cross-Attention in domain shift scenarios. Second, Energy-Based Uncertainty Modeling (EBUM) introduces a contrastive energy objective with a batch-level energy-dispersion signal, stabilizing convergence and yielding a lower Brier score (0.172 to 0.135) in the output-level calibration diagnostic under domain shift. Comprehensive evaluations across major benchmarks confirm that QEGF achieves state-of-the-art performance and superior resilience, laying a pathway toward more interpretable and resilient affective human-machine interaction.

Recent multimodal large language models (MLLMs) have made remarkable progress on fine-grained perception tasks under the "Thinking with Images" (TwI) paradigm by iteratively performing various visual tool operations. However, this paradigm relies heavily on frequent external tool calls and repeated image re-encoding, which leads to substantial computational overhead and inference latency. Furthermore, existing models generally lack capability boundary awareness, making it difficult to adaptively determine when to rely on internal parametric knowledge and when to invoke external tools. To address these issues, we propose Beyond the Eye (BEE), a novel implicit visual tool paradigm centered on self-regulated capability. BEE directly incorporates visual tool invocation behaviors into the training objective and encourages the model to develop a self-regulated invocation mechanism. This design enables the model to adaptively balance internal knowledge and implicit tools, avoiding redundant tool usage while substantially reducing inference latency. Specifically, BEE involves a two-stage training process: (1) Formalized Chain-of-Thought (CoT) Supervised Fine-tuning (SFT). We construct CoT trajectories with structured tool slots and mixed invocation states. This stage activates the model's implicit tool representations and adaptive switching capability. (2) Self-regulated Reward-Driven Alignment. To address redundant tool usage caused by ambiguous cognitive boundaries, we first introduce the Net Tool Gain (NTG) metric to quantify this phenomenon. Based on this observation, we further propose a self-regulated reward mechanism. This mechanism penalizes ineffective tool dependency and encourages the model to perform knowledge routing, ensuring that implicit tools are invoked only when the model's internal knowledge is insufficient. BEE achieves state-of-the-art performance in fine-grained visual perception while remaining competitive in general reasoning tasks and achieving substantial gains in inference efficiency. In particular, under the high-resolution setting of 8192, BEE achieves up to $\sim 86.2\%$ inference acceleration compared with the DeepEyes method. BEE provides a promising direction for developing implicit tool MLLMs that combine strong cognitive ability with extremely low latency and further advances the development of self-regulated MLLMs.

Recent advances in sensor technology have led to an explosion of Spatial High-Dimensional Data (SHD), including spatial transcriptomics and hyperspectral imaging, the two domains studied in this work. These data capture complex interactions between intrinsic sample properties and spatial microenvironments by incorporating location information through accumulation, embedding, or aggregation. Existing methods typically assume homogeneous spatial interaction patterns, overlooking critical asymmetries in SHD: (i) view asymmetry between spatial patterns and high-dimensional attributes, and (ii) spatial asymmetry in the internal pathways of high-dimensional data. To address this gap, we propose DIaS, a data-centric decoupling framework that disentangles spatial-induced, intrinsic, and harmonized information signals to better model the richness of SHD. Specifically, DIaS employs a Gaussian process for spatial embedding, leveraging uncertainty quantification to dynamically infer spatial pathways. During disentanglement, SHD is decomposed into three mutually regularized views to minimize dependence between spatial attributes and intrinsic features. To balance the asymmetric contributions of spatial and intrinsic information, we introduce an asymmetric harmonization mechanism, which adaptively aligns spatial and intrinsic representations to identify salient and diverse multi-view prototypes. Experiments on 18 benchmarks in spatial transcriptomics and hyperspectral imaging, together with downstream task evaluations, demonstrate state-of-the-art performance and the effectiveness of DIaS in alleviating information asymmetry within these two domains. The demo code of this work is publicly available at https://github.com/hexiao0275/DIaS.

Video analytics pipelines (VAPs) have been a paradigm for large-scale video analytics. Due to temporal redundancy in video, frame filtering is widely used in VAPs to reduce analysis workload. While prior works have identified decoding as a significant overhead and proposed pre-decoding filtering to mitigate it, we find that codec reference dependencies still force decoders to process many unselected frames, manifesting an over-decoding trap that bounds end-to-end throughput. This limitation stems from the reference-based design in modern codecs, which requires decoding preceding frames to reconstruct any selected one. As a result, over-decoding has become the practical bottleneck in VAPs using modern decoders, highlighting a critical but under-explored problem. To address this issue, we propose ParaDeco, a high-throughput video analytics framework featuring a novel frame-level parallel generative decoder. Unlike traditional decoders, ParaDeco adopts a decode-what-matters approach with decoupled frame dependencies. To decode arbitrary frames independently, ParaDeco generates frame-wise features as standalone skeletons using compressed video metadata, then predicts pseudo frames maintaining semantic consistency with original frames. Moreover, ParaDeco identifies which frames truly matter for analysis via delicate contribution-based frame filtering. We implement ParaDeco on a cloud server and evaluate it on large-scale real-world surveillance video datasets. Our experimental results show that ParaDeco achieves a 2.76× speedup on averagparse e compared to state-of-the-art VAPs.

This paper presents a learned image compression framework that exploits semantically correlated external images as conditional coding references in the latent domain. Three reference generation methods are investigated: local dictionary retrieval, web-based image search, and diffusion-based image-text-image generation. Based on rate-distortion optimization in the latent domain, we learn a network to synthesize a conditional latent map from these correlated images, providing an effective semantic reference for conditional coding of the input image. These generation processes are synchronized between the encoder and decoder with a very low communication overhead. Our theoretical analysis shows that our approach is robust to perturbations in reference samples, with a subspace recovery error bound that is inversely proportional to the correlation between the reference and original images. This indicates a high tolerance to matching errors of the reference images. Comprehensive experiments on benchmark datasets show that our proposed generative references of conditional latents (GRCL) method achieves significant improvements in coding performance, up to 1.5 dB PSNR gain over state-of-the-art methods with a very small bit-rate overhead, approximately 0.005 bits per pixel. Our work presents a new generic framework to utilize external references for learned image compression.

Generative adversarial networks (GANs) and diffusion probabilistic models (DPMs) have emerged as the most prominent paradigms in generative modeling, and recent works suggest that combining GANs and DPMs could further enhance generative models. While the adversarial training of GANs has been applied to knowledge distillation of DPMs in the literature, the potential of leveraging DPMs to improve GAN training remains under-explored. In this work, we demonstrate the theoretical connections between GANs and DPMs, bridging the two most important paradigms of generative models for the first time. Specifically, we show that the generator supervision from the discriminators is equivalent to the score functions from DPMs. As a result, we propose a universal solution to stabilize GAN training, GraSF, which supervises the generator through backward propagation with gradients computed by pre-trained DPMs in a closed-form manner, with almost no increase in computational cost. Regarding the experimental evidence, we first design an illustrative example to show that our GraSF helps the generator achieve global optimality even if the discriminator suffers from non-optimality. As for real-world datasets, we manage to reach the 1.96 FID on ImageNet 64x64 dataset, achieving the best performance among all GAN-based methods with no pre-trained ImageNet prior, even surpassing some leading DPMs. This significant narrowing of the performance gap between GANs and DPMs confirms the practical value of our theoretical results and the efficacy of our method. Source code has been uploaded to IEEE Xplore Code Ocean.

360° video navigation aims to model users' viewing behavior in immersive videos through continuous viewport selection. This task is challenging because human viewing is not driven by instantaneous responses to local saliency, but by continuous decisions that evolve with semantic events over time, involving attention shifts across events and long-term viewport changes. However, progress remains limited, mainly due to the lack of a well-established benchmark, resulting in limitations in data, annotation, and evaluation. Current datasets are limited in scale and diversity, and mostly consist of short or single-event videos, making cross-event viewing behavior underrepresented. Existing annotation methods rely on frame-level or short-term clustering, which emphasizes local consistency but fails to capture temporal continuity and the rationality of viewport selection during event transitions. Existing metrics focus on spatial overlap or trajectory consistency, ignoring motion-direction agreement between predicted and ground-truth navigation trajectories. As a result, spatially accurate predictions may still follow incorrect motion directions, leading to unreliable evaluation of viewing behavior. To address these issues, we construct a new benchmark for 360° video navigation in multi-event scenarios, with a standardized data and annotation pipeline and a more discriminative evaluation protocol. We further develop an event-level model to capture cross-event viewing behavior and long-term viewport evolution. Extensive experiments show that the proposed benchmark provides more reliable evaluation, and the proposed model achieves superior performance to existing methods.

We propose Self-MAR, a self-supervised paradigm for water-related optical image enhancement that enables input-only restoration via the proposed VAR-inspired Multi-stage AutoRegressive model, in which each restoration stage is conditioned on the preceding restored state. The intricate interactions between light and the water medium introduce residual degradations that are difficult to model explicitly, thereby limiting the effectiveness of existing methods that rely heavily on synthetic data or sparse annotations. Furthermore, acquiring paired ground-truth data in real underwater environments remains challenging, giving rise to the triangular dilemma in underwater image enhancement (UIE). To address these issues, Self-MAR incorporates three key components: 1) We design a multi-stage, coarse-to-fine reconstruction architecture that progressively recovers image structure and suppresses content-coupled underwater object residuals (ORs) at different resolutions, transforming OR removal into structure-aware controllable image generation. 2) We construct an asymmetric phase-amplitude interaction that combines the structural phase of the current observation with the corrected spectral-energy distribution of the preceding restoration estimate. This interaction preserves spatial organization during stage-wise upsampling and propagates corrected color-energy statistics from coarse to fine. 3) We propose the Churning constraint mechanism to preserve the original perceptual features of the input image and prevent semantic drift during the self-supervised training process. Self-MAR derives supervision targets from the inputs themselves: underwater inputs generate OR-removal targets, while in-air inputs generate structural self-reconstruction targets from their own content.Extensive experiments demonstrate competitive full-reference restoration quality, leading performance across complementary no-reference metrics (UIQM, UCIQE, NIQE, and URANKER), efficient inference, and consistent performance across UIEB C60, UCCS, U45, and Z700. Downstream evaluation on SeaThru-D1 further shows the lowest combined depth-consistency error among thecompared state-of-the-art enhancement methods.

Hypergraphs, as advanced mathematical constructs, are adept at modeling complex, multi-faceted relationships across diverse disciplines, including data mining, computer vision, and life sciences. Hypergraph Neural Networks (HGNNs) represent a significant advancement over traditional graph neural networks by efficiently capturing high-order dependencies, which are prevalent in intricate systems such as biological networks, visual data structures, and large-scale relational datasets. This capability enables more precise modeling of interactions involving multiple entities, such as gene-term associations in life sciences, multi-modal data fusion in computer vision, and collaborative patterns in data mining. This article provides a systematic introduction to HGNNs, presenting the latest advancements in their theoretical foundations, architectural designs, and interpretability analysis. We elaborate on how HGNNs address the limitations of traditional graph-based models by leveraging hyperedge-based aggregation mechanisms to encode higher-order semantics. Beyond theoretical frameworks, we highlight their versatile applications across interdisciplinary fields, including drug discovery, protein interaction prediction, visual recognition, and knowledge graph completion. Particular emphasis is placed on identifying critical theoretical challenges, discussing innovative solutions to improve model efficiency and interpretability, and offering practical guidelines for deploying HGNNs in real-world research scenarios. By bridging theoretical rigor with empirical insights, this article aims to serve as a foundational resource for both researchers and practitioners seeking to engage with HGNNs in diverse domains.

Diffusion models (DMs) have steadily demonstrated outstanding performance in high-fidelity, controllable generation across diverse scenarios, following the "pretraining-finetuning" paradigm established in LLMs. Despite these advances, conventional DMs still face critical limitations in sophisticated reasoning, autonomous exploration of generation pathways, and adaptability to complex, human-aligned objectives. Inspired by RLHF's success in LLMs, RL has recently emerged as a pivotal post-training mechanism to iteratively refine DMs through reward-driven trajectory optimization. This paper presents the first comprehensive review of RL-driven DMs and empirical analysis of how RL reshapes the diffusion process towards experience-centric, goal-directed generation. We propose a novel technical panorama elucidating RL's multifaceted roles in diffusion-based language, multimodal, robotic, and scientific models-with granular focus on diffusion trajectory optimization and reward engineering. Through extensive experiments on multiple task-specific benchmarks, we systematically evaluate representative methods to intuitively highlight the advantages of different RL post-training strategies. Furthermore, we reveal and dissect three crucial bottlenecks that currently hinder the real-world deployment of RL-driven DMs: reward design fragility, exploration-exploitation trade-off in high-dimensional generation spaces, and training instability under sparse rewards. Based on these perspectives, we provide a principled roadmap for the integration of RL into DMs and outline promising future research directions for robust, interpretable RL-diffusion systems. We also construct a public repository to continuously track the latest developments in this rapidly evolving field: https://github.com/lyongo/Awesome-RL-for-Diffusion-Models.

Video Moment Retrieval (VMR), the task of identifying a specific moment in an untrimmed video from a natural language query, is a key capability for advanced video understanding. As an emerging research direction, point-supervised VMR offers a cost-effective solution by requiring minimal annotation efforts. In this paper, we explore the information asymmetries within this task: a semantic-positional asymmetry between the clarity of the visual cues and the imprecision of the annotated frame's position, and an instance-duration asymmetry between the certainty of a moment's occurrence and the ambiguity of its duration. To mitigate these asymmetries, we propose a novel framework, Dual-Asymmetry Aligned Network (DAAN), featuring two specialized components: (1) the Center-Guided Calibration module (CGC) which mitigates positional imprecision by dynamically generating a pseudo temporal center and leveraging a decomposed cross-modal attention strategy to recalibrate the feature distribution to align semantic attention with the estimated event centroid; and (2) the Duration-Aware Proposal Discriminator (DPD) which alleviates duration ambiguity by distinguishing the most plausible temporal spans. It infers missing duration information by learning query-conditioned structural compatibility between proposal content and boundary representations. Extensive experiments on TACoS, ActivityNet Captions, and Charades-STA validate the effectiveness of our proposed framework in point-supervised VMR.

Person re-identification (ReID) is a core technology in large-scale intelligent surveillance systems. In practical deployments, surveillance networks evolve continuously as new cameras are incrementally integrated. As the network expands, the system not only encounters new identities but also captures previously known identities from novel camera views. However, existing LReID studies typically assume disjoint identity spaces across stages, failing to model the cumulative evolution of features induced by such expansion. To bridge this gap, we introduce Camera-Incremental Person Re-Identification (CIReID), a challenging setting that requires continuously updating representations from newly added cameras without access to historical raw data. Unlike traditional closed-set forgetting, the primary challenge in CI-ReID lies in feature space distortion and representation fragmentation across domain shifts. To address this, we propose Camera-Incremental Geometric Representation Evolution (CIGRE). First, we introduce a Dynamic Geometric Alignment (DGA) mechanism, which flexibly expands an Equiangular Tight Frame (ETF) classifier to provide a stable and scalable discriminative geometric skeleton for the evergrowing identity set. Building upon this stable structure, we propose a Cross-Camera Identity Consolidation (CCIC) strategy. By maintaining momentum-updated global prototypes, CCIC progressively aggregates the fragmented identity features revealed under novel views. Furthermore, we establish a new CI-ReID benchmark derived from standard datasets. Extensive experiments demonstrate that CIGRE effectively scales with longterm camera expansion and achieves state-of-the-art performance by strictly maintaining cross-camera consistency. Our code and models will be released.

Sharing medical image sets is essential yet challenging due to limited resource and strict regulatory, which significantly hinder the development of medical foundation models and the transfer of medical domain knowledge. In this paper, we utilize medical image dataset distillation techniques to enable effective model construction and knowledge transfer without requiring downstream users to directly access the original data. Trajectory matching has become the mainstream distillation technique, yet our empirical findings on medical image datasets reveal two critical phenomena: instability of distillation optimization and redundancy of synthetic images. To investigate these phenomena, we conduct a theoretical analysis from the perspectives of bi-level optimization and neural tangent kernel, and formally define them as two core challenges: trajectory matching drift and intra-class diversity decay, which underscore the need for more stable and diversity-preserving trajectory matching methods. Guided by the analysis results, we propose a novel progressive and diverse dataset distillation (PD$^{3}$) method, which incorporates (1) a progressive trajectory matching paradigm that progressively unrolls the student training trajectory to extend the matching horizon, while adaptively selecting optimal expert matching trajectory to ensure stable and convergent distillation optimization; and (2) a dynamic diversity enhancing module that alleviates the redundancy of synthetic images by enhancing intra-class distribution and breaking redundancy accumulation. To facilitate fair evaluation, we establish a comprehensive medical image dataset distillation benchmark covering diverse medical image modalities. Extensive experiments show that the proposed method achieves an average absolute improvement of 10.20% and a relative gain of 20.09% over prior state-of-the-art methods, highlighting its effectiveness and practical potential.

In this paper, we propose VistaDreamer, a novel framework to reconstruct a 3D scene from a single-view image. Recent diffusion models enable generating high-quality novel-view images from a single-view input image. However, maintaining consistency across generated novel views remains challenging, often leading to noisy or distorted 3D reconstructions. VistaDreamer addresses this problem by a two-stage pipeline. In the first stage, VistaDreamer builds a global coarse 3D scaffold by zooming out a little step with inpainted boundaries and an estimated depth map. Then, on this global scaffold, iterative RGB-D inpainting generates novel-view images to fill the holes, guided by several geometric consistency constraints. In the second stage, we further enhance the consistency between the generated novel-view images by a novel training-free Multi-view Consistency Sampling (MCS) that introduces multi-view consistency constraints in reverse sampling process of diffusion models. VistaDreamer framework is compatible with various geometric estimators, diffusion models and samplers. Extensive experiments show that, without task-specific training or fine-tuning, it achieves good reconstruction quality and strong generalization across diverse scenes and trajectories, with ablation studies confirming the effectiveness and trends of our designs.

Inductive inference, or extrapolation of general rules from finite instances, is understood to be the foundation of human intelligence. Unfortunately, Deep Neural Networks (DNNs) struggle with inductive inference and thus fail to learn even the simplest algorithms in Algorithmic Reasoning (AR). Existing research efforts on AR with DNNs are limited to those on the architectural design for DNNs. In this study, we investigate the influence of optimization techniques on AR performance. Through toy experiments designed to understand an optimizer's susceptibility to non-generalizing solutions in AR, we reveal that Adam, the naïve choice of optimization, is easily fooled by spurious correlations. To overcome this shortcoming of Adam, we propose a novel optimizer that avoids spurious correlations by balancing gradients of low- and high-frequencies (BGF). We present extensive experiments and analyses to demonstrate the broad and multifaceted advantages of BGF across various architectures and AR tasks. In particular, BGF expands the AR capability of all explored DNN models and even shows the potential to enable learning of tasks that they previously failed at. The observed success of BGF in climbing the Chomsky hierarchy underscores the importance of optimization for developing advanced artificial intelligence with DNNs.