Large Language Models (LLMs) are increasingly explored for mental health assessment and psychiatric decision support, but autoregressive generation remains vulnerable to hallucination, semantic drift, and weak traceability. These risks carry particular consequences in psychiatric settings because decisions depend on structured diagnostic criteria, longitudinal symptom relations, exclusion rules, and clinically interpretable evidence. Retrieval-Augmented Generation (RAG) can improve factual grounding, yet pipelines centered on dense semantic retrieval do not naturally represent the discrete topology, hierarchical organization, and multi-hop dependencies of clinical ontologies and diagnostic rules. We therefore treat neuro-symbolic RAG as an information-fusion problem for explainable mental-health applications. We organize the literature around three fusion loci: pre-retrieval fusion, intra-retrieval alignment and re-ranking, and post-retrieval decoding and decision-level fusion. These loci are compared through source heterogeneity, fusion granularity, constraint strength, and evaluative consequences. The central questions are where neural and symbolic signals meet in the RAG pipeline, which mechanism families recur at each fusion locus, and how those choices shape strengths, risks, and evaluation burdens in mental health applications. To answer them, we draw selectively on adjacent clinical, biomedical, information retrieval, and general RAG literatures. The evaluation lens moves beyond task accuracy to include faithfulness, logical consistency, uncertainty and abstention behavior, and auditability; these dimensions structure a stage-aware evaluation framework. We close by outlining challenges in temporal reasoning, patient-specific graph construction, evaluation standardization, and clinically deployable human-in-the-loop neuro-symbolic systems.
Multi-attribute decision-making (MADM) faces the fundamental challenge of fusing uncertain and often conflicting evidence from diverse sources. The Dempster-Shafer evidence theory provides a rigorous framework for this task, wherein discounting evidence to quantify its reliability is crucial for robust fusion. However, the prevailing decision pathway involving discounted evidence fusion suffers from a dual computational bottleneck: the fusion process itself scales exponentially with the discernment frame (O(2|Θ|)), and optimizing discount factors within this model is equally prohibitive, hindering real-time deployment. We observe that decision-making often requires only a partial order among alternatives. This “order-preserving” requirement opens a new avenue for complexity reduction. First, we derive an equivalent model based on Pignistic Probability, which linearizes the fusion structure. Then, under strict order-consistency, we devise an efficient numerical approximation, reducing complexity to linear in the number of sources (O(|Θ|)). Building on this, we further propose a computationally efficient strategy for optimizing discount factors. Extensive evaluations on 12 public datasets demonstrate that our method outperforms 11 baselines in decision accuracy while also delivering significantly greater computational efficiency relative to other discount factor optimization methods. This work offers a theoretically grounded fast fusion framework for evidence-based reasoning, highlighting its potential to support scalable and efficient multi-attribute decision-makingMADM in future real-time systems.
In emergency disaster response, resource allocation conflicts stemming from functional specialization often trigger noncooperative behaviors that severely compromise group decision-making efficiency. Existing social network group decision-making methods face two critical limitations: static trust models fail to capture the dynamic evolution of decision-makers’ credibility, and penalty mechanisms tend to provoke psychological resistance in high-pressure environments. To address these issues, this study develops an incentive-driven consensus framework. The approach integrates subjective trust relationships with objective behavioral similarity to construct a dynamically updated composite trust scoring mechanism, which enables adaptive allocation of decision makers’ weights. Furthermore, it introduces a harmony degree-based feedback mechanism incorporating a maximum individual cooperation willingness detection model to accurately identify noncooperative decision-makers. The core contribution lies in its weight incentive mechanism, which stimulates cooperation by increasing the decision weights of noncooperators, thereby transforming the traditional penalty paradigm into a reputation-based incentive paradigm. Case studies and comparative analyses demonstrate that the framework not only achieves consensus at lower adjustment but also effectively transforms noncooperative behavior into proactive collaboration while ensuring fairness.
In large-scale and long-term robotic tasks, multi-robot localization and mapping is essential for environmental perception and collaborative navigation. However, under real-world conditions such as GPS-denied environments, unknown initial poses, and environmental changes, establishing stable inter-robot associations remains challenging, often causing accumulated localization errors and reducing global map consistency. To address this, we present Multi-Robot LiDAR-based localization and mapping with rotation-invariant Deep Relocalization (MRDR-SLAM). The system comprises four modules: single-robot odometry, intra-robot loop closure, inter-robot retrieval and matching, and multi-robot back-end optimization and mapping. At the core of the framework is a lightweight rotation-invariant feature extraction network (LRI-Net), which extracts rotation-invariant local geometric features and constructs global descriptors to enable efficient and robust inter-robot association, cross-robot loop detection, and map registration. On the public Snail Radar dataset, LRI-Net improves mean Recall@Top1 by 11.2% over the second-best method across four scenarios. On both self-collected and public multi-robot datasets, MRDR-SLAM achieves best or competitive trajectory accuracy in multiple scenarios, with clear improvements in multi-agent group alignment and global map consistency.
Driven by advances in machine learning and sensors, data-driven state estimation has become a key technology for supporting operational decisions in safety-critical systems. However, a single point estimate fails to reveal the uncertainty and confidence associated with a prediction, which can lead to overconfident decisions; as reliance on such estimates grows, the quantification and management of uncertainty becomes a prerequisite for reliable operation. This comprehensive review analyzes uncertainty management in data-driven state estimation, taking battery state estimation, in which current state estimation and future state prediction coexist, as a representative case study. This paper organizes uncertainty along two axes, its source (Aleatoric/Epistemic) and the mechanism that handles it (memory-based before modeling, model-based within the model), and applies this taxonomy to analyze how the dominant source and the management strategy suited to it shift along the prediction horizon. In doing so, it goes beyond a simple classification of techniques to present the correspondence between uncertainty sources and management strategies, and it summarizes the main open challenges and future research directions. The insights of this paper provide a foundation for reliable and risk-aware decision-making in predictive maintenance and safety-critical systems beyond batteries.
Multimodal learning aims to exploit cross-modal complementarity by integrating heterogeneous modalities such as text, audio, and vision. However, multimodal models often suffer from the imbalanced multimodal learning problem, where the dominant modality monopolizes optimization resources while weaker modalities remain insufficiently learned. Empirical studies show that, unlike unimodal training, gradients across modalities exhibit highly similar evolution trends in multimodal settings, even with existing balancing strategies, indicating that gradient magnitude adjustment alone fails to address the underlying optimization dynamics. In this work, we reanalyze imbalanced multimodal learning from an optimization perspective by incorporating Cognitive Load Theory. We show that, under limited model capacity, jointly optimizing multiple modalities induces a persistent high load state, where gradients from different modalities compete within a constrained update subspace, resulting in gradient conflicts and a persistent imbalanced state. Inspired by human cognitive behavior, we propose a multimodal balanced learning framework that integrates a plug-and-play multimodal alternating learning strategy to alleviate the learning load of simultaneous multimodal optimization and a Gaussian noise based optimizer to prevent overly similar gradient magnitude change trends. Extensive experiments on IEMOCAP, CMU-MOSEI, and AVE demonstrate the rationality and effectiveness of the proposed framework.
The manipulation of articulated objects for part-level motion is crucial due to their prevalence in real-world applications. Although current manipulation methods have improved interaction quality, they share a common issue: neglecting the completeness of motion trajectories. For example, when we use a front-loading drum washing machine, the expected action is to manipulate the door from fully closed to fully open. However, these methods might only result in it being half-open. To tackle this limitation, we introduce a novel framework for optimizing motion trajectories based on multimodal fusion. Specifically, we explicitly model trajectory completeness and propose a motion trajectory construction paradigm (MTCP). This paradigm is applied to a large-scale dataset containing a wide range of articulated objects, generating high-quality motion trajectories for multimodal fusion. Furthermore, to handle trajectory homogeneity, we propose a trajectory enhancement policy (TEP) that enriches the trajectory set by capturing the multimodal distribution of feasible trajectories. Subsequently, to enhance the learning efficiency and task adaptability of Multimodal Large Language Models (MLLMs), we propose a learning strategy for 3D perception inspired by 2D perception (3PI2P), complemented by a progressive reasoning approach. This strategy integrates a dual-branch input design using RGB images and depth maps, combined with six forms of visual question answering tasks, to achieve collaborative reasoning and deep fusion of 2D semantics and 3D geometric information. The robustness and generalizability of the framework are demonstrated through evaluations in both simulation and real-world environments.
Accurate meteorological forecasting in scenarios characterized by sparse observation stations and highly complex spatiotemporal variability remains a long-standing challenge. Traditional deep learning methods often struggle to capture the heterogeneous and dynamically evolving dependencies inherent in atmospheric processes under complex scenarios. To address this limitation, we propose NaSH (Navigating Spatiotemporal Heterogeneity), a graph Transformer with Mixture-of-Experts (MoE), which harnesses spatiotemporal heterogeneity to achieve dynamic adaptation and precise sparse meteorological forecasting in complex scenarios. NaSH integrates three specialized expert networks: temporal expert, static spatiotemporal expert, and dynamic spatiotemporal expert, which respectively model diverse spatiotemporal dependencies across different weather scenarios. To achieve precise scene-to-expert alignment, we then design a Memory-Aware Router (MAR), which incorporates a Meta-node Bank to learn diverse graph structures and employs a memory-matching mechanism to dynamically select the most suitable expert. In addition, a joint loss function is introduced to mitigate routing inertia and enhance expert specialization. Comprehensive experiments on two real-world datasets representing complex atmospheric conditions demonstrate that the proposed method consistently achieves superior forecasting performance across multiple lead times. Further analyses confirm the reliability of the method across diverse scenarios while providing interpretable insights into the learned spatiotemporal patterns. These results highlight the proposed method as a practical and effective solution for sparse meteorological forecasting in regions characterized by complex terrain and heterogeneous climate types.
Mobile edge computing (MEC) has emerged as an effective solution for processing compute-intensive and delay-sensitive tasks in multi-user, multi-server environments. To better orchestrate such complex and heterogeneous scenarios, we integrate a software-defined network to enable centralized coordination of computation offloading and resource allocation. Based on this architecture, a time-variant mixed-integer optimization problem is formulated to minimize system-wide latency and energy cost. Multi-agent reinforcement learning (MARL) offers a theoretical framework for solving this problem, yet its application introduces two major challenges. First, partial observability arises because user equipment (UEs) lack access to global information, which hinders coordinated decision-making. Second, the system involves a hybrid action space comprising discrete offloading decisions and continuous bandwidth allocations, complicating the policy learning process. To address these issues, we propose a novel graph-based asynchronous hierarchical MARL (GAH-MARL) algorithm that enables multi-agent and multi-level information fusion. A graph-communication module allows UEs to selectively share local observations, thereby mitigating partial observability. In addition, a hierarchical learning structure integrates information across decision levels and decouples the optimization of discrete and continuous actions. Furthermore, asynchronous training is incorporated to reduce computational cost and enhance learning efficiency. Simulation results demonstrate that GAH-MARL significantly outperforms existing MARL baselines in reducing latency and energy consumption, particularly under large-scale deployments.
In this paper, we propose CorrFuse, a unified cross-modal dynamic fusion framework for multimodal sentiment analysis (MSA). CorrFuse is designed to address several common challenges in MSA. These include modal asynchrony, signal degradation, limited static fusion strategies, and inconsistent ordinal predictions. First, we design a Cross-modal multi-round Embedding (CE) module. This module iteratively integrates visual and acoustic information using a multi-round cross-attention mechanism. It effectively mitigates the problems of modal asynchrony and signal degradation. Second, we introduce a sentiment-intensity-conditioned dynamic attention mechanism and a correlation alignment loss. These allow real-time adaptive adjustments of fusion weights and dynamically suppress noise interference. Finally, we adopt the Consistent Rank Logits (CORAL) ordinal regression. This is combined with a multi-task collaborative optimization strategy. The discrete and continuous dual constraints effectively guarantee monotonicity and semantic continuity in sentiment prediction. Experimental results on CMU-MOSI, CMU-MOSEI, CH-SIMS and CH-SIMS v2 datasets demonstrate that CorrFuse consistently outperforms existing methods. It shows strong performance and potential practical applicability.
Graph-based electroencephalogram (EEG) emotion recognition methods model inter-channel relationships by constructing connectivity graphs over EEG electrodes. Fixed structural priors may be insufficient to characterize subject- and trial-specific variations, whereas highly adaptive graphs may be sensitive to noise. In addition, temporal and spectral EEG representations are often fused mainly at the feature level, leaving the graph-level relationship between representation domains relatively underexplored. To address these issues, we propose PSAGAN, a Prior-Guided Sample-Adaptive Graph Attention Network with diffusion-based cross-domain graph topology regularization for EEG-based emotion recognition. PSAGAN constructs spatial-spectral and spatial-temporal pathways and introduces a prior-guided sample-adaptive graph attention mechanism within each pathway. This mechanism incorporates distance-based anatomical priors into feature-dependent attention computation, enabling the model to use structural guidance while learning sample-dependent edge weights within anatomically guided neighborhoods. To coordinate the two pathway-specific graphs, PSAGAN introduces a diffusion-based topology regularization strategy that compares multi-scale diffusion signatures and uses category-level contrastive learning to encourage coarse-grained topological coherence while retaining domain-specific learned connectivity patterns. An adaptive integration module then combines the two pathway representations for classification. Experiments on SEED and SEED-IV under both subject-dependent and subject-independent settings, and on DEAP under trial-grouped 10-fold cross-validation, show that PSAGAN achieves competitive performance under the adopted evaluation protocols. On SEED, PSAGAN obtains 98.39% subject-dependent accuracy and 90.18% subject-independent accuracy. On SEED-IV, it achieves 90.57% and 79.36% under the two settings, respectively. On DEAP, it reaches 97.83% and 98.21% for valence and arousal classification, respectively. Ablation studies indicate that prior-guided graph learning, diffusion-based topology regularization, and adaptive integration provide complementary benefits. Additional evaluations on motor imagery, sleep staging, and depression detection suggest potential applicability to broader multi-channel EEG tasks.
Visual–inertial odometry (VIO) can be viewed as a structured multimodal fusion problem, where heterogeneous sensory streams with time-varying reliability must be integrated to estimate ego-motion. Most learning-based VIO methods implicitly assume that all modalities are equally reliable at inference time and perform unconditional or softly gated fusion. However, when reliability becomes asymmetric or intermittently uncertain, such strategies may propagate corrupted information across modalities and degrade estimation stability. In this work, we propose WormVIO, a reliability-aware deep VIO framework motivated by compact sensorimotor decision processes, which formulates adaptive inference as a hierarchical discrete fusion problem. At its core, the Instinct-Bias Module is a hierarchical reliability-aware fusion mechanism that explicitly disentangles two decision factors: (i) fusion participation, which determines whether multimodal aggregation is beneficial at a given time step, and (ii) modality dominance, which determines which modality should guide estimation when reliability is imbalanced. These discrete decisions are implemented using a compact Neural Circuit Policy (NCP) combined with differentiable Gumbel-Softmax sampling, enabling end-to-end optimization of structured fusion control. Experiments on the KITTI benchmark demonstrate that the proposed decoupled fusion mechanism improves trajectory stability under intermittent visual and inertial perturbations while preserving competitive performance under nominal conditions. Furthermore, without fine-tuning, WormVIO transfers more consistently to the EuRoC MAV dataset, indicating that disentangling fusion participation and modality dominance enhances robustness to reliability shifts and cross-domain motion variations. The source code is available at: https://github.com/zRiverBird/worm_vio.git.
Managing non-cooperative behaviors is a critical challenge in the consensus reaching process of large-scale group decision-making (LSGDM). Existing studies usually rely on static social relationships and non-overlapping community structures, which cannot adequately capture the dynamic interactions and multiple community memberships of decision experts. To address this issue, this study proposes a consensus model that fuses dynamic social network relationships and overlapping community structures for non-cooperative behavior management in LSGDM. Firstly, the COPRA algorithm is used to cluster decision experts and effectively identify experts and communities with overlapping structures, thus capturing the complex social interaction patterns among decision experts. Secondly, the PageRank algorithm is used to calculate the weights of decision experts based on the social network structure. On this basis, a consensus model is constructed by fusing dynamic network relationships and overlapping communities, which guides experts to adjust their opinions through feedback mechanisms, thereby reducing the influence of non-cooperative behaviors and promoting consensus reaching. To verify the rationality of the proposed method, the model is applied to a subway line planning decision problem. Finally, simulation experiments and comparative analyses demonstrate that the proposed method is effective in managing non-cooperative behaviors and improving consensus performance compared with existing approaches.
Multimodal Affective Analysis (MAA) has made significant progress, but current methods still lack an explicit mechanism to decide when modality-specific (private) features should be combined with modality-invariant (shared) representations and when they should be ignored. Addressing this limitation is necessary because such private cues can be redundant, conflicting, or noisy. We propose SPriG, a shared–private architecture that adds improvement-guided control over private information. Each modality is mapped to a common feature space and split into shared and private (modality-specific) components. A shared-only fusion path first produces a robust baseline prediction, and a gated private residual then allows each modality to contribute additional information only when it helps relative to this baseline. A set of auxiliary objectives guides the disentanglement and model training behaviour. We evaluate SPriG across five public benchmarks covering key areas of affective computing, including sentiment prediction, conversational emotion classification, and humor detection. Across these datasets, SPriG consistently outperforms recent state-of-the-art methods on both regression and classification metrics, demonstrating the effectiveness of our approach.
Diffusion models have achieved breakthrough progress in single-object video editing. However, existing methods still suffer from two limitations when applied directly to multi-object editing: (1) These methods struggle to understand the spatial relationships and layouts of multi-object in the video, leading to semantic confusion and misalignment. (2) They lack explicit semantic consistency supervision in complex scenarios and editing targets, resulting in generated results deviating from text prompt. To address the above limitations, we propose SemEdit, a novel semantic-aware content alignment scheme for fine-grained multi-object video editing. Specifically, SemEdit comprises two key components: Semantic Adaptive Modulation and Semantic Prior Modeling. The former utilizes region-aware and text-aware attention modulation to achieve feature decoupling and precise object localization by adjusting the spatial semantics distribution. Collaboratively, the latter leverages semantic prior to model high-level content correlations, thus ensuring alignment between the edited video and input prompt. Benefiting from the above designs, our method achieves precise spatial object alignment while maintaining semantic consistency across the entire video. Extensive experiments show that our SemEdit outperforms existing video editing methods in both editing precision and semantic alignment, providing a new perspective for multi-object video editing.
Engineering and intelligent systems increasingly require decision-making under heterogeneous evidence. These sources include multi-source data, predictive models, competing objectives, feasibility constraints, and uncertainty. Evolutionary algorithms (EAs) are widely used in such settings, yet the literature is typically organised by algorithmic lineage, which obscures how and where information is fused within the evolutionary process. This survey reframes EAs as adaptive information fusion architectures. We introduce a fusion-centric taxonomy spanning data-level integration of heterogeneous observations, model/feature-level integration through surrogate and learning components, objective-level integration through multi- and many-objective formulations, constraint handling as feasibility-signal integration, and decision-level integration through ensembles, distributed (island) evolution, and multi-run aggregation. Building on this perspective, we synthesise major EA families according to their dominant fusion mechanisms and review fusion-driven applications across structural and mechanical design, energy and smart grids, robotics and control, communications, healthcare engineering, and neural architecture search. To complement this qualitative synthesis, we propose lightweight quantitative indicators—integration depth, integration diversity, and decision outcome entropy—to characterise algorithm–fusion–domain alignment and to identify recurring success and failure modes, including surrogate bias, over-aggregation, and fusion over-complexity. We conclude with practical design guidelines and discuss emerging directions, including federated evolutionary fusion and reliability considerations in learning- and large language model (LLM)-assisted fusion.
