Timely nowcasting is critical for public safety during fast-evolving storms, where even short delays can trigger cascading failures—as in the October 2024 Spain flash flood that claimed over 90 lives within minutes. While radar offers reliable real-time sensing of atmospheric structure, models that collapse 3D volumes into 2D slices inevitably discard vertical information essential for capturing storm growth, phase transitions, and collapse. We introduce StormMind, a physically grounded framework that forecasts convective evolution by modeling causal interactions across stratified atmospheric layers. StormMind addresses two fundamental challenges:(1) the nonlinear, asynchronous coupling between low-, mid-, and high-level processes; and (2) reflectivity uncertainty, where storms with distinct vertical structures may appear deceptively similar on radar, masking their true phase and intensity. To tackle these issues, StormMind designs: i) a Convection Dynamics Extractor that models storm evolution from two complementary perspectives—horizontal morphology, capturing the spatial organization of physical processes within individual atmospheric layers, and vertical coupling, modeling energy exchanges across layers; and ii) a Convection Manifestation Reconstructor that adaptively fuses intra- and inter-layer signals, conditioned on the evolving storm state, to infer phase transitions (e.g., initiation, intensification, dissipation). Evaluated on the large-scale 3D-NEXRAD dataset (2020–2022, U.S.), StormMind outperforms strong baselines, achieving a 14.71% gain in CSI40. In real-world deployment with the Guangzhou Meteorological Bureau (Mar–May 2025), it improves CSI40 by 9.39% and boosts early-warning accuracy (98.33%)

Magnetic microrobots capable of autonomous operation hold significant promise for critical cell and small creature manipulation tasks, including trapping, transportation, sorting, etc. Although conventional microrobot navigation methods have shown notable performance, they often lack adaptability to novel environments. To address these limitations, we propose a learning-based framework for real-world microrobot navigation and dynamic target tracking. Our approach employs spatial-temporal transformer reinforcement learning (STTRL) with a deterministic velocity obstacle (DVO) that processes historical navigation states and virtual light detection and ranging (LiDAR) scans to predict optimal actions. The key innovation lies in the model's ability to extract and utilize contextual information from observation histories, enabling adaptive behavior even in previously unseen environments. Through large-scale model-free reinforcement learning trained in randomized simulation environments, our method achieves remarkable real-world performance with zero-shot transfer capability. Experimental results demonstrate superior navigation agility with an 89.8% success rate in the base environment, representing a 7.4% improvement over state-of-the-art (SOTA) algorithms. Furthermore, the method exhibits robust generalization in diverse unseen environments, validating its adaptability to different environmental characteristics.

Image denoising is essential in low-level vision applications such as photography and automated driving. Existing methods struggle with distinguishing complex noise patterns in real-world scenes and consume significant computational resources due to reliance on Transformer-based models. In this work, the Context-guided Receptance Weighted Key-Value (CRWKV) model is proposed, combining enhanced multi-view feature integration with efficient sequence modeling. The Context-guided Token Shift (CTS) mechanism is introduced to effectively capture local spatial dependencies and enhance the model's ability to model real-world noise distributions. Also, the Frequency Mix (FMix) module extracting frequency-domain features is designed to isolate noise in high-frequency spectra, and is integrated with spatial representations through a multi-view learning process. To improve computational efficiency, the Bidirectional WKV (BiWKV) mechanism is adopted, enabling full pixel-sequence interaction with linear complexity while overcoming the causal selection constraints. The model is validated on multiple real-world image denoising datasets, outperforming the state-of-the-art methods quantitatively and reducing inference time up to 40%. Qualitative results further demonstrate the ability of our model to restore fine details in various scenes. The code is publicly available at https://github.com/Seeker98/CRWKV.

Unsupervised hashing is applied in large-scale multimodal retrieval by mapping original data from heterogeneous modalities into compact binary codes. Transformer-based retrieval augmented generation possesses significant advantages in retrieval accuracy and context-awareness, yet faces scalability challenges due to the computational overhead of dense embedding. Thus, the integration of hash learning and Transformer provides a feasible improvement scheme, which can achieve efficient retrieval preserving semantic association. This paper proposes a novel Unsupervised Similarity-Fusion Transformer Hashing for multimodal retrieval, denoted as USFTH. Initially, the modal fusion similarity matrix based on Gaussian kernel, sigmoid function, and Laplacian transformation is introduced to construct a discriminative similarity matrix, ensuring that semantic correlation among samples can be captured precisely. Then, cross-modal multiplex joint construction via Transformer-based attention mechanisms is designed, realizing effective integration of heterogeneous modalities in the similarity matrix through multi-path fusion. Furthermore, the consensus fusion strategy is proposed to ensure that hash codes generated under unsupervised conditions possess a uniform distribution and achieve accurate retrieval. In addition, comprehensive experiments on MIRFlickr, NUS-WIDE, and IAPR-TC12 datasets demonstrate the superior performance of USFTH to state-of-the-art hashing approaches.

In the field of Medical image processing, Transformers and CNNs bring filter-based local sensing and attention-based global modelling, which strengthened their dominant position. However, the scales required in different applications vary, and improper usage or selection on models can lead to performance decrease, resulting in more serious consequences in downstream tasks. The recent advanced Structured State Space Model has shown its potential as a baseline model, and is further proved to maintaining implicit duality with the attention mechanism, The corresponding model Mamba2 shows its long-range modelling ability with optimized long-term memory and linear computational complexity.In this paper, we propose SSD-MIR, a novel architecture leveraging the advantages of state space duality mechanism. SSD-MIR includes three parts: input projection and feature representation, deep feature transformation, output projection and feature inversion. With the three components sequentially connected to form our model, it is able to model long-range dependencies and learn the mapping between low-quality input and high-quality output simultaneously. In the main deep feature transformation part, depthwise convolution is used for additional feature mapping as an alternative of 1D causal convolution, and an optimized quad-direction scanning mechanism is implemented to form the overall 2D state space dual module. The corresponding wrapper module is built in MetaFormer style with the usage of residual connection and LayerNorm. We perform extensive experiments on multiple medical image restoration tasks, and the results are superior to the state-of-the-arts, demonstrating the strength and efficacy of our model.

Spatial transcriptomics data provides new methods of exploring gene expressions. However, the results usually do not contain cell-level proportions, leaving a gap between the new technologies and the corresponding analysis. To address the issues, we propose a novel framework, OptiGraph3D, in this paper. Our approach starts with a pivoted sample-level sequential registration, eliminating potential cumulative errors in multiple registrations. The registration results provide sufficient spatial prior for our cell-type deconvolving process, which leverages the power of ADMM method by modelling the cell-type deconvolution task as a matrix inverse problem and proposes efficient solvers with guaranteed convergence. Furthermore, the cell-type deconvolution results can in turn perform as supervision for cluster recognition in 3D reconstruction. Our model introduces a semi-supervision training manner by introducing a specific class token in the multi-headed graph attention network, bridging the gap between gene-level expressions of spatial transcriptomics and presenting new angle for sample-wise analysis. Extensive experiments are performed on the human dorsolateral prefrontal cortex (DLPFC) and the developing human embryonic heart dataset, suggesting our approach's superiority over other methods.

Recent works have shown the promise of learning pre-trained models for 3D molecular representation. However, existing pre-training models focus predominantly on equilibrium data and largely overlook off-equilibrium conformations. It is challenging to extend these methods to off-equilibrium data because their training objective relies on assumptions of conformations being the local energy minima. We address this gap by proposing a force-centric pretraining model for 3D molecular conformations covering both equilibrium and off-equilibrium data. For off-equilibrium data, our model learns directly from their atomic forces. For equilibrium data, we introduce zero-force regularization and forced-based denoising techniques to approximate near-equilibrium forces. We obtain a unified pre-trained model for 3D molecular representation with over 15 million diverse conformations. Experiments show that, with our pre-training objective, we increase forces accuracy by around 3 times compared to the un-pre-trained Equivariant Transformer model. By incorporating regularizations on equilibrium data, we solved the problem of unstable MD simulations in vanilla Equivariant Transformers, achieving state-of-the-art simulation performance with 2.45 times faster inference time than NequIP. As a powerful molecular encoder, our pre-trained model achieves on-par performance with state-of-the-art property prediction tasks.

A wide range of modern science and engineering applications are formulated as optimization problems with a system of partial differential equations (PDEs) as constraints. These PDE-constrained optimization problems are typically solved in a standard discretize-then-optimize approach. In many industry applications that require high-resolution solutions, the discretized constraints can easily have millions or even billions of variables, making it very slow for the standard iterative optimizer to solve the exact gradients. In this work, we propose a general framework to speed up PDE-constrained optimization using online neural synthetic gradients (ONSG) with a novel two-scale optimization scheme. We successfully apply our ONSG framework to computational morphogenesis, a representative and challenging class of PDE-constrained optimization problems. Extensive experiments have demonstrated that our method can significantly speed up computational morphogenesis (also known as topology optimization), and meanwhile maintain the quality of final solution compared to the standard optimizer. On a large-scale 3D optimal design problem with around 1,400,000 design variables, our method achieves up to 7.5x speedup while producing optimized designs with comparable objectives.

Deep learning and symbolic reasoning are complementary techniques for an intelligent system. However, principled combinations of these techniques have limited scalability, rendering them ill-suited for real-world applications. We propose Scallop, a system that builds upon probabilistic deductive databases, to bridge this gap. The key insight underlying Scallop is a provenance framework that introduces a tunable parameter to specify the level of reasoning granularity. Scallop thereby i) generalizes exact probabilistic reasoning, ii) asymptotically reduces computational cost, and iii) provides relative accuracy guarantees. On a suite of tasks that involve mathematical and logical reasoning, Scallop scales significantly better without sacrificing accuracy compared to DeepProbLog, a principled neural logic programming approach. We also create and evaluate on a real-world Visual Question Answering (VQA) benchmark that requires multi-hop reasoning. Scallop outperforms two VQA-tailored models, a Neural Module Networks based and a transformer based model, by 12.42% and 21.66% respectively.

Programs, consisting of semantic and structural information, play an important role in the communication between humans and agents. Towards learning general program executors to unify perception, reasoning, and decision making, we formulate program-guided tasks which require learning to execute a given program on the observed task specification. Furthermore, we propose Program-Guided Transformers (ProTo), which integrates both semantic and structural guidance of a program by leveraging cross-attention and masked self-attention to pass messages between the specification and routines in the program. ProTo executes a program in a learned latent space and enjoys stronger representation ability than previous neural-symbolic approaches. We demonstrate that ProTo significantly outperforms the previous state-of-the-art methods on GQA visual reasoning and 2D Minecraft policy learning datasets. Additionally, ProTo demonstrates better generalization to unseen, complex, and human-written programs.

Software APIs exhibit rich diversity and complexity which not only renders them a common source of programming errors but also hinders program analysis tools for checking them. Such tools either expect a precise API specification, which requires program analysis expertise, or presume that correct API usages follow simple idioms that can be automatically mined from code, which suffers from poor accuracy. We propose a new approach that allows regular programmers to find API misuses. Our approach interacts with the user to classify valid and invalid usages of each target API method. It minimizes user burden by employing an active learning algorithm that ranks API usages by their likelihood of being invalid. We implemented our approach in a tool called ARBITRAR for C/C++ programs, and applied it to check the uses of 18 API methods in 21 large real-world programs, including OpenSSL and Linux Kernel. Within just 3 rounds of user interaction on average per API method, ARBITRAR found 40 new bugs, with patches accepted for 18 of them. Moreover, ARBITRAR finds all known bugs reported by a state-of-the-art tool APISAN in a benchmark suite comprising 92 bugs with a false positive rate of only 51.5% compared to APISAN’s 87.9%.

Retrosynthetic planning is a critical task in organic chemistry which identifies a series of reactions that can lead to the synthesis of a target product. The vast number of possible chemical transformations makes the size of the search space very big, and retrosynthetic planning is challenging even for experienced chemists. However, existing methods either require expensive return estimation by rollout with high variance, or optimize for search speed rather than the quality. In this paper, we propose Retro*, a neural-based A*-like algorithm that finds high-quality synthetic routes efficiently. It maintains the search as an AND-OR tree, and learns a neural search bias with off-policy data. Then guided by this neural network, it performs best-first search efficiently during new planning episodes. Experiments on benchmark USPTO datasets show that, our proposed method outperforms existing state-of-the-art with respect to both the success rate and solution quality, while being more efficient at the same time.