DVM: Real-Time Kernel Generation for Dynamic AI Models

Dynamism is common in AI computation, e.g., the dynamic tensor shapes and the dynamic control flows in models. Due to the long compilation time, existing runtime compilation damages the model efficiency, while the offline compilers either suffer from the long compilation time and device memory footprint to cover all the possible execution instances of a dynamic model, or sacrifice optimization opportunities for usability. In this paper, we rethink the feasibility of runtime compilation for dynamic models and identify that the key for it to work is to speed up the compilation or hide the compilation overhead. To do this, we propose a real-time compiler, DVM. In DVM, we design a runtime operator compiler based on a bytecode virtual machine to perform effective and efficient compilation for each dynamic operator instance given its input. Specifically, instead of compiling programs into machine code, we encode the operator program into bytecode on the CPU and decode the bytecode into virtual instructions for direct execution on the NPU. Based on the runtime operator compiler, we further propose an operator fuser, which performs symbol-deduction-based fusion on static graphs and runtime fusion on dynamic graphs. Both pattern- and stacking-based fusion are supported to increase fusion opportunities. Evaluation on operators, subgraphs, and models shows that, compared with TorchInductor, PyTorch-eager and MindSpore-graph-O0, we are up to 11.77$\times$ better in terms of the operator/model efficiency and up to 5 orders of magnitude faster in terms of the maximum compilation time.

TagLLM: A Fine-Grained Tag Generation Approach for Note Recommendation

Large Language Models (LLMs) have shown promising potential in E-commerce community recommendation. While LLMs and Multimodal LLMs (MLLMs) are widely used to encode notes into implicit embeddings, leveraging their generative capabilities to represent notes with interpretable tags remains unexplored. In the field of tag generation, traditional close-ended methods heavily rely on the design of tag pools, while existing open-ended methods applied directly to note recommendations face two limitations: (1) MLLMs lack guidance during generation, resulting in redundant tags that fail to capture user interests; (2) The generated tags are often coarse and lack fine-grained representation of notes, interfering with downstream recommendations. To address these limitations, we propose TagLLM, a fine-grained tag generation method for note recommendation. TagLLM captures user interests across note categories through a User Interest Handbook and constructs fine-grained tag data using multimodal CoT Extraction. A Tag Knowledge Distillation method is developed to equip small models with competitive generation capabilities, enhancing inference efficiency. In online A/B test, TagLLM increases average view duration per user by 0.31%, average interactions per user by 0.96%, and page view click-through rate in cold-start scenario by 32.37%, demonstrating its effectiveness.

More Than Sum of Its Parts: Deciphering Intent Shifts in Multimodal Hate Speech Detection

Combating hate speech on social media is critical for securing cyberspace, yet relies heavily on the efficacy of automated detection systems. As content formats evolve, hate speech is transitioning from solely plain text to complex multimodal expressions, making implicit attacks harder to spot. Current systems, however, often falter on these subtle cases, as they struggle with multimodal content where the emergent meaning transcends the aggregation of individual modalities. To bridge this gap, we move beyond binary classification to characterize semantic intent shifts where modalities interact to construct implicit hate from benign cues or neutralize toxicity through semantic inversion. Guided by this fine-grained formulation, we curate the Hate via Vision-Language Interplay (H-VLI) benchmark where the true intent hinges on the intricate interplay of modalities rather than overt visual or textual slurs. To effectively decipher these complex cues, we further propose the Asymmetric Reasoning via Courtroom Agent DEbate (ARCADE) framework. By simulating a judicial process where agents actively argue for accusation and defense, ARCADE forces the model to scrutinize deep semantic cues before reaching a verdict. Extensive experiments demonstrate that ARCADE significantly outperforms state-of-the-art baselines on H-VLI, particularly for challenging implicit cases, while maintaining competitive performance on established benchmarks. Our code and data are available at: https://github.com/Sayur1n/H-VLI

Data-driven ensemble prediction of the global ocean

Data-driven models have advanced deterministic ocean forecasting, but extending machine learning to probabilistic global ocean prediction remains an open challenge. Here we introduce FuXi-ONS, the first machine-learning ensemble forecasting system for the global ocean, providing 5-day forecasts on a global 1° grid up to 365 days for sea-surface temperature, sea-surface height, subsurface temperature, salinity and ocean currents. Rather than relying on repeated integration of computationally expensive numerical models, FuXi-ONS learns physically structured perturbations and incorporates an atmospheric encoding module to stabilize long-range forecasts. Evaluated against GLORYS12 reanalysis, FuXi-ONS improves both ensemble-mean skill and probabilistic forecast quality relative to deterministic and noise-perturbed baselines, and shows competitive performance against established seasonal forecast references for SST and Niño3.4 variability, while running orders of magnitude faster than conventional ensemble systems. These results provide a strong example of machine learning advancing a core problem in ocean science, and establish a practical path toward efficient probabilistic ocean forecasting and climate risk assessment.

Fast Low-light Enhancement and Deblurring for 3D Dark Scenes

Novel view synthesis from low-light, noisy, and motion-blurred imagery remains a valuable and challenging task. Current volumetric rendering methods struggle with compound degradation, and sequential 2D preprocessing introduces artifacts due to interdependencies. In this work, we introduce FLED-GS, a fast low-light enhancement and deblurring framework that reformulates 3D scene restoration as an alternating cycle of enhancement and reconstruction. Specifically, FLED-GS inserts several intermediate brightness anchors to enable progressive recovery, preventing noise blow-up from harming deblurring or geometry. Each iteration sharpens inputs with an off-the-shelf 2D deblurrer and then performs noise-aware 3DGS reconstruction that estimates and suppresses noise while producing clean priors for the next level. Experiments show FLED-GS outperforms state-of-the-art LuSh-NeRF, achieving 21$\times$ faster training and 11$\times$ faster rendering.

K-Gen: A Multimodal Language-Conditioned Approach for Interpretable Keypoint-Guided Trajectory Generation

Generating realistic and diverse trajectories is a critical challenge in autonomous driving simulation. While Large Language Models (LLMs) show promise, existing methods often rely on structured data like vectorized maps, which fail to capture the rich, unstructured visual context of a scene. To address this, we propose K-Gen, an interpretable keypoint-guided multimodal framework that leverages Multimodal Large Language Models (MLLMs) to unify rasterized BEV map inputs with textual scene descriptions. Instead of directly predicting full trajectories, K-Gen generates interpretable keypoints along with reasoning that reflects agent intentions, which are subsequently refined into accurate trajectories by a refinement module. To further enhance keypoint generation, we apply T-DAPO, a trajectory-aware reinforcement fine-tuning algorithm. Experiments on WOMD and nuPlan demonstrate that K-Gen outperforms existing baselines, highlighting the effectiveness of combining multimodal reasoning with keypoint-guided trajectory generation.

DSA-SRGS: Super-Resolution Gaussian Splatting for Dynamic Sparse-View DSA Reconstruction

Digital subtraction angiography (DSA) is a key imaging technique for the auxiliary diagnosis and treatment of cerebrovascular diseases. Recent advancements in gaussian splatting and dynamic neural representations have enabled robust 3D vessel reconstruction from sparse dynamic inputs. However, these methods are fundamentally constrained by the resolution of input projections, where performing naive upsampling to enhance rendering resolution inevitably results in severe blurring and aliasing artifacts. Such lack of super-resolution capability prevents the reconstructed 4D models from recovering fine-grained vascular details and intricate branching structures, which restricts their application in precision diagnosis and treatment. To solve this problem, this paper proposes DSA-SRGS, the first super-resolution gaussian splatting framework for dynamic sparse-view DSA reconstruction. Specifically, we introduce a Multi-Fidelity Texture Learning Module that integrates high-quality priors from a fine-tuned DSA-specific super-resolution model, into the 4D reconstruction optimization. To mitigate potential hallucination artifacts from pseudo-labels, this module employs a Confidence-Aware Strategy to adaptively weight supervision signals between the original low-resolution projections and the generated high-resolution pseudo-labels. Furthermore, we develop Radiative Sub-Pixel Densification, an adaptive strategy that leverages gradient accumulation from high-resolution sub-pixel sampling to refine the 4D radiative gaussian kernels. Extensive experiments on two clinical DSA datasets demonstrate that DSA-SRGS significantly outperforms state-of-the-art methods in both quantitative metrics and qualitative visual fidelity.

BAED: a New Paradigm for Few-shot Graph Learning with Explanation in the Loop

The challenges of training and inference in few-shot environments persist in the area of graph representation learning. The quality and quantity of labels are often insufficient due to the extensive expert knowledge required to annotate graph data. In this context, Few-Shot Graph Learning (FSGL) approaches have been developed over the years. Through sophisticated neural architectures and customized training pipelines, these approaches enhance model adaptability to new label distributions. However, compromises in \textcolor{black}{the model's} robustness and interpretability can result in overfitting to noise in labeled data and degraded performance. This paper introduces the first explanation-in-the-loop framework for the FSGL problem, called BAED. We novelly employ the belief propagation algorithm to facilitate label augmentation on graphs. Then, leveraging an auxiliary graph neural network and the gradient backpropagation method, our framework effectively extracts explanatory subgraphs surrounding target nodes. The final predictions are based on these informative subgraphs while mitigating the influence of redundant information from neighboring nodes. Extensive experiments on seven benchmark datasets demonstrate superior prediction accuracy, training efficiency, and explanation quality of BAED. As a pioneer, this work highlights the potential of the explanation-based research paradigm in FSGL.

Rethinking ANN-based Retrieval: Multifaceted Learnable Index for Large-scale Recommendation System

Approximate nearest neighbor (ANN) search is widely used in the retrieval stage of large-scale recommendation systems. In this stage, candidate items are indexed using their learned embedding vectors, and ANN search is executed for each user (or item) query to retrieve a set of relevant items. However, ANN-based retrieval has two key limitations. First, item embeddings and their indices are typically learned in separate stages: indexing is often performed offline after embeddings are trained, which can yield suboptimal retrieval quality-especially for newly created items. Second, although ANN offers sublinear query time, it must still be run for every request, incurring substantial computation cost at industry scale. In this paper, we propose MultiFaceted Learnable Index (MFLI), a scalable, real-time retrieval paradigm that learns multifaceted item embeddings and indices within a unified framework and eliminates ANN search at serving time. Specifically, we construct a multifaceted hierarchical codebook via residual quantization of item embeddings and co-train the codebook with the embeddings. We further introduce an efficient multifaceted indexing structure and mechanisms that support real-time updates. At serving time, the learned hierarchical indices are used directly to identify relevant items, avoiding ANN search altogether. Extensive experiments on real-world data with billions of users show that MFLI improves recall on engagement tasks by up to 11.8\%, cold-content delivery by up to 57.29\%, and semantic relevance by 13.5\% compared with prior state-of-the-art methods. We also deploy MFLI in the system and report online experimental results demonstrating improved engagement, less popularity bias, and higher serving efficiency.

WiSparse: Boosting LLM Inference Efficiency with Weight-Aware Mixed Activation Sparsity

Large Language Models (LLMs) offer strong capabilities but incur high inference costs due to dense computation and memory access. Training-free activation sparsity is a promising approach for efficient LLM inference, yet existing methods often rely solely on activation information and uniform sparsity ratios. This overlooks the critical interplay with weights and inter-block sensitivity variation, leading to suboptimal performance. We identify two key phenomena in modern LLMs: 1) less significant activations may align with highly important weights, and 2) sparsity sensitivity varies non-monotonically across model blocks. We propose Weight-aware Mixed-Granularity Training-free Activation Sparsity (WiSparse), which leverages both activation and weight information for adaptive sparsity allocation. Specifically, we introduce a weight-aware mechanism integrating activation magnitudes with precomputed weight norms to accurately identify salient channels. This is combined with a mixed-granularity allocation scheme: a global budget is distributed across blocks via evolutionary search to protect sensitive regions, then refined within blocks to minimize reconstruction error. We improve sparse kernels and demonstrate effectiveness on three representative models. Notably, at 50% sparsity, WiSparse preserves 97% of Llama3.1's dense performance, surpassing the strongest baseline by 2.23 percentage points while achieving a 21.4% acceleration in end-to-end inference speed. Our research advances the limits of training-free approaches for efficient LLM inference, pushing the boundaries of achievable speedup without training.