Test-Time Scaling (TTS) improves large language models (LLMs) by allocating additional computation during inference, typically through parallel, sequential, or hybrid scaling. However, prior studies often assume fixed collaboration architectures (e.g., topologies) and single-model usage, overlooking that optimal architectures and model combinations can vary across tasks. Therefore, we study the novel problem of searching for compute-optimal model combinations and architectures in TTS under a fixed budget. We formalize it as a multi-LLM collaboration graph, where nodes encode roles and LLM model assignments, and edges capture information flow. This problem is challenging because (i) the combinatorial search space is prohibitively large, and (ii) task-specific requirements demand tailored designs. To address these, we reformulate the problem as probabilistic graph optimization and, through pilot experiments, derive three empirical insights into TTS collaboration graphs. Guided by these insights, we propose Agent-REINFORCE, an LLM-agent-augmented framework that mirrors the REINFORCE pipeline by mapping sampling-gradient-update to sampling-feedback-update, where feedback serves as a textual gradient to update the probabilistic graph and efficiently search for optimal multi-LLM collaboration graphs. Experiments show that Agent-REINFORCE outperforms both traditional and LLM-based baselines in sample efficiency and search performance, and effectively identifies optimal graphs under joint objectives of accuracy and inference latency.

Stock trend prediction is crucial for profitable trading strategies and portfolio management yet remains challenging due to market volatility, complex temporal dynamics and multifaceted inter-stock relationships. Existing methods struggle to effectively capture temporal dependencies and dynamic inter-stock interactions, often neglecting cross-sectional market influences, relying on static correlations, employing uniform treatments of nodes and edges, and conflating diverse relationships. This work introduces MaGNet, a novel Mamba dual-hyperGraph Network for stock prediction, integrating three key innovations: (1) a MAGE block, which leverages bidirectional Mamba with adaptive gating mechanisms for contextual temporal modeling and integrates a sparse Mixture-of-Experts layer to enable dynamic adaptation to diverse market conditions, alongside multi-head attention for capturing global dependencies; (2) Feature-wise and Stock-wise 2D Spatiotemporal Attention modules enable precise fusion of multivariate features and cross-stock dependencies, effectively enhancing informativeness while preserving intrinsic data structures, bridging temporal modeling with relational reasoning; and (3) a dual hypergraph framework consisting of the Temporal-Causal Hypergraph (TCH) that captures fine-grained causal dependencies with temporal constraints, and Global Probabilistic Hypergraph (GPH) that models market-wide patterns through soft hyperedge assignments and Jensen-Shannon Divergence weighting mechanism, jointly disentangling localized temporal influences from instantaneous global structures for multi-scale relational learning. Extensive experiments on six major stock indices demonstrate MaGNet outperforms state-of-the-art methods in both superior predictive performance and exceptional investment returns with robust risk management capabilities. Codes available at: https://github.com/PeilinTime/MaGNet.

flowengineR is an R package designed to provide a modular and extensible framework for building reproducible algorithmic workflows for general-purpose machine learning pipelines. It is motivated by the rapidly evolving field of algorithmic fairness where new metrics, mitigation strategies, and machine learning methods continuously emerge. A central challenge in fairness, but also far beyond, is that existing toolkits either focus narrowly on single interventions or treat reproducibility and extensibility as secondary considerations rather than core design principles. flowengineR addresses this by introducing a unified architecture of standardized engines for data splitting, execution, preprocessing, training, inprocessing, postprocessing, evaluation, and reporting. Each engine encapsulates one methodological task yet communicates via a lightweight interface, ensuring workflows remain transparent, auditable, and easily extensible. Although implemented in R, flowengineR builds on ideas from workflow languages (CWL, YAWL), graph-oriented visual programming languages (KNIME), and R frameworks (BatchJobs, batchtools). Its emphasis, however, is less on orchestrating engines for resilient parallel execution but rather on the straightforward setup and management of distinct engines as data structures. This orthogonalization enables distributed responsibilities, independent development, and streamlined integration. In fairness context, by structuring fairness methods as interchangeable engines, flowengineR lets researchers integrate, compare, and evaluate interventions across the modeling pipeline. At the same time, the architecture generalizes to explainability, robustness, and compliance metrics without core modifications. While motivated by fairness, it ultimately provides a general infrastructure for any workflow context where reproducibility, transparency, and extensibility are essential.

While Vision-language Models (VLMs) have demonstrated strong semantic capabilities, their ability to interpret the underlying geometric structure of visual information is less explored. Pictographic characters, which combine visual form with symbolic structure, provide an ideal test case for this capability. We formulate this visual recognition challenge in the mathematical domain, where each character is represented by an executable program of geometric primitives. This is framed as a program synthesis task, training a VLM to decompile raster images into programs composed of Bézier curves. Our model, acting as a "visual decompiler", demonstrates performance superior to strong zero-shot baselines, including GPT-4o. The most significant finding is that when trained solely on modern Chinese characters, the model is able to reconstruct ancient Oracle Bone Script in a zero-shot context. This generalization provides strong evidence that the model acquires an abstract and transferable geometric grammar, moving beyond pixel-level pattern recognition to a more structured form of visual understanding.

Quarter level cell (QLC) 3D NAND flash memory is emerging as the predominant storage solution in the era of artificial intelligence. QLC 3D NAND flash stores 4 bit per cell to expand the storage density, resulting in narrower read margins. Constrained to read margins, QLC always suffers from lateral charge migration (LCM), which caused by non-uniform charge density across adjacent memory cells. To suppress charge density gap between cells, there are some algorithm in form of intra-page data mapping such as WBVM, DVDS. However, we observe inter-page data arrangements also approach the suppression. Thus, we proposed an intelligent model PDA-LSTM to arrange intra-page data for LCM suppression, which is a physics-knowledge-driven neural network model. PDA-LSTM applies a long-short term memory (LSTM) neural network to compute a data arrangement probability matrix from input page data pattern. The arrangement is to minimize the global impacts derived from the LCM among wordlines. Since each page data can be arranged only once, we design a transformation from output matrix of LSTM network to non-repetitive sequence generation probability matrix to assist training process. The arranged data pattern can decrease the bit error rate (BER) during data retention. In addition, PDA-LSTM do not need extra flag bits to record data transport of 3D NAND flash compared with WBVM, DVDS. The experiment results show that the PDA-LSTM reduces the average BER by 80.4% compared with strategy without data arrangement, and by 18.4%, 15.2% compared respectively with WBVM and DVDS with code-length 64.

Generative AI is reshaping fashion by enabling virtual looks and avatars making it essential to find real products that best match AI-generated styles. We propose an end-to-end product search system that has been deployed in a real-world, internet scale which ensures that AI-generated looks presented to users are matched with the most visually and semantically similar products from the indexed vector space. The search pipeline is composed of four key components: query generation, vectorization, candidate retrieval, and reranking based on AI-generated looks. Recommendation quality is evaluated using human-judged accuracy scores. The system currently serves more than 350,000 AI Looks in production per day, covering diverse product categories across global markets of over 12 million products. In our experiments, we observed that across multiple annotators and categories, CLIP outperformed alternative models by a small relative margin of 3--7\% in mean opinion scores. These improvements, though modest in absolute numbers, resulted in noticeably better user perception matches, establishing CLIP as the most reliable backbone for production deployment.

We introduce [Cosmos-Predict2.5], the latest generation of the Cosmos World Foundation Models for Physical AI. Built on a flow-based architecture, [Cosmos-Predict2.5] unifies Text2World, Image2World, and Video2World generation in a single model and leverages [Cosmos-Reason1], a Physical AI vision-language model, to provide richer text grounding and finer control of world simulation. Trained on 200M curated video clips and refined with reinforcement learning-based post-training, [Cosmos-Predict2.5] achieves substantial improvements over [Cosmos-Predict1] in video quality and instruction alignment, with models released at 2B and 14B scales. These capabilities enable more reliable synthetic data generation, policy evaluation, and closed-loop simulation for robotics and autonomous systems. We further extend the family with [Cosmos-Transfer2.5], a control-net style framework for Sim2Real and Real2Real world translation. Despite being 3.5$\times$ smaller than [Cosmos-Transfer1], it delivers higher fidelity and robust long-horizon video generation. Together, these advances establish [Cosmos-Predict2.5] and [Cosmos-Transfer2.5] as versatile tools for scaling embodied intelligence. To accelerate research and deployment in Physical AI, we release source code, pretrained checkpoints, and curated benchmarks under the NVIDIA Open Model License at https://github.com/nvidia-cosmos/cosmos-predict2.5 and https://github.com/nvidia-cosmos/cosmos-transfer2.5. We hope these open resources lower the barrier to adoption and foster innovation in building the next generation of embodied intelligence.

Understanding why neural networks make certain decisions is pivotal for their use in safety-critical applications. Feature-Guided Analysis (FGA) extracts slices of neural networks relevant to their tasks. Existing feature-guided approaches typically monitor the activation of the neural network neurons to extract the relevant rules. Preliminary results are encouraging and demonstrate the feasibility of this solution by assessing the precision and recall of Feature-Guided Analysis on two pilot case studies. However, the applicability in industrial contexts needs additional empirical evidence. To mitigate this need, this paper assesses the applicability of FGA on a benchmark made by the MNIST and LSC datasets. We assessed the effectiveness of FGA in computing rules that explain the behavior of the neural network. Our results show that FGA has a higher precision on our benchmark than the results from the literature. We also evaluated how the selection of the neural network architecture, training, and feature selection affect the effectiveness of FGA. Our results show that the selection significantly affects the recall of FGA, while it has a negligible impact on its precision.

As the large language models (LLMs) grow in size each day, efficient training and fine-tuning has never been as important as nowadays. This resulted in the great interest in parameter efficient fine-tuning (PEFT), and effective methods including low-rank adapters (LoRA) has emerged. Although the various PEFT methods have been studied extensively in the recent years, the greater part of the subject remains unexplored with the huge degree of freedom. In this paper, we propose FLoRA, a family of fused forward-backward adapters (FFBA) for parameter-efficient fine-tuning of LLMs on downstream tasks. The FFBA combine ideas from the popular LoRA and parallel adapters to improve the overall fine-tuning accuracies. At the same time, latencies are minimized by fusing the forward and backward adapters into existing projection layers of the base model. Experimental results show that the proposed FFB adapters perform significantly better than the popularly used LoRA in both accuracy and latency for a similar parameter budget.

Accurate and robust weather forecasting remains a fundamental challenge due to the inherent spatio-temporal complexity of atmospheric systems. In this paper, we propose a novel self-supervised learning framework that leverages spatio-temporal structures to improve multi-variable weather prediction. The model integrates a graph neural network (GNN) for spatial reasoning, a self-supervised pretraining scheme for representation learning, and a spatio-temporal adaptation mechanism to enhance generalization across varying forecasting horizons. Extensive experiments on both ERA5 and MERRA-2 reanalysis datasets demonstrate that our approach achieves superior performance compared to traditional numerical weather prediction (NWP) models and recent deep learning methods. Quantitative evaluations and visual analyses in Beijing and Shanghai confirm the model's capability to capture fine-grained meteorological patterns. The proposed framework provides a scalable and label-efficient solution for future data-driven weather forecasting systems.