EX ViSiON

This paper introduces a novel framework called "Efficient and Scalable Granular-ball Graph Coarsening Method" to address the high computational overhead of Graph Convolutional Networks (GCNs) on large-scale graph datasets. The proposed method first employs a multi-granularity granular-ball graph coarsening algorithm to reduce the original graph into multiple subgraphs. This coarsening stage boasts linear time complexity, significantly outperforming existing methods. Subsequently, these granular-ball subgraphs are randomly sampled to form minibatches for GCN training. The algorithm adaptively and substantially reduces graph scale, thereby enhancing GCN training efficiency and scalability. Experimental results on various node classification datasets demonstrate superior performance, and the code is made publicly available.

This paper introduces a novel gated active learning framework designed to overcome the limitations of noisy data in autonomous experimental systems, particularly in structure-property learning tasks like Image-to-Spectrum and Spectrum-to-Image translations. Standard active learning often misinterprets noise as uncertainty, leading to the acquisition of poor-quality measurements. The proposed framework combines curiosity-driven sampling with a physics-informed quality control filter, based on Simple Harmonic Oscillator model fits, to automatically exclude low-fidelity data during acquisition. Evaluations on a pre-acquired dataset of band-excitation piezoresponse spectroscopy data from PbTiO3 thin films demonstrate that this method significantly outperforms random sampling, standard active learning, and multitask learning strategies. Furthermore, its effectiveness was validated in real-time experiments on BiFeO3 thin films, showcasing its applicability in real autonomous microscopy. This work advocates for a shift towards hybrid autonomy in self-driving labs, integrating physics-informed quality assessment with active decision-making for more reliable scientific discovery.

HCLSM introduces a novel world model architecture designed for object-centric world modeling, addressing limitations of flat latent representations. It operates on three core principles: object-centric decomposition using slot attention with spatial broadcast decoding, hierarchical temporal dynamics combining selective state space models, sparse transformers, and compressed transformers, and causal structure learning via graph neural networks. A two-stage training protocol ensures slot specialization before dynamics prediction. Trained on the PushT robotic manipulation benchmark, the 68M-parameter model achieves high next-state prediction accuracy. A custom Triton kernel significantly accelerates the SSM scan by 38x, demonstrating practical optimization. This work represents a significant step towards more sophisticated and efficient world models for embodied AI.