Attention mechanisms have made significant strides in graph learning, yet they still exhibit notable limitations: local attention faces challenges in capturing long-range information due to the inherent problems of the message-passing scheme, while global attention cannot reflect the hierarchical neighborhood structure and fails to capture fine-grained local information. In this paper, we propose a novel multihop graph attention mechanism, named Subtree Attention (STA), to address the aforementioned issues. STA seamlessly bridges the fully-attentional structure and the rooted subtree, with theoretical proof that STA approximates the global attention under extreme settings.

The transformer architecture [52], which underlies present-day Large Language Models, has been one of the main drivers of recent advances in machine learning and artificial intelligence. At each layer, the hidden state of the network is updated by sequentially applying two distinct operations: attention modules [3], which capture long-range interactions in the input sequence, and classical MultiLayer Perceptrons (MLPs), acting separately on each element of that sequence. Despite their empirical success, the mechanisms governing information propagation through depth, and the way attention and MLP blocks jointly shape internal representations, remain only partially understood from a theoretical viewpoint. Recent progress has come from viewing transformers in suitable scaling limits as deterministic mean-field interacting particle systems modeling the evolution of N tokens1 through the layers of the neural network architecture (the so-called residual stream dynamics), see, among others, [46, 26, 27, 45]. In these descriptions, depth plays the role of a continuous time variable, and, in the large-context regime (N), the evolution of token representations is encoded by a PDE for their empirical distribution. This viewpoint is closely connected to the literature on scaling laws, where the effect of various scaling exponents controlling the relative size of the network's hyperparameters (e.g., depth, width, context length) on the effective dynamics of the model

A.1 Molecule Design We present more examples of generated molecules by our method and the CNN baseline liGAN. We select 6 molecules with highest binding affinity for each method and each binding site. The 3 additional binding sites are selected randomly from the testing set. By comparing the samples from two methods, we can find that the 3D molecules generated by our method are generally more realistic, while molecules generated by the baseline have more erroneous structures, such as bonds that are too short and angles that are too sharp. Besides, molecules generated by our method are more diverse, while the 3D atom configurations generated by the baseline are often similar.

The InfoNCE loss in contrastive learning depends critically on a temperature parameter, yet its dynamics under fixed versus annealed schedules remain poorly understood. We provide a theoretical analysis by modeling embedding evolution under Langevin dynamics on a compact Riemannian manifold. Under mild smoothness and energy-barrier assumptions, we show that classical simulated annealing guarantees extend to this setting: slow logarithmic inverse-temperature schedules ensure convergence in probability to a set of globally optimal representations, while faster schedules risk becoming trapped in suboptimal minima. Our results establish a link between contrastive learning and simulated annealing, providing a principled basis for understanding and tuning temperature schedules.

To be consistent with accuracy definition, we denote the correctness ofstj for instance t as sim(stj,rt) = ( 2 distance(stj,rt))/ 2 where sim(stj,rt) is in the range [0,1] and distance(stj,rt) is in range [0, 2], 2 is the largest Euclidean distance in the probability simplex. Given a test dataset I, the correctness of a learner SLj on I can be denoted as 2 corrSLj = 1n Pn t=1sim(stj,rt). In this section, we define multiple metrics for consistency, accuracy, and correct-consistency in detail. Figure 1 shows the metrics computation in our experiments. We have created a git repository for this work and will be posted upon the acceptance and publicationofthiswork.

However, local attention limits the receptive field to one-hop neighbors.

Such direct sim2real transfer is not guaranteed to succeed, however, and in cases where it fails, it is unclear how to best utilize the simulator. In this work, we show that in many regimes, while direct sim2real transfer may fail, we can utilize the simulator to learn a set of exploratory policies which enable efficient exploration in the real world.