Junyi An, Xinyu Lu, Yunfei Shi, Peijia Lin, Licheng Xu, Qianwei Tang, Chao Qu, Fenglei Cao 等 9 位作者
原文摘要
Equivariant Graph Neural Networks (GNNs) have significantly advanced the modeling of 3D molecular structure by leveraging group representations. However, message passing built only from exactly equivariant operations can have restricted practical expressiveness at a fixed representation degree. We introduce the Equivariant Spherical Transformer (EST), a plug-and-play module that applies Transformer attention to spherical spatial representations obtained from group representations. EST explicitly mixes information across sampled orientations, unlike prior orientation-wise spherical operations, while retaining a strong approximate equivariant inductive bias through near-uniform spherical sampling. As demonstrated by our experiments on challenging benchmarks like OC20, MPtrj and QM9, EST-based models achieve state-of-the-art performance. For the complex molecular systems within OC20, small models empowered by EST can outperform some larger models and those using additional data. We characterize the ideal closed-domain condition under which the continuous operation is exactly equivariant and empirically quantify the finite-sampling accuracy--equivariance--memory trade-off used in practice.
When the number of training samples is smaller than the dimension, discriminant analysis faces a singular within-class scatter matrix, and fused objectives that combine the total variance C, the between-class scatter S_b and the within-class scatter S_w are one way to handle it. Using the seven variants of Kernelized Linear Principal Component Discriminant Analysis (KLPCDA) as a platform, we derive the structure of such objectives in this small-sample-size (SSS) setting. The training projections split into a class-mean part and a within-class part, the two scatter operators satisfy WB = 0, and all between-class information lies in the null space of S_w. A single fusion coefficient therefore cannot change the solution, and the variants, KPCA, the regularized kernel Fisher discriminant and the null space methods lie on the boundary of one family with two parameters, the fusion ratio t and a ridge parameter mu. On six tasks from five application domains, the regularized members were more accurate than the best of the seven variants chosen on validation data on all six tasks, by 0.8 to 5.7 points, also under added noise, whereas the fusion ratio gave no gain, including in controlled experiments on dimension ratio, class imbalance and class separation. Selecting a variant per task could gain less than one point, while a fixed choice made on other tasks transferred to every held-out task. The analysis thus gives a principled default for SSS kernel discriminant analysis, with diagnostic quantities, the null space dimension and the class-size spectrum, that can be computed before training.
Large Language Models (LLMs) that can continually improve beyond their training budgets are able to solve increasingly difficult problems by adapting at test time, a property we refer to as extrapolation. However, standard reinforcement learning (RL) operates over fixed problem distributions and training budgets, which limits extrapolation amidst distribution shift at test time. To address this, we introduce RC, an iterative decoding algorithm that replaces standard autoregressive decoding during both training and inference. RC exploits an asymmetry between the response generation and summarization capabilities of LLMs to construct reasoning chains that consistently improve across iterations. Models trained to use RC can extrapolate and continually improve over reasoning horizons more than an order of magnitude longer than those seen during training. Empirically, training a 4B model with RC using a 16k-token training budget improves performance on HMMT 2025 from 40% to nearly 70% with 0.5m tokens at test time, outperforming both comparably sized models and many larger reasoning LLMs. Finally, we also show that models trained with RC can more effectively leverage existing scaffolds to further scale test-time performance, due to the improved summary-conditioned generation abilities learned through training.