By applying AlphaZero to Hex, the study reveals performance scales exponentially with board size and compute, enabling prediction of large-scale results from small experiments.

Andy L. Jones

DIG: a comprehensive platform integrating graph generation, self-supervised learning, explainability, and 3D graph modeling, with standardized interfaces and metrics.

Meng Liu, Youzhi Luo, Limei Wang et al.

Proposes a circuit-based gradient estimator for time-varying neural networks using fractional calculus and energy models, extending equilibrium propagation.

Jack Kendall

Proposes SORS, a self-supervised method that densifies rewards to improve sample efficiency in sparse-reward RL tasks.

Farzan Memarian, Wonjoon Goo, Rudolf Lioutikov et al.

Proposes unbalanced minibatch optimal transport (UOT) for domain adaptation, outperforming classical OT with theoretical guarantees and robustness in large-scale noisy data.

Kilian Fatras, Thibault Séjourné, Nicolas Courty et al.

Proposes GPU zero-copy architecture for GCN training, achieving 65-92% speedup in multi-GPU setups.

Seung Won Min, Kun Wu, Sitao Huang et al.

Proposes a Hessian-based machine unlearning algorithm deleting up to O(n/d^{1/4}), surpassing DP's O(n/d^{1/2}).

Ayush Sekhari, Jayadev Acharya, Gautam Kamath et al.

Introduces Message Passing Simplicial Networks (MPSNs), leveraging simplicial complexes to surpass WL test in expressive power for graph and higher-order structure analysis.

Cristian Bodnar, Fabrizio Frasca, Yu Guang Wang et al.

By learning reverse diffusion variances and integrating hybrid objectives, the model achieves high sample quality with fewer steps, improving log-likelihood and sampling speed simultaneously.

Alex Nichol, Prafulla Dhariwal

GNN-based methods encode graph structures for combinatorial optimization, improving solution speed and quality, with demonstrated 15% speedup on TSP instances.

Quentin Cappart, Didier Chételat, Elias Khalil et al.

FrugalMCT framework achieves 90% cost reduction for multi-label classification while maintaining high accuracy.

Lingjiao Chen, Matei Zaharia, James Zou

α-Precision, β-Recall, and Authenticity provide sample-level, domain-agnostic evaluation of fidelity, diversity, and memorization.

Ahmed M. Alaa, Boris van Breugel, Evgeny Saveliev et al.

Proposes a unified theory of neural scaling laws, identifying four regimes via kernel spectrum and data manifold analysis.

Yasaman Bahri, Ethan Dyer, Jared Kaplan et al.

BRECQ pushes PTQ bitwidth to INT2 for the first time, achieving 240x faster model production.

Yuhang Li, Ruihao Gong, Xu Tan et al.

GASP learns continuous function distributions, reaching FID 7.42/19.16 on CelebAHQ while enabling cross-resolution generation.

Emilien Dupont, Yee Whye Teh, Arnaud Doucet

Introduces IAP with efficient Transformer attention algorithms for pixel-level focus, boosting vision-based RL performance.

Krzysztof Marcin Choromanski, Deepali Jain, Wenhao Yu et al.

PaiNN, a rotational equivariant message passing neural network, predicts tensor properties and molecular spectra with 4-5 orders of magnitude speedup.

Kristof T. Schütt, Oliver T. Unke, Michael Gastegger

GraphDF introduces a discrete flow model for molecular graph generation, reducing computational cost and avoiding dequantization bias.

Youzhi Luo, Keqiang Yan, Shuiwang Ji

Switch Transformer simplifies MoE routing with single expert (k=1), enabling trillion-parameter models with 7x faster pretraining.

William Fedus, Barret Zoph, Noam Shazeer

Analyzing intrinsic dimension explains why fine-tuning large models with few parameters achieves near-optimal performance, with empirical evidence showing low intrinsic dimensions (e.g., 200 parameters for RoBERTa on MRPC).

Armen Aghajanyan, Luke Zettlemoyer, Sonal Gupta