Introduces Subequivariant Graph Neural Network, improving physical dynamics prediction accuracy by over 3%.

Jiaqi Han, Wenbing Huang, Hengbo Ma et al.

CTL++ dataset tests neural networks' systematic generalization on unseen function compositions; transformers struggle significantly, with accuracy dropping below 30% in out-of-distribution tests.

Róbert Csordás, Kazuki Irie, Jürgen Schmidhuber

Using symmetry analysis, proves over-parameterization turns spurious minima into saddles, facilitating gradient optimization.

Yossi Arjevani, Michael Field

Using Neural Tangent Kernel (NTK) spectral analysis, the paper reveals robust and non-robust features, enabling training-free adversarial example generation and insights into robustness mechanisms.

Nikolaos Tsilivis, Julia Kempe

metaPNS employs Bayesian meta-learning with set-conditioned generative models to achieve personalized cardiac surrogates with few samples, improving accuracy and efficiency.

Xiajun Jiang, Zhiyuan Li, Ryan Missel et al.

This study compares FNO, ResNet, and U-Net architectures for multi-scale PDE modeling, demonstrating a single surrogate's strong generalization across parameters.

Jayesh K. Gupta, Johannes Brandstetter

Proposes InterFlow, a stochastic interpolant-based continuous-time normalizing flow, optimizing path length and surpassing traditional methods in high-resolution image generation.

Michael S. Albergo, Eric Vanden-Eijnden

ASPiRe accelerates reinforcement learning using adaptive skill priors, significantly improving learning efficiency.

Mengda Xu, Manuela Veloso, Shuran Song

DiGress employs discrete diffusion for graph generation, achieving up to 99% validity on molecular datasets and scaling to 1.3 million molecules.

Clement Vignac, Igor Krawczuk, Antoine Siraudin et al.

Proposes induction heads as the primary mechanism for in-context learning in large transformers, supported by six lines of evidence including training phase shifts and causal ablations.

Catherine Olsson, Nelson Elhage, Neel Nanda et al.

Toy models demonstrate neural superposition, phase transitions, geometric structures, and links to adversarial examples.

Nelson Elhage, Tristan Hume, Catherine Olsson et al.

Using GPT-3's algorithmic fidelity, conditioned responses simulate diverse US social groups, validated through surveys with high correlation to real data.

Lisa P. Argyle, Ethan C. Busby, Nancy Fulda et al.

Using small neural networks (probabilistic transformers) to embed arbitrary metric spaces into Gaussian mixture spaces with low distortion, ensuring bi-Hölder and bi-Lipschitz guarantees.

Anastasis Kratsios, Valentin Debarnot, Ivan Dokmanić

Proposes Interventional Normalizing Flows (INF) for density estimation of potential outcomes, combining bias correction and scalable deep models.

Valentyn Melnychuk, Dennis Frauen, Stefan Feuerriegel

Introduces iPFI algorithm for online feature importance on data streams, validated through experiments.

Fabian Fumagalli, Maximilian Muschalik, Eyke Hüllermeier et al.

Proposes Lyapunov-based diffusion bridges with physical/statistical priors, improving molecule stability and point cloud uniformity.

Lemeng Wu, Chengyue Gong, Xingchao Liu et al.

Adan is an adaptive Nesterov momentum optimizer achieving \(\mathcal{O}(\varepsilon^{-3.5})\) complexity, significantly speeding up deep model training.

Xingyu Xie, Pan Zhou, Huan Li et al.

Introduces GCP-CROWN, combining MIP-generated cuts with bound propagation, achieving complete neural network verification with under 5 seconds per instance.

Huan Zhang, Shiqi Wang, Kaidi Xu et al.

Introduces two industrial process datasets for benchmarking causal discovery algorithms, evaluated with metrics like TP, FDR, and SHD.

Giovanni Menegozzo, Diego Dall'Alba, Paolo Fiorini

PanGu-Coder uses two-stage function-level training, reaching 17.07% HumanEval pass@1 with 317M parameters.

Fenia Christopoulou, Gerasimos Lampouras, Milan Gritta et al.