Spectral Attention Network (SAN) leverages Laplacian spectrum for node positional encoding, outperforming traditional GNNs with full connectivity.

Devin Kreuzer, Dominique Beaini, William L. Hamilton et al.

Neural Monge Map learns unpaired, general-cost optimal maps; CelebA inpainting reaches FID 3.7109 versus WGAN-GP 18.7942.

Jiaojiao Fan, Shu Liu, Shaojun Ma et al.

This paper proves self-attention matrices can approximate arbitrary sparse patterns with input adjustment, requiring hidden size d = O(log L).

Valerii Likhosherstov, Krzysztof Choromanski, Adrian Weller

SketchGen uses Transformer-based sequence modeling to generate CAD sketches, outperforming state-of-the-art methods with improved distribution alignment.

Wamiq Reyaz Para, Shariq Farooq Bhat, Paul Guerrero et al.

An ICNN benchmark shows that tW2s best recovers high-dimensional W2 maps, yet OT accuracy does not guarantee better generation.

Alexander Korotin, Lingxiao Li, Aude Genevay et al.

Proposes Galerkin Transformer with softmax-free attention, grounded in Petrov-Galerkin theory, improving PDE operator learning efficiency.

Shuhao Cao

Using mean-field and continuous limits, the paper proves that deep ResNets with infinite width and depth converge to zero loss via PDE analysis, providing theoretical insight into overparameterization.

Zhiyan Ding, Shi Chen, Qin Li et al.

Proposes a policy gradient algorithm for multi-objective nonlinear concave optimization with sample complexity O(M^4σ^2/(1-γ)^8ε^4), ensuring global convergence.

Qinbo Bai, Mridul Agarwal, Vaneet Aggarwal

AIM leverages Wasserstein-1 distance with a goal-specific quasimetric to efficiently guide goal-conditioned reinforcement learning, accelerating convergence by ~50%.

Ishan Durugkar, Mauricio Tec, Scott Niekum et al.

Unified inference learning methods (Hebbian, Fenchel BP) as recursive bilevel optimization, showing their relation to standard backpropagation.

Christopher Zach

Informed Equation Learning (iEQL) integrates expert knowledge and robust training strategies to support singular functions like log and division, achieving high-accuracy, interpretable models.

Matthias Werner, Andrej Junginger, Philipp Hennig et al.

Proposes spectral descriptor-based functional mapping and manifold regularization for unpaired heterogeneous multimodal data analysis.

Maysam Behmanesh, Peyman Adibi, Jocelyn Chanussot et al.

Diffusion models outperform GANs in image synthesis, achieving FID scores of 2.97, 4.59, and 7.72 on ImageNet 128×128, 256×256, and 512×512 respectively, with efficient sampling and classifier guidance.

Prafulla Dhariwal, Alex Nichol

Using metric entropy, proved RNNs can optimally learn stable linear systems, establishing quantitative bounds on system realization.

Clemens Hutter, Recep Gül, Helmut Bölcskei

Proposes 8 sanity checks for reward functions, revealing widespread flaws in autonomous driving RL reward design.

W. Bradley Knox, Alessandro Allievi, Holger Banzhaf et al.

Proposes importance sampling to accelerate PINNs training, improving convergence by ~30% and reducing training time by 20%.

Mohammad Amin Nabian, Rini Jasmine Gladstone, Hadi Meidani

Reset-Free Reinforcement Learning via Multi-Task Learning achieves complex dexterous manipulation without human intervention.

Abhishek Gupta, Justin Yu, Tony Z. Zhao et al.

Proposes a transferable deep learning framework combining GFNet and Mosaic predictor for unseen PDE domains, achieving 3 orders speedup.

Hengjie Wang, Robert Planas, Aparna Chandramowlishwaran et al.

This study quantifies the energy use and carbon footprint of large NLP models like GPT-3, proposing strategies for greener training.

David Patterson, Joseph Gonzalez, Quoc Le et al.

Proposes a universal algorithm for constructing equivariant MLPs for arbitrary matrix groups, enabling efficient layer design for complex symmetries.

Marc Finzi, Max Welling, Andrew Gordon Wilson