AISTATS 2026 Batch — Oral & Spotlight Papers

Hamidreza Hashempoor • Institute for AI, University of Stuttgart

Accepted oral and spotlight papers from AISTATS 2026. Each entry summarizes the core idea and theory and adds a small, self-contained experiment with a companion notebook.

Papers

• Paper 1

  Synthetic-Regularized Kernel Regression

  Synthetic-data regularization • bias–variance • kernel methods

  A kernel regressor that trades noisy real labels against a synthetic generator $g$ via a penalty $\lambda\|f-g\|_{\mathcal H_K}^2$. Larger $\lambda$ cuts variance but adds bias when $g\neq f_\star$; the best $\lambda$ depends on how well $g$ matches the target — especially in the hard, high-frequency kernel directions. Verified on a 1-D RBF experiment.

  Open paper →
• Paper 2

  EventFlow — Flow Matching for Temporal Point Processes

  temporal point processes • flow matching • generative models

  A non-autoregressive generative model for event sequences. A balanced coupling pairs reference and data sequences of equal length, a linear noisy interpolant defines the flow, and a vector field trained by masked MSE transports a uniform reference TPP onto the data distribution — preserving the event count while pulling event times onto the data clusters. Verified on a clustered synthetic TPP.

  Open paper →
• Paper 3

  A Proof of Learning Rate Transfer Under μP

  μP • learning-rate transfer • infinite-width limit

  Why the optimal learning rate stops moving as a network gets wider. Under the maximal-update parametrization, the optimal one-step learning rate of a deep linear net has a closed-form infinite-width limit $\eta_\infty^{(1)}=\frac{m}{L}\,y^\top K y/\|Ky\|^2$ set by the input Gram matrix, and finite widths approach it at rate $O_P(n^{-1/2})$ — the theory behind transferring a tuned learning rate across scales. Verified on a deep linear MLP.

  Open paper →
• Paper 4

  Spherical Linear Gaussian Process for High-Dimensional BO

  high-dimensional Bayesian optimization • Gaussian processes • spherical projection

  A low-capacity surrogate for high-dimensional low-data BO. A standard linear kernel has prior variance $k(x,x)=\|x\|^2$, so variance-seeking acquisition is dragged to the box boundary. Mapping inputs onto a unit hypersphere via the inverse-stereographic projection $P(z)$ makes the prior variance constant ($k(x,x)=b_0+b_1=1$) while keeping angular structure — removing the boundary pathology. Verified on a $D=100$ sparse-sine problem.

  Open paper →
• Paper 5

  Gaussian Equivalence for Self-Attention Spectra

  random matrix theory • self-attention • Gaussian equivalence • spectra

  The singular-value spectrum of a randomly-initialized softmax self-attention matrix $A$ splits into a rank-one uniform-averaging mode (one large Perron value, $s_1(A)\to1$) plus a diffusive bulk. After centering, that bulk has the same limiting squared-singular-value law as an explicit Gaussian-equivalent linear model $Y_{\text{lin}}^f=\beta\tfrac{S}{\sqrt d}+\sqrt{e^{\beta^2}-1-\beta^2}\,\tfrac{W}{\sqrt d}$ — the softmax survives only through two scalars. Verified on a $d=1000$ experiment.

  Open paper →
• More to come

  Paper y, paper z, …

  Additional oral & spotlight papers will be added here as I work through the batch, each with the same idea-summary-plus-experiment treatment.