Stochasticity in language model fine-tuning, often caused by the small batch sizes typically used in this regime, can destabilize training by introducing large oscillations in generation quality. A popular approach to mitigating this instability is to take an Exponential moving average (EMA) of weights throughout training. While EMA reduces stochasticity, thereby smoothing training, the introduction of bias from old iterates often creates a lag in optimization relative to vanilla training. In this work, we propose the Bias-Corrected Exponential Moving Average (BEMA), a simple and practical augmentation of EMA that retains variance-reduction benefits while eliminating bias. BEMA is motivated by a simple theoretical model wherein we demonstrate provable acceleration of BEMA over both a standard EMA and vanilla training. Through an extensive suite of experiments on Language Models, we show that BEMA leads to significantly improved convergence rates and final performance over both EMA and vanilla training in a variety of standard LM benchmarks, making BEMA a practical and theoretically motivated intervention for more stable and efficient fine-tuning.

We analyze the statistical problem of recovering an atomic signal, modeled as a discrete uniform distribution $μ$, from a binned Poisson convolution model. This question is motivated, among others, by super-resolution laser microscopy applications, where precise estimation of $μ$ provides insights into spatial formations of cellular protein assemblies. Our main results quantify the local minimax risk of estimating $μ$ for a broad class of smooth convolution kernels. This local perspective enables us to sharply quantify optimal estimation rates as a function of the clustering structure of the underlying signal. Moreover, our results are expressed under a multiscale loss function, which reveals that different parts of the underlying signal can be recovered at different rates depending on their local geometry. Overall, these results paint an optimistic perspective on the Poisson deconvolution problem, showing that accurate recovery is achievable under a much broader class of signals than suggested by existing global minimax analyses. Beyond Poisson deconvolution, our results also allow us to establish the local minimax rate of parameter estimation in Gaussian mixture models with uniform weights. We apply our methods to experimental super-resolution microscopy data to identify the location and configuration of individual DNA origamis. In addition, we complement our findings with numerical experiments on runtime and statistical recovery that showcase the practical performance of our estimators and their trade-offs.

HealthBench, a benchmark designed to measure the capabilities of AI systems for health better (Arora et al., 2025), has advanced medical language model evaluation through physician-crafted dialogues and transparent rubrics. However, its reliance on expert opinion, rather than high-tier clinical evidence, risks codifying regional biases and individual clinician idiosyncrasies, further compounded by potential biases in automated grading systems. These limitations are particularly magnified in low- and middle-income settings, where issues like sparse neglected tropical disease coverage and region-specific guideline mismatches are prevalent. The unique challenges of the African context, including data scarcity, inadequate infrastructure, and nascent regulatory frameworks, underscore the urgent need for more globally relevant and equitable benchmarks. To address these shortcomings, we propose anchoring reward functions in version-controlled Clinical Practice Guidelines (CPGs) that incorporate systematic reviews and GRADE evidence ratings. Our roadmap outlines "evidence-robust" reinforcement learning via rubric-to-guideline linkage, evidence-weighted scoring, and contextual override logic, complemented by a focus on ethical considerations and the integration of delayed outcome feedback. By re-grounding rewards in rigorously vetted CPGs, while preserving HealthBench's transparency and physician engagement, we aim to foster medical language models that are not only linguistically polished but also clinically trustworthy, ethically sound, and globally relevant.

This week we welcomed a record-breaking baby to the world. Thaddeus Daniel Pierce, who arrived over the weekend, developed from an embryo that was frozen in storage for 30 and a half years. You could call him the world's oldest baby. His parents, Lindsey and Tim Pierce, were themselves only young children when that embryo was created, all the way back in 1994. Linda Archerd, who donated the embryo, described the experience as "surreal." Stories like this also highlight how reproductive technologies are shaping families.

Certain data are naturally modeled by networks or weighted graphs, be they arterial networks or mobility networks. When there is no canonical labeling of the nodes across the dataset, we talk about unlabeled networks. In this paper, we focus on the question of dimensionality reduction for this type of data. More specifically, we address the issue of interpreting the feature subspace constructed by dimensionality reduction methods. Most existing methods for network-valued data are derived from principal component analysis (PCA) and therefore rely on subspaces generated by a set of vectors, which we identify as a major limitation in terms of interpretability. Instead, we propose to implement the method called barycentric subspace analysis (BSA), which relies on subspaces generated by a set of points. In order to provide a computationally feasible framework for BSA, we introduce a novel embedding for unlabeled networks where we replace their usual representation by equivalence classes of isomorphic networks with that by equivalence classes of cospectral networks. We then illustrate BSA on simulated and real-world datasets, and compare it to tangent PCA.

We propose TPP-SD, a novel approach that accelerates Transformer temporal point process (TPP) sampling by adapting speculative decoding (SD) techniques from language models. By identifying the structural similarities between thinning algorithms for TPPs and speculative decoding for language models, we develop an efficient sampling framework that leverages a smaller draft model to generate multiple candidate events, which are then verified by the larger target model in parallel. TPP-SD maintains the same output distribution as autoregressive sampling while achieving significant acceleration. Experiments on both synthetic and real datasets demonstrate that our approach produces samples from identical distributions as standard methods, but with 2-6$\times$ speedup. Our ablation studies analyze the impact of hyperparameters such as draft length and draft model size on sampling efficiency. TPP-SD bridges the gap between powerful Transformer TPP models and the practical need for rapid sequence sampling.

Are AI systems truly representing human values, or merely averaging across them? Our study suggests a concerning reality: Large Language Models (LLMs) fail to represent diverse cultural moral frameworks despite their linguistic capabilities. We expose significant gaps between AI-generated and human moral intuitions by applying the Moral Foundations Questionnaire across 19 cultural contexts. Comparing multiple state-of-the-art LLMs' origins against human baseline data, we find these models systematically homogenize moral diversity. Surprisingly, increased model size doesn't consistently improve cultural representation fidelity. Our findings challenge the growing use of LLMs as synthetic populations in social science research and highlight a fundamental limitation in current AI alignment approaches. Without data-driven alignment beyond prompting, these systems cannot capture the nuanced, culturally-specific moral intuitions. Our results call for more grounded alignment objectives and evaluation metrics to ensure AI systems represent diverse human values rather than flattening the moral landscape.

Breakthroughs, discoveries, and DIY tips sent every weekday. In October 2017, a single flash of lightning during a thunderstorm streaked across the Great Plains for 515 miles. The flash traveled from eastern Texas all the way to Kansas City--and now into the record books. The World Meteorological Organization (WMO) certified that this megaflash is now the longest single lightning flash in the United States. The massive lightning bolt is detailed in a study published July 31 in the Bulletin of the American Meteorological Society.

Many data-science applications involve detecting a shared signal between two high-dimensional variables. Using random matrix theory methods, we determine when such signal can be detected and reconstructed from sample correlations, despite the background of sampling noise induced correlations. We consider three different covariance matrices constructed from two high-dimensional variables: their individual self covariance, their cross covariance, and the self covariance of the concatenated (joint) variable, which incorporates the self and the cross correlation blocks. We observe the expected Baik, Ben Arous, and Péché detectability phase transition in all these covariance matrices, and we show that joint and cross covariance matrices always reconstruct the shared signal earlier than the self covariances. Whether the joint or the cross approach is better depends on the mismatch of dimensionalities between the variables. We discuss what these observations mean for choosing the right method for detecting linear correlations in data and how these findings may generalize to nonlinear statistical dependencies.

We study algorithmic aspects of finding $n$-dimensional \emph{positive} and \emph{negative} Hopfield ($\pm$Hop) model ground state free energies. This corresponds to classical maximization of random positive/negative semi-definite quadratic forms over binary $\left \{\pm \frac{1}{\sqrt{n}} \right \}^n$ vectors. The key algorithmic question is whether these problems can be computationally efficiently approximated within a factor $\approx 1$. Following the introduction and success of \emph{Controlled Loosening-up} (CLuP-SK) algorithms in finding near ground state energies of closely related Sherrington-Kirkpatrick (SK) models [82], we here propose a CLuP$\pm$Hop counterparts for $\pm$Hop models. Fully lifted random duality theory (fl RDT) [78] is utilized to characterize CLuP$\pm$Hop \emph{typical} dynamics. An excellent agreement between practical performance and theoretical predictions is observed. In particular, for $n$ as small as few thousands CLuP$\pm$Hop achieve $\sim 1.77$ and $\sim 0.33$ as the ground state free energies of the positive and negative Hopfield models. At the same time we obtain on the 6th level of lifting (6-spl RDT) corresponding theoretical thermodynamic ($n\rightarrow\infty$) limits $\approx 1.7784$ and $\approx 0.3281$. This positions determining Hopfield models near ground state energies as \emph{typically} easy problems. Moreover, the very same 6th lifting level evaluations allow to uncover a fundamental intrinsic difference between two models: $+$Hop's near optimal configurations are \emph{typically close} to each other whereas the $-$Hop's are \emph{typically far away}.