Time-series forecasting predicts future values from past data. In real-world settings, some anomalous events have lasting effects and influence the forecast, while others are short-lived and should be ignored. Standard forecasting models fail to make this distinction, often either overreacting to noise or missing persistent shifts. We propose Co-TSF A (Co ntrastive T ime-Series F orecasting with A nomalies), a regularization framework that learns when to ignore anomalies and when to respond. Co-TSFA generates input-only and input-output augmentations to model forecast-irrelevant and forecast-relevant anomalies, and introduces a latent-output alignment loss that ties representation changes to forecast changes. This encourages invariance to irrelevant perturbations while preserving sensitivity to meaningful distributional shifts. Experiments on the Traffic and Electricity benchmarks, as well as on a real-world cash-demand dataset, demonstrate that Co-TSFA improves performance under anomalous conditions while maintaining accuracy on normal data. An anonymized GitHub repository with the implementation of Co-TSFA is provided at this anonymized GitHub repository and will be made public upon acceptance. Sequence 1 shows an input-only anomaly that should not affect the forecast, whereas Sequence 2 shows an input anomaly that persists into the output (forecast-relevant).

Modern astronomy has generated an extensive taxonomy of celestial objects based on their physical characteristics and predicted future state. As theories of the development, expansion, history, and predicted future state of the universe rely on identifying and observing celestial bodies, it is essential to have quick and accurate classification of newly observed objects. Historically, classification was performed manually, but the rapid expansion of modern catalogues of celestial objects - such as the Sloan Digital Sky Survey, which grows at a rate of thousands of entries daily [1] - makes this manual classification impractical. Supervised and semi-supervised machine learning represent the most promising candidates for the desired computational classification. Until recently, the data, hardware, and software required for large-scale training and deployment of these methods were unavailable to the general research community. However, improvements to parallel processing hardware have driven increased success and adoption, resulting in the invention of models capable of equaling or surpassing human-level intelligence in tasks formerly considered intractable to computers. Such improvements have been recognized in facial recognition [2] and combinatorial game theory [3], but despite their meteoric rise in popularity, there is a significant gap in astronomical literature on applying machine learning models to the problem of celestial object classification. In an effort to improve this state, we explore a number of machine learning based models for a simplified celestial object classification problem to assess the performance and potential of these models in the field of astronomy.

LLM-as-judge evaluation has become the de facto standard for scaling model assessment, but the practice is statistically unsound: uncalibrated scores can invert preferences, naive confidence intervals on uncalibrated scores achieve near-0% coverage, and importance-weighted estimators collapse under limited overlap despite high effective sample size (ESS). We introduce Causal Judge Evaluation (CJE), a framework that fixes all three failures. On n=4,961 Chatbot Arena prompts (after filtering from 5k), CJE achieves 99% pairwise ranking accuracy at full sample size (94% averaged across configurations), matching oracle quality, at 14x lower cost (for ranking 5 policies) by calibrating a 16x cheaper judge on just 5% oracle labels (~250 labels). CJE combines three components: (i) AutoCal-R, reward calibration via mean-preserving isotonic regression; (ii) SIMCal-W, weight stabilization via stacking of S-monotone candidates; and (iii) Oracle-Uncertainty Aware (OUA) inference that propagates calibration uncertainty into confidence intervals. We formalize the Coverage-Limited Efficiency (CLE) diagnostic, which explains why IPS-style estimators fail even when ESS exceeds 90%: the logger rarely visits regions where target policies concentrate. Key findings: SNIPS inverts rankings even with reward calibration (38% pairwise, negative Kendall's tau) due to weight instability; calibrated IPS remains near-random (47%) despite weight stabilization, consistent with CLE; OUA improves coverage from near-0% to ~86% (Direct) and ~96% (stacked-DR), where naive intervals severely under-cover.

Many problems in science and engineering involve time-dependent, high dimensional datasets arising from complex physical processes, which are costly to simulate. In this work, we propose WeldNet: Windowed Encoders for Learning Dynamics, a data-driven nonlinear model reduction framework to build a low-dimensional surrogate model for complex evolution systems. Given time-dependent training data, we split the time domain into multiple overlapping windows, within which nonlinear dimension reduction is performed by auto-encoders to capture latent codes. Once a low-dimensional representation of the data is learned, a propagator network is trained to capture the evolution of the latent codes in each window, and a transcoder is trained to connect the latent codes between adjacent windows. The proposed windowed decomposition significantly simplifies propagator training by breaking long-horizon dynamics into multiple short, manageable segments, while the transcoders ensure consistency across windows. In addition to the algorithmic framework, we develop a mathematical theory establishing the representation power of WeldNet under the manifold hypothesis, justifying the success of nonlinear model reduction via deep autoencoder-based architectures. Our numerical experiments on various differential equations indicate that WeldNet can capture nonlinear latent structures and their underlying dynamics, outperforming both traditional projection-based approaches and recently developed nonlinear model reduction methods.

We identify test prediction variance (TPV) -- the first-order sensitivity of model outputs to parameter perturbations around a trained solution -- as a unifying quantity that links several classical observations about generalization in deep networks. TPV is a fully label-free object whose trace form separates the geometry of the trained model from the specific perturbation mechanism, allowing a broad family of parameter perturbations like SGD noise, label noise, finite-precision noise, and other post-training perturbations to be analyzed under a single framework. Theoretically, we show that TPV estimated on the training set converges to its test-set value in the overparameterized limit, providing the first result that prediction variance under local parameter perturbations can be inferred from training inputs alone. Empirically, TPV exhibits a striking stability across datasets and architectures -- including extremely narrow networks -- and correlates well with clean test loss. Finally, we demonstrate that modeling pruning as a TPV perturbation yields a simple label-free importance measure that performs competitively with state-of-the-art pruning methods, illustrating the practical utility of TPV. Code available at github.com/devansharpit/TPV.

For four decades statistical physics has been providing a framework to analyse neural networks. A long-standing question remained on its capacity to tackle deep learning models capturing rich feature learning effects, thus going beyond the narrow networks or kernel methods analysed until now. We positively answer through the study of the supervised learning of a multi-layer perceptron. Importantly, (i) its width scales as the input dimension, making it more prone to feature learning than ultra wide networks, and more expressive than narrow ones or ones with fixed embedding layers; and (ii) we focus on the challenging interpolation regime where the number of trainable parameters and data are comparable, which forces the model to adapt to the task. We consider the matched teacher-student setting. Therefore, we provide the fundamental limits of learning random deep neural network targets and identify the sufficient statistics describing what is learnt by an optimally trained network as the data budget increases. A rich phenomenology emerges with various learning transitions. With enough data, optimal performance is attained through the model's "specialisation" towards the target, but it can be hard to reach for training algorithms which get attracted by sub-optimal solutions predicted by the theory. Specialisation occurs inhomogeneously across layers, propagating from shallow towards deep ones, but also across neurons in each layer. Furthermore, deeper targets are harder to learn. Despite its simplicity, the Bayes-optimal setting provides insights on how the depth, non-linearity and finite (proportional) width influence neural networks in the feature learning regime that are potentially relevant in much more general settings.

In a small room in San Diego last week, a man in a black leather jacket explained to me how to save the world from destruction by AI. Max Tegmark, a notable figure in the AI-safety movement, believes that "artificial general intelligence," or AGI, could precipitate the end of human life. I was in town for NeurIPS, one of the largest AI-research conferences, and Tegmark had invited me, along with five other journalists, to a briefing on an AI-safety index that he would release the next day. No company scored better than a C+. The threat of technological superintelligence is the stuff of science fiction, yet it has become a topic of serious discussion in the past few years.

If language is what makes us human, what does it mean now that large language models have gained "metalinguistic" abilities? Among the myriad abilities that humans possess, which ones are uniquely human? Language has been a top candidate at least since Aristotle, who wrote that humanity was "the animal that has language." Even as large language models such as ChatGPT superficially replicate ordinary speech, researchers want to know if there are specific aspects of human language that simply have no parallels in the communication systems of other animals or artificially intelligent devices. In particular, researchers have been exploring the extent to which language models can reason about language itself.

He created Grand Theft Auto. Dan Houser was one of the masterminds behind revolutionary video game series Grand Theft Auto. Now, after leaving Rockstar Games and launching his own company, he's released a debut novel about a very different type of game. A Better Paradise is a dystopian vision of the near future in which an AI-led computer game goes rogue. Set in a polarised world, it finds Mark Tyburn attempting to create a virtual haven for people to find sanctuary and reconnect within themselves against an all-consuming social media hellscape.

OpenAI turned 10 yesterday, and President Donald Trump incidentally gave the company a very special birthday gift: a sweeping executive order aiming to dismantle and preempt many state-level regulations of artificial intelligence. "There's only going to be one winner here, and it's probably going to be the U.S. or China," Trump said in a press conference announcing the order. And for the United States to win, "we have to be unified. Almost all of the AI industry's biggest players have been pushing for this move. OpenAI has been asking all year for the Trump administration to preempt state-level AI regulations, which the company believes would be burdensome in various ways; Microsoft, Google, Meta, Nvidia, and the major venture-capital firm Andreessen Horowitz have made similar requests.