Reducing the environmental impact of cloud computing requires efficient workload distribution across geographically dispersed Data Center Clusters (DCCs) and simultaneously optimizing liquid and air (HVAC) cooling with time shift of workloads within individual data centers (DC). This paper introduces Green-DCC, which proposes a Reinforcement Learning (RL) based hierarchical controller to optimize both workload and liquid cooling dynamically in a DCC. By incorporating factors such as weather, carbon intensity, and resource availability, Green-DCC addresses realistic constraints and interdependencies. We demonstrate how the system optimizes multiple data centers synchronously, enabling the scope of digital twins, and compare the performance of various RL approaches based on carbon emissions and sustainability metrics while also offering a framework and benchmark simulation for broader ML research in sustainability.

Power grid operation is becoming more complex due to the increase in generation of renewable energy. The recent series of Learning To Run a Power Network (L2RPN) competitions have encouraged the use of artificial agents to assist human dispatchers in operating power grids. However, the combinatorial nature of the action space poses a challenge to both conventional optimizers and learned controllers. Action space factorization, which breaks down decision-making into smaller sub-tasks, is one approach to tackle the curse of dimensionality. In this study, we propose a centrally coordinated multi-agent (CCMA) architecture for action space factorization. In this approach, regional agents propose actions and subsequently a coordinating agent selects the final action. We investigate several implementations of the CCMA architecture, and benchmark in different experimental settings against various L2RPN baseline approaches. The CCMA architecture exhibits higher sample efficiency and superior final performance than the baseline approaches. The results suggest high potential of the CCMA approach for further application in higher-dimensional L2RPN as well as real-world power grid settings.

Training and fine-tuning large language models (LLMs) with hundreds of billions to trillions of parameters requires tens of thousands of GPUs, and a highly scalable software stack. In this work, we present a novel four-dimensional hybrid parallel algorithm implemented in a highly scalable, portable, open-source framework called AxoNN. We describe several performance optimizations in AxoNN to improve matrix multiply kernel performance, overlap non-blocking collectives with computation, and performance modeling to choose performance optimal configurations. These have resulted in unprecedented scaling and peak flop/s (bf16) for training of GPT-style transformer models on Perlmutter (620.1 Petaflop/s), Frontier (1.381 Exaflop/s) and Alps (1.423 Exaflop/s). While the abilities of LLMs improve with the number of trainable parameters, so do privacy and copyright risks caused by memorization of training data, which can cause disclosure of sensitive or private information at inference time. We highlight this side effect of scale through experiments that explore "catastrophic memorization", where models are sufficiently large to memorize training data in a single pass, and present an approach to prevent it. As part of this study, we demonstrate fine-tuning of a 405-billion parameter LLM using AxoNN on Frontier.

While deep generative models~(DGMs) have demonstrated remarkable success in capturing complex data distributions, they consistently fail to learn constraints that encode domain knowledge and thus require constraint integration. Existing solutions to this challenge have primarily relied on heuristic methods and often ignore the underlying data distribution, harming the generative performance. In this work, we propose a probabilistically sound approach for enforcing the hard constraints into DGMs to generate constraint-compliant and realistic data. This is achieved by our proposed gradient estimators that allow the constrained distribution, the data distribution conditioned on constraints, to be differentiably learned. We carry out extensive experiments with various DGM model architectures over five image datasets and three scientific applications in which domain knowledge is governed by linear equality constraints. We validate that the standard DGMs almost surely generate data violating the constraints. Among all the constraint integration strategies, ours not only guarantees the satisfaction of constraints in generation but also archives superior generative performance than the other methods across every benchmark.

Graph neural networks (GNNs) have shown considerable promise in computational chemistry. However, the limited availability of molecular data raises concerns regarding GNNs' ability to effectively capture the fundamental principles of physics and chemistry, which constrains their generalization capabilities. To address this challenge, we introduce a novel self-supervised approach termed Equivariant Masked Position Prediction (EMPP), grounded in intramolecular potential and force theory. Unlike conventional attribute masking techniques, EMPP formulates a nuanced position prediction task that is more well-defined and enhances the learning of quantum mechanical features. EMPP also bypasses the approximation of the Gaussian mixture distribution commonly used in denoising methods, allowing for more accurate acquisition of physical properties. Experimental results indicate that EMPP significantly enhances performance of advanced molecular architectures, surpassing state-of-the-art self-supervised approaches. Graph neural networks (GNNs) have found widespread application in computational chemistry. However, unlike other fields such as natural language processing (NLP), the limited availability of molecular data hampers the development of GNNs in this domain. For example, one of the largest molecular dataset, OC20 (Chanussot et al., 2021), contains only 1.38 million samples, and collecting more molecular data with ab initio calculations is both challenging and expensive. To address this limitation, molecular self-supervised learning has gained increasing attention. This approach enables molecular GNNs to learn more general physical and chemical knowledge, enhancing performance in various computational chemistry tasks, such as drug discovery (Hasselgren & Oprea, 2024) and catalyst design (Chanussot et al., 2021). Current self-supervised methods for molecular learning contain two mainstream categories: masking and denoising. Masking methods (Hu et al., 2020; Hou et al., 2022; Inae et al., 2023) adapt the concept of masked token prediction from natural language processing (NLP) to graph learning, where graph information, such as node attribute, is masked instead of token. However, there are two major limitations: underdetermined reconstruction and lack of deep quantum mechanical (QM) insight.

This study introduces a novel AI microcontroller optimized for cost-effective, battery-powered edge AI applications. Unlike traditional single bit/cell memory configurations, the proposed microcontroller integrates zero-standby power weight memory featuring standard logic compatible 4-bits/cell embedded flash technology tightly coupled to a Near-Memory Computing Unit. This architecture enables efficient and low-power AI acceleration. Advanced state mapping and an overstress-free word line (WL) driver circuit extend verify levels, ensuring robust 16 state cell margin. A ping-pong buffer reduces internal data movement while supporting simultaneous multi-bit processing. The fabricated microcontroller demonstrated high reliability, maintaining accuracy after 160 hours of unpowered baking at 125$^\circ$C.

In recent years, there has been an increasing interest in using deep learning and neural networks to tackle scientific problems, particularly in solving partial differential equations (PDEs). However, many neural network-based methods, such as physics-informed neural networks, depend on automatic differentiation and the sampling of collocation points, which can result in a lack of interpretability and lower accuracy compared to traditional numerical methods. To address this issue, we propose two approaches for learning discontinuous Galerkin solutions to PDEs using small linear convolutional neural networks. Our first approach is supervised and depends on labeled data, while our second approach is unsupervised and does not rely on any training data. In both cases, our methods use substantially fewer parameters than similar numerics-based neural networks while also demonstrating comparable accuracy to the true and DG solutions for elliptic problems.

Congratulations --- the reviewers all agreed that this work was a creative and interesting contribution to the machine learning literature! In your revision, please address any questions from the reviews and promises in your rebuttal.

Over the past few weeks, Donald Trump has positioned himself as an unabashed bull on America's need to dominate AI. Yet the president has also tied this newfound and futuristic priority to a more traditional mission of his: to go big with fossil fuels. A true AI revolution will need "double the energy" that America produces today, Trump said in a recent address to the World Economic Forum, days after declaring a national energy emergency. And he noted a few ways to supply that power: "We have more coal than anybody. We also have more oil and gas than anybody."

Large language models (LLMs) have achieved significant performance gains via scaling up model sizes and/or data. However, recent evidence suggests diminishing returns from such approaches, motivating scaling the computation spent at inference time. Existing inference-time scaling methods, usually with reward models, cast the task as a search problem, which tends to be vulnerable to reward hacking as a consequence of approximation errors in reward models. In this paper, we instead cast inference-time scaling as a probabilistic inference task and leverage sampling-based techniques to explore the typical set of the state distribution of a state-space model with an approximate likelihood, rather than optimize for its mode directly. We propose a novel inference-time scaling approach by adapting particle-based Monte Carlo methods to this task. Our empirical evaluation demonstrates that our methods have a 4-16x better scaling rate over our deterministic search counterparts on various challenging mathematical reasoning tasks. Using our approach, we show that Qwen2.5-Math-1.5B-Instruct can surpass GPT-4o accuracy in only 4 rollouts, while Qwen2.5-Math-7B-Instruct scales to o1 level accuracy in only 32 rollouts. Our work not only presents an effective method to inference-time scaling, but also connects the rich literature in probabilistic inference with inference-time scaling of LLMs to develop more robust algorithms in future work. Code, videos, and further information available at https://probabilistic-inference-scaling.github.io.