Search for: All records

The trajectory of a molecular system undergoing a reversible reaction A ⇌ B and crossing and recrossing a transition state separating the reactant and product consists of loops, i.e., excursions from the transition state to either side and back to the transition state. Motivated by recent experimental observations of loops, here, we discuss some of their statistical properties. In particular, we highlight that the transition-state rate is not only an upper bound on the true reaction rate but also a physical property of the loops. We further find that loops can be unambiguously divided into two sub-ensembles. Those consist of short loops, which are brief excursions from the transition state, and long loops that get trapped in the reactant or product wells before eventually returning to the barrier. Finally, we show that the loop time distribution contains information about both the reaction rate coefficients and their transition-state-theory counterparts. 

Dense document embeddings are central to neural retrieval. The dominant paradigm is to train and construct embeddings by running encoders directly on individual documents. In this work, we argue that these embeddings, while effective, are implicitly out-of-context for targeted use cases of retrieval, and that a contextualized document embedding should take into account both the document and neighboring documents in context - analogous to contextualized word embeddings. We propose two complementary methods for contextualized document embeddings: first, an alternative contrastive learning objective that explicitly incorporates the document neighbors into the intra-batch contextual loss; second, a new contextual architecture that explicitly encodes neighbor document information into the encoded representation. Results show that both methods achieve better performance than biencoders in several settings, with differences especially pronounced out-of-domain. We achieve state-of-the-art results on the MTEB benchmark with no hard negative mining, score distillation, dataset-specific instructions, intra-GPU example-sharing, or extremely large batch sizes. Our method can be applied to improve performance on any contrastive learning dataset and any biencoder. 

Data selection can reduce the amount of training data needed to finetune LLMs; however, the efficacy of data selection scales directly with its compute. Motivated by the practical challenge of compute-constrained finetuning, we consider the setting in which both the cost of selecting data and training are budgeted for. We first formalize the problem of data selection with a cost-aware utility function, and model the data selection problem as trading off initial-selection cost for training gain. We run a comprehensive sweep of experiments across multiple tasks, varying compute budget by scaling finetuning tokens, model sizes, and data selection compute. Interestingly we find that many powerful data selection methods are almost never compute-optimal, and that cheaper data selection alternatives dominate both from a theoretical and empirical perspective. For compute-optimal training, we find that perplexity and gradient data selection require training-to-selection model size ratios of 5x and 10x, respectively 

The demand for computing power has been growing exponentially with the rise of artificial intelligence (AI), machine learning, and the Internet of Things (IoT). This growth requires unconventional computing primitives that prioritize energy efficiency, while also addressing the critical need for scalability. Neuromorphic computing, inspired by the biological brain, offers a transformative paradigm for addressing these challenges. This review paper provides an overview of advancements in 2D spintronics and device architectures designed for neuromorphic applications, with a focus on techniques such as spin-orbit torque, magnetic tunnel junctions, and skyrmions. Emerging van der Waals materials like CrI3, Fe3GaTe2, and graphene-based heterostructures have demonstrated unparalleled potential for integrating memory and logic at the atomic scale. This work highlights technologies with ultra-low energy consumption (0.14 fJ/operation), high switching speeds (sub-nanosecond), and scalability to sub-20 nm footprints. It covers key material innovations and the role of spintronic effects in enabling compact, energy-efficient neuromorphic systems, providing a foundation for advancing scalable, next-generation computing architectures. 

Convective gravity waves are important for the forcing of the quasi biennial oscillation (QBO). There is a wave component that is stationary with respect to the convective cells that is triggered by convection acting like a barrier to the background flow (moving mountain mechanism). Waves from this mechanism have only been observed in a few case studies and are not parameterized in climate models. However, the representation of the whole spectrum of gravity waves is crucial for the simulation of the QBO, especially in the lowermost stratosphere (below 50 hPa) where the QBO amplitudes are under‐estimated in current global circulation models. In this study, we present analysis of convective gravity wave observations from superpressure balloons in boreal winter 2019 and 2021, retrieving phase speeds, momentum fluxes, and drag. We also identify waves generated by the moving mountain mechanism using the theory of the Beres scheme as a basis. These waves do not have a specific period, but are of smaller horizontal scale, on average around 300 km, which is similar to the scale of convective systems. Our results show that gravity waves contribute up to 2/3 to the QBO forcing below 50 hPa and waves from the moving mountain mechanism are responsible for up to 10% of this forcing.