Search

We study the superconducting critical fields of two-dimensional electron systems at (111)$\mathrm{LaAlO_3/KTaO_3}$ interfaces as a function of electrostatic back-gating. Our work reveals inplane critical fields of unprecedented magnitudes at oxide interfaces. By comparing the critical fields in-plane and out-of-plane we discover an extremely anisotropic superconductor with an effective thickness below 1 nm and a 12-fold violation of the Chandrasekhar-Clogston paramagnetic limit. The analysis of magneto-transport indicates that the enhancement of the critical fields is due to an exceptionally thin superconducting layer and to a paramagnetic susceptibility suppressed by spin-orbit scattering.

Statistical heterogeneity is a measure of how skewed the samples of a dataset are. It is a common problem in the study of differential privacy that the usage of a statistically heterogeneous dataset results in a significant loss of accuracy. In federated scenarios, statistical heterogeneity is more likely to happen, and so the above problem is even more pressing. We explore the three most promising ways to measure statistical heterogeneity and give formulae for their accuracy, while simultaneously incorporating differential privacy. We find the optimum privacy parameters via an analytic mechanism, which incorporates root finding methods. We validate the main theorems and related hypotheses experimentally, and test the robustness of the analytic mechanism to different heterogeneity levels. The analytic mechanism in a distributed setting delivers superior accuracy to all combinations involving the classic mechanism and/or the centralized setting. All measures of statistical heterogeneity do not lose significant accuracy when a heterogeneous sample is used., Comment: 26 pages, 6 tables, 1 figure

We construct four-variable $p$-adic $L$-functions for the spin Galois representation of a Siegel modular form of genus 2 twisted by the Galois representation of a cuspidal modular form as the modular forms vary in Coleman families. The main ingredient is the construction of a space of nearly overconvergent modular forms in the coherent cohomology of the Siegel threefold, extending the spaces of overconvergent modular forms appearing in higher Coleman theory. In addition to this, we construct $p$-adic distributions interpolating the Gan-Gross-Prasad automorphic periods for $(\mathrm{GSpin}(5), \mathrm{GSpin}(4))$ which, conditional on the local and global Gan-Gross-Prasad conjectures for this pair of groups, provides a construction of "square-root" $p$-adic $L$-functions for $\mathrm{GSp}(4) \times \mathrm{GL}(2) \times \mathrm{GL}(2)$ as the automorphic forms vary in Coleman families., Comment: 54 pages

This article proposes a new integration of linguistic anthropology and machine learning (ML) around convergent interests in both the underpinnings of language and making language technologies more socially responsible. While linguistic anthropology focuses on interpreting the cultural basis for human language use, the ML field of interpretability is concerned with uncovering the patterns that Large Language Models (LLMs) learn from human verbal behavior. Through the analysis of a conversation between a human user and an LLM-powered chatbot, we demonstrate the theoretical feasibility of a new, conjoint field of inquiry, cultural interpretability (CI). By focusing attention on the communicative competence involved in the way human users and AI chatbots co-produce meaning in the articulatory interface of human-computer interaction, CI emphasizes how the dynamic relationship between language and culture makes contextually sensitive, open-ended conversation possible. We suggest that, by examining how LLMs internally "represent" relationships between language and culture, CI can: (1) provide insight into long-standing linguistic anthropological questions about the patterning of those relationships; and (2) aid model developers and interface designers in improving value alignment between language models and stylistically diverse speakers and culturally diverse speech communities. Our discussion proposes three critical research axes: relativity, variation, and indexicality., Comment: Accepted for publication in Big Data & Society, November 2, 2024

The Legacy Survey of Space and Time (LSST) at Vera C. Rubin Observatory is planned to begin in the Fall of 2025. The LSST survey cadence has been designed via a community-driven process regulated by the Survey Cadence Optimization Committee (SCOC), which recommended up to 3% of the observing time to carry out Target of Opportunity (ToO) observations. Experts from the scientific community, Rubin Observatory personnel, and members of the SCOC were brought together to deliver a recommendation for the implementation of the ToO program during a workshop held in March 2024. Four main science cases were identified: gravitational wave multi-messenger astronomy, high energy neutrinos, Galactic supernovae, and small potentially hazardous asteroids possible impactors. Additional science cases were identified and briefly addressed in the documents, including lensed or poorly localized gamma-ray bursts and twilight discoveries. Trigger prioritization, automated response, and detailed strategies were discussed for each science case. This document represents the outcome of the Rubin ToO 2024 workshop, with additional contributions from members of the Rubin Science Collaborations. The implementation of the selection criteria and strategies presented in this document has been endorsed in the SCOC Phase 3 Recommendations document (PSTN-056). Although the ToO program is still to be finalized, this document serves as a baseline plan for ToO observations with the Rubin Observatory.

When comparing the linguistic capabilities of language models (LMs) with humans using LM probabilities, factors such as the length of the sequence and the unigram frequency of lexical items have a significant effect on LM probabilities in ways that humans are largely robust to. Prior works in comparing LM and human acceptability judgments treat these effects uniformly across models, making a strong assumption that models require the same degree of adjustment to control for length and unigram frequency effects. We propose MORCELA, a new linking theory between LM scores and acceptability judgments where the optimal level of adjustment for these effects is estimated from data via learned parameters for length and unigram frequency. We first show that MORCELA outperforms a commonly used linking theory for acceptability--SLOR (Pauls and Klein, 2012; Lau et al. 2017)--across two families of transformer LMs (Pythia and OPT). Furthermore, we demonstrate that the assumed degrees of adjustment in SLOR for length and unigram frequency overcorrect for these confounds, and that larger models require a lower relative degree of adjustment for unigram frequency, though a significant amount of adjustment is still necessary for all models. Finally, our subsequent analysis shows that larger LMs' lower susceptibility to frequency effects can be explained by an ability to better predict rarer words in context.

A spectral minimal partition of a manifold is a decomposition into disjoint open sets that minimizes a spectral energy functional. While it is known that bipartite minimal partitions correspond to nodal partitions of Courant-sharp Laplacian eigenfunctions, the non-bipartite case is much more challenging. In this paper, we unify the bipartite and non-bipartite settings by defining a modified Laplacian operator and proving that the nodal partitions of its eigenfunctions are exactly the critical points of the spectral energy functional. Moreover, we prove that the Morse index of a critical point equals the nodal deficiency of the corresponding eigenfunction. Some striking consequences of our main result are: 1) in the bipartite case, every local minimum of the energy functional is in fact a global minimum; 2) in the non-bipartite case, every local minimum of the energy functional minimizes within a certain topological class of partitions. Our results are valid for partitions with non-smooth boundaries; this introduces considerable technical challenges, which are overcome using delicate approximation arguments in the Sobolev space $H^{1/2}$.

Class imbalance and label noise are pervasive in large-scale datasets, yet much of machine learning research assumes well-labeled, balanced data, which rarely reflects real world conditions. Existing approaches typically address either label noise or class imbalance in isolation, leading to suboptimal results when both issues coexist. In this work, we propose Conformal-in-the-Loop (CitL), a novel training framework that addresses both challenges with a conformal prediction-based approach. CitL evaluates sample uncertainty to adjust weights and prune unreliable examples, enhancing model resilience and accuracy with minimal computational cost. Our extensive experiments include a detailed analysis showing how CitL effectively emphasizes impactful data in noisy, imbalanced datasets. Our results show that CitL consistently boosts model performance, achieving up to a 6.1% increase in classification accuracy and a 5.0 mIoU improvement in segmentation. Our code is publicly available: CitL., Comment: Under Review

The worlds of Data Science (including big and/or federated data, machine learning, etc) and Astrophysics started merging almost two decades ago. For instance, around 2005, international initiatives such as the Virtual Observatory framework rose to standardize the way we publish and transfer data, enabling new tools such as VOSA (SED Virtual Observatory Analyzer) to come to existence and remain relevant today. More recently, new facilities like the Vera Rubin Observatory, serve as motivation to develop efficient and extremely fast (very often deep learning based) methodologies in order to fully exploit the informational content of the vast Legacy Survey of Space and Time (LSST) dataset. However, fundamental changes in the way we explore and analyze data cannot permeate in the "astrophysical sociology and idiosyncrasy" without adequate training. In this talk, I will focus on one specific initiative that has been extremely successful and is based on "learning by doing": the La Serena School for Data Science. I will also briefly touch on a different successful approach: a series of schools organized by the Spanish Virtual Observatory. The common denominator among the two kinds of schools is to present the students with real scientific problems that benefit from the concepts / methodologies taught. On the other hand, the demographics targeted by both initiatives vary significantly and can represent examples of two "flavours" to be followed by others., Comment: Invited talk IAU-GA 2024, FM 7

We use Magnetospheric Multiscale (MMS) mission data to investigate electron-scale energy transfer due to lower hybrid drift waves during magnetopause reconnection. We analyze waves observed in an electron-scale plasma mixing layer at the edge of the magnetospheric outflow. Using high-resolution 7.5 ms electron moments, we obtain an electron current density with a Nyquist frequency of ~66 Hz, sufficient to resolve most of the lower hybrid drift wave power observed in the event. We then employ wavelet analysis to evaluate dJ.dE, which accounts for the phase differences between the fluctuating quantities. The analysis shows that the energy exchange is localized within the plasma mixing layer, and it is highly fluctuating, with energy bouncing between waves and electrons throughout the analyzed time and frequency range. However, the cumulative sum over time indicates a net energy transfer from the waves to electrons. We observe an anomalous electron flow toward the magnetosphere, consistent with diffusion and electron mixing. These results indicate that waves and electrons interact dynamically to dissipate the excess internal energy accumulated by sharp density gradients. We conclude that the electron temperature profile within the plasma mixing layer is produced by a combination of electron diffusion across the layer, as well as heating by large-scale parallel potential and lower hybrid drift waves., Comment: 23 pages, 6 figures. Submitted to Journal of Geophysical Research: Space Physics

We identify symmetries in a broad class of vector-like confining dark sectors that forbid the leading electromagnetic moments that would ordinarily mediate dark baryon scattering with the Standard Model. The absence of these operators implies dark baryon dark matter has much smaller cross sections for elastic scattering off nuclei, leading to suppressed direct detection signals. In the confined description, we identify an ``$\mathcal{H}$-parity'' symmetry that exists in any dark sector with dark quarks transforming under a vector-like representation of a new confining SU($N_c$) gauge theory as well as a vector-like representation of the electroweak group SU(2)$_L$. The parity is independent of $N_c$ and $N_f$, though it is essential that the dark quarks are neutral under hypercharge. This parity forbids dark hadron electric and magnetic dipole moments, charge radius, and anapole moment, while permitting dimension-7 operators that include polarizability, electroweak loop-induced interactions, as well as lower dimensional electromagnetic $\textit{transition}$ moments between different neutral dark baryon states. We work out an explicit example, $N_c=N_f=3$, that is the most minimal theory with fermionic dark baryons. In this specific model, we use the non-relativistic quark model to show the magnetic dipole moment and charge radius vanish while the transition moments are non-zero, consistent with $\mathcal{H}$-parity. We discuss the implications of a suppressed direct detection signal, emphasizing that this broad class of models provide a well-motivated target for future colliders., Comment: 28+14 pages, 1+1 figures, 2 tables

The ability to resolve photon numbers is crucial in quantum information science and technology, driving the development of detectors with intrinsic photon-number resolving (PNR) capabilities. Although transition edge sensors represent the state-of-the-art in PNR performance, superconducting nanowire single-photon detectors (SNSPDs) offer superior efficiency, speed, noise reduction, and timing precision. Directly inferring photon numbers, however, has only recently become feasible due to advances in readout technology. Despite this, photon-number discrimination remains constrained by the nanowire's electrical properties and readout jitter. In this work, we employ waveguide-integrated SNSPDs and time-resolved measurements to explore how the nanowire kinetic inductance and system jitter affect PNR capabilities. By analyzing the latency time of the photon detection, we can resolve changes in the rising edge of the detection pulse. We find that lower jitter as well as increased kinetic inductance enhances the pulse separation for different photon numbers and improves the PNR capability. Enhancing the kinetic inductance from 165 nH to 872 nH improves PNR quality by 12%, 31% and 23% over the first three photon numbers, though at the cost of reducing the detector's count rate from 165 Mcps to 19 Mcps. Our findings highlight the trade-off between PNR resolution and detector speed., Comment: 21 pages, 7 figures

A strong first-order electroweak phase transition (SFOEWPT) is essential for explaining baryogenesis and for potentially generating observable gravitational waves. This study investigates the potential of a high-energy muon collider to examine the occurrence of SFOEWPT within the context of a Standard Model extended by a real scalar singlet (xSM). We analyzed all possible decay modes of the singlet to constrain the valid parameter space of SFOEWPT, which was extracted numerically at different renormalization scales to account for theoretical uncertainties, thereby determining the sensitivity of a muon collider to the production and decay channels of novel heavy scalar particles that emerge in the xSM. The findings demonstrate that a 3 TeV muon collider can directly examine the nature of electroweak symmetry breaking by efficiently detecting novel scalar particles associated with a first-order electroweak phase transition through jet-rich final states, thus complementing the indirect constraints from gravitational wave experiments.

Trust and interpretability are crucial for the use of Artificial Intelligence (AI) in scientific research, but current models often operate as black boxes offering limited transparency and justifications for their outputs. We introduce AiSciVision, a framework that specializes Large Multimodal Models (LMMs) into interactive research partners and classification models for image classification tasks in niche scientific domains. Our framework uses two key components: (1) Visual Retrieval-Augmented Generation (VisRAG) and (2) domain-specific tools utilized in an agentic workflow. To classify a target image, AiSciVision first retrieves the most similar positive and negative labeled images as context for the LMM. Then the LMM agent actively selects and applies tools to manipulate and inspect the target image over multiple rounds, refining its analysis before making a final prediction. These VisRAG and tooling components are designed to mirror the processes of domain experts, as humans often compare new data to similar examples and use specialized tools to manipulate and inspect images before arriving at a conclusion. Each inference produces both a prediction and a natural language transcript detailing the reasoning and tool usage that led to the prediction. We evaluate AiSciVision on three real-world scientific image classification datasets: detecting the presence of aquaculture ponds, diseased eelgrass, and solar panels. Across these datasets, our method outperforms fully supervised models in low and full-labeled data settings. AiSciVision is actively deployed in real-world use, specifically for aquaculture research, through a dedicated web application that displays and allows the expert users to converse with the transcripts. This work represents a crucial step toward AI systems that are both interpretable and effective, advancing their use in scientific research and scientific discovery.

Quasars are bright active galactic nuclei powered by the accretion of matter around supermassive black holes at the center of galaxies. Their stochastic brightness variability depends on the physical properties of the accretion disk and black hole. The upcoming Rubin Observatory Legacy Survey of Space and Time (LSST) is expected to observe tens of millions of quasars, so there is a need for efficient techniques like machine learning that can handle the large volume of data. Quasar variability is believed to be driven by an X-ray corona, which is reprocessed by the accretion disk and emitted as UV/optical variability. We are the first to introduce an auto-differentiable simulation of the accretion disk and reprocessing. We use the simulation as a direct component of our neural network to jointly model the driving variability and reprocessing to fit simulated LSST 10-year quasar light curves. The driving variability is reconstructed using a latent stochastic differential equation, a physically motivated, generative deep learning method that can model continuous-time stochastic dynamics. By embedding these physical processes into our network, we achieve a model that is more robust and interpretable. We also use transformers to scale our model to tens of millions of parameters. We demonstrate how our model outperforms a Gaussian process regression baseline and can infer accretion disk parameters and time delays between wavebands, even for out-of-distribution driving signals. Our approach provides a powerful and scalable framework that can be adapted to solve other inverse problems in multivariate time series with irregular sampling., Comment: 31 pages, 19 figures

Recent advances in multi-agent reinforcement learning (MARL) have created opportunities to solve complex real-world tasks. Cybersecurity is a notable application area, where defending networks against sophisticated adversaries remains a challenging task typically performed by teams of security operators. In this work, we explore novel MARL strategies for building autonomous cyber network defenses that address challenges such as large policy spaces, partial observability, and stealthy, deceptive adversarial strategies. To facilitate efficient and generalized learning, we propose a hierarchical Proximal Policy Optimization (PPO) architecture that decomposes the cyber defense task into specific sub-tasks like network investigation and host recovery. Our approach involves training sub-policies for each sub-task using PPO enhanced with domain expertise. These sub-policies are then leveraged by a master defense policy that coordinates their selection to solve complex network defense tasks. Furthermore, the sub-policies can be fine-tuned and transferred with minimal cost to defend against shifts in adversarial behavior or changes in network settings. We conduct extensive experiments using CybORG Cage 4, the state-of-the-art MARL environment for cyber defense. Comparisons with multiple baselines across different adversaries show that our hierarchical learning approach achieves top performance in terms of convergence speed, episodic return, and several interpretable metrics relevant to cybersecurity, including the fraction of clean machines on the network, precision, and false positives on recoveries., Comment: 9 pages, 7 figures, AAMAS preprint

Accelerating research on Large Multimodal Models (LMMs) in non-English languages is crucial for enhancing user experiences across broader populations. In this paper, we introduce JMMMU (Japanese MMMU), the first large-scale Japanese benchmark designed to evaluate LMMs on expert-level tasks based on the Japanese cultural context. To facilitate comprehensive culture-aware evaluation, JMMMU features two complementary subsets: (i) culture-agnostic (CA) subset, where the culture-independent subjects (e.g., Math) are selected and translated into Japanese, enabling one-to-one comparison with its English counterpart MMMU; and (ii) culture-specific (CS) subset, comprising newly crafted subjects that reflect Japanese cultural context. Using the CA subset, we observe performance drop in many LMMs when evaluated in Japanese, which is purely attributable to language variation. Using the CS subset, we reveal their inadequate Japanese cultural understanding. Further, by combining both subsets, we identify that some LMMs perform well on the CA subset but not on the CS subset, exposing a shallow understanding of the Japanese language that lacks depth in cultural understanding. We hope this work will not only help advance LMM performance in Japanese but also serve as a guideline to create high-standard, culturally diverse benchmarks for multilingual LMM development. The project page is https://mmmu-japanese-benchmark.github.io/JMMMU/., Comment: Project page: https://mmmu-japanese-benchmark.github.io/JMMMU/

We present the results of a search for gravitational-wave transients associated with core-collapse supernova SN 2023ixf, which was observed in the galaxy Messier 101 via optical emission on 2023 May 19th, during the LIGO-Virgo-KAGRA 15th Engineering Run. We define a five-day on-source window during which an accompanying gravitational-wave signal may have occurred. No gravitational waves have been identified in data when at least two gravitational-wave observatories were operating, which covered $\sim 14\%$ of this five-day window. We report the search detection efficiency for various possible gravitational-wave emission models. Considering the distance to M101 (6.7 Mpc), we derive constraints on the gravitational-wave emission mechanism of core-collapse supernovae across a broad frequency spectrum, ranging from 50 Hz to 2 kHz where we assume the GW emission occurred when coincident data are available in the on-source window. Considering an ellipsoid model for a rotating proto-neutron star, our search is sensitive to gravitational-wave energy $1 \times 10^{-5} M_{\odot} c^2$ and luminosity $4 \times 10^{-5} M_{\odot} c^2/\text{s}$ for a source emitting at 50 Hz. These constraints are around an order of magnitude more stringent than those obtained so far with gravitational-wave data. The constraint on the ellipticity of the proto-neutron star that is formed is as low as $1.04$, at frequencies above $1200$ Hz, surpassing results from SN 2019ejj., Comment: Main paper: 6 pages, 4 figures and 1 table. Total with appendices: 20 pages, 4 figures, and 1 table

Web browsers are a portal to the internet, where much of human activity is undertaken. Thus, there has been significant research work in AI agents that interact with the internet through web browsing. However, there is also another interface designed specifically for machine interaction with online content: application programming interfaces (APIs). In this paper we ask -- what if we were to take tasks traditionally tackled by browsing agents, and give AI agents access to APIs? To do so, we propose two varieties of agents: (1) an API-calling agent that attempts to perform online tasks through APIs only, similar to traditional coding agents, and (2) a Hybrid Agent that can interact with online data through both web browsing and APIs. In experiments on WebArena, a widely-used and realistic benchmark for web navigation tasks, we find that API-based agents outperform web browsing agents. Hybrid Agents out-perform both others nearly uniformly across tasks, resulting in a more than 20.0% absolute improvement over web browsing alone, achieving a success rate of 35.8%, achiving the SOTA performance among task-agnostic agents. These results strongly suggest that when APIs are available, they present an attractive alternative to relying on web browsing alone., Comment: 24 pages, 6 figures

Despite recent advances in multimodal large language models (MLLMs), their development has predominantly focused on English- and western-centric datasets and tasks, leaving most of the world's languages and diverse cultural contexts underrepresented. This paper introduces Pangea, a multilingual multimodal LLM trained on PangeaIns, a diverse 6M instruction dataset spanning 39 languages. PangeaIns features: 1) high-quality English instructions, 2) carefully machine-translated instructions, and 3) culturally relevant multimodal tasks to ensure cross-cultural coverage. To rigorously assess models' capabilities, we introduce PangeaBench, a holistic evaluation suite encompassing 14 datasets covering 47 languages. Results show that Pangea significantly outperforms existing open-source models in multilingual settings and diverse cultural contexts. Ablation studies further reveal the importance of English data proportions, language popularity, and the number of multimodal training samples on overall performance. We fully open-source our data, code, and trained checkpoints, to facilitate the development of inclusive and robust multilingual MLLMs, promoting equity and accessibility across a broader linguistic and cultural spectrum., Comment: 52 pages, 27 figures

Vision-language models (VLMs) have made significant progress in recent visual-question-answering (VQA) benchmarks that evaluate complex visio-linguistic reasoning. However, are these models truly effective? In this work, we show that VLMs still struggle with natural images and questions that humans can easily answer, which we term natural adversarial samples. We also find it surprisingly easy to generate these VQA samples from natural image-text corpora using off-the-shelf models like CLIP and ChatGPT. We propose a semi-automated approach to collect a new benchmark, NaturalBench, for reliably evaluating VLMs with 10,000 human-verified VQA samples. Crucially, we adopt a $\textbf{vision-centric}$ design by pairing each question with two images that yield different answers, preventing blind solutions from answering without using the images. This makes NaturalBench more challenging than previous benchmarks that can be solved with commonsense priors. We evaluate 53 state-of-the-art VLMs on NaturalBench, showing that models like LLaVA-OneVision, Cambrian-1, Llama3.2-Vision, Molmo, Qwen2-VL, and even GPT-4o lag 50%-70% behind human performance (over 90%). We analyze why NaturalBench is hard from two angles: (1) Compositionality: Solving NaturalBench requires diverse visio-linguistic skills, including understanding attribute bindings, object relationships, and advanced reasoning like logic and counting. To this end, unlike prior work that uses a single tag per sample, we tag each NaturalBench sample with 1 to 8 skill tags for fine-grained evaluation. (2) Biases: NaturalBench exposes severe biases in VLMs, as models often choose the same answer regardless of the image. Lastly, we apply our benchmark curation method to diverse data sources, including long captions (over 100 words) and non-English languages like Chinese and Hindi, highlighting its potential for dynamic evaluations of VLMs., Comment: Accepted to NeurIPS 24; We open-source our dataset at: https://huggingface.co/datasets/BaiqiL/NaturalBench ; Project page at: https://linzhiqiu.github.io/papers/naturalbench/

A dedicated gamma-ray burst (GRB) afterglow observing program was performed between 2007 and 2016 with GROND, a seven-channel optical and near-infrared imager at the 2.2m telescope of the Max-Planck Society at ESO/La Silla. In this first of a series of papers, we describe the GRB observing plan, providing first readings of all so far unpublished GRB afterglow measurements and some observing statistics. In total, we observed 514 GRBs with GROND, including 434 Swift-detected GRBs, representing 81\% of the observable Swift sample. For GROND-observations within 30 min of the GRB trigger, the optical/NIR afterglow detection rate is 81\% for long- and 57\% for short-duration GRBs. We report the discovery of ten new GRB afterglows plus one candidate, along with redshift estimates (partly improved) for four GRBs and new host detections for seven GRBs. We identify the (already known) afterglow of GRB 140209A as the sixth GRB exhibiting a 2175 Angstroem dust feature. As a side result, we identified two blazars, with one at a redshift of z=3.8 (in the GRB 131209A field)., Comment: 20 pages, 31 figures, accepted for publication in Astronomy & Astrophysics

Text-rich visual understanding-the ability to process environments where dense textual content is integrated with visuals-is crucial for multimodal large language models (MLLMs) to interact effectively with structured environments. To enhance this capability, we propose synthesizing general multimodal instructions from webpage UIs using text-based large language models (LLMs). Despite lacking direct visual input, text-based LLMs are able to process structured text representations from webpage accessibility trees. These instructions are then paired with UI screenshots to train multimodal models. We introduce MultiUI, a dataset containing 7.3 million samples from 1 million websites, covering diverse multimodal tasks and UI layouts. Models trained on MultiUI not only excel in web UI tasks-achieving up to a 48% improvement on VisualWebBench and a 19.1% boost in element accuracy on a web agent dataset Mind2Web-but also generalize surprisingly well to non-web UI tasks and even to non-UI domains, such as document understanding, OCR, and chart interpretation. These results highlight the broad applicability of web UI data for advancing text-rich visual understanding across various scenarios.

Reduced transmon qubit $T_1$ coherence times have been linked to the amorphous oxide layers formed by thin film capacitors during processing. Because Ta or Ta capped Nb capacitors exhibit overall superior qubit performance to those fabricated with Nb capacitors, it has been hypothesized that the amorphous, non-stoichiometric Ta$_2$O$_{5-x}$ oxide is less lossy than its Nb$_2$O$_{5-x}$ counterpart. The origins of what makes amorphous Ta$_2$O$_{5-x}$ less susceptible to accepted decoherence channels is unknown. Here we establish the microscopic features of amorphous Nb$_2$O$_{5-x}$ and Ta$_2$O$_{5-x}$ using a combination of \textit{ab initio} molecular dynamics and density functional theory calculations. Our simulations establish that oxygen deficiency is less likely to occur in amorphous Ta$_2$O$_{5-x}$ than in Nb$_2$O$_{5-x}$ for $0\le x \le 0.25$ and that at a given level of oxygen deficiency the formation of metal Ta-Ta bonds is enhanced. These bonds, which are accommodated by structural flaws in the amorphous network, capture electrons better than in amorphous Nb$_2$O$_{5-x}$. These thermochemical differences quench or highly suppress magnetic moments in amorphous Ta$_2$O$_{5-x}$ and eliminate a potential source of quasiparticles and magnetic flux noise. Our calculations also show that hyperfine couplings between Nb nuclei and local magnetic moments in Nb$_2$O$_{5-x}$ could form "two-level systems" (TLS) or "two-level fluctuators" (TLF) with energy splittings of 100-1000 MHz or higher. This reveals a new TLS mechanism in amorphous Nb$_2$O$_{5-x}$ oxide layers that is unlikely in Ta$_2$O$_{5-x}$. Our work provides fundamental understanding of the materials chemistry and limitations imposed by native oxides of superconducting qubits, which can be used to guide materials selection and processing., Comment: 8 pages, 4 figures

We carried out a comprehensive crystal structure characterization of Ti-doped lithium ruthenate (Li$_2$Ti$_x$Ru$_{1-x}$O$_3$), to investigate the effect of Ti-doping on the structural phase transition. Experimental tools sensitive to the average structure (X-ray diffraction), as well as those sensitive to local structure (Extended X-ray Absorption Fine Structure, EXAFS; pair distribution function, PDF) are used. We observed non-monotonic dependence of the structural transition temperature on the Ti-doping level. At low doping, the transition temperature slightly increases with doping, while at high doping, the temperature decreases significantly with doping. We note two important observations from our studies. First, Ti K-edge EXAFS data shows persistent Ti-Ru dimerization even with substantial Ti doping. Second, we were able to use the PDF data to estimate the dimer correlation length above the transition temperature, which would correspond to the size of the proposed local `dimer clusters' formed by Ru-Ru and Ti-Ru neighbours. The dimer correlation length is found to be around 10~\AA, which remains robust regardless of doping. Our study therefore suggests that Ti$^{4+}$ with its $d^0$ electronic configuration is a special type of dopant when replacing Ru., Comment: 12 pages, 11 figures

We describe the structure of the singular sets of constant curvature, convex hypersurfaces in hyperbolic space for general convex curvature functions. We apply this result to the study of the ideal Plateau problem in hyperbolic space for such curvature functions., Comment: A brief review, by invitation, for a forthcoming volume in honour of Marcos Dajczer. 16 pages, 6 figures

The magnetar SGR 1935+2154 is the only known Galactic source of fast radio bursts (FRBs). FRBs from SGR 1935+2154 were first detected by CHIME/FRB and STARE2 in 2020 April, after the conclusion of the LIGO, Virgo, and KAGRA Collaborations' O3 observing run. Here we analyze four periods of gravitational wave (GW) data from the GEO600 detector coincident with four periods of FRB activity detected by CHIME/FRB, as well as X-ray glitches and X-ray bursts detected by NICER and NuSTAR close to the time of one of the FRBs. We do not detect any significant GW emission from any of the events. Instead, using a short-duration GW search (for bursts $\leq$ 1 s) we derive 50\% (90\%) upper limits of $10^{48}$ ($10^{49}$) erg for GWs at 300 Hz and $10^{49}$ ($10^{50}$) erg at 2 kHz, and constrain the GW-to-radio energy ratio to $\leq 10^{14} - 10^{16}$. We also derive upper limits from a long-duration search for bursts with durations between 1 and 10 s. These represent the strictest upper limits on concurrent GW emission from FRBs., Comment: 15 pages of text including references, 4 figures, 5 tables

Unique continuation principles are fundamental properties of elliptic partial differential equations, giving conditions that guarantee that the solution to an elliptic equation must be uniformly zero. Since finite-element discretizations are a natural tool to help gain understanding into elliptic equations, it is natural to ask if such principles also hold at the discrete level. In this work, we prove a version of the unique continuation principle for piecewise-linear and -bilinear finite-element discretizations of the Laplacian eigenvalue problem on polygonal domains in $\mathbb{R}^2$. Namely, we show that any solution to the discretized equation $-\Delta u = \lambda u$ with vanishing Dirichlet and Neumann traces must be identically zero under certain geometric and topological assumptions on the resulting triangulation. We also provide a counterexample, showing that a nonzero \emph{inner solution} exists when the topological assumptions are not satisfied. Finally, we give an application to an eigenvalue interlacing problem, where the space of inner solutions makes an explicit appearance.

Modern computing systems are capable of exascale calculations, which are revolutionizing the development and application of high-fidelity numerical models in computational science and engineering. While these systems continue to grow in processing power, the available system memory has not increased commensurately, and electrical power consumption continues to grow. A predominant approach to limit the memory usage in large-scale applications is to exploit the abundant processing power and continually recompute many low-level simulation quantities, rather than storing them. However, this approach can adversely impact the throughput of the simulation and diminish the benefits of modern computing architectures. We present two novel contributions to reduce the memory burden while maintaining performance in simulations based on finite element discretizations. The first contribution develops dictionary-based data compression schemes that detect and exploit the structure of the discretization, due to redundancies across the finite element mesh. These schemes are shown to reduce memory requirements by more than 99 percent on meshes with large numbers of nearly identical mesh cells. For applications where this structure does not exist, our second contribution leverages a recently developed augmented Lagrangian sequential quadratic programming algorithm to enable r-adaptive mesh optimization, with the goal of enhancing redundancies in the mesh. Numerical results demonstrate the effectiveness of the proposed methods to detect, exploit and enhance mesh structure on examples inspired by large-scale applications., Comment: 22 pages

Between 2017 and 2024, the Helioseismic and Magnetic Imager (HMI) onboard the Solar Dynamics Observatory has observed several white-light solar flares. Notably, for the X9.3 flare of September 6, 2017, HMI spectro-polarimetric observations reveal one or more locations within the umbra of the associated active region where the Fe I 6173 {\AA} line goes into full emission, indicating significant heating of the photosphere and lower chromosphere. For these flares, we performed a spectro-polarimetric analysis at the aforementioned locations using HMI 90s cadence Stokes data. At the Fe I emission locations, line-core emission is observed to last for a single 90 s frame and is either concurrent with or followed by increases in the line continuum intensity lasting 90 to 180 seconds. This is followed by a smooth decay to pre-flare conditions over the next three to twenty minutes. For most locations, permanent changes to the Stokes Q, U, and/or V profiles were observed, indicating long-lasting non-transient changes to the photospheric magnetic field. These emissions coincided with local maxima in hard X-ray emission observed by the Konus instrument onboard the Wind spacecraft, as well as local maxima in the time derivative of soft X-ray emission observed by GOES satellites. Comparison of the Fe I 6173 {\AA} line profile synthesis for the ad-hoc heating of the initial empirical VAL-S umbra model and quiescent Sun (VAL-C-like) model indicates that the Fe I 6173 {\AA} line emission in the white-light flare kernels could be explained by the strong heating of initially cool photospheric regions.

Dark matter's existence is known thanks to its gravitational interaction with Standard Model particles, but it remains unknown whether this is the only force present between them. While many searches for such new interactions with dark matter focus on short-range, contact-like interactions, it is also possible that there exist weak, long-ranged forces between dark matter and the Standard Model. In this work, we present two types of constraints on such new interactions. First, we consider constraints arising from the fact that such a force would also induce long range interactions between Standard Model particles themselves, as well as between dark matter particles themselves. Combining the constraints on these individual forces generally sets the strongest constraints available on new Standard Model-dark matter interactions. Second, we consider the possibility of constraining new long-ranged interactions between dark matter and the Standard Model using the effects of dynamical friction in ultrafaint dwarf galaxies, especially Segue I. Such new interactions would accelerate the transfer of kinetic energy from stars to their surrounding dark matter, slowly reducing their orbits; the present-day stellar half-light radius of Segue I therefore allows us to exclude new forces which would have reduced stars' orbital radii below this scale by now., Comment: 23 pages, 10 figures

As large language models (LLMs) become increasingly versatile, numerous large scale benchmarks have been developed to thoroughly assess their capabilities. These benchmarks typically consist of diverse datasets and prompts to evaluate different aspects of LLM performance. However, comprehensive evaluations on hundreds or thousands of prompts incur tremendous costs in terms of computation, money, and time. In this work, we investigate strategies to improve evaluation efficiency by selecting a subset of examples from each benchmark using a learned policy. Our approach models the dependencies across test examples, allowing accurate prediction of the evaluation outcomes for the remaining examples based on the outcomes of the selected ones. Consequently, we only need to acquire the actual evaluation outcomes for the selected subset. We rigorously explore various subset selection policies and introduce a novel RL-based policy that leverages the captured dependencies. Empirical results demonstrate that our approach significantly reduces the number of evaluation prompts required while maintaining accurate performance estimates compared to previous methods.

We demonstrate that when people use large language models (LLMs) to generate recommendations, the LLMs produce responses that reflect both what the user wants and who the user is. While personalized recommendations are often desired by users, it can be difficult in practice to distinguish cases of bias from cases of personalization: we find that models generate racially stereotypical recommendations regardless of whether the user revealed their identity intentionally through explicit indications or unintentionally through implicit cues. We argue that chatbots ought to transparently indicate when recommendations are influenced by a user's revealed identity characteristics, but observe that they currently fail to do so. Our experiments show that even though a user's revealed identity significantly influences model recommendations (p < 0.001), model responses obfuscate this fact in response to user queries. This bias and lack of transparency occurs consistently across multiple popular consumer LLMs (gpt-4o-mini, gpt-4-turbo, llama-3-70B, and claude-3.5) and for four American racial groups.

We propose using highly excited cyclotron states of a trapped electron to detect meV axion and dark photon dark matter, marking a significant improvement over our previous proposal and demonstration [Phys. Rev. Lett. 129, 261801]. When the axion mass matches the cyclotron frequency $\omega_c$, the cyclotron state is resonantly excited, with a transition probability proportional to its initial quantum number, $n_c$. The sensitivity is enhanced by taking $n_c \sim 10^6 \left( \frac{0.1~\text{meV}}{\omega_c} \right)^2$. By optimizing key experimental parameters, we minimize the required averaging time for cyclotron detection to $t_{\text{ave}} \sim 10^{-6} $ seconds, permitting detection of such a highly excited state before its decay. An open-endcap trap design enables the external photon signal to be directed into the trap, rendering our background-free detector compatible with large focusing cavities, such as the BREAD proposal, while capitalizing on their strong magnetic fields. Furthermore, the axion conversion rate can be coherently enhanced by incorporating layers of dielectrics with alternating refractive indices within the cavity. Collectively, these optimizations enable us to probe the QCD axion parameter space from 0.1 meV to 2.3 meV (25-560 GHz), covering a substantial portion of the predicted post-inflationary QCD axion mass range. This sensitivity corresponds to probing the kinetic mixing parameter of the dark photon down to $\epsilon \approx 2 \times 10^{-16}$., Comment: 47 pages, 8 figures

Over the past several decades, time-series photometry of CSPNe has yielded significant results including, but not limited to, discoveries of nearly 100 binary systems, insights into pulsations and winds in young white dwarfs, and studies of stars undergoing very late thermal pulses. We have undertaken a systematic study of optical photometric variability of cataloged CSPNe, using the epochal photometric data from the Zwicky Transient Facility (ZTF). By applying appropriate variability metrics, we arrive at a list of 94 significantly variable CSPNe. Based on the timescales of the light-curve activity, we classify the variables broadly into short- and long-timescale variables. In this first paper in this series, we focus on the former, which is the majority class comprising 83 objects. We infer periods for six sources for the first time, and recover several known periodic variables. Among the aperiodic sources, most exhibit a jitter around a median flux with a stable amplitude, and a few show outbursts. We draw attention to WeSb 1, which shows a different kind of variability: prominent deep and aperiodic dips, resembling transits from a dust/debris disk. We find strong evidence for a binary nature of WeSb 1 (possibly an A- to G-type companion). The compactness of the emission lines and inferred high electron densities make WeSb 1 a candidate for either an EGB 6-type planetary nucleus, or a symbiotic system inside an evolved planetary nebula, both of which are rare objects. To demonstrate further promise with ZTF, we report three additional newly identified periodic sources that do not appear in the list of highly variable sources. Finally, we also introduce a two-dimensional metric space defined by the von Neumann statistics and Pearson Skew and demonstrate its effectiveness in identifying unique variables of astrophysical interest, like WeSb 1., Comment: 19 pages + 8 pages appendix, 5 tables, 17 figures; Submitted to PASP; Comments are welcome!

Elastoinertial turbulence (EIT) is a chaotic state that emerges in the flows of dilute polymer solutions. Direct numerical simulation (DNS) of EIT is highly computationally expensive due to the need to resolve the multi-scale nature of the system. While DNS of 2D EIT typically requires $O(10^6)$ degrees of freedom, we demonstrate here that a data-driven modeling framework allows for the construction of an accurate model with 50 degrees of freedom. We achieve a low-dimensional representation of the full state by first applying a viscoelastic variant of proper orthogonal decomposition to DNS results, and then using an autoencoder. The dynamics of this low-dimensional representation are learned using the neural ODE method, which approximates the vector field for the reduced dynamics as a neural network. The resulting low-dimensional data-driven model effectively captures short-time dynamics over the span of one correlation time, as well as long-time dynamics, particularly the self-similar, nested traveling wave structure of 2D EIT in the parameter range considered.

General-purpose LLM judges capable of human-level evaluation provide not only a scalable and accurate way of evaluating instruction-following LLMs but also new avenues for supervising and improving their performance. One promising way of leveraging LLM judges for supervision is through Minimum Bayes Risk (MBR) decoding, which uses a reference-based evaluator to select a high-quality output from amongst a set of candidate outputs. In the first part of this work, we explore using MBR decoding as a method for improving the test-time performance of instruction-following LLMs. We find that MBR decoding with reference-based LLM judges substantially improves over greedy decoding, best-of-N decoding with reference-free judges and MBR decoding with lexical and embedding-based metrics on AlpacaEval and MT-Bench. These gains are consistent across LLMs with up to 70B parameters, demonstrating that smaller LLM judges can be used to supervise much larger LLMs. Then, seeking to retain the improvements from MBR decoding while mitigating additional test-time costs, we explore iterative self-training on MBR-decoded outputs. We find that self-training using Direct Preference Optimisation leads to significant performance gains, such that the self-trained models with greedy decoding generally match and sometimes exceed the performance of their base models with MBR decoding.

Complex chaotic dynamics, seen in natural and industrial systems like turbulent flows and weather patterns, often span vast spatial domains with interactions across scales. Accurately capturing these features requires a high-dimensional state space to resolve all the time and spatial scales. For dissipative systems the dynamics lie on a finite-dimensional manifold with fewer degrees of freedom. Thus, by building reduced-order data-driven models in manifold coordinates, we can capture the essential behavior of chaotic systems. Unfortunately, these tend to be formulated globally rendering them less effective for large spatial systems. In this context, we present a data-driven low-dimensional modeling approach to tackle the complexities of chaotic motion, Markovian dynamics, multi-scale behavior, and high numbers of degrees of freedom within large spatial domains. Our methodology involves a parallel scheme of decomposing a spatially extended system into a sequence of local `patches', and constructing a set of coupled, local low-dimensional dynamical models for each patch. Here, we choose to construct the set of local models using autoencoders (for constructing the low-dimensional representation) and neural ordinary differential equations, NODE, for learning the evolution equation. Each patch, or local model, shares the same underlying functions (e.g., autoencoders and NODEs) due to the spatial homogeneity of the underlying systems we consider. We apply this method to the Kuramoto-Sivashinsky equation and 2D Kolmogorov flow, and reduce state dimension by up to two orders of magnitude while accurately capturing both short-term dynamics and long-term statistics.

While experiments and simulations have provided a rich picture of the dynamic heterogeneity in glasses at constant temperature or under steady shear, the dynamics of glasses under oscillatory shear remain comparatively less explored. Recent work has shown that oscillatory shear protocols can embed a ``memory'' into a glass's structure, whereby the material will exhibit dynamics that are encoded by the oscillatory shear protocol applied. However, most of the computational work studying the memory effect has been performed in the zero temperature limit, and the effects of thermalization are poorly characterized. In this work, we use nonequilibrium molecular dynamics simulations to study the dynamics of a model two-dimensional glass former at low, non-zero temperatures under oscillatory shear. While we show that the systems' dynamics are independent of sample preparation for either small or larger strain amplitudes, the dynamics become distinct near the yield point when the deformation is applied at finite temperature. We then characterize the dynamic heterogeneity using two metrics, one derived from the vibrational modes and one that exploits machine learning to identify regions prone to rearrangement. This analysis provides evidence that the dynamics below and above yield emerge from distinct structural origins that may be important for developing improved constitutive models that can predict memory in disordered solids.

Accurate diagnosis of ocular surface diseases is critical in optometry and ophthalmology, which hinge on integrating clinical data sources (e.g., meibography imaging and clinical metadata). Traditional human assessments lack precision in quantifying clinical observations, while current machine-based methods often treat diagnoses as multi-class classification problems, limiting the diagnoses to a predefined closed-set of curated answers without reasoning the clinical relevance of each variable to the diagnosis. To tackle these challenges, we introduce an innovative multi-modal diagnostic pipeline (MDPipe) by employing large language models (LLMs) for ocular surface disease diagnosis. We first employ a visual translator to interpret meibography images by converting them into quantifiable morphology data, facilitating their integration with clinical metadata and enabling the communication of nuanced medical insight to LLMs. To further advance this communication, we introduce a LLM-based summarizer to contextualize the insight from the combined morphology and clinical metadata, and generate clinical report summaries. Finally, we refine the LLMs' reasoning ability with domain-specific insight from real-life clinician diagnoses. Our evaluation across diverse ocular surface disease diagnosis benchmarks demonstrates that MDPipe outperforms existing standards, including GPT-4, and provides clinically sound rationales for diagnoses., Comment: Accepted to MICCAI 2024. Project Webpage: https://danielchyeh.github.io/MDPipe/

Large Language Models (LLMs) have shown remarkable ability to generalize effectively across numerous industry domains while executing a range of tasks. Many of these competencies are obtained from the data utilized during the pre-training phase of the Language Models (LMs). However, these models exhibit limitations when tasked with performing in specialized or low-resource industry domains. More recent approaches use LLMs for generating domain-specific synthetic data but most often they lack in truthfulness and complexity. Alternatively, in cases where domain data is available like healthcare and finance most of the LMs are proprietary necessitating the need for a scalable method to curate real world industry specific pre-training data. In this work, we propose an automated and scalable framework - DoPAMine:Domain-specific Pre-training Adaptation from seed-guided data Mining, to mine domain specific training data from a large data corpus for domain adaptation of a LM. The framework leverages the parametric knowledge of a LLM to generate diverse and representative seed data tailored to a specific domain which is then used to mine real world data from a large data corpus like Common Crawl. We evaluated our framework's performance in the continual pre-training (CPT) setting by training two domain specific 7B parameter LMs in healthcare and finance with data mined via DoPAMine. Our experiments show that DoPAMine boosts the performance of pre-trained LLMs on average by 4.9% and 5.1% in zero-shot and 5-shot settings respectively on healthcare tasks from MMLU, MedQA, MedMCQA and PubMedQA datasets, and 2.9% and 6.7% for zero-shot and 5-shot settings respectively on finance tasks from FiQA-SA, FPB and Headlines datasets when compared to the baseline.

Large Language Models have advanced clinical Natural Language Generation, creating opportunities to manage the volume of medical text. However, the high-stakes nature of medicine requires reliable evaluation, which remains a challenge. In this narrative review, we assess the current evaluation state for clinical summarization tasks and propose future directions to address the resource constraints of expert human evaluation.

This paper presents the development of a new entropy-based feature selection method for identifying and quantifying impacts. Here, impacts are defined as statistically significant differences in spatio-temporal fields when comparing datasets with and without an external forcing in Earth system models. Temporal feature selection is performed by first computing the cross-fuzzy entropy to quantify similarity of patterns between two datasets and then applying changepoint detection to identify regions of statistically constant entropy. The method is used to capture temperate north surface cooling from a 9-member simulation ensemble of the Mt. Pinatubo volcanic eruption, which injected 10 Tg of SO2 into the stratosphere. The results estimate a mean difference decrease in near surface air temperature of -0.560 K with a 99% confidence interval between -0.864 K and -0.257 K between April and November of 1992, one year following the eruption. A sensitivity analysis with decreasing SO2 injection revealed that the impact is statistically significant at 5 Tg but not at 3 Tg. Using identified features, a dependency graph model composed of 68 nodes and 229 edges directly connecting initial aerosol optical depth changes in the tropics to solar flux and temperature changes before the temperate north surface cooling is presented.

LLMs can now act as autonomous agents that interact with digital environments and complete specific objectives (e.g., arranging an online meeting). However, accuracy is still far from satisfactory, partly due to a lack of large-scale, direct demonstrations for digital tasks. Obtaining supervised data from humans is costly, and automatic data collection through exploration or reinforcement learning relies on complex environmental and content setup, resulting in datasets that lack comprehensive coverage of various scenarios. On the other hand, there is abundant knowledge that may indirectly assist task completion, such as online tutorials that were created for human consumption. In this work, we present Synatra, an approach that effectively transforms this indirect knowledge into direct supervision at scale. We define different types of indirect knowledge, and carefully study the available sources to obtain it, methods to encode the structure of direct demonstrations, and finally methods to transform indirect knowledge into direct demonstrations. We use 100k such synthetically-created demonstrations to finetune a 7B CodeLlama, and demonstrate that the resulting agent surpasses all comparably sized models on three web-based task benchmarks Mind2Web, MiniWoB++ and WebArena, as well as surpassing GPT-3.5 on WebArena and Mind2Web. In addition, while synthetic demonstrations prove to be only 3% the cost of human demonstrations (at $0.031 each), we show that the synthetic demonstrations can be more effective than an identical number of human demonstrations collected from limited domains.

Completions of metric spaces are usually constructed using Cauchy sequences. However, this does not work for general uniform spaces, where Cauchy filters or nets must be used instead. The situation in pointfree topology is more straightforward: the correct completion of uniform locales can indeed be obtained as a quotient of a locale of Cauchy sequences., Comment: 17 pages

Bulk AlN possesses high thermal conductivity due to long phonon mean-free-paths, high group velocity, and long lifetimes. However, the thermal transport scenario becomes very different in a thin AlN film due to phonon-defect and phonon-boundary scattering. Herein, we report experimental observation of ballistic to diffusive transition in a series of AlN thin films (1.6 - 2440 nm) grown on sapphire substrates. The ballistic transport is characterized by constant thermal resistance as a function of film thickness due to phonon scattering by defects and boundaries. In this transport regime, phonons possess very small group velocities and lifetimes. The lifetime of the optical phonons increases by more than an order of magnitude in the diffusive regime, however, remains nearly constant afterwards. Our study is important for understanding the details of nano and microscale thermal transport in a highly conductive material.

Neural operators (NOs) employ deep neural networks to learn mappings between infinite-dimensional function spaces. Deep operator network (DeepONet), a popular NO architecture, has demonstrated success in the real-time prediction of complex dynamics across various scientific and engineering applications. In this work, we introduce a random sampling technique to be adopted during the training of DeepONet, aimed at improving the generalization ability of the model, while significantly reducing the computational time. The proposed approach targets the trunk network of the DeepONet model that outputs the basis functions corresponding to the spatiotemporal locations of the bounded domain on which the physical system is defined. Traditionally, while constructing the loss function, DeepONet training considers a uniform grid of spatiotemporal points at which all the output functions are evaluated for each iteration. This approach leads to a larger batch size, resulting in poor generalization and increased memory demands, due to the limitations of the stochastic gradient descent (SGD) optimizer. The proposed random sampling over the inputs of the trunk net mitigates these challenges, improving generalization and reducing memory requirements during training, resulting in significant computational gains. We validate our hypothesis through three benchmark examples, demonstrating substantial reductions in training time while achieving comparable or lower overall test errors relative to the traditional training approach. Our results indicate that incorporating randomization in the trunk network inputs during training enhances the efficiency and robustness of DeepONet, offering a promising avenue for improving the framework's performance in modeling complex physical systems.

We present Agglomerative Token Clustering (ATC), a novel token merging method that consistently outperforms previous token merging and pruning methods across image classification, image synthesis, and object detection & segmentation tasks. ATC merges clusters through bottom-up hierarchical clustering, without the introduction of extra learnable parameters. We find that ATC achieves state-of-the-art performance across all tasks, and can even perform on par with prior state-of-the-art when applied off-the-shelf, i.e. without fine-tuning. ATC is particularly effective when applied with low keep rates, where only a small fraction of tokens are kept and retaining task performance is especially difficult., Comment: ECCV 2024. Project webpage at https://vap.aau.dk/atc/