Search

Recent advances in AI-generated steganography highlight its potential for safeguarding the privacy of vulnerable democratic actors, including aid workers, journalists, and whistleblowers operating in oppressive regimes. In this work, we address current limitations and establish the foundations for large-throughput generative steganography. We introduce a novel approach that enables secure and efficient steganography within latent diffusion models. We show empirically that our methods perform well across a variety of open-source latent diffusion models, particularly in generative image and video tasks.

We present follow-up 21cm HI line observations made with the Nan\c{c}ay Radio Telescope (NRT) of 99 weak or potential detections of galaxies from the EZOA catalogue in the northern Zone of Avoidance (ZoA), which were extracted from the shallow version of the EBHIS blind HI survey performed with the Effelsberg radio telescope. The new NRT observations are on average almost three times as sensitive as those from EBHIS. Of the 99 observed sources, we confirmed 72, while three yielded inconclusive results. We find that the quality assessment of the EZOA catalogue entries correlates well with the NRT recovery rate; for instance, only four of the 22 potential detections could be confirmed. Due to the higher sensitivity as well as the large north-south extent of the NRT beam, the NRT observations also yielded five serendipitous detections, which we include here. We updated the EZOA catalogue with the improved HI parameters and detections. To test mitigation of Radio Frequency Interference signals, we also observed selected sources using a dedicated receiver and data processing system., Comment: 24 pages, 6 figures; accepted by A&A; includes all tables (electronic versions are available through CDS or by request from the main author)

In the IceCube Neutrino Observatory, a signal of astrophysical neutrinos is obscured by backgrounds from atmospheric neutrinos and muons produced in cosmic-ray interactions. IceCube event selections used to isolate the astrophysical neutrino signal often focus on t/he morphology of the light patterns recorded by the detector. The analyses presented here use the new IceCube Enhanced Starting Track Event Selection (ESTES), which identifies events likely generated by muon neutrino interactions within the detector geometry, focusing on neutrino energies of 1-500 TeV with a median angular resolution of 1.4{\deg}. Selecting for starting track events filters out not only the atmospheric-muon background, but also the atmospheric-neutrino background in the southern sky. This improves IceCube's muon neutrino sensitivity to southern-sky neutrino sources, especially for Galactic sources that are not expected to produce a substantial flux of neutrinos above 100 TeV. In this work, the ESTES sample was applied for the first time to searches for astrophysical sources of neutrinos, including a search for diffuse neutrino emission from the Galactic plane. No significant excesses were identified from any of the analyses; however, constraining limits are set on the hadronic emission from TeV gamma-ray Galactic plane objects and models of the diffuse Galactic plane neutrino flux., Comment: 23 pages, 8 figures, 4 tables. Submitted to ApJ

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 3,000 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai., Comment: 25 pages, 6 figures

Large language models (LLMs) have been positioned to revolutionize HCI, by reshaping not only the interfaces, design patterns, and sociotechnical systems that we study, but also the research practices we use. To-date, however, there has been little understanding of LLMs' uptake in HCI. We address this gap via a systematic literature review of 153 CHI papers from 2020-24 that engage with LLMs. We taxonomize: (1) domains where LLMs are applied; (2) roles of LLMs in HCI projects; (3) contribution types; and (4) acknowledged limitations and risks. We find LLM work in 10 diverse domains, primarily via empirical and artifact contributions. Authors use LLMs in five distinct roles, including as research tools or simulated users. Still, authors often raise validity and reproducibility concerns, and overwhelmingly study closed models. We outline opportunities to improve HCI research with and on LLMs, and provide guiding questions for researchers to consider the validity and appropriateness of LLM-related work., Comment: This is a preprint version of the paper conditionally accepted to CHI'25

Systematic reviews are time-consuming endeavors. Historically speaking, knowledgeable humans have had to screen and extract data from studies before it can be analyzed. However, large language models (LLMs) hold promise to greatly accelerate this process. After a pilot study which showed great promise, we investigated the use of freely available LLMs for extracting data for systematic reviews. Using three different LLMs, we extracted 24 types of data, 9 explicitly stated variables and 15 derived categorical variables, from 112 studies that were included in a published scoping review. Overall we found that Gemini 1.5 Flash, Gemini 1.5 Pro, and Mistral Large 2 performed reasonably well, with 71.17%, 72.14%, and 62.43% of data extracted being consistent with human coding, respectively. While promising, these results highlight the dire need for a human-in-the-loop (HIL) process for AI-assisted data extraction. As a result, we present a free, open-source program we developed (AIDE) to facilitate user-friendly, HIL data extraction with LLMs.

The advent of high-average-power, ultrafast ytterbium-based lasers allows us to generate laser filaments at repetition rates ranging from 10s of kHz up to 100s of kHz. At such high repetition rates, the inter-pulse time lies below the time required for the total diffusion of the deposited heat by each laser pulse, leading to cumulative hydrodynamic effects that have so far been rarely studied. Here, we present, to the best of our knowledge, the first experimental time-resolved measurements of these dynamics in air for laser repetition rates between 1 kHz and 100 kHz. We measure the change in the air refractive index caused by the localized heat deposition and the length of the filament-generated plasma channel, with which we can infer the corresponding change in air density. We observe that at repetition rates above 10 kHz, stationary density depletions with vanishing dynamics emerge. Our findings are of wide relevance for the fields of high-repetition-rate laser filamentation and its applications, as well as THz generation from laser-induced plasma sources., Comment: 8 pages, 4 figurees

We report a search for high-energy astrophysical neutrino multiplets, detections of multiple neutrino clusters in the same direction within 30 days, based on an analysis of 11.4 years of IceCube data. A new search method optimized for transient neutrino emission with a monthly time scale is employed, providing a higher sensitivity to neutrino fluxes. This result is sensitive to neutrino transient emission, reaching per-flavor flux of approximately $10^{-10}\ {\rm erg}\ {\rm cm}^{-2}\ {\rm sec}^{-1}$ from the Northern sky in the energy range $E\gtrsim 50$~TeV. The number of doublets and triplets identified in this search is compatible with the atmospheric background hypothesis, which leads us to set limits on the nature of neutrino transient sources with emission timescales of one month.

In regimes of low signal strengths and therefore a small signal-to-noise ratio, standard data analysis methods often fail to accurately estimate system properties. We present a method based on Monte Carlo simulations to effectively restore robust parameter estimates from large sets of undersampled data. This approach is illustrated through the analysis of photoluminescence excitation spectroscopy data for optical linewidth characterization of a nitrogen-vacancy color center in diamond. We evaluate the quality of parameter prediction using standard statistical data analysis methods, such as the median, and the Monte Carlo method. Depending on the signal strength, we find that the median can be precise (narrow confidence intervals) but very inaccurate. A detailed analysis across a broad range of parameters allows to identify the experimental conditions under which the median provides a reliable predictor of the quantum emitter's linewidth. We also explore machine learning to perform the same task, forming a promising addition to the parameter estimation toolkit. Finally, the developed method offers a broadly applicable tool for accurate parameter prediction from low signal data, opening new experimental regimes previously deemed inaccessible., Comment: Main part: 7 pages incl. references, 5 figures

The restaurant industry is currently facing a challenging socio-economic situation caused by the rise of delivery services, inflation, and typically low margins. Often, technological opportunities for process optimization or customer retention are not fully utilized. In our design case study, we investigate which technologies are already being used to improve the customer experience in restaurants and explore a novel new approach to this issue. We designed, implemented, and evaluated a platform with customers and restaurateurs to increase visibility and emotional connection to nearby restaurants through their dishes. Some of our key findings include the enormous potential of combining location-based systems and conversational agents, but also the difficulties in creating content for such platforms. We contribute to the field of Human-Food Interaction by (1) identifying promising design spaces as well as customer and restaurateur requirements for technology in this domain, (2) presenting an innovative design case study to improve the user experience, and (3) exploring the broader implications of our design case study findings for approaching a real-world metaverse., Comment: 32 pages, 8 figures

Cosmic strings represent an attractive source of gravitational waves (GWs) from the early Universe. However, in contrast to other primordial GW sources whose signal may be modeled by simple analytic expressions, computing the GW signal from cosmic strings requires the numerical evaluation of complicated integral and sum expressions, which can become computationally costly in large parameter scans. This shortcoming motivates us to rederive the GW signal from a network of local stable cosmic strings in the Nambu-Goto approximation and based on the velocity-dependent one-scale model from a "pedestrian" perspective. That is, we derive purely analytical expressions for the total GW spectrum, which remain exact wherever possible and whose error can be tracked and reduced in a controlled way in crucial situations in which we are forced to introduce approximations. In this way, we obtain powerful formulas that, unlike existing results in the literature, are valid across the entire frequency spectrum and across the entire conceivable range of cosmic-string tensions. We provide an in-depth discussion of the GW spectra thus obtained, including their characteristic break frequencies and approximate power-law behaviors, comment on the effect of changes in the effective number of degrees of freedom during radiation domination, and conclude with a concise summary of our main formulas that can readily be used in future studies., Comment: 55 pages, 15 figures, summary box at the end of the paper

A presented analytical formulation of (optical)phonon-mediated and momentum-resonant Franck-Condon emission of photoexcited electrons from polar semiconductors is shown to be very consistent with (i) the observed emission properties of a Cesiated GaAs(001) photocathode at 808nm [J. Phys. D: Appl. Phys. 54, 205301 (2021)] and (ii) the measured spectral emission properties of a GaN(0001) photocathode from just below its bandgap energy to 5eV. The theoretical analysis in the parabolic band approximation predicts the form of both the quantum efficiency and mean transverse energy of photoemission as a function of the photocathode's electron affinity and the electron temperature in the vicinity of its emission face. The good agreement between theory and experimental data also suggests that sub-10nm rms surface roughness effects are not significant for polar semiconductor photocathodes., Comment: 14 pages including supplemental material, 3 figures, submitted to Physical Review X

Deep Reinforcement learning has shown to be a powerful tool for developing policies in environments where an optimal solution is unclear. In this paper, we attempt to apply Twin Delayed Deep Deterministic Policy Gradients to train a neural network to act as a velocity controller for a quadcopter. The quadcopter's objective is to quickly fly through a gate while avoiding crashing into the gate. We transfer our trained policy to the real world by deploying it on a quadcopter in a laboratory environment. Finally, we demonstrate that the trained policy is able to navigate the drone to the gate in the real world.

As LLM-based applications reach millions of customers, ensuring their scalability and continuous quality improvement is critical for success. However, the current workflows for developing, maintaining, and operating (DevOps) these applications are predominantly manual, slow, and based on trial-and-error. With this paper we introduce the Generative AI Toolkit, which automates essential workflows over the whole life cycle of LLM-based applications. The toolkit helps to configure, test, continuously monitor and optimize Generative AI applications such as agents, thus significantly improving quality while shortening release cycles. We showcase the effectiveness of our toolkit on representative use cases, share best practices, and outline future enhancements. Since we are convinced that our Generative AI Toolkit is helpful for other teams, we are open sourcing it on and hope that others will use, forward, adapt and improve, Comment: 16 pages, 6 figures. For source code see https://github.com/awslabs/generative-ai-toolkit

We employ polarization tomography to characterize the modal properties of a dielectric nanocavity with sub-wavelength mode confinement. Our analysis of reflection spectra shows that the Fano-lineshape depends strongly on the polarization in a confocal configuration, and that the lineshape can be transformed into a Lorentzian-like peak for a certain polarization. For this polarization setting, the background is almost fully suppressed in a finite range of frequencies. This enables us to identify another resonance that has not yet been experimentally reported for these nanocavities. Lastly, we use symmetry-forbidden polarizations and show that, surprisingly, the modal resonance features of the system remain visible.

Large Language Models (LLMs) have become increasingly powerful and ubiquitous, but their stochastic nature poses challenges to the reliability of their outputs. While deterministic settings can improve consistency, they do not guarantee reliability, as a single sample from the model's probability distribution can still be misleading. Building upon the concept of LLM-as-a-judge, we introduce a novel framework for rigorously evaluating the reliability of LLM judgments, leveraging McDonald's omega. We evaluate the reliability of LLMs when judging the outputs of other LLMs on standard single-turn and multi-turn benchmarks, simultaneously investigating the impact of temperature on reliability. By analyzing these results, we demonstrate the limitations of fixed randomness and the importance of considering multiple samples, which we show has significant implications for downstream applications. Our findings highlight the need for a nuanced understanding of LLM reliability and the potential risks associated with over-reliance on single-shot evaluations. This work provides a crucial step towards building more trustworthy and reliable LLM-based systems and applications.

Constructing confidence intervals (CIs) for the average treatment effect (ATE) from patient records is crucial to assess the effectiveness and safety of drugs. However, patient records typically come from different hospitals, thus raising the question of how multiple observational datasets can be effectively combined for this purpose. In our paper, we propose a new method that estimates the ATE from multiple observational datasets and provides valid CIs. Our method makes little assumptions about the observational datasets and is thus widely applicable in medical practice. The key idea of our method is that we leverage prediction-powered inferences and thereby essentially `shrink' the CIs so that we offer more precise uncertainty quantification as compared to na\"ive approaches. We further prove the unbiasedness of our method and the validity of our CIs. We confirm our theoretical results through various numerical experiments. Finally, we provide an extension of our method for constructing CIs from combinations of experimental and observational datasets.

We present measurements of the neutral atomic hydrogen (HI) mass function (HIMF) and cosmic HI density ($\Omega_{\rm HI}$) at $0 \leq z \leq 0.088$ from the Looking at the Distant Universe with MeerKAT Array (LADUMA) survey. Using LADUMA Data Release 1 (DR1), we analyze the HIMF via a new "recovery matrix" (RM) method that we benchmark against a more traditional Modified Maximum Likelihood (MML) method. Our analysis, which implements a forward modeling approach, corrects for survey incompleteness and uses extensive synthetic source injections to ensure robust estimates of the HIMF parameters and their associated uncertainties. This new method tracks the recovery of sources in mass bins different from those in which they were injected and incorporates a Poisson likelihood in the forward modeling process, allowing it to correctly handle uncertainties in bins with few or no detections. The application of our analysis to a high-purity subsample of the LADUMA DR1 spectral line catalog in turn mitigates any possible biases that could result from the inconsistent treatment of synthetic and real sources. For the surveyed redshift range, the recovered Schechter function normalization, low-mass slope, and "knee" mass are $\phi_\ast = 3.56_{-1.92}^{+0.97} \times 10^{-3}$ Mpc$^{-3}$ dex$^{-1}$, $\alpha = -1.18_{-0.19}^{+0.08}$, and $\log(M_\ast/M_\odot) = 10.01_{-0.12}^{+0.31}$, respectively, which together imply a comoving cosmic HI density of $\Omega_{\rm HI}=3.09_{-0.47}^{+0.65}\times 10^{-4}$. Our results show consistency between RM and MML methods and with previous low-redshift studies, giving confidence that the cosmic volume probed by LADUMA, even at low redshifts, is not an outlier in terms of its HI content., Comment: 25 pages, 10 figures, Accepted to ApJ

The MAP problem in Bayesian networks is notoriously intractable, even when approximated. In an earlier paper we introduced the Most Frugal Explanation heuristic approach to solving MAP, by partitioning the set of intermediate variables (neither observed nor part of the MAP variables) into a set of relevant variables, which are marginalized out, and irrelevant variables, which will be assigned a sampled value from their domain. In this study we explore whether knowledge about which variables are relevant for a particular query (i.e., domain knowledge) speeds up computation sufficiently to beat both exact MAP as well as approximate MAP while giving reasonably accurate results. Our results are inconclusive, but also show that this probably depends on the specifics of the MAP query, most prominently the number of MAP variables., Comment: 16 pages, 7 figures

We analyzed the 7.92$\times 10^{11}$ cosmic-ray-induced muon events collected by the IceCube Neutrino Observatory from May 13, 2011, when the fully constructed experiment started to take data, to May 12, 2023. This dataset provides an up-to-date cosmic-ray arrival direction distribution in the Southern Hemisphere with unprecedented statistical accuracy covering more than a full period length of a solar cycle. Improvements in Monte Carlo event simulation and better handling of year-to-year differences in data processing significantly reduce systematic uncertainties below the level of statistical fluctuations compared to the previously published results. We confirm the observation of a change in the angular structure of the cosmic-ray anisotropy between 10 TeV and 1 PeV, more specifically in the 100-300 TeV energy range.

Enabling effective collaboration among LLMs is a crucial step toward developing autonomous systems capable of solving complex problems. While LLMs are typically used as single-model generators, where humans critique and refine their outputs, the potential for jointly-trained collaborative models remains largely unexplored. Despite promising results in multi-agent communication and debate settings, little progress has been made in training models to work together on tasks. In this paper, we present a first step toward "Multi-agent LLM training" (MALT) on reasoning problems. Our approach employs a sequential multi-agent setup with heterogeneous LLMs assigned specialized roles: a generator, verifier, and refinement model iteratively solving problems. We propose a trajectory-expansion-based synthetic data generation process and a credit assignment strategy driven by joint outcome based rewards. This enables our post-training setup to utilize both positive and negative trajectories to autonomously improve each model's specialized capabilities as part of a joint sequential system. We evaluate our approach across MATH, GSM8k, and CQA, where MALT on Llama 3.1 8B models achieves relative improvements of 14.14%, 7.12%, and 9.40% respectively over the same baseline model. This demonstrates an early advance in multi-agent cooperative capabilities for performance on mathematical and common sense reasoning questions. More generally, our work provides a concrete direction for research around multi-agent LLM training approaches., Comment: Preliminary work

Energy-based models (EBMs) offer a flexible framework for probabilistic modelling across various data domains. However, training EBMs on data in discrete or mixed state spaces poses significant challenges due to the lack of robust and fast sampling methods. In this work, we propose to train discrete EBMs with Energy Discrepancy, a loss function which only requires the evaluation of the energy function at data points and their perturbed counterparts, thus eliminating the need for Markov chain Monte Carlo. We introduce perturbations of the data distribution by simulating a diffusion process on the discrete state space endowed with a graph structure. This allows us to inform the choice of perturbation from the structure of the modelled discrete variable, while the continuous time parameter enables fine-grained control of the perturbation. Empirically, we demonstrate the efficacy of the proposed approaches in a wide range of applications, including the estimation of discrete densities with non-binary vocabulary and binary image modelling. Finally, we train EBMs on tabular data sets with applications in synthetic data generation and calibrated classification., Comment: To appear in the proceedings of NeurIPS 2024

As AI systems become increasingly powerful, the need for safe AI has become more pressing. Humans are an attractive model for AI safety: as the only known agents capable of general intelligence, they perform robustly even under conditions that deviate significantly from prior experiences, explore the world safely, understand pragmatics, and can cooperate to meet their intrinsic goals. Intelligence, when coupled with cooperation and safety mechanisms, can drive sustained progress and well-being. These properties are a function of the architecture of the brain and the learning algorithms it implements. Neuroscience may thus hold important keys to technical AI safety that are currently underexplored and underutilized. In this roadmap, we highlight and critically evaluate several paths toward AI safety inspired by neuroscience: emulating the brain's representations, information processing, and architecture; building robust sensory and motor systems from imitating brain data and bodies; fine-tuning AI systems on brain data; advancing interpretability using neuroscience methods; and scaling up cognitively-inspired architectures. We make several concrete recommendations for how neuroscience can positively impact AI safety., Comment: 133 pages, 19 figures

Scientists and engineers use simulators to model empirically observed phenomena. However, tuning the parameters of a simulator to ensure its outputs match observed data presents a significant challenge. Simulation-based inference (SBI) addresses this by enabling Bayesian inference for simulators, identifying parameters that match observed data and align with prior knowledge. Unlike traditional Bayesian inference, SBI only needs access to simulations from the model and does not require evaluations of the likelihood-function. In addition, SBI algorithms do not require gradients through the simulator, allow for massive parallelization of simulations, and can perform inference for different observations without further simulations or training, thereby amortizing inference. Over the past years, we have developed, maintained, and extended $\texttt{sbi}$, a PyTorch-based package that implements Bayesian SBI algorithms based on neural networks. The $\texttt{sbi}$ toolkit implements a wide range of inference methods, neural network architectures, sampling methods, and diagnostic tools. In addition, it provides well-tested default settings but also offers flexibility to fully customize every step of the simulation-based inference workflow. Taken together, the $\texttt{sbi}$ toolkit enables scientists and engineers to apply state-of-the-art SBI methods to black-box simulators, opening up new possibilities for aligning simulations with empirically observed data.

Modern laser plasma accelerators (LPAs) often require meter-scale plasma waveguides to propagate a high-intensity drive laser pulse. Tapering the longitudinal gas density profile in meter-scale gas jets could allow for single stage laser plasma acceleration well beyond 10 GeV with current petawatt-class laser systems. Via simulation and interferometry measurements, we show density control by longitudinally adjusting the throat width and jet angle. Density profiles appropriate for tapering were calculated analytically and via particle-in-cell (PIC) simulations, and were matched experimentally. These simulations show that tapering can increase electron beam energy using 19 J laser energy from $\sim$9 GeV to $>$12 GeV in a 30 cm plasma, and the accelerated charge by an order of magnitude.

Addressing data integrity challenges, such as unlearning the effects of data poisoning after model training, is necessary for the reliable deployment of machine learning models. State-of-the-art influence functions, such as EK-FAC, often fail to accurately attribute abnormal model behavior to the specific poisoned training data responsible for the data poisoning attack. In addition, traditional unlearning algorithms often struggle to effectively remove the influence of poisoned samples, particularly when only a few affected examples can be identified. To address these challenge, we introduce $\Delta$-Influence, a novel approach that leverages influence functions to trace abnormal model behavior back to the responsible poisoned training data using as little as just one poisoned test example. $\Delta$-Influence applies data transformations that sever the link between poisoned training data and compromised test points without significantly affecting clean data. This allows $\Delta$-Influence to detect large negative shifts in influence scores following data transformations, a phenomenon we term as influence collapse, thereby accurately identifying poisoned training data. Unlearning this subset, e.g. through retraining, effectively eliminates the data poisoning. We validate our method across three vision-based poisoning attacks and three datasets, benchmarking against four detection algorithms and five unlearning strategies. We show that $\Delta$-Influence consistently achieves the best unlearning across all settings, showing the promise of influence functions for corrective unlearning. Our code is publicly available at: \url{https://github.com/andyisokay/delta-influence}, Comment: Accepted at NeurIPS Workshop on Attributing Model Behavior at Scale (ATTRIB @ NeurIPS 2024)

Semantic segmentation is a crucial task in medical imaging. Although supervised learning techniques have proven to be effective in performing this task, they heavily depend on large amounts of annotated training data. The recently introduced Segment Anything Model (SAM) enables prompt-based segmentation and offers zero-shot generalization to unfamiliar objects. In our work, we leverage SAM's abstract object understanding for medical image segmentation to provide pseudo labels for semi-supervised learning, thereby mitigating the need for extensive annotated training data. Our approach refines initial segmentations that are derived from a limited amount of annotated data (comprising up to 43 cases) by extracting bounding boxes and seed points as prompts forwarded to SAM. Thus, it enables the generation of dense segmentation masks as pseudo labels for unlabelled data. The results show that training with our pseudo labels yields an improvement in Dice score from $74.29\,\%$ to $84.17\,\%$ and from $66.63\,\%$ to $74.87\,\%$ for the segmentation of bones of the paediatric wrist and teeth in dental radiographs, respectively. As a result, our method outperforms intensity-based post-processing methods, state-of-the-art supervised learning for segmentation (nnU-Net), and the semi-supervised mean teacher approach. Our Code is available on GitHub., Comment: Presented at MICCAI Workshop on Advancing Data Solutions in Medical Imaging AI 2024; Code and data: https://github.com/multimodallearning/SamCarriesTheBurden

Education is poised for a transformative shift with the advent of neurosymbolic artificial intelligence (NAI), which will redefine how we support deeply adaptive and personalized learning experiences. NAI-powered education systems will be capable of interpreting complex human concepts and contexts while employing advanced problem-solving strategies, all grounded in established pedagogical frameworks. This will enable a level of personalization in learning systems that to date has been largely unattainable at scale, providing finely tailored curricula that adapt to an individual's learning pace and accessibility needs, including the diagnosis of student understanding of subjects at a fine-grained level, identifying gaps in foundational knowledge, and adjusting instruction accordingly. In this paper, we propose a system that leverages the unique affordances of pedagogical agents -- embodied characters designed to enhance learning -- as critical components of a hybrid NAI architecture. To do so, these agents can thus simulate nuanced discussions, debates, and problem-solving exercises that push learners beyond rote memorization toward deep comprehension. We discuss the rationale for our system design and the preliminary findings of our work. We conclude that education in the era of NAI will make learning more accessible, equitable, and aligned with real-world skills. This is an era that will explore a new depth of understanding in educational tools.

We introduce a family of boundary conditions and point constraints for conformal immersions that increase the controllability of surfaces defined as minimizers of conformal variational problems. Our free boundary conditions fix the metric on the boundary, up to a global scale, and admit a discretization compatible with discrete conformal equivalence. We also introduce constraints on the conformal scale factor, enforcing rigidity of the geometry in regions of interest, and describe how in the presence of point constraints the conformal class encodes knot points of the spline that can be directly manipulated. To control the tangent planes, we introduce flux constraints balancing the internal material stresses. The collection of these point constraints provide intuitive controls for exploring a subspace of conformal immersions interpolating a fixed set of points in space. We demonstrate the applicability of our framework to geometric modeling, mathematical visualization, and form finding., Comment: to appear in Differential Geom. Appl., 26 pages

Laser-driven (LD) ion acceleration has been explored in a newly constructed short focal length beamline at the BELLA petawatt facility (interaction point 2, iP2). For applications utilizing such LD ion beams, a beam transport system is required, which for reasons of compactness be ideally contained within 3 m. While they are generated from a micron-scale source, large divergence and energy spread of LD ion beams present a unique challenge to transporting them compared to beams from conventional accelerators. This study gives an overview of proposed compact transport designs using permanent magnets satisfying different requirements depending on the application for the iP2 proton beamline such as radiation biology, material science, and high energy density science. These designs are optimized for different parameters such as energy spread and peak proton density according to the application's need. The various designs consist solely of permanent magnet elements, which can provide high magnetic field gradients on a small footprint. While the field strengths are fixed, we have shown that the beam size is able to be tuned effectively by varying the placement of the magnets. The performance of each design was evaluated based on high order particle tracking simulations of typical LD proton beams. We also examine the ability of certain configurations to tune and select beam energies, critical for specific applications. A more detailed investigation was carried out for a design to deliver 10 MeV LD accelerated ions for radiation biology applications. With these transport system designs, the iP2 beamline is ready to house various application experiments., Comment: 12 pages, 9 figures, to be submitted to Physical Review Accelerators and Beams

Game-theoretic characterizations of process equivalences traditionally form a central topic in concurrency; for example, most equivalences on the classical linear-time / branching-time spectrum come with such characterizations. Recent work on so-called graded semantics has led to a generic behavioural equivalence game that covers the mentioned games on the linear-time~/ branching-time spectrum and moreover applies in coalgebraic generality, and thus instantiates also to equivalence games on systems with non-relational branching type (probabilistic, weighted, game-based etc.). In the present work, we generalize this approach to cover other types of process comparison beyond equivalence, such as behavioural preorders or pseudometrics. At the most general level, we abstract such notions of behavoiural conformance in terms of topological categories, and later specialize to conformances presented as relational structures to obtain a concrete syntax. We obtain a sound and complete generic game for behavioural conformances in this sense. We present a number of instantiations, obtaining game characterizations of, e.g., trace inclusion, probabilistic trace distance, bisimulation topologies, and simulation distances on metric labelled transition systems.

Wavefront reconstruction is a critical component in various optical systems, including adaptive optics, interferometry, and phase contrast imaging. Traditional reconstruction methods often employ either the Cartesian (pixel) basis or the Zernike polynomial basis. While the Cartesian basis is adept at capturing high-frequency features, it is susceptible to overfitting and inefficiencies due to the high number of degrees of freedom. The Zernike basis efficiently represents common optical aberrations but struggles with complex or non-standard wavefronts such as optical vortices, Bessel beams, or wavefronts with sharp discontinuities. This paper introduces a novel approach to wavefront reconstruction using an over-complete phase dictionary combined with sparse representation techniques. By constructing a dictionary that includes a diverse set of basis functions - ranging from Zernike polynomials to specialized functions representing optical vortices and other complex modes - we enable a more flexible and efficient representation of complex wavefronts. Furthermore, a trainable affine transform is implemented to account for misalignment. Utilizing principles from compressed sensing and sparse coding, we enforce sparsity in the coefficient space to avoid overfitting and enhance robustness to noise.

We present the detection of directional muon beams produced using a PW laser at the Lawrence Berkeley National Laboratory. The muon source is a multi-GeV electron beam generated in a 30 cm laser plasma accelerator interacting with a high-Z converter target. The GeV photons resulting from the interaction are converted into a high-flux, directional muon beam via pair production. By employing scintillators to capture delayed events, we were able to identify the produced muons and characterize the source. Using theoretical knowledge of the muon production process combined with simulations that show outstanding agreement with the experiments, we demonstrate that the multi-GeV electron beams produce muon beams with GeV energies and fluxes, at a few meters from the source, up to 4 orders of magnitude higher than cosmic ray muons. Laser-plasma-accelerator-based muon sources can therefore enhance muon imaging applications thanks to their compactness, directionality, and high fluxes which reduce the exposure time by orders of magnitude compared to cosmic ray muons. Using the Geant4-based simulation code we developed to gain insight into the experimental results, we can design future experiments and applications based on LPA-generated muons.

Ultrafast transmission electron microscopy (UTEM) has emerged as a versatile technique for the time-resolved imaging of nanoscale dynamics on timescales down to few-hundred attoseconds but the temporal and spatial resolutions are still limited by the coherence properties of pulsed electron sources. Here, we report the development of a novel laser-driven linear cold-field electron emitter integrated in a state-of-the-art UTEM system. Illuminating the sharp tungsten emitter tip with a UV light pulse generates ultrashort femtosecond electron pulses of 220 fs pulse duration, with energy widths as low as 360 meV. The photoelectron emitter demonstrates exceptional spatial coherence, achieving focal spot sizes down to 2 $\mathring {\mathrm A}$ and a peak normalized brightness exceeding 6.7 $\times 10^{13}$ A/m$^2$sr. With an order-of-magnitude improvement compared to previously employed laser-driven Schottky field emitters, the present development opens up the field of ultrafast atomic-scale electron probing., Comment: 20 pages, Manuscript + Supporting Information

We present a systematic review of peer-reviewed research into ways to mitigate manipulative information about climate change on social media. Such information may include disinformation, harmful influence campaigns, or the unintentional spread of misleading information. We find that commonly recommended approaches to addressing manipulation about climate change include corrective information sharing and education campaigns targeting media literacy. However, most relevant research fails to test the approaches and interventions it proposes. We locate research gaps that include the lack of attention to large commercial and political entities involved in generating and disseminating manipulation, video- and image-focused platforms, and computational methods to collect and analyze data. Evidence drawn from many studies demonstrates an emerging consensus about policies required to promote reliable knowledge about climate change and resist manipulation.

Topic models allow researchers to extract latent factors from text data and use those variables in downstream statistical analyses. However, these methodologies can vary significantly due to initialization differences, randomness in sampling procedures, or noisy data. Reliability of these methods is of particular concern as many researchers treat learned topic models as ground truth for subsequent analyses. In this work, we show that the standard practice for quantifying topic model reliability fails to capture essential aspects of the variation in two widely-used topic models. Drawing from a extensive literature on measurement theory, we provide empirical and theoretical analyses of three other metrics for evaluating the reliability of topic models. On synthetic and real-world data, we show that McDonald's $\omega$ provides the best encapsulation of reliability. This metric provides an essential tool for validation of topic model methodologies that should be a standard component of any topic model-based research.

In this paper, we derive a priori error estimates for variational inequalities of the first kind in an abstract framework. This is done by combining the first Strang Lemma and the Falk Theorem. The main application consists in the derivation of a priori error estimates for Galerkin methods, in which "variational crimes" may perturb the underlying variational inequality. Different types of perturbations are incorporated into the abstract framework and discussed by various examples. For instance, the perturbation caused by an inexact quadrature is examined in detail for the Laplacian obstacle problem. For this problem, guaranteed rates for the approximation error resulting from the use of higher-order finite elements are derived. In numerical experiments, the influence of the number of quadrature points on the approximation error and on the quadrature-related error itself is studied for several discretization methods.

Aluminum scandium nitride (Al$_{1-x}$Sc$_x$N) is a promising material for ferroelectric devices due to its large remanent polarization, scalability, and compatibility with semiconductor technology. By doping AlN with Sc, the bonds in the polar AlN structure are weakened, which enables ferroelectric switching below the dielectric breakdown field. However, one disadvantage of Sc doping is that it increases the material's tendency towards oxidation. In the present study, the oxidation process of tungsten-capped and uncapped Al$_{0.83}$Sc$_{0.17}$N thin films is investigated by hard X-ray photoelectron spectroscopy (HAXPES). The samples had been exposed to air for either two weeks or 6 months. HAXPES spectra indicate the replacement of nitrogen by oxygen, and the tendency of oxygen to favor oxidation with Sc rather than Al. The appearance of an N$_2$ spectral feature thus can be directly related to the oxidation process. We present an oxidation model that mimics these spectroscopic results of the element-specific oxidation processes within Al$_{1-x}$Sc$_x$N. Finally, in operando HAXPES data of uncapped and capped AlScN-capacitor stacks are interpreted using the proposed model.

One of the most challenging forms of misinformation involves the out-of-context (OOC) use of images paired with misleading text, creating false narratives. Existing AI-driven detection systems lack explainability and require expensive fine-tuning. We address these issues with MAD-Sherlock: a Multi-Agent Debate system for OOC Misinformation Detection. MAD-Sherlock introduces a novel multi-agent debate framework where multimodal agents collaborate to assess contextual consistency and request external information to enhance cross-context reasoning and decision-making. Our framework enables explainable detection with state-of-the-art accuracy even without domain-specific fine-tuning. Extensive ablation studies confirm that external retrieval significantly improves detection accuracy, and user studies demonstrate that MAD-Sherlock boosts performance for both experts and non-experts. These results position MAD-Sherlock as a powerful tool for autonomous and citizen intelligence applications.

As the signatories of this manifesto, we denounce the attention economy as inhumane and a threat to our sociopolitical and ecological well-being. We endorse policymakers' efforts to address the negative consequences of the attention economy's technology, but add that these approaches are often limited in their criticism of the systemic context of human attention. Starting from Buddhist philosophy, we advocate a broader approach: an ecology of attending, that centers on conceptualizing, designing, and using attention (1) in an embedded way and (2) focused on the alleviating of suffering. With 'embedded' we mean that attention is not a neutral, isolated mechanism but a meaning-engendering part of an 'ecology' of bodily, sociotechnical and moral frameworks. With 'focused on the alleviation of suffering' we explicitly move away from the (often implicit) conception of attention as a tool for gratifying desires., Comment: 21 pages, 1 figure

Prediction Rule Ensembles (PREs) are robust and interpretable statistical learning techniques with potential for predictive analytics, yet their efficacy in the presence of missing data is untested. This study uses multiple imputation to fill in missing values, but uses a data stacking approach instead of a traditional model pooling approach to combine the results. We perform a simulation study to compare imputation methods under realistic conditions, focusing on sample sizes of $N=200$ and $N=400$ across 1,000 replications. Evaluated techniques include multiple imputation by chained equations with predictive mean matching (MICE PMM), MICE with Random Forest (MICE RF), Random Forest imputation with the ranger algorithm (missRanger), and imputation using extreme gradient boosting (MIXGBoost), with results compared to listwise deletion. Because stacking multiple imputed datasets can overly complicate models, we additionally explore different coarsening levels to simplify and enhance the interpretability and performance of PRE models. Our findings highlight a trade-off between predictive performance and model complexity in selecting imputation methods. While MIXGBoost and MICE PMM yield high rule recovery rates, they also increase false positives in rule selection. In contrast, MICE RF and missRanger promote rule sparsity. MIXGBoost achieved the greatest MSE reduction, followed by MICE PMM, MICE RF, and missRanger. Avoiding too-course rounding of variables helps to reduce model size with marginal loss in performance. Listwise deletion has an adverse impact on model validity. Our results emphasize the importance of choosing suitable imputation techniques based on research goals and of advancing methods for handling missing data in statistical learning., Comment: *Contributed equally and share the first authorship

While behavioural equivalences among systems of the same type, such as Park/Milner bisimilarity of labelled transition systems, are an established notion, a systematic treatment of relationships between systems of different type is currently missing. We provide such a treatment in the framework of universal coalgebra, in which the type of a system (nondeterministic, probabilistic, weighted, game-based etc.) is abstracted as a set functor: We introduce relational connectors among set functors, which induce notions of heterogeneous (bi)simulation among coalgebras of the respective types. We give a number of constructions on relational connectors. In particular, we identify composition and converse operations on relational connectors; we construct corresponding identity relational connectors, showing that the latter generalize the standard Barr extension of weak-pullback-preserving functors; and we introduce a Kantorovich construction in which relational connectors are induced from relations between modalities. For Kantorovich relational connectors, one has a notion of dual-purpose modal logic interpreted over both system types, and we prove a corresponding Hennessy-Milner-type theorem stating that generalized (bi)similarity coincides with theory inclusion on finitely-branching systems. We apply these results to a number of example scenarios involving labelled transition systems with different label alphabets, probabilistic systems, and input/output conformances.

Generic notions of bisimulation for various types of systems (nondeterministic, probabilistic, weighted etc.) rely on identity-preserving (normal) lax extensions of the functor encapsulating the system type, in the paradigm of universal coalgebra. It is known that preservation of weak pullbacks is a sufficient condition for a functor to admit a normal lax extension (the Barr extension, which in fact is then even strict); in the converse direction, nothing is currently known about necessary (weak) pullback preservation conditions for the existence of normal lax extensions. In the present work, we narrow this gap by showing on the one hand that functors admitting a normal lax extension preserve 1/4-iso pullbacks, i.e. pullbacks in which at least one of the projections is an isomorphism. On the other hand, we give sufficient conditions, showing that a functor admits a normal lax extension if it weakly preserves either 1/4-iso pullbacks and 4/4-epi pullbacks (i.e. pullbacks in which all morphisms are epic) or inverse images. We apply these criteria to concrete examples, in particular to functors modelling neighbourhood systems and weighted systems., Comment: Full version of STACS 2025 paper

This work presents a novel lattice-based methodology for incorporating multidimensional constraints into continuous decision variables within a genetic algorithm (GA) framework. The proposed approach consolidates established transcription techniques for crossover of continuous decision variables, aiming to leverage domain knowledge and guide the search process towards feasible regions of the design space. This work offers a robust and general purpose lattice-based GA that is applicable to a broad range of optimization problems. Monte Carlo analysis demonstrates that lattice-based methods find solutions two orders of magnitude closer to optima in fewer generations. The effectiveness of the lattice-based approach is showcased through two illustrative multi-objective design problems: (1) optimal telescope placement for astrophotography and (2) optimal design of a satellite constellation for maximizing ground station access. The optimal telescope placement example shows that lattice-based methods converge to the Pareto front in 15% fewer generations than traditional methods. The orbit design example shows that lattice-based methods discover an order of magnitude more Pareto-optimal solutions than traditional methods in a highly constrained design space. Overall, the results show that the lattice-based method exhibits enhanced exploration capabilities, traversing the solution space more comprehensively and achieving faster convergence compared to conventional GAs.

Selecting cost-effective optimal sensor configurations for subsequent inference of parameters in black-box stochastic systems faces significant computational barriers. We propose a novel and robust approach, modelling the joint distribution over input parameters and solution with a joint energy-based model, trained on simulation data. Unlike existing simulation-based inference approaches, which must be tied to a specific set of point evaluations, we learn a functional representation of parameters and solution. This is used as a resolution-independent plug-and-play surrogate for the joint distribution, which can be conditioned over any set of points, permitting an efficient approach to sensor placement. We demonstrate the validity of our framework on a variety of stochastic problems, showing that our method provides highly informative sensor locations at a lower computational cost compared to conventional approaches., Comment: 23 pages

Reliable estimation of treatment effects from observational data is important in many disciplines such as medicine. However, estimation is challenging when unconfoundedness as a standard assumption in the causal inference literature is violated. In this work, we leverage arbitrary (potentially high-dimensional) instruments to estimate bounds on the conditional average treatment effect (CATE). Our contributions are three-fold: (1) We propose a novel approach for partial identification through a mapping of instruments to a discrete representation space so that we yield valid bounds on the CATE. This is crucial for reliable decision-making in real-world applications. (2) We derive a two-step procedure that learns tight bounds using a tailored neural partitioning of the latent instrument space. As a result, we avoid instability issues due to numerical approximations or adversarial training. Furthermore, our procedure aims to reduce the estimation variance in finite-sample settings to yield more reliable estimates. (3) We show theoretically that our procedure obtains valid bounds while reducing estimation variance. We further perform extensive experiments to demonstrate the effectiveness across various settings. Overall, our procedure offers a novel path for practitioners to make use of potentially high-dimensional instruments (e.g., as in Mendelian randomization).

Sparse autoencoders (SAEs) are a recent technique for decomposing neural network activations into human-interpretable features. However, in order for SAEs to identify all features represented in frontier models, it will be necessary to scale them up to very high width, posing a computational challenge. In this work, we introduce Switch Sparse Autoencoders, a novel SAE architecture aimed at reducing the compute cost of training SAEs. Inspired by sparse mixture of experts models, Switch SAEs route activation vectors between smaller "expert" SAEs, enabling SAEs to efficiently scale to many more features. We present experiments comparing Switch SAEs with other SAE architectures, and find that Switch SAEs deliver a substantial Pareto improvement in the reconstruction vs. sparsity frontier for a given fixed training compute budget. We also study the geometry of features across experts, analyze features duplicated across experts, and verify that Switch SAE features are as interpretable as features found by other SAE architectures., Comment: Code available at https://github.com/amudide/switch_sae

A key challenge in interpretability is to decompose model activations into meaningful features. Sparse autoencoders (SAEs) have emerged as a promising tool for this task. However, a central problem in evaluating the quality of SAEs is the absence of ground truth features to serve as an evaluation gold standard. Current evaluation methods for SAEs are therefore confronted with a significant trade-off: SAEs can either leverage toy models or other proxies with predefined ground truth features; or they use extensive prior knowledge of realistic task circuits. The former limits the generalizability of the evaluation results, while the latter limits the range of models and tasks that can be used for evaluations. We introduce SAGE: Scalable Autoencoder Ground-truth Evaluation, a ground truth evaluation framework for SAEs that scales to large state-of-the-art SAEs and models. We demonstrate that our method can automatically identify task-specific activations and compute ground truth features at these points. Compared to previous methods we reduce the training overhead by introducing a novel reconstruction method that allows to apply residual stream SAEs to sublayer activations. This eliminates the need for SAEs trained on every task-specific activation location. Then we validate the scalability of our framework, by evaluating SAEs on novel tasks on Pythia70M, GPT-2 Small, and Gemma-2-2. Our framework therefore paves the way for generalizable, large-scale evaluations of SAEs in interpretability research.

The rapid proliferation of AI-manipulated or generated audio deepfakes poses serious challenges to media integrity and election security. Current AI-driven detection solutions lack explainability and underperform in real-world settings. In this paper, we introduce novel explainability methods for state-of-the-art transformer-based audio deepfake detectors and open-source a novel benchmark for real-world generalizability. By narrowing the explainability gap between transformer-based audio deepfake detectors and traditional methods, our results not only build trust with human experts, but also pave the way for unlocking the potential of citizen intelligence to overcome the scalability issue in audio deepfake detection.

Qualitative researchers use tools to collect, sort, and analyze their data. Should qualitative researchers use large language models (LLMs) as part of their practice? LLMs could augment qualitative research, but it is unclear if their use is appropriate, ethical, or aligned with qualitative researchers' goals and values. We interviewed twenty qualitative researchers to investigate these tensions. Many participants see LLMs as promising interlocutors with attractive use cases across the stages of research, but wrestle with their performance and appropriateness. Participants surface concerns regarding the use of LLMs while protecting participant interests, and call attention to an urgent lack of norms and tooling to guide the ethical use of LLMs in research. Given the importance of qualitative methods to human-computer interaction, we use the tensions surfaced by our participants to outline guidelines for researchers considering using LLMs in qualitative research and design principles for LLM-assisted qualitative data analysis tools.