Search

Climate studies often rely on remotely sensed images to retrieve two-dimensional maps of cloud properties. To advance volumetric analysis, we focus on recovering the three-dimensional (3D) heterogeneous extinction coefficient field of shallow clouds using multiview remote sensing data. Climate research requires large-scale worldwide statistics. To enable scalable data processing, previous deep neural networks (DNNs) can infer at spaceborne remote sensing downlink rates. However, prior methods are limited to a fixed solar illumination direction. In this work, we introduce the first scalable DNN-based system for 3D cloud retrieval that accommodates varying camera poses and solar directions. By integrating multiview cloud intensity images with camera poses and solar direction data, we achieve greater flexibility in recovery. Training of the DNN is performed by a novel two-stage scheme to address the high number of degrees of freedom in this problem. Our approach shows substantial improvements over previous state-of-the-art, particularly in handling variations in the sun's zenith angle., Comment: 4 pages, 4 figures

Energetic feedback processes associated with accreting supermassive black holes can expel gas from massive haloes and significantly alter various measures of clustering on ~Mpc scales, potentially biasing the values of cosmological parameters inferred from analyses of large-scale structure (LSS) if not modelled accurately. Here we use the state-of-the-art FLAMINGO suite of cosmological hydrodynamical simulations to gauge the impact of feedback on large-scale structure by comparing to Planck + ACT stacking measurements of the kinetic Sunyaev-Zel'dovich (kSZ) effect of SDSS BOSS galaxies. We make careful like-with-like comparisons to the observations, aided by high precision KiDS and DES galaxy-galaxy lensing measurements of the BOSS galaxies to inform the selection of the simulated galaxies. In qualitative agreement with several recent studies using dark matter only simulations corrected for baryonic effects, we find that the kSZ effect measurements prefer stronger feedback than predicted by simulations which have been calibrated to reproduce the gas fractions of low redshift X-ray-selected groups and clusters. We find that the increased feedback can help to reduce the so-called S8 tension between the observed and CMB-predicted clustering on small scales as probed by cosmic shear (although at the expense of agreement with the X-ray group measurements). However, the increased feedback is only marginally effective at reducing the reported offsets between the predicted and observed clustering as probed by the thermal SZ (tSZ) effect power spectrum and tSZ effect--weak lensing cross-spectrum, both of which are sensitive to higher halo masses than cosmic shear., Comment: 20 pages, 10 figures, submitted to MNRAS

NLP benchmarks rely on standardized datasets for training and evaluating models and are crucial for advancing the field. Traditionally, expert annotations ensure high-quality labels; however, the cost of expert annotation does not scale well with the growing demand for larger datasets required by modern models. While crowd-sourcing provides a more scalable solution, it often comes at the expense of annotation precision and consistency. Recent advancements in large language models (LLMs) offer new opportunities to enhance the annotation process, particularly for detecting label errors in existing datasets. In this work, we consider the recent approach of LLM-as-a-judge, leveraging an ensemble of LLMs to flag potentially mislabeled examples. Through a case study of four datasets from the TRUE benchmark, covering different tasks and domains, we empirically analyze the labeling quality of existing datasets, and compare expert, crowd-sourced, and our LLM-based annotations in terms of agreement, label quality, and efficiency, demonstrating the strengths and limitations of each annotation method. Our findings reveal a substantial number of label errors, which, when corrected, induce a significant upward shift in reported model performance. This suggests that many of the LLMs so-called mistakes are due to label errors rather than genuine model failures. Additionally, we discuss the implications of mislabeled data and propose methods to mitigate them in training to improve model performance.

The effect of baryon physics associated with galaxy formation onto the large-scale matter distribution of the Universe is a key uncertainty in the theoretical modelling required for the interpretation of Stage IV cosmology surveys. We use the FLAMINGO suite of simulations to study the baryon response due to galaxy formation of the total matter power spectrum. We find that it is only well converged for simulation volumes in excess of $(200~Mpc)^3$. We report results for simulations of varying feedback intensity, which either match the X-ray inferred gas fractions in clusters and the z=0 stellar mass function, or shifted versions of the data, as well as for different implementations of AGN feedback. We package our results in the form of a Gaussian process emulator which can rapidly reproduce all the simulations' predictions to better than 1% up to the comoving wavenumber $k = 10~h/Mpc$ and up to z=2 for all the feedback models present in the FLAMINGO suite. We find that the response becomes stronger, the range of scales affected increases, and the position of the minimum of the response moves to smaller scales as the redshift decreases. We find that lower gas fractions in groups and clusters lead to a stronger response and that the use of collimated jets instead of thermally driven winds for AGN feedback enhances the effect. Lowering the stellar masses at fixed cluster gas fractions also increases the magnitude of the response. We find only a small (1% at $k<10~h/Mpc$) dependence of our results on the background cosmology, but a wider range of cosmology variations will be needed to confirm this result. The response we obtain for our strongest feedback models is compatible with some of the recent analyses combining weak lensing with external data. Such a response is, however, in strong tension with the X-ray inferred gas fractions in clusters used to calibrate the FLAMINGO model., Comment: 14 pages, 13 figures, submitted to MNRAS, emulator will be released publicly with the final version of the paper, comments welcome

Sum of squares (SOS) optimization is a powerful technique for solving problems where the positivity of a polynomials must be enforced. The common approach to solve an SOS problem is by relaxation to a Semidefinite Program (SDP). The main advantage of this transormation is that SDP is a convex problem for which efficient solvers are readily available. However, while considerable progress has been made in recent years, the standard approaches for solving SDPs are still known to scale poorly. Our goal is to devise an approach that can handle larger, more complex problems than is currently possible. The challenge indeed lies in how SDPs are commonly solved. State-Of-The-Art approaches rely on the interior point method, which requires the factorization of large matrices. We instead propose an approach inspired by polynomial neural networks, which exhibit excellent performance when optimized using techniques from the deep learning toolbox. In a somewhat counter-intuitive manner, we replace the convex SDP formulation with a non-convex, unconstrained, and \emph{over parameterized} formulation, and solve it using a first order optimization method. It turns out that this approach can handle very large problems, with polynomials having over four million coefficients, well beyond the range of current SDP-based approaches. Furthermore, we highlight theoretical and practical results supporting the experimental success of our approach in avoiding spurious local minima, which makes it amenable to simple and fast solutions based on gradient descent. In all the experiments, our approach had always converged to a correct global minimum, on general (non-sparse) polynomials, with running time only slightly higher than linear in the number of polynomial coefficients, compared to higher than quadratic in the number of coefficients for SDP-based methods.

Large Language Models (LLMs) show significant potential in economic and strategic interactions, where communication via natural language is often prevalent. This raises key questions: Do LLMs behave rationally? Can they mimic human behavior? Do they tend to reach an efficient and fair outcome? What is the role of natural language in the strategic interaction? How do characteristics of the economic environment influence these dynamics? These questions become crucial concerning the economic and societal implications of integrating LLM-based agents into real-world data-driven systems, such as online retail platforms and recommender systems. While the ML community has been exploring the potential of LLMs in such multi-agent setups, varying assumptions, design choices and evaluation criteria across studies make it difficult to draw robust and meaningful conclusions. To address this, we introduce a benchmark for standardizing research on two-player, sequential, language-based games. Inspired by the economic literature, we define three base families of games with consistent parameterization, degrees of freedom and economic measures to evaluate agents' performance (self-gain), as well as the game outcome (efficiency and fairness). We develop an open-source framework for interaction simulation and analysis, and utilize it to collect a dataset of LLM vs. LLM interactions across numerous game configurations and an additional dataset of human vs. LLM interactions. Through extensive experimentation, we demonstrate how our framework and dataset can be used to: (i) compare the behavior of LLM-based agents to human players in various economic contexts; (ii) evaluate agents in both individual and collective performance measures; and (iii) quantify the effect of the economic characteristics of the environments on the behavior of agents.

Large language models (LLMs) often produce errors, including factual inaccuracies, biases, and reasoning failures, collectively referred to as "hallucinations". Recent studies have demonstrated that LLMs' internal states encode information regarding the truthfulness of their outputs, and that this information can be utilized to detect errors. In this work, we show that the internal representations of LLMs encode much more information about truthfulness than previously recognized. We first discover that the truthfulness information is concentrated in specific tokens, and leveraging this property significantly enhances error detection performance. Yet, we show that such error detectors fail to generalize across datasets, implying that -- contrary to prior claims -- truthfulness encoding is not universal but rather multifaceted. Next, we show that internal representations can also be used for predicting the types of errors the model is likely to make, facilitating the development of tailored mitigation strategies. Lastly, we reveal a discrepancy between LLMs' internal encoding and external behavior: they may encode the correct answer, yet consistently generate an incorrect one. Taken together, these insights deepen our understanding of LLM errors from the model's internal perspective, which can guide future research on enhancing error analysis and mitigation.

Will a Visual Language Model (VLM)-based bot warn us about slipping if it detects a wet floor? Recent VLMs have demonstrated impressive capabilities, yet their ability to infer outcomes and causes remains underexplored. To address this, we introduce NL-Eye, a benchmark designed to assess VLMs' visual abductive reasoning skills. NL-Eye adapts the abductive Natural Language Inference (NLI) task to the visual domain, requiring models to evaluate the plausibility of hypothesis images based on a premise image and explain their decisions. NL-Eye consists of 350 carefully curated triplet examples (1,050 images) spanning diverse reasoning categories: physical, functional, logical, emotional, cultural, and social. The data curation process involved two steps - writing textual descriptions and generating images using text-to-image models, both requiring substantial human involvement to ensure high-quality and challenging scenes. Our experiments show that VLMs struggle significantly on NL-Eye, often performing at random baseline levels, while humans excel in both plausibility prediction and explanation quality. This demonstrates a deficiency in the abductive reasoning capabilities of modern VLMs. NL-Eye represents a crucial step toward developing VLMs capable of robust multimodal reasoning for real-world applications, including accident-prevention bots and generated video verification.

We revisit the recently introduced Local Glivenko-Cantelli setting, which studies distribution-dependent uniform convegence rates of the Maximum Likelihood Estimator (MLE). In this work, we investigate generalizations of this setting where arbitrary estimators are allowed rather than just the MLE. Can a strictly larger class of measures be learned? Can better risk decay rates be obtained? We provide exhaustive answers to these questions -- which are both negative, provided the learner is barred from exploiting some infinite-dimensional pathologies. On the other hand, allowing such exploits does lead to a strictly larger class of learnable measures.

The close link between cognitive decline and language has fostered long-standing collaboration between the NLP and medical communities in dementia research. To examine this, we reviewed over 200 papers applying NLP to dementia related efforts, drawing from medical, technological, and NLP-focused literature. We identify key research areas, including dementia detection, linguistic biomarker extraction, caregiver support, and patient assistance, showing that half of all papers focus solely on dementia detection using clinical data. However, many directions remain unexplored: artificially degraded language models, synthetic data, digital twins, and more. We highlight gaps and opportunities around trust, scientific rigor, applicability, and cross-community collaboration, and showcase the diverse datasets encountered throughout our review: recorded, written, structured, spontaneous, synthetic, clinical, social media based, and more. This review aims to inspire more creative approaches to dementia research within the medical and NLP communities.

The masses of galaxy clusters are commonly measured from X-ray observations under the assumption of hydrostatic equilibrium (HSE). This technique is known to underestimate the true mass systematically. The fiducial FLAMINGO cosmological hydrodynamical simulation predicts the median hydrostatic mass bias to increase from $b_\text{HSE} \equiv (M_\text{HSE,500c}-M_\text{500c})/M_\text{500c} \approx -0.1$ to -0.2 when the true mass increases from group to cluster mass scales. However, the bias is nearly independent of the hydrostatic mass. The scatter at fixed true mass is minimum for $M_\text{500c}\sim 10^{14}~\text{M}_\odot$, where $\sigma(b_\text{HSE})\approx 0.1$, but increases rapidly towards lower and higher masses. At a fixed true mass, the hydrostatic masses increase (decrease) with redshift on group (cluster) scales, and the scatter increases. The bias is insensitive to the choice of analytic functions assumed to represent the density and temperature profiles, but it is sensitive to the goodness of fit, with poorer fits corresponding to a stronger median bias and a larger scatter. The bias is also sensitive to the strength of stellar and AGN feedback. Models predicting lower gas fractions yield more (less) biased masses for groups (clusters). The scatter in the bias at fixed true mass is due to differences in the pressure gradients rather than in the temperature at $R_\text{500c}$. The total kinetic energies within $r_\text{500c}$ in low- and high-mass clusters are sub- and super-virial, respectively, though all become sub-virial when external pressure is accounted for. Analyses of the terms in the virial and Euler equations suggest that non-thermal motions, including rotation, account for most of the hydrostatic mass bias. However, we find that the mass bias estimated from X-ray luminosity weighted profiles strongly overestimates the deviations from hydrostatic equilibrium., Comment: 19 pages, 14 figures (including appendix). Submitted to MNRAS

Pricing decisions stand out as one of the most critical tasks a company faces, particularly in today's digital economy. As with other business decision-making problems, pricing unfolds in a highly competitive and uncertain environment. Traditional analyses in this area have heavily relied on game theory and its variants. However, an important drawback of these approaches is their reliance on common knowledge assumptions, which are hardly tenable in competitive business domains. This paper introduces an innovative personalized pricing framework designed to assist decision-makers in undertaking pricing decisions amidst competition, considering both buyer's and competitors' preferences. Our approach (i) establishes a coherent framework for modeling competition mitigating common knowledge assumptions; (ii) proposes a principled method to forecast competitors' pricing and customers' purchasing decisions, acknowledging major business uncertainties; and, (iii) encourages structured thinking about the competitors' problems, thus enriching the solution process. To illustrate these properties, in addition to a general pricing template, we outline two specifications - one from the retail domain and a more intricate one from the pension fund domain.

Galaxy cluster counts have historically been important for the measurement of cosmological parameters and upcoming surveys will greatly reduce the statistical errors. To exploit the potential of current and future cluster surveys, theoretical uncertainties on the predicted abundance must be smaller than the statistical errors. Models used to predict cluster counts typically combine a model for the dark matter only (DMO) halo mass function (HMF) with an observable - mass relation that is assumed to be a power-law with lognormal scatter. We use the FLAMINGO suite of cosmological hydrodynamical simulations to quantify the biases in the cluster counts and cosmological parameters resulting from the different ingredients of conventional models. For the observable mass proxy we focus on the Compton-Y parameter quantifying the thermal Sunyaev-Zel'dovich effect, which is expected to result in cluster samples that are relatively close to mass-selected samples. We construct three mock surveys based on existing (Planck and SPT) and upcoming (Simons Observatory) surveys. We ignore measurement uncertainties and compare the biases in the counts and inferred cosmological parameters to each survey's Poisson errors. We find that widely used models for the DMO HMF differ significantly from each other and from the DMO version of FLAMINGO, leading to significant biases for all three surveys. For upcoming surveys, dramatic improvements are needed for all additional model ingredients, i.e. the functional forms of the fits to the observable-mass scaling relation and the associated scatter, the priors on the scaling relation and the prior on baryonic effects associated with feedback processes on the HMF., Comment: 17 pages, 8 figures. Submitted to MNRAS

We present Jamba-1.5, new instruction-tuned large language models based on our Jamba architecture. Jamba is a hybrid Transformer-Mamba mixture of experts architecture, providing high throughput and low memory usage across context lengths, while retaining the same or better quality as Transformer models. We release two model sizes: Jamba-1.5-Large, with 94B active parameters, and Jamba-1.5-Mini, with 12B active parameters. Both models are fine-tuned for a variety of conversational and instruction-following capabilties, and have an effective context length of 256K tokens, the largest amongst open-weight models. To support cost-effective inference, we introduce ExpertsInt8, a novel quantization technique that allows fitting Jamba-1.5-Large on a machine with 8 80GB GPUs when processing 256K-token contexts without loss of quality. When evaluated on a battery of academic and chatbot benchmarks, Jamba-1.5 models achieve excellent results while providing high throughput and outperforming other open-weight models on long-context benchmarks. The model weights for both sizes are publicly available under the Jamba Open Model License and we release ExpertsInt8 as open source., Comment: Webpage: https://www.ai21.com/jamba

The evolution of open quantum systems is a fundamental topic in various scientific fields. During time propagation, the environment occasionally makes measurements, forcing the system's wave function to collapse randomly. The von Neumann density matrix incorporates the statistics involved in these random processes, and its time development is often described by Markovian quantum master equations that incorporate a dissipator. For large systems, the complexity of the dissipator grows with the increasing number of possible measurements, posing conceptual and severe computational challenges. This article introduces a stochastic representation of the dissipator, using bundled measurement operators to address this complexity. Taking the Morse oscillator as an example, we demonstrate that small samples of bundled operators capture the system's dynamics. This stochastic bundling is different from the stochastic unraveling and the jump operator formalism and offers a new way of understanding quantum dissipation and decoherence.

This work addresses the challenge of quantifying originality in text-to-image (T2I) generative diffusion models, with a focus on copyright originality. We begin by evaluating T2I models' ability to innovate and generalize through controlled experiments, revealing that stable diffusion models can effectively recreate unseen elements with sufficiently diverse training data. Then, our key insight is that concepts and combinations of image elements the model is familiar with, and saw more during training, are more concisly represented in the model's latent space. We hence propose a method that leverages textual inversion to measure the originality of an image based on the number of tokens required for its reconstruction by the model. Our approach is inspired by legal definitions of originality and aims to assess whether a model can produce original content without relying on specific prompts or having the training data of the model. We demonstrate our method using both a pre-trained stable diffusion model and a synthetic dataset, showing a correlation between the number of tokens and image originality. This work contributes to the understanding of originality in generative models and has implications for copyright infringement cases., Comment: GenLaw ICML 2024

Shortcuts to adiabaticity are strategies for conserving adiabatic invariants under non-adiabatic (i.e. fast-driving) conditions. Here, we show how to extend classical, Hamiltonian shortcuts to adiabaticity to allow the crossing of a phase-space separatrix -- a situation in which a corresponding adiabatic protocol does not exist. Specifically, we show how to construct a time-dependent Hamiltonian that evolves one energy shell to another energy shell across a separatrix. Leveraging this method, we design an erasure procedure whose energy cost and fidelity do not depend on the protocol's duration.

We study the supersymmetric structure of the spin current in four dimensional $\mathcal{N}=1$ supersymmetric theories. By coupling the stress tensor multiplet to a vierbein multiplet we identify a spin current with the Hodge dual of the bottom component of the stress tensor multiplet in a wide range of theories. This implies that in holographic theories the Hodge dual of the $R$ current may serve as a spin current, paving the way for holographic studies of theories with background torsion.

We present a full-size engineering model of GALI - The GAmma-ray burst Localizing Instrument, composed of 362 CsI(Tl) small cubic scintillators, distributed within a small volume of $\sim2$l, and read out by silicon photo-multipliers. GALI can provide directional information about GRBs with high angular accuracy from angle-dependent mutual obstruction between its scintillators. Here, we demonstrate GALI's laboratory experiments with an $^{241}$Am source, which achieved directional reconstruction of $<$3$^\circ$ accuracy, in agreement with our Monte-Carlo simulations. GALI has a wide field view of the unobstructed sky. With its current cubic configuration, GALI's effective area varies between 97 cm$^2$ (face on) and 138 cm$^2$ (from the corners at 45$^\circ$), which is verified in the current experiment.

Recent advancements in NLP systems, particularly with the introduction of LLMs, have led to widespread adoption of these systems by a broad spectrum of users across various domains, impacting decision-making, the job market, society, and scientific research. This surge in usage has led to an explosion in NLP model interpretability and analysis research, accompanied by numerous technical surveys. Yet, these surveys often overlook the needs and perspectives of explanation stakeholders. In this paper, we address three fundamental questions: Why do we need interpretability, what are we interpreting, and how? By exploring these questions, we examine existing interpretability paradigms, their properties, and their relevance to different stakeholders. We further explore the practical implications of these paradigms by analyzing trends from the past decade across multiple research fields. To this end, we retrieved thousands of papers and employed an LLM to characterize them. Our analysis reveals significant disparities between NLP developers and non-developer users, as well as between research fields, underscoring the diverse needs of stakeholders. For example, explanations of internal model components are rarely used outside the NLP field. We hope this paper informs the future design, development, and application of methods that align with the objectives and requirements of various stakeholders.

Context. Mock galaxy catalogues are essential for correctly interpreting current and future generations of galaxy surveys. Despite their significance in galaxy formation and cosmology, little to no work has been done to validate the predictions of these mocks for high-order clustering statistics. Aims. We compare the predicting power of the latest generation of empirical models used in the creation of mock galaxy catalogues: a 13-parameter Halo Occupation Distribution (HOD) and an extension of the SubHalo Abundance Matching technique (SHAMe). Methods. We build GalaxyEmu-Planck, an emulator that makes precise predictions for the two-point correlation function, galaxy-galaxy lensing (restricted to distances greater than 1 $h^{-1} {\rm Mpc}$ to avoid baryonic effects), and other high-order statistics resulting from the evaluation of SHAMe and HOD models. Results. We evaluate the precision of GalaxyEmu-Planck using two galaxy samples extracted from the FLAMINGO hydrodynamical simulation that mimic the properties of DESI-BGS and BOSS galaxies, finding that the emulator reproduces all the predicted statistics precisely. The HOD showed comparable performance when fitting galaxy clustering and galaxy-galaxy lensing. In contrast, the SHAMe model showed better predictions for higher-order statistics, especially regarding the galaxy assembly bias. We also tested the performance of the models after removing some of their extensions, finding that we can withdraw two of the HOD parameters without a loss of performance. Conclusions. The results of this paper validate the current generation of empirical models as a way to reproduce galaxy clustering, galaxy-galaxy lensing and other high-order statistics. The excellent performance of the SHAMe model with a small number of free parameters suggests that it is a valid method to extract cosmological constraints from galaxy clustering., Comment: 22 pages, 12 figures, submitted to A&A. Comments welcomed!

This study concerns an inventory of the gravitational force and tidal field induced by filaments, walls, cluster nodes and voids on Megaparsec scales and how they assemble and shape the Cosmic Web. The study is based on a N$_{\rm Part}=512^3$ $\Lambda$CDM dark matter only N-body simulation in a (300$h^{-1}$~Mpc)$^3$ box at $z=0$. We invoke the density field NEXUS+ multiscale morphological procedure to assign the appropriate morphological feature to each location. We then determine the contribution by each of the cosmic web components to the local gravitational and tidal forces. We find that filaments are, by far, the dominant dynamical component in the interior of filaments, in the majority of underdense void regions and in all wall regions. The gravitational influence of cluster nodes is limited, and they are only dominant in their immediate vicinity. The force field induced by voids is marked by divergent outflowing patterns, yielding the impression of a segmented volume in which voids push matter towards their boundaries. Voids manifest themselves strongly in the tidal field as a cellular tapestry that is closely linked to the multiscale cosmic web. However, even within the interior of voids, the dynamical influence of the surrounding filaments is stronger than the outward push by voids. Therefore, the dynamics of voids cannot be understood without taking into account the influence of the environment. We conclude that filaments constitute the overpowering gravitational agent of the cosmic web, while voids are responsible for the cosmic web's spatial organisation and hence of its spatial connectivity., Comment: 33 pages, 21 figures (Including the appendix). Submitted to MNRAS

Measuring the effect of gravity on antimatter is a longstanding problem in physics that has significant implications for our understanding of the fundamental nature of the universe. Here, we present a technique to measure the gravitational mass of atoms, motivated by a recent measurement of antimatter atoms in CERN [1]. We demonstrate the results on ultra-cold atoms by measuring the surviving fraction of atoms gradually released from a quadrupole magnetic trap, which is tilted due to gravitational potential. We compare our measurements with a Monte Carlo simulation to extract the value of the gravitational constant. The difference between the literature value for g, the local acceleration due to gravity, and the measured value is $(-1.9 \pm 12^{stat} \pm 5^{syst}) \times 10^{- 4} g$. We demonstrate the importance of various design parameters in the experiment setup, and estimate their contribution to the achievable accuracy in future experiments. Our method demonstrates simplicity, precision, and reliability, paving the way for future precision studies of the gravitational force on antimatter. It also enables a precise calibration of atom traps based on the known gravitational attraction of normal matter to Earth.

Credit attribution is crucial across various fields. In academic research, proper citation acknowledges prior work and establishes original contributions. Similarly, in generative models, such as those trained on existing artworks or music, it is important to ensure that any generated content influenced by these works appropriately credits the original creators. We study credit attribution by machine learning algorithms. We propose new definitions--relaxations of Differential Privacy--that weaken the stability guarantees for a designated subset of $k$ datapoints. These $k$ datapoints can be used non-stably with permission from their owners, potentially in exchange for compensation. Meanwhile, the remaining datapoints are guaranteed to have no significant influence on the algorithm's output. Our framework extends well-studied notions of stability, including Differential Privacy ($k = 0$), differentially private learning with public data (where the $k$ public datapoints are fixed in advance), and stable sample compression (where the $k$ datapoints are selected adaptively by the algorithm). We examine the expressive power of these stability notions within the PAC learning framework, provide a comprehensive characterization of learnability for algorithms adhering to these principles, and propose directions and questions for future research., Comment: 16 pages, 1 figure. NeurIPS 2024 camera-ready version

This study empirically tests the $\textit{Narrative Economics}$ hypothesis, which posits that narratives (ideas that are spread virally and affect public beliefs) can influence economic fluctuations. We introduce two curated datasets containing posts from X (formerly Twitter) which capture economy-related narratives (Data will be shared upon paper acceptance). Employing Natural Language Processing (NLP) methods, we extract and summarize narratives from the tweets. We test their predictive power for $\textit{macroeconomic}$ forecasting by incorporating the tweets' or the extracted narratives' representations in downstream financial prediction tasks. Our work highlights the challenges in improving macroeconomic models with narrative data, paving the way for the research community to realistically address this important challenge. From a scientific perspective, our investigation offers valuable insights and NLP tools for narrative extraction and summarization using Large Language Models (LLMs), contributing to future research on the role of narratives in economics.

Blockchains implement monetary systems operated by committees of nodes. The robustness of established blockchains presents an opportunity to leverage their infrastructure for creating expansion chains. Expansion chains can provide additional functionality to the primary chain they leverage or implement separate functionalities, while benefiting from the primary chain's security and the stability of its tokens. Indeed, tools like Ethereum's EigenLayer enable nodes to stake (deposit collateral) on a primary chain to form a committee responsible for operating an expansion chain. But here is the rub. Classical protocols assume correct, well-behaved nodes stay correct indefinitely. Yet in our case, the stake incentivizes correctness--it will be slashed (revoked) if its owner deviates. Once a node withdraws its stake, there is no basis to assume its correctness. To address the new challenge, we present Aegis, an expansion chain based on primary-chain stake, assuming a bounded primary-chain write time. Aegis uses references from Aegis blocks to primary blocks to define committees, checkpoints on the primary chain to perpetuate decisions, and resets on the primary chain to establish a new committee if the previous one becomes obsolete. It ensures safety at all times and rapid progress when latency among Aegis nodes is low.

PLATO (PLAnetary Transits and Oscillations of stars) is ESA's M3 mission designed to detect and characterise extrasolar planets and perform asteroseismic monitoring of a large number of stars. PLATO will detect small planets (down to <2 R_(Earth)) around bright stars (<11 mag), including terrestrial planets in the habitable zone of solar-like stars. With the complement of radial velocity observations from the ground, planets will be characterised for their radius, mass, and age with high accuracy (5 %, 10 %, 10 % for an Earth-Sun combination respectively). PLATO will provide us with a large-scale catalogue of well-characterised small planets up to intermediate orbital periods, relevant for a meaningful comparison to planet formation theories and to better understand planet evolution. It will make possible comparative exoplanetology to place our Solar System planets in a broader context. In parallel, PLATO will study (host) stars using asteroseismology, allowing us to determine the stellar properties with high accuracy, substantially enhancing our knowledge of stellar structure and evolution. The payload instrument consists of 26 cameras with 12cm aperture each. For at least four years, the mission will perform high-precision photometric measurements. Here we review the science objectives, present PLATO's target samples and fields, provide an overview of expected core science performance as well as a description of the instrument and the mission profile at the beginning of the serial production of the flight cameras. PLATO is scheduled for a launch date end 2026. This overview therefore provides a summary of the mission to the community in preparation of the upcoming operational phases.

Galaxy clusters provide an avenue to expand our knowledge of cosmology and galaxy evolution. Because it is difficult to accurately measure the total mass of a large number of individual clusters, cluster samples are typically selected using an observable proxy for mass. Selection effects are therefore a key problem in understanding galaxy cluster statistics. We make use of the $(2.8~\rm{Gpc})^3$ FLAMINGO hydrodynamical simulation to investigate how selection based on X-ray luminosity, thermal Sunyaev-Zeldovich effect or galaxy richness influences the halo mass distribution. We define our selection cuts based on the median value of the observable at a fixed mass and compare the resulting samples to a mass-selected sample. We find that all samples are skewed towards lower mass haloes. For X-ray luminosity and richness cuts below a critical value, scatter dominates over the trend with mass and the median mass becomes biased increasingly low with respect to a mass-selected sample. At $z\leq0.5$, observable cuts corresponding to median halo masses between $M_\text{500c}=10^{14}$ and $10^{15}~\rm{M_{\odot}}$ give nearly unbiased median masses for all selection methods, but X-ray selection results in biased medians for higher masses. For cuts corresponding to median masses $<10^{14}$ at $z\leq0.5$ and for all masses at $z\geq1$, only Compton-Y selection yields nearly unbiased median masses. Importantly, even when the median mass is unbiased, the scatter is not because for each selection the sample is skewed towards lower masses than a mass-selected sample. Each selection leads to a different bias in secondary quantities like cool-core fraction, temperature and gas fraction., Comment: 19 pages, 12 figures (Including the appendix). Accepted for publication in MNRAS. Minor changes w.r.t. version 1

Located at the bottom of the main sequence, ultracool dwarf stars are widespread in the solar neighbourhood. Nevertheless, their extremely low luminosity has left their planetary population largely unexplored, and only one of them, TRAPPIST-1, has so far been found to host a transiting planetary system. In this context, we present the SPECULOOS project's detection of an Earth-sized planet in a 17 h orbit around an ultracool dwarf of M6.5 spectral type located 16.8 pc away. The planet's high irradiation (16 times that of Earth) combined with the infrared luminosity and Jupiter-like size of its host star make it one of the most promising rocky exoplanet targets for detailed emission spectroscopy characterization with JWST. Indeed, our sensitivity study shows that just ten secondary eclipse observations with the Mid-InfraRed Instrument/Low-Resolution Spectrometer on board JWST should provide strong constraints on its atmospheric composition and/or surface mineralogy.

We study the effects of correlations in a random environment on a random walker. The dependence of its asymptotic speed on the correlations is a nonperturbative effect as it is not captured by a homogeneous version of the same environment. For a slowly cooling environment, the buildup of correlations modifies the walker's speed and, by so, realizes acceleration. We remark on the possible relevance in the discussion of cosmic acceleration as traditionally started from the Friedmann equations, which, from a statistical mechanical point of view, would amount to a mean-field approximation. Our environment is much simpler though, with transition rates sampled from the one-dimensional Ising model and allowing exact results and detailed velocity characteristics.

Using drones to track multiple individuals simultaneously in their natural environment is a powerful approach for better understanding group primate behavior. Previous studies have demonstrated that it is possible to automate the classification of primate behavior from video data, but these studies have been carried out in captivity or from ground-based cameras. To understand group behavior and the self-organization of a collective, the whole troop needs to be seen at a scale where behavior can be seen in relation to the natural environment in which ecological decisions are made. This study presents a novel dataset from drone videos for baboon detection, tracking, and behavior recognition. The baboon detection dataset was created by manually annotating all baboons in drone videos with bounding boxes. A tiling method was subsequently applied to create a pyramid of images at various scales from the original 5.3K resolution images, resulting in approximately 30K images used for baboon detection. The tracking dataset is derived from the detection dataset, where all bounding boxes are assigned the same ID throughout the video. This process resulted in half an hour of very dense tracking data. The behavior recognition dataset was generated by converting tracks into mini-scenes, a video subregion centered on each animal; each mini-scene was manually annotated with 12 distinct behavior types, resulting in over 20 hours of data. Benchmark results show mean average precision (mAP) of 92.62\% for the YOLOv8-X detection model, multiple object tracking precision (MOTA) of 63.81\% for the BotSort tracking algorithm, and micro top-1 accuracy of 63.97\% for the X3D behavior recognition model. Using deep learning to classify wildlife behavior from drone footage facilitates non-invasive insight into the collective behavior of an entire group., Comment: Dataset will be published shortly

When large language models are aligned via supervised fine-tuning, they may encounter new factual information that was not acquired through pre-training. It is often conjectured that this can teach the model the behavior of hallucinating factually incorrect responses, as the model is trained to generate facts that are not grounded in its pre-existing knowledge. In this work, we study the impact of such exposure to new knowledge on the capability of the fine-tuned model to utilize its pre-existing knowledge. To this end, we design a controlled setup, focused on closed-book QA, where we vary the proportion of the fine-tuning examples that introduce new knowledge. We demonstrate that large language models struggle to acquire new factual knowledge through fine-tuning, as fine-tuning examples that introduce new knowledge are learned significantly slower than those consistent with the model's knowledge. However, we also find that as the examples with new knowledge are eventually learned, they linearly increase the model's tendency to hallucinate. Taken together, our results highlight the risk in introducing new factual knowledge through fine-tuning, and support the view that large language models mostly acquire factual knowledge through pre-training, whereas fine-tuning teaches them to use it more efficiently., Comment: Accepted as a long paper at EMNLP 2024

The resonant conversion, within the inter-galactic medium, of regular photons into dark photons amplifies the anisotropy observed in the CMB, thereby imposing stringent constraints on the existence of light dark photons. In this study, we investigate the impact of light dark photons, with masses in the range $3\times 10^{-15} ~\rm{eV} < m_{A'} < 3\times 10^{-12}~\rm{eV}$ on the power spectrum of temperature anisotropies within the cosmic microwave background (CMB) radiation utilizing the state-of-the-art large-volume FLAMINGO cosmological simulations. Our results show that using full Planck data, one can expect the existing constraints on the dark photon mixing parameter in this mass range to improve by an order of magnitude., Comment: 16 pages, 8 figures

Folding can transform mundane objects such as napkins into stunning works of art. However, finding new folding transformations for sheet materials is a challenging problem that requires expertise and real-world experimentation. In this paper, we present Modal Folding -- an automated approach for discovering energetically optimal folding transformations, i.e., large deformations that require little mechanical work. For small deformations, minimizing internal energy for fixed displacement magnitudes leads to the well-known elastic eigenmodes. While linear modes provide promising directions for bending, they cannot capture the rotational motion required for folding. To overcome this limitation, we introduce strain-space modes -- nonlinear analogues of elastic eigenmodes that operate on per-element curvatures instead of vertices. Using strain-space modes to determine target curvatures for bending elements, we can generate complex nonlinear folding motions by simply minimizing the sheet's internal energy. Our modal folding approach offers a systematic and automated way to create complex designs. We demonstrate the effectiveness of our method with simulation results for a range of shapes and materials, and validate our designs with physical prototypes., Comment: 9 pages, SIGGRAPH 2024 Conference

Automatic conversion of free-text radiology reports into structured data using Natural Language Processing (NLP) techniques is crucial for analyzing diseases on a large scale. While effective for tasks in widely spoken languages like English, generative large language models (LLMs) typically underperform with less common languages and can pose potential risks to patient privacy. Fine-tuning local NLP models is hindered by the skewed nature of real-world medical datasets, where rare findings represent a significant data imbalance. We introduce SMP-BERT, a novel prompt learning method that leverages the structured nature of reports to overcome these challenges. In our studies involving a substantial collection of Crohn's disease radiology reports in Hebrew (over 8,000 patients and 10,000 reports), SMP-BERT greatly surpassed traditional fine-tuning methods in performance, notably in detecting infrequent conditions (AUC: 0.99 vs 0.94, F1: 0.84 vs 0.34). SMP-BERT empowers more accurate AI diagnostics available for low-resource languages.

In recent years, the challenge of extracting information from business documents has emerged as a critical task, finding applications across numerous domains. This effort has attracted substantial interest from both industry and academy, highlighting its significance in the current technological landscape. Most datasets in this area are primarily focused on Key Information Extraction (KIE), where the extraction process revolves around extracting information using a specific, predefined set of keys. Unlike most existing datasets and benchmarks, our focus is on discovering key-value pairs (KVPs) without relying on predefined keys, navigating through an array of diverse templates and complex layouts. This task presents unique challenges, primarily due to the absence of comprehensive datasets and benchmarks tailored for non-predetermined KVP extraction. To address this gap, we introduce KVP10k , a new dataset and benchmark specifically designed for KVP extraction. The dataset contains 10707 richly annotated images. In our benchmark, we also introduce a new challenging task that combines elements of KIE as well as KVP in a single task. KVP10k sets itself apart with its extensive diversity in data and richly detailed annotations, paving the way for advancements in the field of information extraction from complex business documents., Comment: accepted ICDAR2024

We study the formation and properties of perfectly-flat zero energy bands in a multi-layer graphene systems in the chiral limit. Employing the degrees of freedoms of the multi-layer system, such as relative twist-angle and relative shifts, in a way that preserves a set of symmetries, we define a two-dimensional parameter plane that hosts lines of two and four flat bands. This plane enables adiabatic continuation of multi-layer chiral systems to weakly coupled bi- and tri-layer systems, and through that mapping provides tools for calculating the Chern numbers of the flat bands. We show that a flat band of Chern number $C$ can be spanned by $C$ effective Landau levels, all experiencing an effective flux of $1/C$ flux quantum per unit cell, and each carrying its own intra-unit-cell wave function. The flat bands do not disperse under the effect of a perpendicular magnetic field, and the gap to the dispersive bands closes when the externally applied flux cancels the $1/C$ effective flux., Comment: 16 pages, 11 figures

Background: Recent advancements in Artificial Intelligence (AI) contributed significantly to suicide assessment, however, our theoretical understanding of this complex behavior is still limited. Objective: This study aimed to harness AI methodologies to uncover hidden risk factors that trigger or aggravate suicide behaviors. Method: The primary dataset included 228,052 Facebook postings by 1,006 users who completed the gold-standard Columbia Suicide Severity Rating Scale. This dataset was analyzed using a bottom-up research pipeline without a-priory hypotheses and its findings were validated using a top-down analysis of a new dataset. This secondary dataset included responses by 1,062 participants to the same suicide scale as well as to well-validated scales measuring depression and boredom. Results: An almost fully automated, AI-guided research pipeline resulted in four Facebook topics that predicted the risk of suicide, of which the strongest predictor was boredom. A comprehensive literature review using APA PsycInfo revealed that boredom is rarely perceived as a unique risk factor of suicide. A complementing top-down path analysis of the secondary dataset uncovered an indirect relationship between boredom and suicide, which was mediated by depression. An equivalent mediated relationship was observed in the primary Facebook dataset as well. However, here, a direct relationship between boredom and suicide risk was also observed. Conclusions: Integrating AI methods allowed the discovery of an under-researched risk factor of suicide. The study signals boredom as a maladaptive 'ingredient' that might trigger suicide behaviors, regardless of depression. Further studies are recommended to direct clinicians' attention to this burdening, and sometimes existential experience.

Stochastic orbital techniques offer reduced computational scaling and memory requirements to describe ground and excited states at the cost of introducing controlled statistical errors. Such techniques often rely on two basic operations, stochastic trace estimation and stochastic resolution of identity, both of which lead to statistical errors that scale with the number of stochastic realizations ($N_{\xi}$) as $\sqrt{N_{\xi}^{-1}}$. Reducing the statistical errors without significantly increasing $N_{\xi}$ has been challenging and is central to the development of efficient and accurate stochastic algorithms. In this work, we build upon recent progress made to improve stochastic trace estimation based on the ubiquitous Hutchinson's algorithm and propose a two-step approach for the stochastic resolution of identity, in the spirit of the Hutch++ method. Our approach is based on employing a randomized low-rank approximation followed by a residual calculation, resulting in statistical errors that scale much better than $\sqrt{N_{\xi}^{-1}}$. We implement the approach within the second-order Born approximation for the self-energy in the computation of neutral excitations and discuss three different low-rank approximations for the two-body Coulomb integrals. Tests on a series of hydrogen dimer chains with varying lengths demonstrate that the Hutch++-like approximations are computationally more efficient than both deterministic and purely stochastic (Hutchinson) approaches for low error thresholds and intermediate system sizes. Notably, for arbitrarily large systems, the Hutchinson-like approximation outperforms both deterministic and Hutch++-like methods.

The relativistic Sunyaev-Zel'dovich (SZ) effect can be used to measure intracluster gas temperatures independently of X-ray spectroscopy. Here, we use the large-volume FLAMINGO simulation suite to determine whether SZ $y$-weighted temperatures lead to more accurate hydrostatic mass estimates in massive ($M_{\rm 500c} > 7.5\times 10^{14}\,{\rm M}_{\odot}$) clusters than when using X-ray spectroscopic-like temperatures. We find this to be the case, on average. The median bias in the SZ mass at redshift zero is $\left< b \right> \equiv 1-\left< M_{\rm 500c,hse}/M_{\rm 500c,true} \right> = -0.05 \pm 0.01$, over 4 times smaller in magnitude than the X-ray spectroscopic-like case, $\left< b \right> = 0.22 \pm 0.01$. However, the scatter in the SZ bias, $\sigma_{b} \approx 0.2$, is around 40 per cent larger than for the X-ray case. We show that this difference is strongly affected by clusters with large pressure fluctuations, as expected from shocks in ongoing mergers. Selecting the clusters with the best-fitting generalized NFW pressure profiles, the median SZ bias almost vanishes, $\left< b \right> = -0.009 \pm 0.005$, and the scatter is halved to $\sigma_{b} \approx 0.1$. We study the origin of the SZ/X-ray difference and find that, at $R_{\rm 500c}$ and in the outskirts, SZ weighted gas better reflects the hot, hydrostatic atmosphere than the X-ray weighted gas. The SZ/X-ray temperature ratio increases with radius, a result we find to be insensitive to variations in baryonic physics, cosmology and numerical resolution., Comment: 19 pages, 22 figures, submitted to MNRAS. Comments welcome

We analyze the sample complexity of full-batch Gradient Descent (GD) in the setup of non-smooth Stochastic Convex Optimization. We show that the generalization error of GD, with common choice of hyper-parameters, can be $\tilde \Theta(d/m + 1/\sqrt{m})$, where $d$ is the dimension and $m$ is the sample size. This matches the sample complexity of \emph{worst-case} empirical risk minimizers. That means that, in contrast with other algorithms, GD has no advantage over naive ERMs. Our bound follows from a new generalization bound that depends on both the dimension as well as the learning rate and number of iterations. Our bound also shows that, for general hyper-parameters, when the dimension is strictly larger than number of samples, $T=\Omega(1/\epsilon^4)$ iterations are necessary to avoid overfitting. This resolves an open problem by Schlisserman et al.23 and Amir er Al.21, and improves over previous lower bounds that demonstrated that the sample size must be at least square root of the dimension.

The advent of Generative Artificial Intelligence (GenAI) models, including GitHub Copilot, OpenAI GPT, and Stable Diffusion, has revolutionized content creation, enabling non-professionals to produce high-quality content across various domains. This transformative technology has led to a surge of synthetic content and sparked legal disputes over copyright infringement. To address these challenges, this paper introduces a novel approach that leverages the learning capacity of GenAI models for copyright legal analysis, demonstrated with GPT2 and Stable Diffusion models. Copyright law distinguishes between original expressions and generic ones (Sc\`enes \`a faire), protecting the former and permitting reproduction of the latter. However, this distinction has historically been challenging to make consistently, leading to over-protection of copyrighted works. GenAI offers an unprecedented opportunity to enhance this legal analysis by revealing shared patterns in preexisting works. We propose a data-driven approach to identify the genericity of works created by GenAI, employing "data-driven bias" to assess the genericity of expressive compositions. This approach aids in copyright scope determination by utilizing the capabilities of GenAI to identify and prioritize expressive elements and rank them according to their frequency in the model's dataset. The potential implications of measuring expressive genericity for copyright law are profound. Such scoring could assist courts in determining copyright scope during litigation, inform the registration practices of Copyright Offices, allowing registration of only highly original synthetic works, and help copyright owners signal the value of their works and facilitate fairer licensing deals. More generally, this approach offers valuable insights to policymakers grappling with adapting copyright law to the challenges posed by the era of GenAI., Comment: Presented at ACM CSLAW 2024

Large-scale structure surveys have reported measurements of the density of matter, $\Omega_\mathrm{m}$, and the amplitude of clustering, $\sigma_8$, that are in tension with the values inferred from observations of the cosmic microwave background. While this may be a sign of new physics that slows the growth of structure at late times, strong astrophysical feedback processes could also be responsible. In this work, we argue that astrophysical processes are not independent of cosmology and that their coupling naturally leads to stronger baryonic feedback in cosmological models with suppressed structure formation or when combined with a mechanism that removes dark matter from halos. We illustrate this with two well-motivated extensions of the Standard Model known to suppress structure formation: massive neutrinos and decaying dark matter. Our results, based on the FLAMINGO suite of hydrodynamical simulations, show that the combined effect of baryonic and non-baryonic suppression mechanisms is greater than the sum of its parts, particularly for decaying dark matter. We also show that the dependence of baryonic feedback on cosmology can be modelled as a function of the ratio $f_\mathrm{b}/c^2_\mathrm{v}\sim f_\mathrm{b}/(\Omega_\mathrm{m}\sigma_8)^{1/4}$ of the universal baryon fraction, $f_\mathrm{b}$, to a velocity-based definition of halo concentration, $c^2_\mathrm{v}$, giving an accurate fitting formula for the baryonic suppression of the matter power spectrum. Although the combination of baryonic and non-baryonic suppression mechanisms can resolve the tension, the models with neutrinos and decaying dark matter are challenged by constraints on the expansion history., Comment: 16 pages, 9 figures, submitted to MNRAS

Significant uncertainty in climate prediction and cloud physics is tied to observational gaps relating to shallow scattered clouds. Addressing these challenges requires remote sensing of their three-dimensional (3D) heterogeneous volumetric scattering content. This calls for passive scattering computed tomography (CT). We design a learning-based model (ProbCT) to achieve CT of such clouds, based on noisy multi-view spaceborne images. ProbCT infers - for the first time - the posterior probability distribution of the heterogeneous extinction coefficient, per 3D location. This yields arbitrary valuable statistics, e.g., the 3D field of the most probable extinction and its uncertainty. ProbCT uses a neural-field representation, making essentially real-time inference. ProbCT undergoes supervised training by a new labeled multi-class database of physics-based volumetric fields of clouds and their corresponding images. To improve out-of-distribution inference, we incorporate self-supervised learning through differential rendering. We demonstrate the approach in simulations and on real-world data, and indicate the relevance of 3D recovery and uncertainty to precipitation and renewable energy.

Motivated by the recent JWST discovery of galaxy overdensities during the Epoch of Reionzation, we examine the physical properties of high-$z$ protoclusters and their evolution using the FLAMINGO simulation suite. We investigate the impact of the apertures used to define protoclusters, because the heterogeneous apertures used in the literature have limited our understanding of the population. Our results are insensitive to the uncertainties of the subgrid models at a given resolution, whereas further investigation into the dependence on numerical resolution is needed. When considering galaxies more massive than $M_\ast\,{\simeq}\,10^8\,{\rm M_\odot}$, the FLAMINGO simulations predict a dominant contribution from progenitors similar to those of the Coma cluster to the cosmic star-formation rate density during the reionization epoch. Our results indicate the onset of suppression of star formation in the protocluster environments as early as $z\,{\simeq}\,5$. The galaxy number density profiles are similar to NFW at $z\,{\lesssim}\,1$ while showing a steeper slope at earlier times before the formation of the core. Different from most previous simulations, the predicted star-formation history for individual protoclusters is in good agreement with observations. We demonstrate that, depending on the aperture, the integrated physical properties including the total (dark matter and baryonic) mass can be biased by a factor of 2 to 5 at $z\,{=}\,5.5$--$7$, and by an order of magnitude at $z\,{\lesssim}\,4$. This correction suffices to remove the ${\simeq}\,3\,\sigma$ tensions with the number density of structures found in recent JWST observations., Comment: 22 pages, 15 figures, accepted for publication in MNRAS

Queer youth face increased mental health risks, such as depression, anxiety, and suicidal ideation. Hindered by negative stigma, they often avoid seeking help and rely on online resources, which may provide incompatible information. Although access to a supportive environment and reliable information is invaluable, many queer youth worldwide have no access to such support. However, this could soon change due to the rapid adoption of Large Language Models (LLMs) such as ChatGPT. This paper aims to comprehensively explore the potential of LLMs to revolutionize emotional support for queers. To this end, we conduct a qualitative and quantitative analysis of LLM's interactions with queer-related content. To evaluate response quality, we develop a novel ten-question scale that is inspired by psychological standards and expert input. We apply this scale to score several LLMs and human comments to posts where queer youth seek advice and share experiences. We find that LLM responses are supportive and inclusive, outscoring humans. However, they tend to be generic, not empathetic enough, and lack personalization, resulting in nonreliable and potentially harmful advice. We discuss these challenges, demonstrate that a dedicated prompt can improve the performance, and propose a blueprint of an LLM-supporter that actively (but sensitively) seeks user context to provide personalized, empathetic, and reliable responses. Our annotated dataset is available for further research.