Search

As large language models (LLMs) demonstrate increasingly advanced capabilities, aligning their behaviors with human values and preferences becomes crucial for their wide adoption. While previous research focuses on general alignment to principles such as helpfulness, harmlessness, and honesty, the need to account for individual and diverse preferences has been largely overlooked, potentially undermining customized human experiences. To address this gap, we train LLMs that can ''interact to align'', essentially cultivating the meta-skill of LLMs to implicitly infer the unspoken personalized preferences of the current user through multi-turn conversations, and then dynamically align their following behaviors and responses to these inferred preferences. Our approach involves establishing a diverse pool of 3,310 distinct user personas by initially creating seed examples, which are then expanded through iterative self-generation and filtering. Guided by distinct user personas, we leverage multi-LLM collaboration to develop a multi-turn preference dataset containing 3K+ multi-turn conversations in tree structures. Finally, we apply supervised fine-tuning and reinforcement learning to enhance LLMs using this dataset. For evaluation, we establish the ALOE (ALign With CustOmized PrEferences) benchmark, consisting of 100 carefully selected examples and well-designed metrics to measure the customized alignment performance during conversations. Experimental results demonstrate the effectiveness of our method in enabling dynamic, personalized alignment via interaction., Comment: The code and dataset are made public at https://github.com/ShujinWu-0814/ALOE

Megathrust earthquakes are the largest on Earth, capable of causing strong ground shaking and generating tsunamis. Physical models used to understand megathrust earthquake hazard are limited by existing uncertainties about material properties and governing processes in subduction zones. A key quantity in megathrust hazard assessment is the distance between the updip and downdip rupture limits. The thermal structure of a subduction zone exerts a first-order control on the extent of rupture. We simulate temperature for profiles of the Cascadia, Nankai and Hikurangi subduction zones using a 2D coupled kinematic-dynamic thermal model. We then build reduced-order models (ROMs) for temperature using the interpolated Proper Orthogonal Decomposition (iPOD). The resulting ROMs are data-driven, model agnostic, and computationally cheap to evaluate. Using the ROMs, we can efficiently investigate the sensitivity of temperature to input parameters, physical processes, and modeling choices. We find that temperature, and by extension the potential rupture extent, is most sensitive to variability in parameters that describe shear heating on the slab interface, followed by parameters controlling the thermal structure of the incoming lithosphere and coupling between the slab and the mantle. We quantify the effect of using steady-state vs. time-dependent models, and of uncertainty in the choice of isotherm representing the downdip rupture limit. We show that variability in input parameters translates to significant differences in estimated moment magnitude. Our analysis highlights the strong effect of variability in the apparent coefficient of friction, with previously published ranges resulting in pronounced variability in estimated rupture limit depths.

So far, even the highest resolution galaxy formation simulations with gravitational softening have failed to reproduce realistic life cycles of star clusters. We present the first star-by-star galaxy models of star cluster formation to account for hydrodynamics, star formation, stellar evolution and collisional gravitational interactions between stars and compact remnants using the updated SPHGAL+KETJU code, part of the GRIFFIN-project. Gravitational dynamics in the vicinity of $>3$ M$_\odot$ stars and their remnants are solved with a regularised integrator (KETJU) without gravitational softening. Comparisons of idealised star cluster evolution with SPHGAL+KETJU and direct N-body show broad agreement and the failure of simulations that use gravitational softening. In the hydrodynamical dwarf galaxy simulations run with SPHGAL+KETJU, clusters up to $\sim900$ M$_\odot$ are formed compact (effective radii $0.1-1$ pc) and their sizes increase by up to a factor of ten in agreement with previous N-body simulations and the observed sizes of exposed star clusters. The sizes increase rapidly once the clusters become exposed due to photoionising radiation. On average $63\%$ of the gravitationally bound clusters disrupt during the first $100$ Myr of evolution in the galactic tidal field. The addition of collisional dynamics reduces the fraction of supernovae in bound clusters by a factor of $\sim 2.6$, however the global star formation and outflow histories change by less than $30\%$. We demonstrate that the accurate treatment of gravitational encounters with massive stars enables more realistic star cluster life cycles from the earliest stages of cluster formation until disruption in simulated low-mass galaxies., Comment: 20 pages, 12 figures, submitted to MNRAS, comments welcome

We study real and positive tropicalizations of the varieties of low rank symmetric matrices over real or complex Puiseux series. We show that real tropicalization coincides with complex tropicalization for rank two and corank one cases. We also show that the two notions of positive tropicalization introduced by Speyer and Williams coincide for symmetric rank two matrices, but they differ for symmetric corank one matrices.

The dual phosphorylation network provides an essential component of intracellular signaling, affecting the expression of phenotypes and cell metabolism. For particular choices of kinetic parameters, this system exhibits multistationarity, a property that is relevant in the decision-making of cells. Determining which reaction rate constants correspond to monostationarity and which produce multistationarity is an open problem. The system's monostationarity is linked to the nonnegativity of a specific polynomial. A previous study by Feliu et al. provides a sufficient condition for monostationarity via a decomposition of this polynomial into nonnegative circuit polynomials. However, this decomposition is not unique. We extend their work by a systematic approach to classifying such decompositions in the dual phosphorylation network. Using this result classification, we provide a qualitative comparison of the decompositions into nonnegative circuit polynomials via empirical experiments and improve on previous conditions for the region of monostationarity., Comment: 20 pages, 9 figures, 2 tables

Current instruction-tuned language models are exclusively trained with textual preference data and thus are often not aligned with the unique requirements of other modalities, such as speech. To better align language models with the speech domain, we explore (i) prompting strategies grounded in radio-industry best practices and (ii) preference learning using a novel speech-based preference data of 20K samples, generated with a wide spectrum of prompts that induce varying dimensions of speech-suitability and labeled by annotators who listen to response pairs. Both human and automatic evaluation show that both prompting and preference learning increase the speech-suitability of popular instruction-tuned LLMs. Interestingly, we find that prompting and preference learning can be additive; combining them achieves the best win rates in head-to-head comparison, resulting in responses that are preferred or tied to the base model in 76.2% of comparisons on average. Lastly, we share lexical, syntactical, and qualitative analyses to showcase how each method contributes to improving the speech-suitability of generated responses., Comment: EMNLP2024

Propaganda plays a critical role in shaping public opinion and fueling disinformation. While existing research primarily focuses on identifying propaganda techniques, it lacks the ability to capture the broader motives and the impacts of such content. To address these challenges, we introduce propainsight, a conceptual framework grounded in foundational social science research, which systematically dissects propaganda into techniques, arousal appeals, and underlying intent. propainsight offers a more granular understanding of how propaganda operates across different contexts. Additionally, we present propagaze, a novel dataset that combines human-annotated data with high-quality synthetic data generated through a meticulously designed pipeline. Our experiments show that off-the-shelf LLMs struggle with propaganda analysis, but training with propagaze significantly improves performance. Fine-tuned Llama-7B-Chat achieves 203.4% higher text span IoU in technique identification and 66.2% higher BertScore in appeal analysis compared to 1-shot GPT-4-Turbo. Moreover, propagaze complements limited human-annotated data in data-sparse and cross-domain scenarios, showing its potential for comprehensive and generalizable propaganda analysis., Comment: 8 pages

Flavor development in the food industry is increasingly challenged by the need for rapid innovation and precise flavor profile creation. Traditional flavor research methods typically rely on iterative, subjective testing, which lacks the efficiency and scalability required for modern demands. This paper presents three contributions to address the challenges. Firstly, we define a new problem domain for scientific agents in flavor science, conceptualized as the generation of hypotheses for flavor profile sourcing and understanding. To facilitate research in this area, we introduce the FoodPuzzle, a challenging benchmark consisting of 978 food items and 1,766 flavor molecules profiles. We propose a novel Scientific Agent approach, integrating in-context learning and retrieval augmented techniques to generate grounded hypotheses in the domain of food science. Experimental results indicate that our model significantly surpasses traditional methods in flavor profile prediction tasks, demonstrating its potential to transform flavor development practices.

We study the constant $\mathscr{C}_p$ defined as the smallest constant $C$ such that $|f(0)|^p \leq C\|f\|_p^p$ holds for every function $f$ in the Paley-Wiener space $PW^p$. Brevig, Chirre, Ortega-Cerd\`a, and Seip have recently shown that $\mathscr{C}_p

2$. We improve this bound for $2

We study the implications of relaxing the requirement for ultralight axions to account for all dark matter in the Universe by examining mixed dark matter (MDM) cosmologies with axion fractions $f \leq 0.3$ within the fuzzy dark matter (FDM) window $10^{-25}$ eV $\lesssim m \lesssim 10^{-23}$ eV. Our simulations, using a new MDM gravity solver implemented in AxiREPO, capture wave dynamics across various scales with high accuracy down to redshifts $z\approx 1$. We identify halos with Rockstar using the CDM component and find good agreement of inferred halo mass functions (HMFs) and concentration-mass relations with theoretical models across redshifts $z=1-10$. This justifies our halo finder approach a posteriori as well as the assumptions underlying the MDM halo model AxionHMcode. Using the inferred axion halo mass - cold halo mass relation $M_{\text{a}}(M_{\text{c}})$ and calibrating a generalised smoothing parameter $\alpha$ to our MDM simulations, we present a new version of AxionHMcode. The code exhibits excellent agreement with simulations on scales $k< 20 \ h$ cMpc$^{-1}$ at redshifts $z=1-3.5$ for $f\leq 0.1$ around the fiducial axion mass $m = 10^{-24.5}$ eV $ = 3.16\times 10^{-25}$ eV, with maximum deviations remaining below 10%. For axion fractions $f\leq 0.3$, the model maintains accuracy with deviations under 20% at redshifts $z\approx 1$ and scales $k< 10 \ h$ cMpc$^{-1}$, though deviations can reach up to 30% for higher redshifts when $f=0.3$. Reducing the run-time for a single evaluation of AxionHMcode to below $1$ minute, these results highlight the potential of AxionHMcode to provide a robust framework for parameter sampling across MDM cosmologies in Bayesian constraint and forecast analyses., Comment: 20 pages, 10 figures, 5 Tables, comments welcome

We propose fractional anomalous Hall crystals (FAHCs) as possible ground states of strongly interacting electrons in parent bands with Berry curvature. FAHCs are exotic states of matter that spontaneously break continuous translation symmetry to form a fractional Chern insulator. We construct a unified family of variational wavefunctions that describe FAHCs and their competing states in the presence of uniform parent Berry curvature. We calculate their variational energy with Coulomb interactions semi-analytically in the thermodynamic limit. Our analysis reveals that FAHCs can be energetically favorable over both Wigner crystals and integer anomalous Hall crystals for sufficiently strong interactions or flat dispersion.

Elastodynamic Green's functions are an essential ingredient in seismology as they form the connection between direct observations of seismic waves and the earthquake source. They are also fundamental to various seismological techniques including physics-based ground motion prediction and kinematic or dynamic source inversions. In regions with established 3D models of the Earth's elastic structure, 3D Green's functions can be computed using numerical simulations of seismic wave propagation. However, such simulations are computationally expensive which poses challenges for real-time ground motion prediction. Here, we use a reduced-order model (ROM) approach that enables the rapid evaluation of approximate Green's functions. The ROM technique developed approximates three-component surface velocity wavefields obtained from numerical simulations of seismic wave propagation. We apply our ROM approach to a 50 km x 40 km area in the greater Los Angeles area accounting for topography, site effects, 3D subsurface velocity structure, and viscoelastic attenuation. The ROM constructed for this region enables rapid computation (0.001 CPU hours) of complete, high-resolution, 0.5 Hz surface velocity wavefields that are accurate for a shortest wavelength of 1.0 km. Using leave-one-out cross validation, we measure the accuracy of our Green's functions in both the time-domain and frequency-domain. Averaged across all sources and receivers, the error in the rapid seismograms is less than 0.01 cm/s. We demonstrate that the ROM can accurately and rapidly reproduce simulated seismograms for generalized moment tensor sources in our region, as well as kinematic sources by using a finite fault model of the 1987 Mw 5.9 Whittier Narrows earthquake as an example. We envision that our rapid, approximate Green's functions will be useful for constructing rapid ground motion synthetics with high spatial resolution.

We study transits of several ``hot Jupiter'' systems - including WASP-12 b, WASP-43 b, WASP-103 b, HAT-P-23 b, KELT-16 b, WD 1856+534 b, and WTS-2 b - with the goal of detecting tidal orbital decay and extending the baselines of transit times. We find no evidence of orbital decay in any of the observed systems except for that of the extensively studied WASP-12 b. Although the orbit of WASP-12 b is unequivocally decaying, we find no evidence for acceleration of said orbital decay, with measured $\ddot{P} = (-7 \pm 8) \times 10^{-14} \rm ~s^{-1}$, against the expected acceleration decay of $\ddot{P} \approx -10^{-23} \rm ~s^{-1}$. In the case of WD 1856+534 b, there is a tentative detection of orbital growth with $\dot{P} = (5.0 \pm 1.5) \times 10^{-10}$. While statistically significant, we err on the side of caution and wait for longer follow-up observations to consider the measured $\dot{P}$ real. For most systems, we provide a 95\%-confidence lower limit on the tidal quality factor, $Q_\star'$. The possibility of detecting orbital decay in hot Jupiters via long-term radial velocity (RV) measurements is also explored. We find that $\sim 1 \rm ~m~s^{-1}$ precision in RVs will be required to detect orbital decay of WASP-12 b with only 3 yr of observations. Currently available RV measurements and precision are unable to detect orbital decay in any of the systems studied here., Comment: 14 pages, 11 figures; Accepted in MNRAS on 2024 August 30. Received 2024 August 29; in original form 2024 February 13

The integration of Large Language Models (LLMs) into the drug discovery and development field marks a significant paradigm shift, offering novel methodologies for understanding disease mechanisms, facilitating drug discovery, and optimizing clinical trial processes. This review highlights the expanding role of LLMs in revolutionizing various stages of the drug development pipeline. We investigate how these advanced computational models can uncover target-disease linkage, interpret complex biomedical data, enhance drug molecule design, predict drug efficacy and safety profiles, and facilitate clinical trial processes. Our paper aims to provide a comprehensive overview for researchers and practitioners in computational biology, pharmacology, and AI4Science by offering insights into the potential transformative impact of LLMs on drug discovery and development.

Linear RNN architectures, like Mamba, can be competitive with Transformer models in language modeling while having advantageous deployment characteristics. Given the focus on training large-scale Transformer models, we consider the challenge of converting these pretrained models for deployment. We demonstrate that it is feasible to distill large Transformers into linear RNNs by reusing the linear projection weights from attention layers with academic GPU resources. The resulting hybrid model, which incorporates a quarter of the attention layers, achieves performance comparable to the original Transformer in chat benchmarks and outperforms open-source hybrid Mamba models trained from scratch with trillions of tokens in both chat benchmarks and general benchmarks. Moreover, we introduce a hardware-aware speculative decoding algorithm that accelerates the inference speed of Mamba and hybrid models. Overall we show how, with limited computation resources, we can remove many of the original attention layers and generate from the resulting model more efficiently. Our top-performing model, distilled from Llama3-8B-Instruct, achieves a 29.61 length-controlled win rate on AlpacaEval 2 against GPT-4 and 7.35 on MT-Bench, surpassing the best instruction-tuned linear RNN model., Comment: Code is open-sourced at https://github.com/jxiw/MambaInLlama

Neutron beam, being electrically neutral and highly penetrating, offers unique advantages for irradiation of biological species such as plants, seeds, and microorganisms. We comprehensively investigate the potential of neutron irradiation for inducing genetic mutations using simulations of J-PARC BL10, JRR-3 TNRF, and KUANS for spallation, reactor, and compact neutron sources. We analyze neutron flux, energy deposition rates, and Linear Energy Transfer (LET) distributions. KUANS demonstrated the highest dose rate of 17 Gy/h, significantly surpassing BL10, due to the large solid angle by the optimal sample placement. The findings highlight KUANS's suitability for efficient genetic mutations and neutron breeding, particularly for inducing targeted mutations in biological samples. The LET range of KUANS is concentrated in 20-70 keV/{\mu}m, which is potentially ideal for inducing specific genetic mutations. The importance of choosing neutron sources based on LET requirements to maximize mutation induction efficiency is emphasized. This research shows the potential of compact neutron sources like KUANS for effective biological irradiation and neutron breeding, offering a viable alternative to larger facilities. The neutron filters used in BL10 and TNRF effectively excluded low-energy neutrons with keeping the high LET component. The neutron capture reaction, 14N(n,p)14C, was found to be the main dose under thermal neutron-dominated conditions., Comment: 18 pages, 6 Figures, 1 Table Category: physics.acc-ph (Accelerator Physics)

SPIDER is a balloon-borne instrument designed to map the cosmic microwave background at degree-angular scales in the presence of Galactic foregrounds. SPIDER has mapped a large sky area in the Southern Hemisphere using more than 2000 transition-edge sensors (TESs) during two NASA Long Duration Balloon flights above the Antarctic continent. During its first flight in January 2015, SPIDER observed in the 95 GHz and 150 GHz frequency bands, setting constraints on the B-mode signature of primordial gravitational waves. Its second flight in the 2022-23 season added new receivers at 280 GHz, each using an array of TESs coupled to the sky through feedhorns formed from stacks of silicon wafers. These receivers are optimized to produce deep maps of polarized Galactic dust emission over a large sky area, providing a unique data set with lasting value to the field. In this work, we describe the instrument's performance during SPIDER's second flight., Comment: Submitted to SPIE Astronomical Telescopes + Instrumentation 2024, JATIS

Hydrogen-deficient Carbon (HdC) stars are rare, low-mass, chemically peculiar, supergiant variables believed to be formed by a double white dwarf (DWD) merger, specifically of a Carbon/Oxygen- (CO-) and a Helium-white dwarf (He-WD). They consist of two subclasses -- the dust-producing R Coronae Borealis (RCB) variables and their dustless counterparts the dustless HdCs (dLHdCs). Additionally, there is another, slightly cooler set of potentially related carbon stars, the DY Persei type variables which have some, but not conclusive, evidence of Hydrogen-deficiency. Recent works have begun to explore the relationship between these three classes of stars, theorizing that they share an evolutionary pathway (a DWD merger) but come from different binary populations, specifically different total masses (M$_{\rm tot}$) and mass ratios ($q$). In this work, we use the MESA modelling framework that has previously been used to model RCB stars and vary the merger parameters, M$_{\rm tot}$ and $q$, to explore how those parameters affect the abundances, temperatures, and luminosities of the resultant post-merger stars. We find that lower M$_{\rm tot}$ and larger $q$'s both decrease the luminosity and temperatures of post-merger models to the region of the Hertzsprung-Russell Diagram populated by the dLHdCs. These lower M$_{\rm tot}$ and larger $q$ models also have smaller oxygen isotopic ratios ($^{16}$O/$^{18}$O) which is consistent with recent observations of dLHdCs compared to RCBs. None of the models generated in this work can explain the existence of the DY Persei type variables, however this may arise from the assumed metallicity of the models., Comment: 10 pages, 6 figures, re-submitted to MNRAS, has undergone one round of positive review

The design and optimization of laser-Compton x-ray systems based on compact distributed charge accelerator structures can enable micron-scale imaging of disease and the concomitant production of beams of Very High Energy Electrons (VHEEs) capable of producing FLASH-relevant dose rates. The physics of laser-Compton x-ray scattering ensures that the scattered x-rays follow exactly the trajectory of the incident electrons, thus providing a route to image-guided, VHEE FLASH radiotherapy. The keys to a compact architecture capable of producing both laser-Compton x-rays and VHEEs are the use of X-band RF accelerator structures which have been demonstrated to operate with over 100 MeV/m acceleration gradients. The operation of these structures in a distributed charge mode in which each radiofrequency (RF) cycle of the drive RF pulse is filled with a low-charge, high-brightness electron bunch is enabled by the illumination of a high-brightness photogun with a train of UV laser pulses synchronized to the frequency of the underlying accelerator system. The UV pulse trains are created by a patented pulse synthesis approach which utilizes the RF clock of the accelerator to phase and amplitude modulate a narrow band continuous wave (CW) seed laser. In this way it is possible to produce up to 10 {\mu}A of average beam current from the accelerator. Such high current from a compact accelerator enables production of sufficient x-rays via laser-Compton scattering for clinical imaging and does so from a machine of "clinical" footprint. At the same time, the production of 1000 or greater individual micro-bunches per RF pulse enables > 10 nC of charge to be produced in a macrobunch of < 100 ns. The design, construction, and test of the 100-MeV class prototype system in Irvine, CA is also presented., Comment: 28 pages, 13 figures, 2 tables

Large language models (LLMs) can be used to analyze cyber threat intelligence (CTI) data from cybercrime forums, which contain extensive information and key discussions about emerging cyber threats. However, to date, the level of accuracy and efficiency of LLMs for such critical tasks has yet to be thoroughly evaluated. Hence, this study assesses the performance of an LLM system built on the OpenAI GPT-3.5-turbo model [8] to extract CTI information. To do so, a random sample of more than 700 daily conversations from three cybercrime forums - XSS, Exploit_in, and RAMP - was extracted, and the LLM system was instructed to summarize the conversations and predict 10 key CTI variables, such as whether a large organization and/or a critical infrastructure is being targeted, with only simple human-language instructions. Then, two coders reviewed each conversation and evaluated whether the information extracted by the LLM was accurate. The LLM system performed well, with an average accuracy score of 96.23%, an average precision of 90% and an average recall of 88.2%. Various ways to enhance the model were uncovered, such as the need to help the LLM distinguish between stories and past events, as well as being careful with verb tenses in prompts. Nevertheless, the results of this study highlight the relevance of using LLMs for cyber threat intelligence.

A solid understanding of fractions is the cornerstone for acquiring proficiency with rational numbers and paves the way for learning advanced mathematical concepts, such as algebra. Fraction difficulties limit not only students' educational and vocational opportunities but also their ability to solve everyday problems. Students who exit 6th grade with inadequate understanding of fractions may experience far-reaching repercussions that lead to lifelong avoidance of mathematics. This paper presents the results of a randomized controlled trial (RCT) focusing on the first two cohorts of a larger efficacy investigation aimed at building fraction sense in students with mathematics difficulties. Teachers implemented an evidence-informed fraction sense intervention (FSI) within their 6th-grade intervention classrooms. The lessons draw from research in cognitive science as well as mathematics education research. Employing random assignment at the classroom level, multilevel modeling revealed a significant effect of the intervention on posttest fractions scores, after controlling for pretest fractions scores, working memory, vocabulary, proportional reasoning, and classroom attentive behavior. Students in the FSI group outperformed their counterparts in the control group with noteworthy effect sizes on most fraction measures. Challenges associated with carrying out school-based intervention research are addressed. [This is the online first version of an article published in "Journal of Experimental Child Psychology."]

In early 2020, the global COVID-19 pandemic closed schools for the rest of the 2019-20 school year (fiscal year 2020 or FY20). The United States Congress deployed funds to help K-12 schools adjust and plan for reopening via the Coronavirus Aid, Relief, and Economic Security (CARES) Act, including the initial $13.2 billion installment through the Elementary and Secondary School Emergency Relief Fund (ESSER I) and initial $2.95 billion installment through the Governor's Emergency Education Relief (GEER I) Fund. Private sector and non-profit organizations, including charter schools, were also eligible for loans through the Small Business Administration's Paycheck Protection Program (PPP), most of which were partially or fully forgiven. In this report, we extend our analysis of school funding during FY20 from the 2023 report, "Charter School Funding: Little Progress Toward Equity in the City." In that report, we found that, on average, charter schools receive about 30 percent ($7,147) less funding per pupil compared to traditional public schools (TPS). That analysis excluded COVID relief funding. Here we compare the initial emergency COVID relief funds received by TPS and charter schools in 18 US Cities: Atlanta, Georgia; Boston, Massachusetts; Camden, New Jersey; Chicago, Illinois; Denver, Colorado; Detroit, Michigan; Houston, Texas; Indianapolis, Indiana; Little Rock, Arkansas; Memphis, Tennessee; New Orleans, Louisiana; New York City, New York; Oakland, California; Phoenix, Arizona; San Antonio, Texas; Tulsa, Oklahoma; and Washington, DC. We use data from federal and state sources to address the following research questions: (1) did emergency COVID relief funds allocated to publicly-funded schools in FY20 widen or narrow the preexisting charter school funding gap? and (2) was initial COVID relief funding (allocated for FY20) distributed equitably relative to student poverty?

Charter schools are public schools that operate free from some government regulations in return for a commitment to achieve a set of student outcomes specified in their charter. Nearly 8,000 public charter schools enrolled 3.7 million students in the U.S. in 2020-21. In major cities, charter schools receive less funding per pupil compared to traditional public schools (TPS). Charter schools also use their funding more efficiently, achieving better short- and long-term outcomes per dollar invested, relative to TPS. In this study, the authors reexamine the productivity of publicly funded schools, using funding data from the charter school revenue report "Charter School Funding: Little Progress Towards Equity in the City." The authors also use achievement data from the Center for Research on Educational Outcomes' (CREDO's) city and national studies, the NAEP Data Explorer, and wage data from the Bureau of Labor Statistics. The authors also have access to complete data for nine cities: Camden, New Jersey; Denver, Colorado; Houston, Texas; Indianapolis, Indiana; Memphis, Tennessee; New Orleans, Louisiana; New York City, New York; San Antonio, Texas; and Washington, DC. They found that charter schools demonstrate an approximately 40 percent higher level of cost-effectiveness than TPS on average across nine cities.

Polygenic risk scores (PRSs) can significantly enhance breast cancer risk prediction when combined with clinical risk factor data. While many studies have explored the value-add of PRSs, little is known about the potential impact of gene-by-gene or gene-by-environment interactions towards enhancing the risk discrimination capabilities of multi-modal models combining PRSs with clinical data. In this study, we integrated data on 318 individual genotype variants along with clinical data in a neural network to explore whether gene-by-gene (i.e., between individual variants) and/or gene-by-environment (between clinical risk factors and variants) interactions could be leveraged jointly during training to improve breast cancer risk prediction performance. We benchmarked our approach against a baseline model combining traditional univariate PRSs with clinical data in a logistic regression model and ran an interpretability analysis to identify feature interactions. While our model did not demonstrate improved performance over the baseline, we discovered 248 (<1%) statistically significant gene-by-gene and gene-by-environment interactions out of the ~53.6k possible feature pairs, the most contributory of which included rs6001930 (MKL1) and rs889312 (MAP3K1), with age and menopause being the most heavily interacting non-genetic risk factors. We also modeled the significant interactions as a network of highly connected features, suggesting that potential higher-order interactions are captured by the model. Although gene-by-environment (or gene-by-gene) interactions did not enhance breast cancer risk prediction performance in neural networks, our study provides evidence that these interactions can be leveraged by these models to inform their predictions. This study represents the first application of neural networks to screen for interactions impacting breast cancer risk using real-world data.

Tectonic deformation crucially shapes the Earth's surface, with strain localization resulting in the formation of shear zones and faults that accommodate significant tectonic displacement. Earthquake dynamic rupture models, which provide valuable insights into earthquake mechanics and seismic ground motions, rely on initial conditions such as pre-stress states and fault geometry. However, these are often inadequately constrained due to observational limitations. To address these challenges, we develop a new method that loosely couples 3D geodynamic models to 3D dynamic rupture simulations, providing a mechanically consistent framework for earthquake analysis. Our approach does not prescribe fault geometry but derives it from the underlying lithospheric rheology and tectonic velocities using the medial axis transform. We perform three long-term geodynamics models of a strike-slip geodynamic system, each involving different continental crust rheology. We link these with nine dynamic rupture models, in which we investigate the role of varying fracture energy and plastic strain energy dissipation in the dynamic rupture behavior. These simulations suggest that for our fault, long-term rheology, and geodynamic system, a plausible critical linear slip weakening distance falls within Dc in [0.6,1.5]. Our results indicate that the long-term 3D stress field favors slip on fault segments better aligned with the regional plate motion and that minor variations in the long-term 3D stress field can strongly affect rupture dynamics, providing a physical mechanism for arresting earthquake propagation. Our geodynamically informed earthquake models highlight the need for detailed 3D fault modeling across time scales for a comprehensive understanding of earthquake mechanics., Comment: 38 pages, 14 figures, 3 tables

Vision-based robot policy learning, which maps visual inputs to actions, necessitates a holistic understanding of diverse visual tasks beyond single-task needs like classification or segmentation. Inspired by this, we introduce Theia, a vision foundation model for robot learning that distills multiple off-the-shelf vision foundation models trained on varied vision tasks. Theia's rich visual representations encode diverse visual knowledge, enhancing downstream robot learning. Extensive experiments demonstrate that Theia outperforms its teacher models and prior robot learning models using less training data and smaller model sizes. Additionally, we quantify the quality of pre-trained visual representations and hypothesize that higher entropy in feature norm distributions leads to improved robot learning performance. Code and models are available at https://github.com/bdaiinstitute/theia.

Following the rapid progress in natural language processing (NLP) models, language models are applied to increasingly more complex interactive tasks such as negotiations and conversation moderations. Having human evaluators directly interact with these NLP models is essential for adequately evaluating the performance on such interactive tasks. We develop BotEval, an easily customizable, open-source, evaluation toolkit that focuses on enabling human-bot interactions as part of the evaluation process, as opposed to human evaluators making judgements for a static input. BotEval balances flexibility for customization and user-friendliness by providing templates for common use cases that span various degrees of complexity and built-in compatibility with popular crowdsourcing platforms. We showcase the numerous useful features of BotEval through a study that evaluates the performance of various chatbots on their effectiveness for conversational moderation and discuss how BotEval differs from other annotation tools., Comment: ACL 2024 SDT, 10 pages

This paper investigates the capability of LLMs in storytelling, focusing on narrative development and plot progression. We introduce a novel computational framework to analyze narratives through three discourse-level aspects: i) story arcs, ii) turning points, and iii) affective dimensions, including arousal and valence. By leveraging expert and automatic annotations, we uncover significant discrepancies between the LLM- and human- written stories. While human-written stories are suspenseful, arousing, and diverse in narrative structures, LLM stories are homogeneously positive and lack tension. Next, we measure narrative reasoning skills as a precursor to generative capacities, concluding that most LLMs fall short of human abilities in discourse understanding. Finally, we show that explicit integration of aforementioned discourse features can enhance storytelling, as is demonstrated by over 40% improvement in neural storytelling in terms of diversity, suspense, and arousal.

Observing exoplanets through transmission spectroscopy supplies detailed information on their atmospheric composition, physics, and chemistry. Prior to JWST, these observations were limited to a narrow wavelength range across the near-ultraviolet to near-infrared, alongside broadband photometry at longer wavelengths. To understand more complex properties of exoplanet atmospheres, improved wavelength coverage and resolution are necessary to robustly quantify the influence of a broader range of absorbing molecular species. Here we present a combined analysis of JWST transmission spectroscopy across four different instrumental modes spanning 0.5-5.2 micron using Early Release Science observations of the Saturn-mass exoplanet WASP-39b. Our uniform analysis constrains the orbital and stellar parameters within sub-percent precision, including matching the precision obtained by the most precise asteroseismology measurements of stellar density to-date, and further confirms the presence of Na, K, H$_2$O, CO, CO$_2$, and SO$_2$ atmospheric absorbers. Through this process, we also improve the agreement between the transmission spectra of all modes, except for the NIRSpec PRISM, which is affected by partial saturation of the detector. This work provides strong evidence that uniform light curve analysis is an important aspect to ensuring reliability when comparing the high-precision transmission spectra provided by JWST., Comment: 34 pages, 11 figures, 3 tables. Nat Astron (2024)