Search

Visual document understanding (VDU) is a challenging task that involves understanding documents across various modalities (text and image) and layouts (forms, tables, etc.). This study aims to enhance generalizability of small VDU models by distilling knowledge from LLMs. We identify that directly prompting LLMs often fails to generate informative and useful data. In response, we present a new framework (called DocKD) that enriches the data generation process by integrating external document knowledge. Specifically, we provide an LLM with various document elements like key-value pairs, layouts, and descriptions, to elicit open-ended answers. Our experiments show that DocKD produces high-quality document annotations and surpasses the direct knowledge distillation approach that does not leverage external document knowledge. Moreover, student VDU models trained with solely DocKD-generated data are not only comparable to those trained with human-annotated data on in-domain tasks but also significantly excel them on out-of-domain tasks., Comment: Accepted to EMNLP 2024

A robust pose estimation algorithm based on an extended Kalman filter using measurements from accelerometers, rate gyroscopes, monocular vision and ultra-wideband radar is presented. The sensor fusion and pose estimation algorithm incorporates Mahalonobis distance-based outlier rejection and under-weighting of measurements for robust filter performance in the case of sudden range measurements led by the absence of measurements due to range limitations of radar transceivers. The estimator is further validated through an experimental analysis using low-cost radar, IMU and camera sensors. The pose estimate is utilized to perform proximity operations and docking of Transforming Proximity Operations and Docking Service (TPODS) satellite modules with a fixed target.

The scaling of large language models to encode all the world's knowledge in model parameters is unsustainable and has exacerbated resource barriers. Retrieval-Augmented Generation (RAG) presents a potential solution, yet its application to vision-language models (VLMs) is under explored. Existing methods focus on models designed for single tasks. Furthermore, they're limited by the need for resource intensive pre training, additional parameter requirements, unaddressed modality prioritization and lack of clear benefit over non-retrieval baselines. This paper introduces RAVEN, a multitask retrieval augmented VLM framework that enhances base VLMs through efficient, task specific fine-tuning. By integrating retrieval augmented samples without the need for additional retrieval-specific parameters, we show that the model acquires retrieval properties that are effective across multiple tasks. Our results and extensive ablations across retrieved modalities for the image captioning and VQA tasks indicate significant performance improvements compared to non retrieved baselines +1 CIDEr on MSCOCO, +4 CIDEr on NoCaps and nearly a +3\% accuracy on specific VQA question types. This underscores the efficacy of applying RAG approaches to VLMs, marking a stride toward more efficient and accessible multimodal learning.

The orange carotenoid protein (OCP) is the water-soluble mediator of non-photochemical quenching in cyanobacteria, a crucial photoprotective mechanism in response to excess illumination. OCP converts from a globular, inactive state (OCPo) to an extended, active conformation (OCPr) under high-light conditions, resulting in a concomitant redshift in the absorption of the bound carotenoid. Here, OCP was trapped in either the active or inactive state by fixing each protein conformation in trehalose-sucrose glass. Glass-encapsulated OCPo did not convert under intense illumination and OCPr did not convert in darkness, allowing the optical properties of each conformation to be determined at room temperature. We measured pump wavelength-dependent transient absorption of OCPo in glass films and found that initial OCP photoproducts are still formed, despite the glass preventing completion of the photocycle. By comparison to the pump wavelength dependence of the OCPo to OCPr photoconversion yield in buffer, we show that the long-lived carotenoid singlet-like feature (S*) is associated with ground-state heterogeneity within OCPo, rather than triggering OCP photoconversion.

The observed accelerated expansion of the Universe at present epoch can be explained by some of the $f(R)$ models without invoking the existence of dark energy or any other such exotic component in cosmic fluid. The $f(R)$ models in Palatini formalism is relatively less explored in recent times with respect to their counterpart in metric formalism. We study seven $f(R)$ models in Palatini formalism: Hu-Sawicki (two cases), Starobinsky, exponential, Tsujikawa, $f(R) = R -\beta /R^ n$, and $f(R)= R + \alpha \ln(R) - \beta$. Following standard statistical procedure and utilizing data sets: type Ia supernovae data, cosmic chronometer observations, baryonic acoustic oscillations data, data from H \textsc{ii} starburst galaxies, local measurements of the \emph{Hubble} parameter ($H_{0}$), and distance priors of cosmic microwave background radiation data, we obtain constraints on the model parameters. When compared with the standard `lambda-cold dark matter model', for many data set combinations, the support for $f(R)$ models is significant. We obtain the relevant quantities for characterizing the accelerated expansion of the Universe, and these quantities are consistent with those obtained in a model-independent way by others. The curve of effective/total equation-of-state parameter, obtained from parameter constraints, clearly shows correct phases of the expansion history: the radiation-dominated epochs and the matter-dominated epochs, of the past, and the current accelerated expansion epoch eventually evolving to de-Sitter phase in the distant future. Overall, our results advocate in favour of pursuing $f(R)$ models in Palatini formalism as a potential alternative for explaining accelerated expansion of the Universe.

Understanding the limits of rocky planet habitability is one of the key goals of current and future exoplanet characterization efforts. An intrinsic concept of rocky planet habitability is the Habitable Zone. To date, the most widely used estimates of the Habitable Zone are based on cloud-free, one-dimensional (vertical) radiative-convective climate model calculations. However, recent three-dimensional global climate modeling efforts have revealed that rocky planet habitability is strongly impacted by radiative cloud feedbacks, where computational expense and model limitations can prevent these tools from exploring the limits of habitability across the full range of parameter space. We leverage a patchy cloud one-dimensional radiative-convective climate model with parameterized cloud microphysics to investigate Inner Edge limits to the Habitable Zone for main sequence stars ($T_{\rm eff}$ = 2600 -7200K). We find that Inner Edge limits to the Habitable Zone can be 3.3 and 4.7 times closer than previous cloud-free estimates for Earth- and super-Earth-sized worlds, respectively, depending on bulk cloud parameters (e.g., fractional cloudiness and sedimentation efficiency). These warm, moist Inner Edge climates are expected to have extensive cloud decks that could mute deep atmosphere spectral features. To aid in rocky planet characterization studies, we identify the potential of using $\rm{CO_{\rm 2}}$ absorption features in transmission spectroscopy as a means of quantifying cloud deck height and cloud sedimentation efficiency. Moist greenhouse climates may represent key yet poorly understood states of habitable planets for which continued study will uncover new insights into the search and characterization of habitable worlds., Comment: 33 pages, 12 figures, submitted, community comments welcome

In this work, we explore the possibility that Early Dark Energy (EDE) is dynamical in nature and study its effect on cosmological observables. We introduce a parameterization of the equation of state allowing for an equation of state $w$ differing considerably from cosmological constant (cc, $w={-1}$) and vary both the initial $w_i$ as well final $w_f$ equation of state of the EDE fluid. This idea is motivated by the fact that in many models of EDE, the scalar field may have some kinetic energy when it starts to behave like EDE before the CMB decoupling. We find that the present data have a mild preference for non-cc early dark energy $( w_i= -0.78)$ using Planck+BAO+Pantheon+S$H_0$ES data sets, leading to $\Delta \chi^2_{\rm min}$ improvement of -2.5 at the expense of one more parameter. However, $w_i$ is only weakly constrained, with $w_i < -0.56$ at $1\sigma$. We argue that allowing for $w_i\neq -1$ can play a role in decreasing the $\sigma_8$ parameter. Yet, in practice the decrease is only $\sim0.4\sigma$ and $\sigma_8$ is still larger than weak lensing measurements. We conclude that while promising, a dynamical EDE cannot resolve both $H_0$ and $\sigma_8$ tensions simultaneously., Comment: 8 figures, Comments are most welcome

We present a new formulation for structured information extraction (SIE) from visually rich documents. It aims to address the limitations of existing IOB tagging or graph-based formulations, which are either overly reliant on the correct ordering of input text or struggle with decoding a complex graph. Instead, motivated by anchor-based object detectors in vision, we represent an entity as an anchor word and a bounding box, and represent entity linking as the association between anchor words. This is more robust to text ordering, and maintains a compact graph for entity linking. The formulation motivates us to introduce 1) a DOCument TRansformer (DocTr) that aims at detecting and associating entity bounding boxes in visually rich documents, and 2) a simple pre-training strategy that helps learn entity detection in the context of language. Evaluations on three SIE benchmarks show the effectiveness of the proposed formulation, and the overall approach outperforms existing solutions.

Modified theories of gravity encompass a class of $f(R)$-models that seek to elucidate the observed late time accelerated expansion of the universe. In this study, we examine a set of viable $f(R)$ models (Hu-Sawicki: two cases, Satrobinsky, Tsujikawa, exponential and arcTanh models) in metric formalism, using recent cosmological data sets: type Ia supernovae data, cosmic chronometer observations, baryonic acoustic oscillations data, data from H\textsc{ii} starburst galaxies, and local measurements of the Hubble parameter $H_0$. The model parameters are constrained using a Bayesian analysis with the Monte Carlo Markov Chain method. We employ statistical tools such as the Akaike Information Criterion, Bayesian Information Criterion, and reduced chi-square statistics to conduct a comparative investigation of these models. We determine the transition redshift, the evolution of total equation-of-state (EoS) parameter, and the EoS for the component responsible for current accelerated expansion to characterize the expansion's evolution. Taking into account the ``Hubble tension," we perform the study with and without a Gaussian prior for $H_0$ from local measurements. Our findings are as follows: (i) in many cases the $f(R)$ models are strongly favored over the standard $\Lambda$CDM model, (ii) the deviation parameter ($b$) significantly deviates from zero in several cases, (iii) the inclusion of local $H_0$ not only increases the fitted value of $H_0$ (as expected) but also affects the gap between predictions of $f(R)$ models and the $\Lambda$CDM model, and (iv) the relevant quantities characterizing the (accelerated) expansion of the universe obtained in our models are consistent with those obtained in a model-independent way by others. Our investigation and results present a compelling case for pursuing further research on $f(R)$ models with future observations to come.