Search

The development of LLMs has greatly enhanced the intelligence and fluency of question answering, while the emergence of retrieval enhancement has enabled models to better utilize external information. However, the presence of noise and errors in retrieved information poses challenges to the robustness of LLMs. In this work, to evaluate the model's performance under multiple interferences, we first construct a dataset based on machine reading comprehension datasets simulating various scenarios, including critical information absence, noise, and conflicts. To address the issue of model accuracy decline caused by noisy external information, we propose a data augmentation-based fine-tuning method to enhance LLM's robustness against noise. Additionally, contrastive learning approach is utilized to preserve the model's discrimination capability of external information. We have conducted experiments on both existing LLMs and our approach, the results are evaluated by GPT-4, which indicates that our proposed methods improve model robustness while strengthening the model's discrimination capability., Comment: This paper has been accepted by NLPCC-2024

We present a low-resource and robust optical implementation of the four-dimensional Grover coin, a four-port linear-optical scatterer that augments the low dimensionality of a regular beam-splitter. While prior realizations of the Grover coin required a potentially unstable ring-cavity to be formed, this version of the scatterer does not exhibit any internal interference. When this Grover coin is placed in another system, it can be used for interferometry with a higher-dimensional set of optical field modes. In this case, we formed a Grover-Michelson interferometer, which results when the traditional beam-splitter of a Michelson interferometer is replaced with a four-port Grover coin. This replacement has been shown to remove a phase parameter redundancy in the original Michelson system, now allowing continuous tuning of the shape and slope of the interference pattern. We observed an intensity interferogram with $97\%$ visibility and a phase sensitivity more than an order of magnitude larger than a regular Michelson interferometer. Because this device is readily formed with nearly the same number of optomechanical resources as a Michelson interferometer, but can outperform it drastically in phase delay evaluation, it has a great potential to improve many interferometric sensing and control systems., Comment: 12 pages, 6 figures

Using symmetry considerations, we derive a unitary scattering matrix for a three-port optical Y-coupler or Y-branch. The result is shown to be unique up to external phase shifts. Unlike traditional passive linear-optical one-way splitters, coupling light into the conventional output ports of the Y-coupler results in strong coherent back-reflections, making the device a hybrid between feed-forward devices like the beam splitter, which do not reverse the direction of light, and a recently considered class of directionally unbiased multiport scatterers (with dimension greater than two) which do. While the device could immediately find use as a novel scattering vertex for the implementation of quantum walks, we also design a few simple but nonetheless useful optical systems that can be constructed by taking advantage of the symmetry of the scattering process. This includes an interference-free, resource-efficient implementation of the Grover four-port and a higher-dimensional Fabry-Perot interferometer with tunable finesse. Symmetry-breaking generalizations are also considered.

Endoscopic surgery relies on two-dimensional views, posing challenges for surgeons in depth perception and instrument manipulation. While Monocular Visual Simultaneous Localization and Mapping (MVSLAM) has emerged as a promising solution, its implementation in endoscopic procedures faces significant challenges due to hardware limitations, such as the use of a monocular camera and the absence of odometry sensors. This study presents BodySLAM, a robust deep learning-based MVSLAM approach that addresses these challenges through three key components: CycleVO, a novel unsupervised monocular pose estimation module; the integration of the state-of-the-art Zoe architecture for monocular depth estimation; and a 3D reconstruction module creating a coherent surgical map. The approach is rigorously evaluated using three publicly available datasets (Hamlyn, EndoSLAM, and SCARED) spanning laparoscopy, gastroscopy, and colonoscopy scenarios, and benchmarked against four state-of-the-art methods. Results demonstrate that CycleVO exhibited competitive performance with the lowest inference time among pose estimation methods, while maintaining robust generalization capabilities, whereas Zoe significantly outperformed existing algorithms for depth estimation in endoscopy. BodySLAM's strong performance across diverse endoscopic scenarios demonstrates its potential as a viable MVSLAM solution for endoscopic applications., Comment: 16 pages, 7 figures

Large language models (LLMs) demonstrate extraordinary abilities in a wide range of natural language processing (NLP) tasks. In this paper, we show that, beyond text understanding capability, LLMs are capable of processing text layouts that are denoted by spatial markers. They are able to answer questions that require explicit spatial perceiving and reasoning, while a drastic performance drop is observed when the spatial markers from the original data are excluded. We perform a series of experiments with the GPT-3.5, Baichuan2, Llama2 and ChatGLM3 models on various types of layout-sensitive datasets for further analysis. The experimental results reveal that the layout understanding ability of LLMs is mainly introduced by the coding data for pretraining, which is further enhanced at the instruction-tuning stage. In addition, layout understanding can be enhanced by integrating low-cost, auto-generated data approached by a novel text game. Finally, we show that layout understanding ability is beneficial for building efficient visual question-answering (VQA) systems., Comment: This paper has been accepted by ECAI-2024

4$d$ transition metal oxide (TMO) offers an intriguing puzzle for their electronic and magnetic ground state. They are in the cross-over regime of strong spin orbit interaction (SOI) and electron-electron correlation ($U$) with quenched orbital angular momentum. Our work unravels the electronic and magnetic ground state of the less investigated 4$d^{3}$ double perovskite ruthenates A$_{2}$LaRuO$_6$ (A = Ca, Ba). The negligible effect of SOI is evident from the bulk magnetic, specific heat measurements and density functional theory (DFT) calculations, indicating a classical spin-only magnetic ground state (${S}$ = 3/2) for the materials. Magnetization measurements show that both materials have long range antiferromagnetic order with high degree of magnetic frustration ($f$ $\approx$13 -15). Interestingly, a near $T^2$- behavior is observed in low-$T$ magnetic heat capacity measurement, indicating the presence of low-dimensional spin-wave exciation and magnetic frustration in both materials. The temperature dependent resistivity measurements and electronic band structure calculations confirm a conventional Mott insulating ground state in these two systems. Moreover, our experimental investigation and DFT calculations highlight the reason for the nonexistence of Sr$_2$LaRuO$_6$., Comment: 11 pages, 12 figures, 5 tables

We present a finite-element approach for computing the aggregate scattering matrix of a network of linear coherent scatterers. These might be optical scatterers or more general scattering coins studied in quantum walk theory. While techniques exist for two-dimensional lattices of feed-forward scatterers, the present approach is applicable to any network configuration of any collection of scatterers. Unlike traditional finite-element methods in optics, this method does not directly solve Maxwell's equations; instead it is used to assemble and solve a linear, coupled scattering problem that emerges after Maxwell's equations are abstracted within the scattering matrix method. With this approach, a global unitary is assembled corresponding to one time step of the quantum walk on the network. After applying the relevant boundary conditions to this global matrix, the problem becomes non-unitary, and possesses a steady-state solution which is the output scattering state. We provide an algorithm to obtain this steady-state solution exactly using a matrix inversion, yielding the scattering state without requiring a direct calculation of the eigenspectrum. The approach is then numerically validated on a coupled-cavity interferometer example that possesses a known, closed-form solution. Finally, the method is shown to be a generalization of the Redheffer star product, which describes scatterers on one-dimensional lattices (2-regular graphs) and is often applied to the design of thin-film optics, making the current approach an invaluable tool for the design and validation of high-dimensional phase-reprogrammable optical devices and study of quantum walks on arbitrary graphs.

We determine the maximal dimension of compact subvarieties of $\mathcal{A}_g$, the moduli space of complex principally polarized abelian varieties of dimension $g$, and the maximal dimension of a compact subvariety through a very general point of $\mathcal{A}_g$. This also allows us to draw some conclusions for compact subvarieties of the moduli space of complex curves of compact type., Comment: 36 pages, comments welcome!

Visual prompt tuning (VPT) is a promising solution incorporating learnable prompt tokens to customize pre-trained models for downstream tasks. However, VPT and its variants often encounter challenges like prompt initialization, prompt length, and subpar performance in self-supervised pretraining, hindering successful contextual adaptation. This study commences by exploring the correlation evolvement between prompts and patch tokens during proficient training. Inspired by the observation that the prompt tokens tend to share high mutual information with patch tokens, we propose initializing prompts with downstream token prototypes. The strategic initialization, a stand-in for the previous initialization, substantially improves performance in fine-tuning. To refine further, we optimize token construction with a streamlined pipeline that maintains excellent performance with almost no increase in computational expenses compared to VPT. Exhaustive experiments show our proposed approach outperforms existing methods by a remarkable margin. For instance, it surpasses full fine-tuning in 19 out of 24 tasks, using less than 0.4% of learnable parameters on the FGVC and VTAB-1K benchmarks. Notably, our method significantly advances the adaptation for self-supervised pretraining, achieving impressive task performance gains of at least 10% to 30%. Besides, the experimental results demonstrate the proposed SPT is robust to prompt lengths and scales well with model capacity and training data size. We finally provide an insightful exploration into the amount of target data facilitating the adaptation of pre-trained models to downstream tasks. The code is available at https://github.com/WangYZ1608/Self-Prompt-Tuning., Comment: Accepted by ICML2024, See https://github.com/WangYZ1608/Self-Prompt-Tuning

Resonant intermediate states have been proposed to increase the efficiency of entangled two-photon absorption (ETPA). Although resonance-enhanced ETPA (r-ETPA) has been demonstrated in atomic systems using bright squeezed vacuum, it has not been studied in organic molecules. We investigate for the first time r-ETPA in an organic molecular dye, indocyanine green (ICG), when excited by broadband entangled photons in near-IR. Similar to many reported virtual state mediated ETPA (v-ETPA) measurements, no r-ETPA signals are measured, with an experimental upper bound for the cross section placed at $6 \times 10^{-23}$ cm$^2$/molecule. In addition, the classical resonance-enhanced two-photon absorption (r-TPA) cross section of ICG at 800 nm is measured for the first time to be $20(\pm13)$ GM, suggesting that having a resonant intermediate state does not significantly enhance two-photon processes in ICG. The spectrotemporally resolved emission signatures of ICG excited by entangled photons are also presented to support this conclusion.

We study various naturally defined subvarieties of the moduli space ${\mathcal A}_g$ of complex principally polarized abelian varieties (ppav) in a neighborhood of the locus of products of $g$ elliptic curves. In this neighborhood, we obtain a local description for the locus of hyperelliptic curves, reproving the recent result of Shepherd-Barron that the hyperelliptic locus is locally given by tridiagonal matrices. We further reprove and generalize to arbitrary genus the recent result of Agostini and Chua showing that the locus of Jacobians of genus 5 curves with a theta-null is an irreducible component of the locus of ppav with a theta-null such that the singular locus of the theta divisor at the corresponding two-torsion point has tangent cone of rank at most 3. We further show that the locus of ppav such that the gradient vanishes, for some odd theta characteristic, locally has codimension $g$ near the diagonal. Finally, we obtain new results on the locus where the rank of the Hessian of the theta function at a two-torsion point that lies on the theta divisor is equal to 2., Comment: v2: only corrected TeX, to disable showkeys

X-ray Fluorescence Ghost Imaging (XRF-GI) was recently demonstrated for x-ray lab sources. It has the potential to reduce acquisition time and deposited dose by choosing their trade-off with spatial resolution, while alleviating the focusing constraints of the probing beam. Here, we demonstrate the realization of synchrotron-based XRF-GI: We present both an adapted experimental setup and its corresponding required computational technique to process the data. This extends the above-mentioned potential advantages of GI to synchrotron XRF imaging. In addition, it enables new strategies to improve resilience against drifts at all scales, and the study of previously inaccessible samples, such as liquids., Comment: 4 pages, 5 figures

Knowledge distillation (KD) exploits a large well-trained model (i.e., teacher) to train a small student model on the same dataset for the same task. Treating teacher features as knowledge, prevailing methods of knowledge distillation train student by aligning its features with the teacher's, e.g., by minimizing the KL-divergence between their logits or L2 distance between their intermediate features. While it is natural to believe that better alignment of student features to the teacher better distills teacher knowledge, simply forcing this alignment does not directly contribute to the student's performance, e.g., classification accuracy. In this work, we propose to align student features with class-mean of teacher features, where class-mean naturally serves as a strong classifier. To this end, we explore baseline techniques such as adopting the cosine distance based loss to encourage the similarity between student features and their corresponding class-means of the teacher. Moreover, we train the student to produce large-norm features, inspired by other lines of work (e.g., model pruning and domain adaptation), which find the large-norm features to be more significant. Finally, we propose a rather simple loss term (dubbed ND loss) to simultaneously (1) encourage student to produce large-\emph{norm} features, and (2) align the \emph{direction} of student features and teacher class-means. Experiments on standard benchmarks demonstrate that our explored techniques help existing KD methods achieve better performance, i.e., higher classification accuracy on ImageNet and CIFAR100 datasets, and higher detection precision on COCO dataset. Importantly, our proposed ND loss helps the most, leading to the state-of-the-art performance on these benchmarks. The source code is available at \url{https://github.com/WangYZ1608/Knowledge-Distillation-via-ND}., Comment: 16 pages, 8 figures, 6 tables

Fluorescence lifetime experiments are a standard approach for measuring excited state dynamics and local environment effects. Here, we show that entangled photon pairs produced from a continuous-wave (CW) laser diode can replicate pulsed laser experiments without phase modulation. As a proof of principle, picosecond fluorescence lifetimes of indocyanine green are measured in multiple environments. The use of entangled photons has three unique advantages. First, low power CW laser diodes and entangled photon source design lead to straightforward on-chip integration for a direct path to distributable fluorescence lifetime measurements. Second, the entangled pair wavelength is easily tuned by temperature or electric field, allowing a single source to cover octave bandwidths. Third, femtosecond temporal resolutions can be reached without requiring major advances in source technology or external phase modulation. Entangled photons could therefore provide increased accessibility to time-resolved fluorescence while also opening new scientific avenues in photosensitive and inherently quantum systems.

Entities and events are crucial to natural language reasoning and common in procedural texts. Existing work has focused either exclusively on entity state tracking (e.g., whether a pan is hot) or on event reasoning (e.g., whether one would burn themselves by touching the pan), while these two tasks are often causally related. We propose CREPE, the first benchmark on causal reasoning of event plausibility and entity states. We show that most language models, including GPT-3, perform close to chance at .35 F1, lagging far behind human at .87 F1. We boost model performance to .59 F1 by creatively representing events as programming languages while prompting language models pretrained on code. By injecting the causal relations between entities and events as intermediate reasoning steps in our representation, we further boost the performance to .67 F1. Our findings indicate not only the challenge that CREPE brings for language models, but also the efficacy of code-like prompting combined with chain-of-thought prompting for multihop event reasoning., Comment: In Findings of EACL 2023

Tchebycheffian splines are smooth piecewise functions whose pieces are drawn from (possibly different) Tchebycheff spaces, a natural generalization of algebraic polynomial spaces. They enjoy most of the properties known in the polynomial spline case. In particular, under suitable assumptions, Tchebycheffian splines admit a representation in terms of basis functions, called Tchebycheffian B-splines (TB-splines), completely analogous to polynomial B-splines. A particularly interesting subclass consists of Tchebycheffian splines with pieces belonging to null-spaces of constant-coefficient linear differential operators. They grant the freedom of combining polynomials with exponential and trigonometric functions with any number of individual shape parameters. Moreover, they have been recently equipped with efficient evaluation and manipulation procedures. In this paper, we consider the use of TB-splines with pieces belonging to null-spaces of constant-coefficient linear differential operators as an attractive substitute for standard polynomial B-splines and rational NURBS in isogeometric Galerkin methods. We discuss how to exploit the large flexibility of the geometrical and analytical features of the underlying Tchebycheff spaces according to problem-driven selection strategies. TB-splines offer a wide and robust environment for the isogeometric paradigm beyond the limits of the rational NURBS model., Comment: 35 pages, 18 figures