Search

Expert-designed close-ended benchmarks serve as vital tools in assessing the knowledge capacity of large language models (LLMs). Despite their widespread use, concerns have mounted regarding their reliability due to limited test scenarios and an unavoidable risk of data contamination. To rectify this, we present PertEval, a toolkit devised for in-depth probing of LLMs' knowledge capacity through knowledge-invariant perturbations. These perturbations employ human-like restatement techniques to generate on-the-fly test samples from static benchmarks, meticulously retaining knowledge-critical content while altering irrelevant details. Our toolkit further includes a suite of transition analyses that compare performance on raw vs. perturbed test sets to precisely assess LLMs' genuine knowledge capacity. Six state-of-the-art LLMs are re-evaluated using PertEval. Results reveal significantly inflated performance of the LLMs on raw benchmarks, including an absolute 21% overestimation for GPT-4. Additionally, through a nuanced response pattern analysis, we discover that PertEval retains LLMs' uncertainty to specious knowledge, potentially being resolved through rote memorization and leading to inflated performance. We also find that the detailed transition analyses by PertEval could illuminate weaknesses in existing LLMs' knowledge mastery and guide the development of refinement. Given these insights, we posit that PertEval can act as an essential tool that, when applied alongside any close-ended benchmark, unveils the true knowledge capacity of LLMs, marking a significant step toward more trustworthy LLM evaluation., Comment: 23 pages, 12 figures, 10 tables

This paper presents EasyAnimate, an advanced method for video generation that leverages the power of transformer architecture for high-performance outcomes. We have expanded the DiT framework originally designed for 2D image synthesis to accommodate the complexities of 3D video generation by incorporating a motion module block. It is used to capture temporal dynamics, thereby ensuring the production of consistent frames and seamless motion transitions. The motion module can be adapted to various DiT baseline methods to generate video with different styles. It can also generate videos with different frame rates and resolutions during both training and inference phases, suitable for both images and videos. Moreover, we introduce slice VAE, a novel approach to condense the temporal axis, facilitating the generation of long duration videos. Currently, EasyAnimate exhibits the proficiency to generate videos with 144 frames. We provide a holistic ecosystem for video production based on DiT, encompassing aspects such as data pre-processing, VAE training, DiT models training (both the baseline model and LoRA model), and end-to-end video inference. Code is available at: https://github.com/aigc-apps/EasyAnimate. We are continuously working to enhance the performance of our method., Comment: 8 pages, 6 figures

The electronic structure evolutions of few-layer black phosphorus (BP) under pressure shows a wealth of phenomena, such as the nonmonotonic change of direct gap at the {\Gamma} point, the layer-number dependence, and the distinct responses to normal and hydrostatic pressures. A full and unified understanding to these rich phenomena remains lacking. Here, we provide a unified understanding from the competition between interlayer quasi-bonding (QB) interactions and intralayer chemical bonding interactions. The former decreases while the latter increases the band gap under pressure and the origin can be correlated to different combinations of inter- and intra-layer antibonding or bonding interactions at the band edges. More interestingly, the interlayer QB interactions are a coexistence of two categories of interactions, namely, the coexistence of interactions between bands of the same occupancy (occupied-occupied and empty-empty interactions) and of different occupancies (occupied-empty interaction); and, the overall effect is a four-level interaction, which explains the anomalous interlayer-antibonding feature of the conduction band edge of bilayer BP. Our current study lay the foundation for the electronic structure tuning of two-dimensional (2D) BP, and, our analysis method for multi-energy-level interactions can be applied to other 2D semiconductor homo- and hetero-structures that have occupied-empty interlayer interactions.

As the task of 2D-to-3D reconstruction has gained significant attention in various real-world scenarios, it becomes crucial to be able to generate high-quality point clouds. Despite the recent success of deep learning models in generating point clouds, there are still challenges in producing high-fidelity results due to the disparities between images and point clouds. While vision transformers (ViT) and diffusion models have shown promise in various vision tasks, their benefits for reconstructing point clouds from images have not been demonstrated yet. In this paper, we first propose a neat and powerful architecture called DiffPoint that combines ViT and diffusion models for the task of point cloud reconstruction. At each diffusion step, we divide the noisy point clouds into irregular patches. Then, using a standard ViT backbone that treats all inputs as tokens (including time information, image embeddings, and noisy patches), we train our model to predict target points based on input images. We evaluate DiffPoint on both single-view and multi-view reconstruction tasks and achieve state-of-the-art results. Additionally, we introduce a unified and flexible feature fusion module for aggregating image features from single or multiple input images. Furthermore, our work demonstrates the feasibility of applying unified architectures across languages and images to improve 3D reconstruction tasks.

Until recently, the question of the effective inductive bias of deep models on tabular data has remained unanswered. This paper investigates the hypothesis that arithmetic feature interaction is necessary for deep tabular learning. To test this point, we create a synthetic tabular dataset with a mild feature interaction assumption and examine a modified transformer architecture enabling arithmetical feature interactions, referred to as AMFormer. Results show that AMFormer outperforms strong counterparts in fine-grained tabular data modeling, data efficiency in training, and generalization. This is attributed to its parallel additive and multiplicative attention operators and prompt-based optimization, which facilitate the separation of tabular samples in an extended space with arithmetically-engineered features. Our extensive experiments on real-world data also validate the consistent effectiveness, efficiency, and rationale of AMFormer, suggesting it has established a strong inductive bias for deep learning on tabular data. Code is available at https://github.com/aigc-apps/AMFormer., Comment: 11 pages, 8 figures, to be published to AAAI2024

Urban warming is evident in numerous cities. On especially hot days, building surfaces warm up, leading to buoyancy-driven flows adjacent to these surfaces. The dynamics of the flow structures are largely influenced by the interplay between incoming wind and the buoyancy-driven flows. In this study, we used large eddy simulations and wind tunnel measurements to investigate the flow field around an isolated cubic building when different surfaces of the building are warmed. Under conditions of low wind speeds, ranging from 0.5 to 2 m/s, the surface temperatures of the scaled building were maintained between 20 and 95 {\deg}C. As the Richardson number (Ri) varied from 0 to 4.00, the flow, initially dominated by forced convection, shifted to being primarily steered by mixed convection. At low wind speeds and high Ri values, the thermal effect led to noticeable changes in the reattachment and recirculation region lengths, with reductions of up to 48.3% in some cases. At pedestrian levels, thermally induced airflows often created localized hot spots, particularly around building corners and wall sections. This study offers insights into architectural designs that can potentially enhance wind-thermal comfort and improve pollutant dispersion around buildings.

Stable Diffusion web UI (SD-WebUI) is a comprehensive project that provides a browser interface based on Gradio library for Stable Diffusion models. In this paper, We propose a novel WebUI plugin called EasyPhoto, which enables the generation of AI portraits. By training a digital doppelganger of a specific user ID using 5 to 20 relevant images, the finetuned model (according to the trained LoRA model) allows for the generation of AI photos using arbitrary templates. Our current implementation supports the modification of multiple persons and different photo styles. Furthermore, we allow users to generate fantastic template image with the strong SDXL model, enhancing EasyPhoto's capabilities to deliver more diverse and satisfactory results. The source code for EasyPhoto is available at: https://github.com/aigc-apps/sd-webui-EasyPhoto. We also support a webui-free version by using diffusers: https://github.com/aigc-apps/EasyPhoto. We are continuously enhancing our efforts to expand the EasyPhoto pipeline, making it suitable for any identification (not limited to just the face), and we enthusiastically welcome any intriguing ideas or suggestions., Comment: 7 pages, 7 figures

Ferroelectricity can exist in elemental phases as a result of charge transfers between atoms occupying inequivalent Wyckoff positions. We investigate the emergence of ferroelectricity in two-dimensional elemental materials with buckled honeycomb lattices. Various multi-bilayer structures hosting ferroelectricity are designed by stacking-engineering. Ferroelectric materials candidates formed by group IV and V elements are predicted theoretically. Ultrathin Bi films show layer-stacking-dependent physical properties of ferroelectricity, topology, and metallicity. The two-bilayer Bi film with a polar stacking sequence is found to be an elemental topological ferroelectric material. Three and four bilayers Bi films with polar structures are ferroelectric-like elemental polar metals with topological nontrivial edge states. For Ge and Sn, trivial elemental polar metals are predicted. Our work reveals the possibility of design two-dimensional elemental topological ferroelectrics and polar metals by stacking-engineering., Comment: 18 pages, 6 figures

Momentum-matched type II van der Waals heterostructures (vdWHs) have been designed by assembling layered two-dimensional semiconductors (2DSs) with special band-structure combinations - that is, the valence band edge at the Gamma point (the Brillouin-zone center) for one 2DS and the conduction band edge at the Gamma point for the other [Ubrig et al., Nat. Mater. 19, 299 (2020)]. However, the band offset sizes, band-alignment types, and whether momentum matched or not, all are affected by the interfacial effects between the component 2DSs, such as the quasichemical-bonding (QB) interaction between layers and the electrical dipole moment formed around the vdW interface. Here, based on density-functional theory calculations, first we probe the interfacial effects (including different QBs for valence and conduction bands, interface dipole, and, the synergistic effects of these two aspects) on band-edge evolution in energy and valley (location in the Brillouin zone) and the resulting changes in band alignment and momentum matching for a typical vdWH of monolayer InSe and bilayer WS2, in which the band edges of subsystems satisfy the special band-structure combination for a momentum-matched type II vdWH. Then, based on the conclusions of the studied interfacial effects, we propose a practical screening method for robust momentum-matched type II vdWHs. This practical screening method can also be applied to other band alignment types. Our current study opens a way for practical screening and designing of vdWHs with robust momentum-matching and band alignment type.

Multiferroics are materials with a coexistence of magnetic and ferroelectric order allowing the manipulation of magnetism by applications of an electric field through magnetoelectric coupling effects. Here we propose an idea to design a class of multiferroics with a $d^9$ configuration using the magnetic order in copper-oxygen layers appearing in copper oxide high-temperature superconductors by inducing ferroelectricity. Copper-based charge transfer multiferroics SnCuO2 and PbCuO2 having the inversion symmetry breaking $P4mm$ polar space group are predicted to be such materials. The active inner s electrons in Sn and Pb hybridize with O $2p$ states leading the buckling in copper-oxygen layers and thus induces ferroelectricity, which is known as the lone pair mechanism. As a result of the $d^9$ configuration, SnCuO2 and PbCuO2 are charge transfer insulators with the antiferromagnetic ground state of the moment on Cu retaining some strongly correlated physical properties of parent compounds of copper oxide high-temperature superconductors. Our work reveals the possibility of designing multiferroics based on copper oxide high-temperature superconductors., Comment: 18 pages, 5 figures, 1 table

Video-and-language understanding has a variety of applications in the industry, such as video question answering, text-video retrieval, and multi-label classification. Existing video-and-language understanding methods generally adopt heavy multi-modal encoders and feature fusion modules, which consume high computational costs. Specially, they have difficulty dealing with dense video frames or long text prevalent in industrial applications. This paper proposes MuLTI, a highly accurate and efficient video-and-language understanding model that achieves efficient and effective feature fusion and rapid adaptation to downstream tasks. Specifically, we design a Text-Guided MultiWay-Sampler based on adapt-pooling residual mapping and self-attention modules to sample long sequences and fuse multi-modal features, which reduces the computational costs and addresses performance degradation caused by previous samplers. Therefore, MuLTI can handle longer sequences with limited computational costs. Then, to further enhance the model's performance and fill in the lack of pretraining tasks in the video question answering, we propose a new pretraining task named Multiple Choice Modeling. This task bridges the gap between pretraining and downstream tasks and improves the model's ability to align video and text features. Benefiting from the efficient feature fusion module and the new pretraining task, MuLTI achieves state-of-the-art performance on multiple datasets. Implementation and pretrained models will be released.

Motivated by the recently discovered superconductivity in Sr-doped nickelate oxides NdNiO$_2$, we predict a material BiNiO$_2$ that provides an opportunity to study the intertwined ferroelectricity, metallicity, and magnetism in a crystal with very simple atomic structures. There is a ferroelectric structural transition from the nonpolar phase with the P4/mmm space group to the polar phase with the P4mm space group, which is driven by the lone pair on Bi. Calculations based on the Heyd-Scuseria-Ernzerhof hybrid density functional reveal that both the nonmagnetic and ferromagnetic states are metallic for nonpolar and polar phases, while the lowest energy ground-state for polar BiNiO$_2$ is a Hubbard Mott insulator with the G-type antiferromagnetic spin configurations. As a ferroelectric material with an electric polarization of 0.49 C/m2, it may be possible to control the magnetic order in BiNiO$_2$ by an applied electric field. The replacement of Nd by Bi serves as a connecting link between a high-temperature superconductor and a Mott multiferroic. Our work supports a route towards strongly correlated ferroelectrics., Comment: 20 pages, 9 figures, 3 tables

In this paper, we propose the task of consecutive question generation (CQG), which generates a set of logically related question-answer pairs to understand a whole passage, with a comprehensive consideration of the aspects including accuracy, coverage, and informativeness. To achieve this, we first examine the four key elements of CQG, i.e., question, answer, rationale, and context history, and propose a novel dynamic multitask framework with one main task generating a question-answer pair, and four auxiliary tasks generating other elements. It directly helps the model generate good questions through both joint training and self-reranking. At the same time, to fully explore the worth-asking information in a given passage, we make use of the reranking losses to sample the rationales and search for the best question series globally. Finally, we measure our strategy by QA data augmentation and manual evaluation, as well as a novel application of generated question-answer pairs on DocNLI. We prove that our strategy can improve question generation significantly and benefit multiple related NLP tasks., Comment: Findings of EMNLP 2022

The Datta-Das-type spin field-effect transistor, using a two-dimensional electron gas in a semiconductor heterostructure as a channel, plays a key role in spintronics. Here, we theoretically present a type of spin field-effect transistor based on massless birefringent Dirac fermions in polar Dirac semimetals. The manipulation of spin arises from the existence of the strong spin-orbit coupling, polar space groups, and Dirac cones in a single phase. The oscillatory channel conductance can be controlled by the sign of gate voltage in addition to its magnitude due to the gapless band structures of polar Dirac semimetals. Such spin field-effect transistor provides guidance for the further design of spintronic devices., Comment: 14 pages, 3 figures

Polar metals based on binary and ternary compounds have been demonstrated in literature. Here, we propose a design principle for ferroelectric-like elemental polar metals and relate it to real materials. The design principle is that, to be an elemental polar metal, atoms should occupy at least two inequivalent Wyckoff positions in a crystal with a polar space group, where inversion symmetry is spontaneously broken. According to this rule, we propose the first class of potential ferroelectric-like elemental polar metals in a distorted {\alpha}-La-like structure with a polar space group P63mc in which two inequivalent Wyckoff positions 2a (0, 0, z) and 2b (1/3, 2/3, z) are occupied by group-V elements (phosphorus, arsenic, antimony, and bismuth). Analyses based on first-principles calculations indicate that the dynamically stable polar phase results from a lone pair driven polar distortion of the nonploar phase in P63/mmc symmetry where two inequivalent Wyckoff positions 2a (0, 0, 0) and 2c (1/3, 2/3, 1/4) are occupied. This ferroelectric-like transition involves a transition from a metallic state to a semimetallic state. These predicted polar phases are metastable with respect to their corresponding ground phases. Moreover, ionic bonding characters are found due to the inequivalence in Wyckoff positions between group-V atoms. Our work opens a route to single-element parity-breaking phases., Comment: 11 pages, 4 figures

The Rashba and Dresselhaus types of spin-orbit coupling are two typical linear coupling forms. We establish the fundamental physics of a model which can be viewed as the double version of the Rashba and Dresselhaus spin-orbit coupling. This model describes the low energy physics of a class of massless Dirac fermions in spin-orbit systems. The physical properties of the massless Dirac fermions are determined by the mathematical relations of spin-orbit coefficients. For equal Rashba and Dresselhauss coupling constants, k-independent eigenspinors and a persistent spin helix combined with massless birefringent Dirac fermions emerge in this model. The spin-orbit coupled systems described by this model have potential technological applications from spintronics to quantum computation., Comment: 12 pages, 4 figures

Recent studies have revealed that the interlayer interaction in two-dimensional (2D) layered materials is not simply of van der Waals character but could coexist with quasi-bonding character. Here we classify the interlayer quasi-bonding interactions into two main categories (I: homo-occupancy interaction, II: hetero-occupancy interaction) according to the occupancy of the involved energy bands near the Fermi level. Then we investigate the quasi-bonding-interaction-induced band structure evolution of several representative 2D materials based on density functional theory calculations. Further calculations confirm that this classification is applicable to generic 2D layered materials and provides a unified understanding of the total strength of interlayer interaction, which is a synergetic effect of the van der Waals attraction and the quasi-bonding interaction. The latter is stabilizing in main category II and destabilizing in main category I. Thus, the total interlayer interaction strength is relatively stronger in category II and weaker in category I.

The phase transitions among polymorphic two-dimensional (2D) transition metal dichalcogenides (TMDs) have attracted increasing attention for their potential in enabling distinct functionalities in the same material for making integrated devices. Electron-injection to TMDs has been proved to be a feasible way to drive structural phase transition from the semiconducting H-phase to the semimetal dT-phase. In this contribution, based on density-functional theory (DFT) calculations, firstly we demonstrate that hole-injection drives the transition of the H-phase more efficiently to the metallic T-phase than to the semimetallic dT-phase for group VI-B TMDs (MoS2, WS2, and MoSe2, etc.). The origin can be attributed to the smaller work function of the T-phase than that of the dT-phase. Our work function analysis can distinguish the T and dT phases quantitatively while it is challenging for the commonly used crystal field splitting analysis. In addition, our analysis provides a unified understanding for both hole- and electron-injection induced phase transitions for 2D materials beyond TMDs, such as the newly synthesized MoSi2N4 family. Moreover, the hole-driven T-phase transition mechanism can explain the recent experiment of WS2 phase transition by hole-doping with yttrium (Y) atoms. Using 1/3 Y-doped WS2 and MoSe2 as examples, we show that the Mo and W valency increases to 5+. These above findings open up an avenue to obtain the metallic T-phase, which expands the possible stable phases of 2D materials.

We investigate why neural machine translation (NMT) systems assign high probability to empty translations. We find two explanations. First, label smoothing makes correct-length translations less confident, making it easier for the empty translation to finally outscore them. Second, NMT systems use the same, high-frequency EoS word to end all target sentences, regardless of length. This creates an implicit smoothing that increases zero-length translations. Using different EoS types in target sentences of different lengths exposes and eliminates this implicit smoothing., Comment: 6 pages

This paper describes DiDi AI Labs' submission to the WMT2020 news translation shared task. We participate in the translation direction of Chinese->English. In this direction, we use the Transformer as our baseline model, and integrate several techniques for model enhancement, including data filtering, data selection, back-translation, fine-tuning, model ensembling, and re-ranking. As a result, our submission achieves a BLEU score of $36.6$ in Chinese->English., Comment: Accepted at WMT 2020

We create a new task-oriented dialog platform (MEEP) where agents are given considerable freedom in terms of utterances and API calls, but are constrained to work within a push-button environment. We include facilities for collecting human-human dialog corpora, and for training automatic agents in an end-to-end fashion. We demonstrate MEEP with a dialog assistant that lets users specify trip destinations., Comment: 10 pages

Structured information extraction from document images usually consists of three steps: text detection, text recognition, and text field labeling. While text detection and text recognition have been heavily studied and improved a lot in literature, text field labeling is less explored and still faces many challenges. Existing learning based methods for text labeling task usually require a large amount of labeled examples to train a specific model for each type of document. However, collecting large amounts of document images and labeling them is difficult and sometimes impossible due to privacy issues. Deploying separate models for each type of document also consumes a lot of resources. Facing these challenges, we explore one-shot learning for the text field labeling task. Existing one-shot learning methods for the task are mostly rule-based and have difficulty in labeling fields in crowded regions with few landmarks and fields consisting of multiple separate text regions. To alleviate these problems, we proposed a novel deep end-to-end trainable approach for one-shot text field labeling, which makes use of attention mechanism to transfer the layout information between document images. We further applied conditional random field on the transferred layout information for the refinement of field labeling. We collected and annotated a real-world one-shot field labeling dataset with a large variety of document types and conducted extensive experiments to examine the effectiveness of the proposed model. To stimulate research in this direction, the collected dataset and the one-shot model will be released1., Comment: 9 pages

Analytic window query is a commonly used query in the relational databases. It answers the aggregations of data over a sliding window. For example, to get the average prices of a stock for each day. However, it is not supported in the spatial databases. Because the spatial data are not in a one-dimension space, there is no straightforward way to extend the original analytic window query to spatial databases. But these queries are useful and meaningful. For example, to find the average number of visits for all the POIs in the circle with a fixed radius for each POI as the centre. In this paper, we define the aggregate analytic window query over spatial data and propose algorithms for grid index and tree-index. We also analyze the complexity of the algorithms to prove they are efficient and practical.

Analytic function, also called window function, is to query the aggregation of data over a sliding window. For example, a simple query over the online stock platform is to return the average price of a stock of the last three days. These functions are commonly used features in SQL databases. They are supported in most of the commercial databases. With the increasing usage of cloud data infra and machine learning technology, the frequency of queries with analytic window functions rises. Some analytic functions only require const space in memory to store the state, such as SUM, AVG, while others require linear space, such as MIN, MAX. When the window is extremely large, the memory space to store the state may be too large. In this case, we need to spill the state to disk, which is a heavy operation. In this paper, we proposed an algorithm to manipulate the state data in the disk to reduce the disk I/O to make spill available and efficienct. We analyze the complexity of the algorithm with different data distribution., Comment: 6 pages, conference

Two-dimensional materials have been gaining great attention as they displayed a broad series of electronic properties that ranging from superconductivity to topology. Among them, those which possess magnetism are most desirable, enabling us to manipulate charge and spin simultaneously. Here, based on first-principles calculation, we demonstrate monolayer chromium iodide arsenide (CrIAs), an undiscovered stable two-dimensional material, is an intrinsic ferromagnetic semiconductor with out-of-plane spin magnetization. The indirect bandgaps are predicted to be 0.32 eV for majority spin and 3.31 eV for minority spin, large enough to preserve semiconducting features at room temperature. Its Curie temperature, estimated by Heisenberg model with magnetic anisotropic energy using Monte Carlo method, is as high as 655 K that well above the room temperature, owing to strong direct exchange interaction between chromium d and iodine p orbitals. This work offers the affirmative answer of whether there exists two-dimensional ferromagnetic semiconductor at room temperature. And the practical realization of quantum spintronic devices, which have been suppressed because of lacking suitable room temperature magnetic materials, would embrace a great opportunity.