Search

RAG has become a key technique for enhancing LLMs by reducing hallucinations, especially in domain expert systems where LLMs may lack sufficient inherent knowledge. However, developing these systems in low-resource settings introduces several challenges: (1) handling heterogeneous data sources, (2) optimizing retrieval phase for trustworthy answers, and (3) evaluating generated answers across diverse aspects. To address these, we introduce a data generation pipeline that transforms raw multi-modal data into structured corpus and Q&A pairs, an advanced re-ranking phase improving retrieval precision, and a reference matching algorithm enhancing answer traceability. Applied to the automotive engineering domain, our system improves factual correctness (+1.94), informativeness (+1.16), and helpfulness (+1.67) over a non-RAG baseline, based on a 1-5 scale by an LLM judge. These results highlight the effectiveness of our approach across distinct aspects, with strong answer grounding and transparency.

Creating games together is both a playful and effective way to develop skills in computational thinking, collaboration, and more. However, game development can be challenging for younger developers who lack formal training. While teenage developers frequently turn to online communities for peer support, their experiences may vary. To better understand the benefits and challenges teens face within online developer communities, we conducted interviews with 18 teenagers who created games or elements in Roblox and received peer support from one or more online Roblox developer communities. Our findings show that developer communities provide teens with valuable resources for technical, social, and career growth. However, teenagers also struggle with inter-user conflicts and a lack of community structure, leading to difficulties in handling complex issues that may arise, such as financial scams. Based on these insights, we propose takeaways for creating positive and safe online spaces for teenage game creators., Comment: 18 page, 4 figures, conference

We present cosmological constraints from the Dark Energy Camera All Data Everywhere (DECADE) cosmic shear analysis. This work uses shape measurements for 107 million galaxies measured through Dark Energy Camera (DECam) imaging of $5,\!412$ deg$^2$ of sky that is outside the Dark Energy Survey (DES) footprint. We derive constraints on the cosmological parameters $S_8 = 0.791^{+0.027}_{-0.032}$ and $\Omega_{\rm m} =0.269^{+0.034}_{-0.050}$ for the $\Lambda$CDM model, which are consistent with those from other weak lensing surveys and from the cosmic microwave background. We combine our results with cosmic shear results from DES Y3 at the likelihood level, since the two datasets span independent areas on the sky. The combined measurements, which cover $\approx\! 10,\!000$ deg$^2$, prefer $S_8 = 0.791 \pm 0.023$ and $\Omega_{\rm m} = 0.277^{+0.034}_{-0.046}$ under the $\Lambda$CDM model. These results are the culmination of a series of rigorous studies that characterize and validate the DECADE dataset and the associated analysis methodologies (Anbajagane et. al 2025a,b,c). Overall, the DECADE project demonstrates that the cosmic shear analysis methods employed in Stage-III weak lensing surveys can provide robust cosmological constraints for fairly inhomogeneous datasets. This opens the possibility of using data that have been previously categorized as ``unusable'' for cosmic shear analyses, thereby increasing the statistical power of upcoming weak lensing surveys., Comment: 27 pages, 13 figures

We present the Dark Energy Camera All Data Everywhere (DECADE) weak lensing dataset: a catalog of 107 million galaxies observed by the Dark Energy Camera (DECam) in the northern Galactic cap. This catalog was assembled from public DECam data including survey and standard observing programs. These data were consistently processed with the Dark Energy Survey Data Management pipeline as part of the DECADE campaign and serve as the basis of the DECam Local Volume Exploration survey (DELVE) Early Data Release 3 (EDR3). We apply the Metacalibration measurement algorithm to generate and calibrate galaxy shapes. After cuts, the resulting cosmology-ready galaxy shape catalog covers a region of $5,\!412 \,\,{\rm deg}^2$ with an effective number density of $4.59\,\, {\rm arcmin}^{-2}$. The coadd images used to derive this data have a median limiting magnitude of $r = 23.6$, $i = 23.2$, and $z = 22.6$, estimated at ${\rm S/N} = 10$ in a 2 arcsecond aperture. We present a suite of detailed studies to characterize the catalog, measure any residual systematic biases, and verify that the catalog is suitable for cosmology analyses. In parallel, we build an image simulation pipeline to characterize the remaining multiplicative shear bias in this catalog, which we measure to be $m = (-2.454 \pm 0.124) \times10^{-2}$ for the full sample. Despite the significantly inhomogeneous nature of the data set, due to it being an amalgamation of various observing programs, we find the resulting catalog has sufficient quality to yield competitive cosmological constraints.

We present the photometric redshift characterization and calibration for the Dark Energy Camera All Data Everywhere (DECADE) weak lensing dataset: a catalog of 107 million galaxies observed by the Dark Energy Camera (DECam) in the northern Galactic cap. The redshifts are estimated from a combination of wide-field photometry, deep-field photometry with associated redshift estimates, and a transfer function between the wide field and deep field that is estimated using a source injection catalog. We construct four tomographic bins for the galaxy catalog, and estimate the redshift distribution, $n(z)$, within each one using the Self-organizing Map Photo-Z (SOMPZ) methodology. Our estimates include the contributions from sample variance, zeropoint calibration uncertainties, and redshift biases, as quantified for the deep-field dataset. The total uncertainties on the mean redshifts are $\sigma_{\langle z \rangle} \approx 0.01$. The SOMPZ estimates are then compared to those from the clustering redshift method, obtained by cross-correlating our source galaxies with galaxies in spectroscopic surveys, and are shown to be consistent with each other.

Materials synthesis is vital for innovations such as energy storage, catalysis, electronics, and biomedical devices. Yet, the process relies heavily on empirical, trial-and-error methods guided by expert intuition. Our work aims to support the materials science community by providing a practical, data-driven resource. We have curated a comprehensive dataset of 17K expert-verified synthesis recipes from open-access literature, which forms the basis of our newly developed benchmark, AlchemyBench. AlchemyBench offers an end-to-end framework that supports research in large language models applied to synthesis prediction. It encompasses key tasks, including raw materials and equipment prediction, synthesis procedure generation, and characterization outcome forecasting. We propose an LLM-as-a-Judge framework that leverages large language models for automated evaluation, demonstrating strong statistical agreement with expert assessments. Overall, our contributions offer a supportive foundation for exploring the capabilities of LLMs in predicting and guiding materials synthesis, ultimately paving the way for more efficient experimental design and accelerated innovation in materials science., Comment: under review

This manuscript describes an innovative method to tag the muons using the baseline information of the Flash ADC (FADC) waveform of PMTs in the JSNS1 (J-PARC Sterile Neutrino Search at J-PARC Spallation Neutron Source) experiment. This experiment is designed for the search for sterile neutrinos, and a muon tagging is an essential key component for the background rejection since the detector of the experiment is located over-ground, where is the 3rd floor of the J-PARC Material and Life experimental facility (MLF). Especially, stopping muons inside the detector create the Michel electrons, and they are important background to be rejected. Utilizing this innovative method, more than 99.8% of Michel electrons can be rejected even without a detector veto region. This technique can be employed for any experiments which uses the similar detector configurations., Comment: 7 pages, 6 figures

Self-supervised monocular depth estimation (SSMDE) aims to predict the dense depth map of a monocular image, by learning depth from RGB image sequences, eliminating the need for ground-truth depth labels. Although this approach simplifies data acquisition compared to supervised methods, it struggles with reflective surfaces, as they violate the assumptions of Lambertian reflectance, leading to inaccurate training on such surfaces. To tackle this problem, we propose a novel training strategy for an SSMDE by leveraging triplet mining to pinpoint reflective regions at the pixel level, guided by the camera geometry between different viewpoints. The proposed reflection-aware triplet mining loss specifically penalizes the inappropriate photometric error minimization on the localized reflective regions while preserving depth accuracy in non-reflective areas. We also incorporate a reflection-aware knowledge distillation method that enables a student model to selectively learn the pixel-level knowledge from reflective and non-reflective regions. This results in robust depth estimation across areas. Evaluation results on multiple datasets demonstrate that our method effectively enhances depth quality on reflective surfaces and outperforms state-of-the-art SSMDE baselines., Comment: Accepted at ICLR 2025

Recent studies have combined Large Language Models (LLMs) with Knowledge Graphs (KGs) to enhance reasoning, improving inference accuracy without additional training while mitigating hallucination. However, existing frameworks are often rigid, struggling to adapt to KG or task changes. They also rely heavily on powerful LLMs for reliable (i.e., trustworthy) reasoning. To address this, We introduce R2-KG, a plug-and-play, dual-agent framework that separates reasoning into two roles: an Operator (a low-capacity LLM) that gathers evidence and a Supervisor (a high-capacity LLM) that makes final judgments. This design is cost-efficient for LLM inference while still maintaining strong reasoning accuracy. Additionally, R2-KG employs an Abstention mechanism, generating answers only when sufficient evidence is collected from KG, which significantly enhances reliability. Experiments across multiple KG-based reasoning tasks show that R2-KG consistently outperforms baselines in both accuracy and reliability, regardless of the inherent capability of LLMs used as the Operator. Further experiments reveal that the single-agent version of R2-KG, equipped with a strict self-consistency strategy, achieves significantly higher-than-baseline reliability while reducing inference cost. However, it also leads to a higher abstention rate in complex KGs. Our findings establish R2-KG as a flexible and cost-effective solution for KG-based reasoning. It reduces reliance on high-capacity LLMs while ensuring trustworthy inference.

Ultrafast light-matter interactions inspire potential functionalities in picosecond optoelectronic applications. However, achieving directional carrier dynamics in metals remains challenging due to strong carrier scattering within a multiband environment, typically expected to isotropic carrier relaxation. In this study, we demonstrate epitaxial RuO2/TiO2 (110) heterostructures grown by hybrid molecular beam epitaxy to engineer polarization-selectivity of ultrafast light-matter interactions via anisotropic strain engineering. Combining spectroscopic ellipsometry, X-ray absorption spectroscopy, and optical pump-probe spectroscopy, we revealed the strong anisotropic transient optoelectronic response of strain-engineered RuO2/TiO2 (110) heterostructures along both in-plane [001] and [1-10] crystallographic directions. Theoretical analysis identifies strain-induced modifications in band nesting as the underlying mechanism for enhanced anisotropic carrier relaxation. These findings establish epitaxial strain engineering as a powerful tool for tuning anisotropic optoelectronic responses in metallic systems, paving the way for next-generation polarization-sensitive ultrafast optoelectronic devices., Comment: 18 pages

We address the convergence problem in learning the Optimal Transport (OT) map, where the OT Map refers to a map from one distribution to another while minimizing the transport cost. Semi-dual Neural OT, a widely used approach for learning OT Maps with neural networks, often generates fake solutions that fail to transfer one distribution to another accurately. We identify a sufficient condition under which the max-min solution of Semi-dual Neural OT recovers the true OT Map. Moreover, to address cases when this sufficient condition is not satisfied, we propose a novel method, OTP, which learns both the OT Map and the Optimal Transport Plan, representing the optimal coupling between two distributions. Under sharp assumptions on the distributions, we prove that our model eliminates the fake solution issue and correctly solves the OT problem. Our experiments show that the OTP model recovers the optimal transport map where existing methods fail and outperforms current OT-based models in image-to-image translation tasks. Notably, the OTP model can learn stochastic transport maps when deterministic OT Maps do not exist, such as one-to-many tasks like colorization., Comment: 18 pages, 10 figures

We report limits on WIMP dark matter derived from three years of data collected by the COSINE-100 experiment with NaI(Tl) crystals, achieving an improved energy threshold of 0.7 keV. This lowered threshold enhances sensitivity in the sub-GeV mass range, extending the reach for direct detection of low-mass dark matter. Although no excess of WIMP-like events was observed, the increased sensitivity enabled a model-independent comparison between the expected WIMP signal rate-based on mass limits from our data-and DAMA's reported modulation amplitude. Our findings strongly disfavor the DAMA signal as originating from WIMP interactions, fully excluding DAMA/LIBRA 3$\sigma$ allowed regions and providing enhanced WIMP mass limits by an order of magnitude in the spin-independent model compared to previous results. In the spin-dependent model, cross-section upper limits were obtained in the mass range [0.1-5.0] GeV/c$^2$, with additional sensitivity to sub-GeV WIMPs through the inclusion of the Migdal effect. These results represent substantial progress in low-mass dark matter exploration and reinforce constraints on the longstanding DAMA claim.

The anomalous Hall effect (AHE), a hallmark of time-reversal symmetry breaking, has been reported in rutile RuO2, a debated metallic altermagnetic candidate. Previously, AHE in RuO2 was observed only in strain-relaxed thick films under extremely high magnetic fields (~50 T). Yet, in ultrathin strained films with distinctive anisotropic electronic structures, there are no reports, likely due to disorder and defects suppressing metallicity thus hindering its detection. Here, we demonstrate that ultrathin, fully-strained 2 nm TiO2/t nm RuO2/TiO2 (110) heterostructures, grown by hybrid molecular beam epitaxy, retain metallicity and exhibit a sizeable AHE at a significantly lower magnetic field (< 9 T). Density functional theory calculations reveal that epitaxial strain stabilizes a non-compensated magnetic ground state and reconfigures magnetic ordering in RuO2 (110) thin films. These findings establish ultrathin RuO2 as a platform for strain-engineered magnetism and underscore the transformative potential of epitaxial design in advancing spintronic technologies., Comment: 23 pages

As chatbots become increasingly integrated into everyday tasks, designing systems that accommodate diverse user populations is crucial for fostering trust, engagement, and inclusivity. This study investigates the ability of contemporary Large Language Models (LLMs) to generate African American Vernacular English (AAVE) and evaluates the impact of AAVE usage on user experiences in chatbot applications. We analyze the performance of three LLM families (Llama, GPT, and Claude) in producing AAVE-like utterances at varying dialect intensities and assess user preferences across multiple domains, including healthcare and education. Despite LLMs' proficiency in generating AAVE-like language, findings indicate that AAVE-speaking users prefer Standard American English (SAE) chatbots, with higher levels of AAVE correlating with lower ratings for a variety of characteristics, including chatbot trustworthiness and role appropriateness. These results highlight the complexities of creating inclusive AI systems and underscore the need for further exploration of diversity to enhance human-computer interactions.

Scientific figure captioning is a complex task that requires generating contextually appropriate descriptions of visual content. However, existing methods often fall short by utilizing incomplete information, treating the task solely as either an image-to-text or text summarization problem. This limitation hinders the generation of high-quality captions that fully capture the necessary details. Moreover, existing data sourced from arXiv papers contain low-quality captions, posing significant challenges for training large language models (LLMs). In this paper, we introduce a framework called Multi-LLM Collaborative Figure Caption Generation (MLBCAP) to address these challenges by leveraging specialized LLMs for distinct sub-tasks. Our approach unfolds in three key modules: (Quality Assessment) We utilize multimodal LLMs to assess the quality of training data, enabling the filtration of low-quality captions. (Diverse Caption Generation) We then employ a strategy of fine-tuning/prompting multiple LLMs on the captioning task to generate candidate captions. (Judgment) Lastly, we prompt a prominent LLM to select the highest quality caption from the candidates, followed by refining any remaining inaccuracies. Human evaluations demonstrate that informative captions produced by our approach rank better than human-written captions, highlighting its effectiveness. Our code is available at https://github.com/teamreboott/MLBCAP, Comment: Accepted to AAAI 2025 AI4Research Workshop

Using $20.3~\mathrm{fb}^{-1}$ of $e^+e^-$ collision data collected at a center-of-mass energy of 3.773 GeV with the BESIII detector operating at the BEPCII collider, the branching fractions of three hadronic charm meson decays, $D^+\to \phi\pi^+\pi^+\pi^-$, $D^+\to K^0_SK^+\pi^+\pi^-\pi^0$, and $D^+\to K^0_SK^+\omega$, are measured for the first time to be $(0.54\pm0.19\pm0.02)\times 10^{-4}$, $(2.51\pm0.34\pm0.14)\times 10^{-4}$, and $(2.02\pm0.35\pm0.10)\times 10^{-4}$, respectively. Futhermore, the branching fractions of $D^+\to K^+K^-\pi^+\pi^+\pi^-$ and $D^+\to K^0_SK^+\eta$ are measured with improved precision, yielding values of $(0.66\pm0.11\pm0.03)\times 10^{-4}$ and $(2.27\pm0.22\pm0.05)\times 10^{-4}$, respectively., Comment: 11 pages, 3 figures

A Bier sphere is a simplicial sphere obtained as the deleted join of a simplicial complex and its combinatorial Alexander dual. We focus on particular classes of full subcomplexes of Bier spheres, and determine their topological types. As applications, we explicitly describe the cohomology of real toric spaces associated with Bier spheres., Comment: 12 pages, 2 figures

Graph processing requires irregular, fine-grained random access patterns incompatible with contemporary off-chip memory architecture, leading to inefficient data access. This inefficiency makes graph processing an extremely memory-bound application. Because of this, existing graph processing accelerators typically employ a graph tiling-based or processing-in-memory (PIM) approach to relieve the memory bottleneck. In the tiling-based approach, a graph is split into chunks that fit within the on-chip cache to maximize data reuse. In the PIM approach, arithmetic units are placed within memory to perform operations such as reduction or atomic addition. However, both approaches have several limitations, especially when implemented on current memory standards (i.e., DDR). Because the access granularity provided by DDR is much larger than that of the graph vertex property data, much of the bandwidth and cache capacity are wasted. PIM is meant to alleviate such issues, but it is difficult to use in conjunction with the tiling-based approach, resulting in a significant disadvantage. Furthermore, placing arithmetic units inside a memory chip is expensive, thereby supporting multiple types of operation is thought to be impractical. To address the above limitations, we present Piccolo, an end-to-end efficient graph processing accelerator with fine-grained in-memory random scatter-gather. Instead of placing expensive arithmetic units in off-chip memory, Piccolo focuses on reducing the off-chip traffic with non-arithmetic function-in-memory of random scatter-gather. To fully benefit from in-memory scatter-gather, Piccolo redesigns the cache and MHA of the accelerator such that it can enjoy both the advantage of tiling and in-memory operations. Piccolo achieves a maximum speedup of 3.28$\times$ and a geometric mean speedup of 1.62$\times$ across various and extensive benchmarks., Comment: HPCA 2025

We introduce Paralinguistic Speech Captions (ParaSpeechCaps), a large-scale dataset that annotates speech utterances with rich style captions. While rich abstract tags (e.g. guttural, nasal, pained) have been explored in small-scale human-annotated datasets, existing large-scale datasets only cover basic tags (e.g. low-pitched, slow, loud). We combine off-the-shelf text and speech embedders, classifiers and an audio language model to automatically scale rich tag annotations for the first time. ParaSpeechCaps covers a total of 59 style tags, including both speaker-level intrinsic tags and utterance-level situational tags. It consists of 342 hours of human-labelled data (PSC-Base) and 2427 hours of automatically annotated data (PSC-Scaled). We finetune Parler-TTS, an open-source style-prompted TTS model, on ParaSpeechCaps, and achieve improved style consistency (+7.9% Consistency MOS) and speech quality (+15.5% Naturalness MOS) over the best performing baseline that combines existing rich style tag datasets. We ablate several of our dataset design choices to lay the foundation for future work in this space. Our dataset, models and code are released at https://github.com/ajd12342/paraspeechcaps .

We report a precise measurement of the ratio of branching fractions $\mathcal{B}(\Lambda_c^+\to p K_S^0 \pi^0)/\mathcal{B}(\Lambda_c^+\to p K^- \pi^+)$ using 980 fb$^{-1}$ of $e^+e^-$ data from the Belle experiment. We obtain a value of $\mathcal{B}(\Lambda_c^+\to p K_S^0 \pi^0)/\mathcal{B}(\Lambda_c^+\to p K^- \pi^+)=0.339\pm 0.002\pm 0.009$, where the first and second uncertainties are statistical and systematic, respectively. This Belle result is consistent with the previous measurement from the CLEO experiment but has a fivefold improvement in precision. By combining our result with the world average $\mathcal{B}(\Lambda_c^+\to p K^- \pi^+)$, we obtain the absolute branching fraction $\mathcal{B}(\Lambda_c^+\to p K_S^0 \pi^0)=(2.12\pm 0.01\pm 0.05 \pm 0.10)\%$, where the uncertainties are statistical, systematic, and the uncertainty in the absolute branching fraction scale $\mathcal{B}(\Lambda_c^+\to p K^- \pi^+)$, respectively. This measurement can shed light on hadronic decay mechanisms in charmed baryon decays., Comment: 20 pages, 7 figures

Emergent symmetry breakings in condensed matter systems are often intimately linked to collective excitations. For example, the intertwined spin-charge stripe order in cuprate superconductors is associated with spin and charge excitations. While the collective behavior of spin excitations is well established, the nature of charge excitations remains to be understood. Here we present a high-resolution resonant inelastic x-ray scattering (RIXS) study of charge excitations in the stripe-ordered cuprate La$_{1.675}$Eu$_{0.2}$Sr$_{0.125}$CuO$_4$. The RIXS spectra consist of both charge and phonon excitations around the charge ordering wave vector. By modeling the momentum-dependent phonon intensity, the charge-excitation spectral weight is extracted for a wide range of energy. As such, we reveal the highly dispersive nature of the charge excitations, with an energy scale comparable to the spin excitations. Since charge order and superconductivity in cuprates are possibly driven by the same electronic correlations, determining the interaction strength underlying charge order is essential to establishing a comprehensive microscopic model of high-temperature superconductivity., Comment: Supplementary Information available upon request

Implicit Neural Representation for Videos (NeRV) has introduced a novel paradigm for video representation and compression, outperforming traditional codecs. As model size grows, however, slow encoding and decoding speed and high memory consumption hinder its application in practice. To address these limitations, we propose a new video representation and compression method based on 2D Gaussian Splatting to efficiently handle video data. Our proposed deformable 2D Gaussian Splatting dynamically adapts the transformation of 2D Gaussians at each frame, significantly reducing memory cost. Equipped with a multi-plane-based spatiotemporal encoder and a lightweight decoder, it predicts changes in color, coordinates, and shape of initialized Gaussians, given the time step. By leveraging temporal gradients, our model effectively captures temporal redundancy at negligible cost, significantly enhancing video representation efficiency. Our method reduces GPU memory usage by up to 78.4%, and significantly expedites video processing, achieving 5.5x faster training and 12.5x faster decoding compared to the state-of-the-art NeRV methods.

The growing adoption of Large Language Models (LLMs) across various domains has driven the demand for efficient and scalable AI-serving solutions. Deploying LLMs requires optimizations to manage their significant computational and data demands. The prefill stage processes large numbers of input tokens in parallel, increasing computational load, while the decoding stage relies heavily on memory bandwidth due to the auto-regressive nature of LLMs. Current hardware, such as GPUs, often fails to balance these demands, leading to inefficient utilization. While batching improves hardware efficiency, it delays response times, degrading Quality-of-Service (QoS). This disconnect between vendors, who aim to maximize resource efficiency, and users, who prioritize low latency, highlights the need for a better solution. To address this, we propose ADOR, a framework that automatically identifies and recommends hardware architectures tailored to LLM serving. By leveraging predefined architecture templates specialized for heterogeneous dataflows, ADOR optimally balances throughput and latency. It efficiently explores design spaces to suggest architectures that meet the requirements of both vendors and users. ADOR demonstrates substantial performance improvements, achieving 2.51x higher QoS and 4.01x better area efficiency compared to the A100 at high batch sizes, making it a robust solution for scalable and cost-effective LLM serving., Comment: 11pages, 17 figures

In this paper we consider Brill-Noether loci of rank-two, stable vector bundles of given degree $d$, with speciality $2$ and $3$, on a general $\nu$-gonal curve $C$ of genus $g$ with the aim of studying their irreducible components in the whole range of interest for $d$, namely $2g-2 \leq d \leq 4g-4$. For speciality $2$, either we prove that such a Brill-Noether locus is empty or we completely classify all of its irreducible components, giving also some extra information about their dimensions (exhibiting both regular and superabundant components), about their birational geometry as well as giving precise description of their general point. Similarly, in speciality $3$, either we prove that such a Brill-Noether locus is empty or we exhibit irreducible components of several types according to their regularity or superabundance but also according to the precise descriptions of their general points, proving e.g. that in some degrees $d$ there are more than one superabundant component together with the presence of also a regular component. We moreover give extra information about the birational geometry of the constructed components as well as about their local behavior. As a by-product of our general results, we deduce also some consequences on Brill-Noether loci in rank two with fixed general determinant line bundle instead of fixed degree., Comment: Authors acknowledge support from funds "AAAGH" (P.I. G. Pareschi) during visit of first and third author at the Dept. Mathematics-Univ. Rome "Tor Vergata" in November 2023, when collaboration began; from MIUR Excellence Proj. MatMod@TOV 2023-27, and from the Organizers of "Workshop in Classical Algebraic Geometry", Daejeon-South Korea Dec. 2023(D. S. Hwang, J.-M. Hwang, Y. Lee)

How capable are large language models (LLMs) in the domain of taxation? Although numerous studies have explored the legal domain in general, research dedicated to taxation remain scarce. Moreover, the datasets used in these studies are either simplified, failing to reflect the real-world complexities, or unavailable as open source. To address this gap, we introduce PLAT, a new benchmark designed to assess the ability of LLMs to predict the legitimacy of additional tax penalties. PLAT is constructed to evaluate LLMs' understanding of tax law, particularly in cases where resolving the issue requires more than just applying related statutes. Our experiments with six LLMs reveal that their baseline capabilities are limited, especially when dealing with conflicting issues that demand a comprehensive understanding. However, we found that enabling retrieval, self-reasoning, and discussion among multiple agents with specific role assignments, this limitation can be mitigated., Comment: 5 pages

The objective of this work is to align asynchronous subtitles in sign language videos with limited labelled data. To achieve this goal, we propose a novel framework with the following contributions: (1) we leverage fundamental grammatical rules of British Sign Language (BSL) to pre-process the input subtitles, (2) we design a selective alignment loss to optimise the model for predicting the temporal location of signs only when the queried sign actually occurs in a scene, and (3) we conduct self-training with refined pseudo-labels which are more accurate than the heuristic audio-aligned labels. From this, our model not only better understands the correlation between the text and the signs, but also holds potential for application in the translation of sign languages, particularly in scenarios where manual labelling of large-scale sign data is impractical or challenging. Extensive experimental results demonstrate that our approach achieves state-of-the-art results, surpassing previous baselines by substantial margins in terms of both frame-level accuracy and F1-score. This highlights the effectiveness and practicality of our framework in advancing the field of sign language video alignment and translation.

In the age of increasingly realistic generative AI, robust deepfake detection is essential for mitigating fraud and disinformation. While many deepfake detectors report high accuracy on academic datasets, we show that these academic benchmarks are out of date and not representative of real-world deepfakes. We introduce Deepfake-Eval-2024, a new deepfake detection benchmark consisting of in-the-wild deepfakes collected from social media and deepfake detection platform users in 2024. Deepfake-Eval-2024 consists of 45 hours of videos, 56.5 hours of audio, and 1,975 images, encompassing the latest manipulation technologies. The benchmark contains diverse media content from 88 different websites in 52 different languages. We find that the performance of open-source state-of-the-art deepfake detection models drops precipitously when evaluated on Deepfake-Eval-2024, with AUC decreasing by 50\% for video, 48\% for audio, and 45\% for image models compared to previous benchmarks. We also evaluate commercial deepfake detection models and models finetuned on Deepfake-Eval-2024, and find that they have superior performance to off-the-shelf open-source models, but do not yet reach the accuracy of deepfake forensic analysts. The dataset is available at https://github.com/nuriachandra/Deepfake-Eval-2024.

This paper argues that a dataset's legal risk cannot be accurately assessed by its license terms alone; instead, tracking dataset redistribution and its full lifecycle is essential. However, this process is too complex for legal experts to handle manually at scale. Tracking dataset provenance, verifying redistribution rights, and assessing evolving legal risks across multiple stages require a level of precision and efficiency that exceeds human capabilities. Addressing this challenge effectively demands AI agents that can systematically trace dataset redistribution, analyze compliance, and identify legal risks. We develop an automated data compliance system called NEXUS and show that AI can perform these tasks with higher accuracy, efficiency, and cost-effectiveness than human experts. Our massive legal analysis of 17,429 unique entities and 8,072 license terms using this approach reveals the discrepancies in legal rights between the original datasets before redistribution and their redistributed subsets, underscoring the necessity of the data lifecycle-aware compliance. For instance, we find that out of 2,852 datasets with commercially viable individual license terms, only 605 (21%) are legally permissible for commercialization. This work sets a new standard for AI data governance, advocating for a framework that systematically examines the entire lifecycle of dataset redistribution to ensure transparent, legal, and responsible dataset management.

We present a minimal setup within the framework of Horndeski gravity that can describe a nonpathological Genesis scenario. Our setup allows for a fully stable transition to the kination epoch, during which General Relativity (GR) is restored. This Genesis scenario circumvents the no-go theorem at the cost of encountering the risk of strong coupling in the past. Interestingly, our scenario admits two different regimes for the background solution for Hubble parameter at the Genesis stage: power-law behavior and manifestly non-power-law behavior. We explicitly show that, in both regimes, our model remains within unitarity bounds. In most cases, the tensor spectrum is blue-tilted. Then, we adopt a mechanism with a spectator field that allows for a red-tilted scalar power spectrum. We also suggest a deformation of the model that enables us to achieve sufficiently small values for the r ratio. Finally, we discuss the geodesic (in)completeness of the current model., Comment: 58 pages, 17 figures

Misinformation surrounding emerging outbreaks poses a serious societal threat, making robust countermeasures essential. One promising approach is stance detection (SD), which identifies whether social media posts support or oppose misleading claims. In this work, we finetune classifiers on COVID-19 misinformation SD datasets consisting of claims and corresponding tweets. Specifically, we test controllable misinformation generation (CMG) using large language models (LLMs) as a method for data augmentation. While CMG demonstrates the potential for expanding training datasets, our experiments reveal that performance gains over traditional augmentation methods are often minimal and inconsistent, primarily due to built-in safeguards within LLMs. We release our code and datasets to facilitate further research on misinformation detection and generation.

Using language models to scalably approximate human preferences on text quality (LLM-as-a-judge) has become a standard practice applicable to many tasks. A judgment is often extracted from the judge's textual output alone, typically with greedy decoding. However, LLM judges naturally provide distributions over judgment tokens, inviting a breadth of inference methods for extracting fine-grained preferences. We find that taking the mean of the judgment distribution consistently outperforms taking the mode (i.e. greedy decoding) in all evaluation settings (i.e. pointwise, pairwise, and listwise). We further explore novel methods of deriving preferences from judgment distributions, and find that methods incorporating risk aversion often improve performance. Lastly, we analyze LLM-as-a-judge paired with chain-of-thought (CoT) prompting, showing that CoT can collapse the spread of the judgment distribution, often harming performance. Our findings suggest leveraging distributional output can improve LLM-as-a-judge, as opposed to using the text interface alone.

Personalizing large language models (LLMs) is important for aligning outputs with diverse user preferences, yet existing methods struggle with flexibility and generalization. We propose CoPL (Collaborative Preference Learning), a graph-based collaborative filtering framework that models user-response relationships to enhance preference estimation, particularly in sparse annotation settings. By integrating a mixture of LoRA experts, CoPL efficiently fine-tunes LLMs while dynamically balancing shared and user-specific preferences. Additionally, an optimization-free adaptation strategy enables generalization to unseen users without fine-tuning. Experiments on UltraFeedback-P demonstrate that CoPL outperforms existing personalized reward models, effectively capturing both common and controversial preferences, making it a scalable solution for personalized LLM alignment., Comment: 13pages, 4 figures, 6tables

A zoned neutral atom architecture achieves exceptional fidelity by segregating the execution spaces of 1- and 2-qubit gates, being a promising candidate for high-accuracy quantum systems. Unfortunately, naively applying programs designed for static qubit topologies to zoned architectures may result in most execution time being consumed by inter-zone travels of atoms. To address this, we introduce Mantra (Minimizing trAp movemeNts for aTom aRray Architectures), which rewrites quantum programs to reduce the interleaving of single- and two-qubit gates. Mantra incorporates three strategies: (i) a fountain-shaped controlled-Z (CZ) chain, (ii) ZZ-interaction protocol without a 1-qubit gate, and (iii) preemptive gate scheduling. Mantra reduces inter-zone movements by 68%, physical gate counts by 35%, and improves circuit fidelities by 17% compared to the standard executions., Comment: 17 pages, 16 figures, to be published in The 2025 International Symposium on Code Generation and Optimization (CGO '25)

Using $e^+e^-$ collision data taken at the center-of-mass energy of 3.773 GeV with the BESIII detector, corresponding to an integrated luminosity of 20.3 fb$^{-1}$, we report the first amplitude and angular analyses of the semileptonic decays $D^{+(0)}\to K^-\pi^+\pi^{0(-)} e^+\nu_e$. From the amplitude analysis, we determine for the first time the hadronic form factors of the semileptonic $D$ decays into the axial-vector meson $\bar{K}_1(1270)$ to be $r_A=(-11.2\pm1.0\pm0.9)\times10^{-2}$ and $r_V = (-4.3\pm 1.0\pm2.4)\times 10^{-2}$. The angular analysis yields an up-down asymmetry $\mathcal{A}^\prime_{ud} = 0.01\pm0.11$, which is consistent with the Standard Model prediction., Comment: 15 pages, 6 figures, submitted to PRL

Scientific discoveries are often made by finding a pattern or object that was not predicted by the known rules of science. Oftentimes, these anomalous events or objects that do not conform to the norms are an indication that the rules of science governing the data are incomplete, and something new needs to be present to explain these unexpected outliers. The challenge of finding anomalies can be confounding since it requires codifying a complete knowledge of the known scientific behaviors and then projecting these known behaviors on the data to look for deviations. When utilizing machine learning, this presents a particular challenge since we require that the model not only understands scientific data perfectly but also recognizes when the data is inconsistent and out of the scope of its trained behavior. In this paper, we present three datasets aimed at developing machine learning-based anomaly detection for disparate scientific domains covering astrophysics, genomics, and polar science. We present the different datasets along with a scheme to make machine learning challenges around the three datasets findable, accessible, interoperable, and reusable (FAIR). Furthermore, we present an approach that generalizes to future machine learning challenges, enabling the possibility of large, more compute-intensive challenges that can ultimately lead to scientific discovery., Comment: 18 pages 6 figures to be submitted to Nature Communications

Recent advancements in large language models (LLMs) based on transformer architectures have sparked significant interest in understanding their inner workings. In this paper, we introduce a novel approach to modeling transformer architectures using highly flexible non-autonomous neural ordinary differential equations (ODEs). Our proposed model parameterizes all weights of attention and feed-forward blocks through neural networks, expressing these weights as functions of a continuous layer index. Through spectral analysis of the model's dynamics, we uncover an increase in eigenvalue magnitude that challenges the weight-sharing assumption prevalent in existing theoretical studies. We also leverage the Lyapunov exponent to examine token-level sensitivity, enhancing model interpretability. Our neural ODE transformer demonstrates performance comparable to or better than vanilla transformers across various configurations and datasets, while offering flexible fine-tuning capabilities that can adapt to different architectural constraints., Comment: ICLR 2025

Stacking excessive layers in DNN results in highly underdetermined system when training samples are limited, which is very common in medical applications. In this regard, we present a framework capable of deriving an efficient high-dimensional space with reasonable increase in model size. This is done by utilizing a transform (i.e., convolution) that leverages scale-space theory with covariance structure. The overall model trains on this transform together with a downstream classifier (i.e., Fully Connected layer) to capture the optimal multi-scale representation of the original data which corresponds to task-specific components in a dual space. Experiments on neuroimaging measures from Alzheimer's Disease Neuroimaging Initiative (ADNI) study show that our model performs better and converges faster than conventional models even when the model size is significantly reduced. The trained model is made interpretable using gradient information over the multi-scale transform to delineate personalized AD-specific regions in the brain., Comment: ISBI 2023

The recent development of Neural Operator (NeurOp) learning for solutions to the elastic wave equation shows promising results and provides the basis for fast large-scale simulations for different seismological applications. In this paper, we investigate the capabilities of the Fourier Neural Operator (FNO) to directly solve the 3D Helmholtz wave equation for fast forward seismic simulations on different frequencies, and show the frequency bias of the FNO model, i.e. it learns the lower frequencies better comparing to the higher frequencies. To reduce the frequency bias, we adopt the multi-stage FNO training, i.e., after training a 1st stage FNO model for estimating the ground motion, we use a second FNO model as the 2nd stage to learn from the residual, which greatly reduced the errors on the higher frequencies. By adopting this multi-stage training, the FNO models show reduced biases on higher frequencies, which enhanced the overall results of the ground motion simulations. Thus the multi-stage training FNO improves the accuracy and realism of the ground motion simulations.

We explore neuro-symbolic approaches to generalize actionable knowledge, enabling embodied agents to tackle complex tasks more effectively in open-domain environments. A key challenge for embodied agents is the generalization of knowledge across diverse environments and situations, as limited experiences often confine them to their prior knowledge. To address this issue, we introduce a novel framework, NeSyC, a neuro-symbolic continual learner that emulates the hypothetico-deductive model by continually formulating and validating knowledge from limited experiences through the combined use of Large Language Models (LLMs) and symbolic tools. Specifically, we devise a contrastive generality improvement scheme within NeSyC, which iteratively generates hypotheses using LLMs and conducts contrastive validation via symbolic tools. This scheme reinforces the justification for admissible actions while minimizing the inference of inadmissible ones. Additionally, we incorporate a memory-based monitoring scheme that efficiently detects action errors and triggers the knowledge refinement process across domains. Experiments conducted on diverse embodied task benchmarks-including ALFWorld, VirtualHome, Minecraft, RLBench, and a real-world robotic scenario-demonstrate that NeSyC is highly effective in solving complex embodied tasks across a range of open-domain environments., Comment: Accepted at ICLR 2025. Project site with code: https://pjw971022.github.io/nesyc/

Broad exploration of references is critical in the visual design process. While text-to-image (T2I) models offer efficiency and customization of exploration, they often limit support for divergence in exploration. We conducted a formative study (N=6) to investigate the limitations of current interaction with the T2I model for broad exploration and found that designers struggle to articulate exploratory intentions and manage iterative, non-linear workflows. To address these challenges, we developed Expandora. Users can specify their exploratory intentions and desired diversity levels through structured input, and using an LLM-based pipeline, Expandora generates tailored prompt variations. The results are displayed in a mindmap-like interface that encourages non-linear workflows. A user study (N=8) demonstrated that Expandora significantly increases prompt diversity, the number of prompts users tried within a given time, and user satisfaction compared to the baseline. Nonetheless, its limitations in supporting convergent thinking suggest opportunities for holistically improving creative processes., Comment: Accepted to CHI'25 LBW

CeTa$_7$O$_{19}$ is a recently found two-dimensional triangular lattice antiferromagnet without showing magnetic order. We grew high-quality CeTa$_7$O$_{19}$ single crystals and studied the low-temperature magnetic susceptibility, specific heat and thermal conductivity. The dc magnetic susceptibility and magnetization reveal its nature of effective spin-1/2, easy axis anisotropy, and antiferromagnetic spin coupling. The ultralow-temperature ac susceptibility and specific heat data indicate the absence of any phase transition down to 20 mK. The ultralow-temperature thermal conductivity ($\kappa$) at zero magnetic field exhibits a non-zero residual term $\kappa_0/T =$ 0.0056 W/K$^2$m. Although the magnetic field dependence of $\kappa$ is rather weak, the 14 T thermal conductivity shows an essential zero residual term. All these results point to a possible ground state of quantum spin liquid., Comment: 8 pages, 6 figures, accepted for publication in Phys. Rev. B

Quantum circuit simulations are essential for the verification of quantum algorithms on behalf of real quantum devices. However, the memory requirements for such simulations grow exponentially with the number of qubits involved in quantum programs. Moreover, a substantial number of computations in quantum circuit simulations cause low locality data accesses, as they require extensive computations across the entire table of the full state vector. These characteristics lead to significant latency and energy overheads during data transfers between the CPU and main memory. Processing-in-Memory (PIM), which integrates computational logic near DRAM banks, could present a promising solution to address these challenges. In this paper, we introduce PIMutation (PIM framework for qUanTum circuit simulATION) for achieving fast and energy-efficient quantum circuit simulation. PIMutation is the first attempt to leverage UPMEM, a publicly available PIM-integrated DIMM, to implement quantum circuit simulations. PIMutation incorporates three optimization strategies to overcome the overhead of quantum circuit simulation using the real PIM system: (i) gate merging, (ii) row swapping, and (iii) vector partitioning. Our evaluations show that PIMutation achieves an average speedup of 2.99x and 16.51x with a reduction of energy of 25.23% and 75.29% over the QuEST simulator on CPU in 16- and 32-qubit benchmarks, respectively., Comment: 8 pages, 9 figures, to be published in 30th Asia and South Pacific Design Automation Conference (ASPDAC '25)

We describe updated scientific goals for the wide-field, millimeter-wave survey that will be produced by the Simons Observatory (SO). Significant upgrades to the 6-meter SO Large Aperture Telescope (LAT) are expected to be complete by 2028, and will include a doubled mapping speed with 30,000 new detectors and an automated data reduction pipeline. In addition, a new photovoltaic array will supply most of the observatory's power. The LAT survey will cover about 60% of the sky at a regular observing cadence, with five times the angular resolution and ten times the map depth of Planck. The science goals are to: (1) determine the physical conditions in the early universe and constrain the existence of new light particles; (2) measure the integrated distribution of mass, electron pressure, and electron momentum in the late-time universe, and, in combination with optical surveys, determine the neutrino mass and the effects of dark energy via tomographic measurements of the growth of structure at $z < 3$; (3) measure the distribution of electron density and pressure around galaxy groups and clusters, and calibrate the effects of energy input from galaxy formation on the surrounding environment; (4) produce a sample of more than 30,000 galaxy clusters, and more than 100,000 extragalactic millimeter sources, including regularly sampled AGN light-curves, to study these sources and their emission physics; (5) measure the polarized emission from magnetically aligned dust grains in our Galaxy, to study the properties of dust and the role of magnetic fields in star formation; (6) constrain asteroid regoliths, search for Trans-Neptunian Objects, and either detect or eliminate large portions of the phase space in the search for Planet 9; and (7) provide a powerful new window into the transient universe on time scales of minutes to years, concurrent with observations from Rubin of overlapping sky., Comment: 44 pages, 7 figures; abstract slightly abridged. Author contributions to this paper are available at https://simonsobservatory.org/wp-content/uploads/2025/02/Author-contribution-statement-20250228.pdf

Single ferroelectric memcapacitor-based time-domain (TD) content-addressable memory (CAM) is proposed and experimentally demonstrated for high reliability and density. The proposed TD CAM features the symmetric capacitance-voltage characteristics of a ferroelectric memcapacitor with a gated p-i-n diode structure. This CAM performs search operations based on the variable capacitance of cells. The propagation delay of the TD CAM output signal is linearly correlated with the Hamming distance (HD) between input and output vectors. The proposed TD CAM array exhibits exceptional reliability in HD computation and in-memory search tasks owing to this linearity, considerably outperforming the conventional nonlinear voltage-domain CAM.

This monograph is on convex real projective structures on strongly tame n-orbifolds with some appropriate conditions on ends., Comment: xiii + 419 pages, 7 figures

Retrieval-augmented generation (RAG) enhances LLMs by integrating external knowledge, but generation remains fragile due to the uncertain placement of relevant chunks and retrieval-induced information overload, leading to hallucinations. We propose Ext2Gen, a novel extract-then-generate model that enhances RAG robustness by first extracting query-relevant sentences before generating answers. To optimize this model, we employ preference alignment through pairwise feedback learning, enabling the model to generate robust answers regardless of variations in retrieval results. Extensive experiments demonstrate that Ext2Gen effectively identifies query-relevant sentences with high precision and recall, leading to highly reliable answers. Furthermore, deploying our model in a RAG environment reveals that it not only boosts the performance of the base LLM but also synergizes with advanced retrieval strategies like query expansion. The dataset and model will be released soon.

Electrical control of quantum magnetic states is essential in spintronic science. Initial studies on the ferromagnetic state control were extended to collinear antiferromagnets and, more recently, noncollinear antiferromagnets. However, electrical control mechanisms of such exotic magnetic states remain poorly understood. Here, we report the first experimental and theoretical example of the current control of helical antiferromagnets, arising from the competition between collinear antiferromagnetic exchange and interlayer Dzyaloshinskii-Moriya interaction in new van-der-Waals (vdW) material Ni1/3NbS2. Due to the intrinsic broken inversion symmetry, an in-plane current generates spin-orbit torque that, in turn, interacts directly with the helical antiferromagnetic order. Our theoretical analyses indicate that a weak ferromagnetic order coexists due to the Dzyaloshinskii-Moriya interaction, mediating the spin-orbit torque to collectively rotate the helical antiferromagnetic order. Our Ni1/3NbS2 nanodevice experiments produce current-dependent resistance change consistent with the theoretical prediction. This work widens our understanding of the electrical control of helical antiferromagnets and promotes vdW quantum magnets as interesting material platforms for electrical control., Comment: Accepted by Physical Review Letters; 41 pages, 4 main figures, 12 supporting figures

In this work, we study a family of vector bundles on the moduli space of curves constructed from representations of $\text{Vir}_{2k+1,2}$, a family of vertex operator algebras derived from the Virasoro Lie algebra. Using the relationship between rank and degree, we characterize their asymptotic behavior, demonstrating that their first Chern classes are nef on $\overline{\rm{M}}_{g,n}$ in many cases. This is the first nontrivial example of divisors arising from vertex operator algebras that are uniformly positive for all genera. Furthermore, for $g = 1$, these divisors form a $\mathbb{Q}$-basis of the Picard group of $\overline{\rm{M}}_{1,n}$, with several desirable functorial properties. Using this basis, we characterize line bundles on certain contractions of $\overline{\rm{M}}_{1,n}$. We also propose conjectures regarding the conformal blocks of Virasoro VOAs and potential generalizations. In particular, by introducing a generalization of conformal block divisors, we provide a nonlinear nef interpolation between affine and Virasoro conformal block divisors., Comment: 37 pages. Comments Welcome!