arXiv subjects

Kuan-Wei Huang

Publications and source records attributed to Kuan-Wei Huang.

9 recordsLinked to original sources

arXiv2025

InterpDetect: Interpretable Signals for Detecting Hallucinations in Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) integrates external knowledge to mitigate hallucinations, yet models often generate outputs inconsistent with retrieved content. Accurate hallucination detection requires disentangling the contributions of external context and parametric knowledge, which prior methods typically conflate. We investigate the mechanisms underlying RAG hallucinations and find they arise when later-layer FFN modules disproportionately inject parametric knowledge into the residual stream. To address this, we explore a mechanistic detection approach based on external context scores and parametric knowledge scores. Using Qwen3-0.6b, we compute these scores across layers and attention heads and train regression-based classifiers to predict hallucinations. Our method is evaluated against state-of-the-art LLMs (GPT-5, GPT-4.1) and detection baselines (RAGAS, TruLens, RefChecker). Furthermore, classifiers trained on Qwen3-0.6b signals generalize to GPT-4.1-mini responses, demonstrating the potential of proxy-model evaluation. Our results highlight mechanistic signals as efficient, generalizable predictors for hallucination detection in RAG systems.

cs.CL↗

arXiv2025

FRED: Financial Retrieval-Enhanced Detection and Editing of Hallucinations in Language Models

Hallucinations in large language models pose a critical challenge for applications requiring factual reliability, particularly in high-stakes domains such as finance. This work presents an effective approach for detecting and editing factually incorrect content in model-generated responses based on the provided context. Given a user-defined domain-specific error taxonomy, we construct a synthetic dataset by inserting tagged errors into financial question-answering corpora and then fine-tune four language models, Phi-4, Phi-4-mini, Qwen3-4B, and Qwen3-14B, to detect and edit these factual inaccuracies. Our best-performing model, fine-tuned Phi-4, achieves an 8% improvement in binary F1 score and a 30% gain in overall detection performance compared to OpenAI-o3. Notably, our fine-tuned Phi-4-mini model, despite having only 4 billion parameters, maintains competitive performance with just a 2% drop in binary detection and a 0.1% decline in overall detection compared to OpenAI-o3. Our work provides a practical solution for detecting and editing factual inconsistencies in financial text generation while introducing a generalizable framework that can enhance the trustworthiness and alignment of large language models across diverse applications beyond finance. Our code and data are available at https://github.com/pegasi-ai/shield.

cs.CL↗

arXiv2023

LenSiam: Self-Supervised Learning on Strong Gravitational Lens Images

Self-supervised learning has been known for learning good representations from data without the need for annotated labels. We explore the simple siamese (SimSiam) architecture for representation learning on strong gravitational lens images. Commonly used image augmentations tend to change lens properties; for example, zoom-in would affect the Einstein radius. To create image pairs representing the same underlying lens model, we introduce a lens augmentation method to preserve lens properties by fixing the lens model while varying the source galaxies. Our research demonstrates this lens augmentation works well with SimSiam for learning the lens image representation without labels, so we name it LenSiam. We also show that a pre-trained LenSiam model can benefit downstream tasks. We open-source our code and datasets at https://github.com/kuanweih/LenSiam .

astro-ph.IM↗

arXiv2022

Strong Gravitational Lensing Parameter Estimation with Vision Transformer

Quantifying the parameters and corresponding uncertainties of hundreds of strongly lensed quasar systems holds the key to resolving one of the most important scientific questions: the Hubble constant ($H_{0}$) tension. The commonly used Markov chain Monte Carlo (MCMC) method has been too time-consuming to achieve this goal, yet recent work has shown that convolution neural networks (CNNs) can be an alternative with seven orders of magnitude improvement in speed. With 31,200 simulated strongly lensed quasar images, we explore the usage of Vision Transformer (ViT) for simulated strong gravitational lensing for the first time. We show that ViT could reach competitive results compared with CNNs, and is specifically good at some lensing parameters, including the most important mass-related parameters such as the center of lens $θ_{1}$ and $θ_{2}$, the ellipticities $e_1$ and $e_2$, and the radial power-law slope $γ'$. With this promising preliminary result, we believe the ViT (or attention-based) network architecture can be an important tool for strong lensing science for the next generation of surveys. The open source of our code and data is in \url{https://github.com/kuanweih/strong_lensing_vit_resnet}.

astro-ph.CO↗

arXiv2021

Identifying RR Lyrae in the ZTF DR3 dataset

We present a RR Lyrae (RRL) catalogue based on the combination of the third data release of the Zwicky Transient Facility (ZTF DR3) and \textit{Gaia} EDR3. We use a multi-step classification pipeline relying on the Fourier decomposition fitting to the multi-band ZTF light curves and random forest classification. The resulting catalogue contains 71,755 RRLs with period and light curve parameter measurements and has completeness of 0.92 and purity of 0.92 with respect to the SOS \textit{Gaia} DR2 RRLs. The catalogue covers the Northern sky with declination $\geq -28^\circ$, its completeness is $\gtrsim 0.8$ for heliocentric distance $\leq 80$~kpc, and the most distant RRL at 132~kpc. Compared with several other RRL catalogues covering the Northern sky, our catalogue has more RRLs around the Galactic halo and is more complete at low Galactic latitude areas. Analysing the spatial distribution of RRL in the catalogue reveals the previously known major over-densities of the Galactic halo, such as the Virgo over-density and the Hercules-Aquila Cloud, with some evidence of an association between the two. We also analyse the Oosterhoff fraction differences throughout the halo, comparing it with the density distribution, finding increasing Oosterhoff I fraction at the elliptical radii between 16 and 32 kpc and some evidence of different Oosterhoff fractions across various halo substructures.

astro-ph.GA↗

arXiv2020

Search for globular clusters associated with the Milky Way dwarf galaxies using Gaia DR2

We report the result of searching for globular clusters (GCs) around 55 Milky Way satellite dwarf galaxies within the distance of 450 kpc from the Galactic Center except for the Large and Small Magellanic Clouds and the Sagittarius dwarf. For each dwarf, we analyze the stellar distribution of sources in Gaia DR2, selected by magnitude, proper motion, and source morphology. Using the kernel density estimation of stellar number counts, we identify eleven possible GC candidates. Crossed-matched with existing imaging data, all eleven objects are known either GCs or galaxies and only Fornax GC 1-6 among them are associated with the targeted dwarf galaxy. Using simulated GCs, we calculate the GC detection limit $M_{\rm V}^{\rm lim}$ that spans the range from $M_{\rm V}^{\rm lim} \sim -7$ for distant dwarfs to $M_{\rm V}^{\rm lim} \sim 0$ for nearby systems. Assuming a Gaussian GC luminosity function, we compute that the completeness of the GC search is above 90 percent for most dwarf galaxies. We construct the 90 percent credible intervals/upper limits on the GC specific frequency $S_{\rm N}$ of the MW dwarf galaxies: $12 < S_{\rm N} < 47$ for Fornax, $S_{\rm N} < 20$ for the dwarfs with $-12 < M_{\rm V} < -10$, $S_{\rm N} < 30$ for the dwarfs with $-10 < M_{\rm V} < -7$, and $S_{\rm N} < 90$ for the dwarfs with $M_{\rm V} > -7$. Based on $S_{\rm N}$, we derive the probability of galaxies hosting GCs given their luminosity, finding that the probability of galaxies fainter than $M_{\rm V} = -9$ to host GCs is lower than 0.1.

astro-ph.GA↗

arXiv2020

The early growth of supermassive black holes in cosmological hydrodynamic simulations with constrained Gaussian realizations

The paper examines the early growth of supermassive black holes (SMBHs) in cosmological hydrodynamic simulations with different BH seeding scenarios. Employing the constrained Gaussian realization, we reconstruct the initial conditions in the large-volume BlueTides simulation and run them to $z=6$ to cross-validate that the method reproduces the first quasars and their environments. Our constrained simulations in a volume of $(15\, h^{-1}{\rm Mpc})^3$ successfully recover the evolution of large-scale structure and the stellar and BH masses in the vicinity of a $\sim10^{12}\, M_{\odot}$ halo which we identified in BlueTides at $z\sim7$ hosting a $\sim10^9\, M_{\odot}$ SMBH. Among our constrained simulations, only the ones with a low-tidal field and high-density peak in the initial conditions induce the fastest BH growth required to explain the $z>6$ quasars. We run two sets of simulations with different BH seed masses of $5\times10^3$, $5\times10^4$, and $5\times10^5\, h^{-1}M_{\odot}$, (a) with the same ratio of halo to BH seed mass and (b) with the same halo threshold mass. At $z=6$, all the SMBHs converge in mass to $\sim10^9\, M_{\odot}$ except for the one with the smallest seed in (b) undergoing critical BH growth and reaching $10^8$ -- $10^9\, M_{\odot}$, albeit with most of the growth in (b) delayed compared to set (a). The finding of eight BH mergers in the small-seed scenario (four with masses $10^4$ -- $10^6\, M_{\odot}$ at $z>12$), six in the intermediate-seed scenario, and zero in the large-seed scenario suggests that the vast BHs in the small-seed scenario merge frequently during the early phases of the growth of SMBHs. The increased BH merger rate for the low-mass BH seed and halo threshold scenario provides an exciting prospect for discriminating BH formation mechanisms with the advent of multi-messenger astrophysics and next-generation gravitational wave facilities.

astro-ph.GA↗

arXiv2018

BlueTides simulation: establishing black hole-galaxy relations at high-redshift

The scaling relations between the mass of supermassive black holes ($M_{\bullet}$) and host galaxy properties (stellar mass, $M_{\star}$, and velocity dispersion, $σ$), provide a link between the growth of black holes (BHs) and that of their hosts. Here we investigate if and how the BH-galaxy relations are established in the high-$z$ universe using \textsc{BlueTides}, a high-resolution large volume cosmological hydrodynamic simulation. We find the $M_{\bullet}-M_{\star}$ and $M_{\bullet}-σ$ relations at $z=8$: $\log_{10}(M_{\bullet}) = 8.25 + 1.10 \ \log_{10}(M_{\star}/10^{11}M_{\odot})$ and $\log_{10}(M_{\bullet}) = 8.35 + 5.31 \ \log_{10}(σ/200kms^{-1})$ at $z=8$, both fully consistent with the local measurements. The slope of the $M_{\bullet}-σ$ relation is slightly steeper for high star formation rate and $M_{\star}$ galaxies while it remains unchanged as a function of Eddington accretion rate onto the BH. The intrinsic scatter in $M_{\bullet}-σ$ relation in all cases ($ε\sim 0.4$) is larger at these redshifts than inferred from observations and larger than in $M_{\bullet}-M_{\star}$ relation ($ε\sim 0.14$). We find the gas-to-stellar ratio $f=M_{\rm gas}/M_{\star}$ in the host (which can be very high at these redshifts) to have the most significant impact setting the intrinsic scatter of $M_{\bullet}-σ$. The scatter is significantly reduced when galaxies with high gas fractions ($ε= 0.28$ as $f<10$) are excluded (making the sample more comparable to low-$z$ galaxies); these systems have the largest star formation rates and black hole accretion rates, indicating that these fast-growing systems are still moving toward the relation at these high redshifts. Examining the evolution (from $z=10$ to 8) of high mass black holes in $M_{\bullet}-σ$ plane confirms this trend.

astro-ph.GA↗

arXiv2016

Contrasting Galaxy Formation from Quantum Wave Dark Matter, $ψ$DM, with $Λ$CDM, using Planck and Hubble Data

The newly established luminosity functions of high-z galaxies at $4 \lesssim z \lesssim 10$ can provide a stringent check on dark matter models that aim to explain the core properties of dwarf galaxies. The cores of dwarf spheroidal galaxies are understood to be too large to be accounted for by free streaming of warm dark matter without overly suppressing the formation of such galaxies. Here we demonstrate with cosmological simulations that wave dark matter, $ψ$DM, appropriate for light bosons such as axions, does not suffer this problem, given a boson mass of $m_ψ \ge 1.2 \times 10^{-22}{\,\rm eV}$ ($2σ$). In this case, the halo mass function is suppressed below $\sim 10^{10}{\,M_\odot}$ at a level that is consistent with the high-z luminosity functions, while simultaneously generating the kpc-scale cores in dwarf galaxies arising from the solitonic ground state in $ψ$DM. We demonstrate that the reionization history in this scenario is consistent with the Thomson optical depth recently reported by Planck, assuming a reasonable ionizing photon production rate. We predict that the luminosity function should turn over slowly around an intrinsic UV luminosity of $M_{\rm UV} \gtrsim -16$ at $z \gtrsim 4$. We also show that for galaxies magnified $\mathord{>}10\times$ in the Hubble Frontier Fields, $ψ$DM predicts an order of magnitude fewer detections than cold dark matter at $z \gtrsim 10$ down to $M_{\rm UV} \sim -15$, allowing us to distinguish between these very different interpretations for the observed coldness of dark matter.

astro-ph.GA↗