arXiv subjects

Minh Ngoc Le

Publications and source records attributed to Minh Ngoc Le.

11 recordsLinked to original sources

arXiv2026

LIGHTS. Limitations of Far-Infrared Tracers for Galactic Cirrus Mitigation in the Era of Ultra-Deep Imaging

Current and forthcoming deep optical imaging surveys are reaching surface brightness levels fainter than 30 mag/arcsec$^2$, where Galactic cirrus emission becomes a significant contaminant. This poses a major challenge for the study of extremely low-surface-brightness extragalactic structures. We investigate the relationship between far-infrared emission, as traced by the Herschel 250 $μ$m band, and optical surface brightness in the Sloan \textit{g}- and \textit{r}-bands. We find that, for Galactic cirrus with $μ_g\gtrsim$~27 mag/arcsec$^2$, a linear relation is consistent with the data, assuming optically thin dust, with optical cirrus emission amounting to $\sim$0.1\% of the far-infrared flux. Using Herschel far-infrared data to trace Galactic cirrus enables its identification over wide areas, but at the cost of significantly reducing spatial resolution. This limitation introduces source confusion and restricts the effective surface brightness depth that can be robustly probed in optical data. We show that, when cirrus emission is modeled using Herschel 250 $μ$m data, the limiting surface brightness is $μ_V\sim$~28 mag/arcsec$^2$, substantially shallower than the nominal depth of surveys such as LIGHTS or the 10-year coadd LSST ($μ_V\sim$~30.5 mag/arcsec$^2$). These results highlight a fundamental limitation in the use of currently available far-infrared data for cirrus mitigation in ultra-deep optical imaging, and emphasize the need for higher-resolution cirrus tracers to fully exploit the depth of upcoming surveys.

astro-ph.GA↗

arXiv2026

SCOPD: Sparse-Context On-Policy Self-Distillation for Efficient Vision-Language Models

Reasoning vision-language models (VLMs) process images and videos as long sequences of visual tokens, making inference expensive. Training-free token pruning reduces this cost, but aggressive compression can sharply degrade performance, often attributed to irreversible loss of task-relevant visual information. We show that this explanation is incomplete. In a fixed-context Pass@K analysis, repeated sampling from the same pruned visual representation recovers many examples missed by greedy decoding, indicating that useful visual evidence can remain accessible but be used unreliably. We call this the representation-utilization gap. Motivated by this observation, we introduce SCOPD, a sparse-context on-policy self-distillation framework in which a student generates reasoning trajectories from pruned visual tokens while a privileged full-context teacher supervises the same on-policy prefixes. SCOPD requires no ground-truth responses, architectural changes, or additional inference-time computation. We further introduce SCOPD+, which uses a small visual-budget intervention to identify visually sensitive response positions and selectively distill them. At 10% visual-token retention, the Vanilla model retains 86.37% of its unpruned performance across 13 benchmarks. SCOPD raises this to 90.49%, while SCOPD+ further improves it to 92.43%. Across token budgets, benchmarks, and pruning operators, our results show that efficient reasoning depends not only on which visual information survives pruning, but also on how reliably the model learns to use it.

cs.CV↗

arXiv2026

Rubin LSST DP2 unveils almost-dark galaxies in the Virgo Cluster

Galaxies with the faintest surface brightness are currently known only in the Local Group. Similar objects should exist beyond our vicinity and are crucial for understanding galaxy evolution, structure, and dark matter content, yet surveys have not reached the depth required to detect them systematically. We present a population of seven almost-dark galaxies identified in Data Preview 2 of the Rubin Legacy Survey of Space and Time. They surround M49 in the vicinity of the Virgo Cluster, and exhibit central surface brightnesses in the range of $26.9 - 28.5 \, \mathrm{mag \, arcsec^{-2}}$ in the $g$-band, with half-light radii of $0.6 - 4.6 \, \mathrm{kpc}$ at the distance of Virgo, and stellar masses of $10^6 - 10^7 \, \mathrm{M_\odot}$. Their characteristics are analogous to those of the faintest and low-mass galaxies identified among satellite galaxies And XXI, And XXIII, and And XXV in the Local Group. This discovery demonstrates the power of the forthcoming Rubin LSST 10-year survey to uncover extremely faint galaxies at scale, promising the large statistical samples needed to constrain the faint-end luminosity function and the nature of dark matter.

astro-ph.GA↗

arXiv2026

Serendipitous discovery of an almost-dark galaxy in the Virgo Cluster

Analogues of extreme Local Group galaxies with very low surface brightnesses and large effective radii must exist elsewhere, still hidden due to current detection limits. We report the serendipitous discovery of the low-mass almost-dark galaxy TTT J1237327+143535 in the Virgo Cluster, observed using the Two-meter Twin Telescope at the Teide Observatory. TTT J1237327+143535 has a central surface brightness of $27.9 \pm 0.2 \, \mathrm{mag\, arcsec}^{-2}$ in the $g'$ band, an effective radius of $0.9 \pm 0.1 \mathrm{kpc}$, and a total stellar mass of $(2.2 \pm 0.4) \times 10^6 \mathrm{M_\odot}$. Its effective radius and absolute magnitude are similar to those of galaxies And XXI and And XXIII. The discovery of this extremely low-surface-brightness, extended, low-mass galaxy suggests the existence of a significant population of almost-dark galaxies in the Virgo Cluster. An in-depth analysis of the Next Generation Virgo Cluster Survey, as well as upcoming Rubin data releases and the 10-year Rubin Legacy Survey of Space and Time, is expected to reveal large samples of this extreme galaxy population, which will offer insights into galaxy formation in extreme conditions.

astro-ph.GA↗

arXiv2026

RAPTOR: Role-Aware Private Training for Mixture-of-Experts

Differentially private (DP) fine-tuning methods treat sparse Mixture-of-Experts (MoE) models as a single dense block, ignoring that shared layers see all data while experts only see routed records. We identify and formally characterize three resulting failure modes: global clipping suppresses expert gradients, batch-level normalization dilutes sparse expert updates, and fixed privacy noise degrades signal-to-noise ratio on low-load experts. We introduce RAPTOR - a Role-Aware Private Training framework, which alternates shared and expert optimization and targets each failure directly, using expert-specific clipping and noise together with a public expected-owner denominator and a count-independent update schedule that avoids conditioning on private, realized expert counts. We prove the resulting mechanism satisfies $(\varepsilon,δ)$-DP: because each record is assigned to exactly one owner expert, per-expert mechanisms within a layer compose in parallel, so updating all $E$ experts costs no more, in privacy terms, than updating one, with shared and expert streams composing sequentially across training. We further derive a bias-variance decomposition of the public-denominator estimator showing its bias grows predictably with routing imbalance, yielding a privacy-free rule for selecting which layer to protect from routing entropy measured on a small public corpus. Experiments on Switch Transformer and OLMoE fine-tuning across GLUE tasks, and on DeepSeek-VL2-Tiny, show consistent gains over standard DP baselines across several privacy levels ($\varepsilon$), with the largest margins typically at the tightest budgets. Code and models are publicly available: https://github.com/leduckhai/RAPTOR

cs.LG↗

arXiv2026

AVIS: Adaptive Test-Time Scaling for Vision-Language Models

Modern Vision-Language Models (VLMs) benefit from chain-of-thought prompting and test-time scaling, but these gains often come with prohibitive inference cost due to large visual contexts and long decoding chains. We view this cost through two coupled axes: Visual Context Scaling (VCS), which controls how much visual evidence is passed to the language model, and Visual Reasoning Scaling (VRS), which controls how much inference-time reasoning search is performed. Existing methods typically optimize one axis at a time, leaving the joint allocation of compute across these axes underexplored. We introduce Adaptive Visual Inference Scaling (AVIS), a lightweight policy that adapts both VCS and VRS per query. AVIS realizes VCS through Key Diversity Visual (KDV) pruning, a training-free $O(N)$ key-based rule for removing redundant visual tokens before prefilling, and realizes VRS through adaptive self-consistency, using a learned difficulty predictor to select the number of reasoning rollouts. AVIS is deployment-friendly and compatible with shared-prefill inference, where all rollouts reuse a single prefilling pass and KV cache. Across diverse image and video reasoning benchmarks, AVIS improves the accuracy--compute trade-off relative to VCS-only and VRS-only baselines, and remains effective on top of RL post-trained VLMs while keeping compute and latency low.

cs.CV↗

arXiv2026

GEAR: Genetic AutoResearch for Agentic Code Evolution

Autonomous research agents can already run machine learning experiments without human supervision, but many rely on a narrow search strategy: they repeatedly modify one program and keep changes only when they improve the current best result. This can cause them to discard useful partial ideas, alternative promising directions, and insights from failed or incomplete experiments. GEAR, or Genetic AutoResearch, replaces this single-path search with a population-based search over multiple research states. It keeps a set of strong candidate solutions, selects parents based on productivity, novelty, and coverage, and explores new ideas through mutation and crossover. Each research state stores its code changes, reflections, and performance data, allowing future decisions to build on past discoveries. The paper studies three versions of GEAR: one controlled through prompting, one using a fixed programmatic search controller, and one where the controller itself can evolve during the run. Under the same compute budget and environment, all three versions outperform the AutoResearch baseline. More importantly, while the baseline tends to settle into one local optimum, GEAR continues finding improvements over longer runs. Overall, the results suggest that autonomous research agents become more effective when they maintain multiple promising directions and can adapt their search strategy over time.

cs.NE↗

arXiv2026

LIGHTS. The Thin Encircling Stellar Stream of NGC 3938

We present a stellar stream found in images of the nearby, nearly face-on, late-type galaxy, NGC 3938 obtained for the LBT Imaging of Galactic Halos and Tidal Structures (LIGHTS) survey that is thin, has very low mean surface brightness ($\langleμ_g\rangle \approx$ 28.7 mag arcsec$^{-2}$ and $\langleμ_r\rangle \approx$ 28.1 mag arcsec$^{-2}$), appears to lie nearly on the plane of the sky, and wraps more than half way around a host galaxy that is otherwise apparently isolated. We estimate that the progenitor had a stellar mass of $\sim 3.7\times 10^7$ M$_\odot$. Despite an intriguing apparent offset between the centroid of the host galaxy and the apparent center of the stream orbit, we find that we can reproduce the morphology, including this apparent off-centering, with simple models and standard assumptions about the host (thin disk centered within a canonical spherical dark matter halo) and the progenitor satellite orbit. We identify a number of detailed features of the stream, such as changes in curvature and density, that will require more complex models to reproduce. Even this rather simple system provides a rich set of constraints with which to explore the accretion history and gravitational potential of an otherwise unremarkable late-type galaxy. Given the depth of the LIGHTS images, this system is an example of the types of stellar stream that could be found in a majority of nearby giant galaxies with the 10-year stack of Rubin/LSST data.

astro-ph.GA↗

arXiv2025

Few-Shot-Based Modular Image-to-Video Adapter for Diffusion Models

Diffusion models (DMs) have recently achieved impressive photorealism in image and video generation. However, their application to image animation remains limited, even when trained on large-scale datasets. Two primary challenges contribute to this: the high dimensionality of video signals leads to a scarcity of training data, causing DMs to favor memorization over prompt compliance when generating motion; moreover, DMs struggle to generalize to novel motion patterns not present in the training set, and fine-tuning them to learn such patterns, especially using limited training data, is still under-explored. To address these limitations, we propose Modular Image-to-Video Adapter (MIVA), a lightweight sub-network attachable to a pre-trained DM, each designed to capture a single motion pattern and scalable via parallelization. MIVAs can be efficiently trained on approximately ten samples using a single consumer-grade GPU. At inference time, users can specify motion by selecting one or multiple MIVAs, eliminating the need for prompt engineering. Extensive experiments demonstrate that MIVA enables more precise motion control while maintaining, or even surpassing, the generation quality of models trained on significantly larger datasets.

cs.CV↗

arXiv2025

LIGHTS. A robust technique to identify galaxy edges

The LIGHTS survey is imaging galaxies at a depth and spatial resolution comparable to what the Legacy Survey of Space and Time (LSST) will produce in 10 years (i.e., $\sim$31 mag/arcsec$^2$; 3$σ$ in areas equivalent to 10$^{\prime\prime}$$\times$ 10$^{\prime\prime}$). This opens up the possibility of probing the edge of galaxies, as the farthest location of in-situ star formation, with a precision that we have been unable to achieve in the past. Traditionally, galaxy edges have been analyzed in one-dimension through ellipse averaging or visual inspection. Our approach allows for a two-dimensional exploration of galaxy edges, which is crucial for understanding deviations from disc symmetry and the environmental effects on galaxy growth. In this paper, we propose a novel method using the second derivative of the surface mass density map of a galaxy to determine its edges. This offers a robust quantitative alternative to traditional edge-detection methods when deep imaging is available. Our technique incorporates Wiener-Hunt deconvolution to remove the effect of the Point Spread Function (PSF) by the galaxy itself. By applying our methodology to the LIGHTS galaxy NGC 3486, we identify the edge at 205$^{\prime\prime}$ $\pm$ 5$^{\prime\prime}$. At this radius, the stellar surface mass density is $\sim$1 M$_\odot$/pc$^2$, supporting a potential connection between galaxy edges and a threshold for in-situ star formation. Our two-dimensional analysis on NGC 3486 reveals an edge asymmetry of $\sim$5$\%$. These techniques will be of paramount importance for a physically motivated determination of the sizes of galaxies in ultra-deep surveys such as LSST, Euclid and Roman.

astro-ph.GA↗

arXiv2024

Globular Cluster Counts around 700 Nearby Galaxies

Empirically, the total number (or total mass) of globular clusters bound in a single galactic system correlates with the viral mass of the system. The form of this relation and its intrinsic scatter are potentially valuable constraints on theories of globular cluster formation and galaxy evolution. In this work, we use the DESI Legacy Imaging Survey to make a large-scale, homogeneous estimate of GC abundance around 707 galaxies at distances $\lesssim 30\,\mathrm{Mpc}$ with luminosities $8 \leq \log_{10}L/\mathrm{L}_\odot \leq 11.5$. The combination of depth and sky coverage in DESI-LS allow us to extend the techniques used by previous ground-based photometric GC surveys to a larger and potentially more representative sample of galaxies. We find average GC counts and radial profiles that are broadly consistent with the literature on individual galaxies, including good agreement with the distribution of GCs in the Milky Way, demonstrating the viability of DESI-LS images for this purpose. We find a relation between GC counts and virial mass in agreement with previous estimates based on heterogenous datasets, except at the lowest masses we probe, where we find a larger scatter in the number of cluster candidates and a slightly higher average count.

astro-ph.GA↗