搜索结果：Redux

Low-rank adaptation (LoRA) has emerged as the de facto standard for parameter-efficient fine-tuning (PEFT) of foundation models, enabling the adaptation of billion-parameter networks with minimal computational and memory overhead. Despite its empirical success and rapid proliferation of variants, it remains elusive which architectural choices, optimization techniques, and deployment constraints should guide practical method selection. This overview revisits LoRA through the lens of signal processing (SP), bridging modern adapter designs with classical low-rank modeling tools and inverse problems, as well as highlighting how SP principles can inform principled advances of fine-tuning approaches. Rather than providing a comprehensive enumeration and empirical comparisons of LoRA variants, emphasis is placed on the technical mechanisms underpinning these approaches to justify their effectiveness. These advances are categorized into three complementary axes: architectural design, efficient optimization, and pertinent applications. The first axis builds on singular value decomposition (SVD)-based factorization, rank-augmentation constructions, and cross-layer tensorization, while the se

We fix the notion of parity complex by a judicious selection from among the axioms originally considered by Street. We show that parity complexes so defined, together with the morphisms of parity complexes defined by Verity, form a category equivalent to the category of strong Steiner complexes (nés augmented directed complexes with strongly loop-free unital bases). To this end, we isolate the purely combinatorial structure possessed by the bases of free augmented directed complexes. This analysis reveals the essential advantage of Steiner's formalism to be that the role of subsets in Street's formalism is played instead by multisets.

The development of Large Language Models (LLMs) requires robust benchmarks that encompass not only academic domains but also industrial fields to effectively evaluate their applicability in real-world scenarios. In this paper, we introduce two Korean expert-level benchmarks. KMMLU-Redux, reconstructed from the existing KMMLU, consists of questions from the Korean National Technical Qualification exams, with critical errors removed to enhance reliability. KMMLU-Pro is based on Korean National Professional Licensure exams to reflect professional knowledge in Korea. Our experiments demonstrate that these benchmarks comprehensively represent industrial knowledge in Korea. We release our dataset publicly available.

The Noether procedure carries an inherent ambiguity due to the necessary local extension, no longer a symmetry, of the global symmetry. The gauging should fix the ambiguity once and for all, however, and, for translations, the general covariance demands us to use the Lie derivative. We argue that, with this alone and without any further tweaking, the Noether energy-momentum $\hat{\mathbb{T}}$ must equal the symmetric counterpart, $T$, inevitably and show the equality explicitly for general tensors. For spinors, a subtlety with the Lie derivative itself enters the issue and leads us to the Kosmann lift, often unnoticed by the physics community, from which $T=\hat{\mathbb{T}}$ again emerges straightforwardly and in a naturally symmetric form. Finally, we address how the same Kosmann lift affects the anomaly computations and show that the diffeomorphism anomaly from the seminal paper must be halved, while the venerable anomaly polynomials themselves stand unaffected. We discuss the ramifications of these findings.

The BlueWalker 3 satellite is now fainter than during the first months after deployment. The greatest improvement is that the average maximum luminosity near zenith has been reduced from magnitude 1.0 to 2.2. However, the spacecraft is still usually bright enough to interfere with astronomical research.

In the study of the Political Resource Curse (Brollo et al.,2013), the authors identified a new channel to investigate whether the windfalls of resources are unambiguously beneficial to society, both with theory and empirical evidence. This paper revisits the framework with a new dataset. Specifically, we implemented a regression discontinuity design and difference-in-difference specification

Conic programming has well-documented merits in a gamut of signal processing and machine learning tasks. This contribution revisits a recently developed first-order conic descent (CD) solver, and advances it in three aspects: intuition, theory, and algorithmic implementation. It is found that CD can afford an intuitive geometric derivation that originates from the dual problem. This opens the door to novel algorithmic designs, with a momentum variant of CD, momentum conic descent (MOCO) exemplified. Diving deeper into the dual behavior CD and MOCO reveals: i) an analytically justified stopping criterion; and, ii) the potential to design preconditioners to speed up dual convergence. Lastly, to scale semidefinite programming (SDP) especially for low-rank solutions, a memory efficient MOCO variant is developed and numerically validated.

We reconsider the ellipsoid method for linear inequalities. Using the ellipsoid representation of Burrell and Todd, we show the method can be viewed as coordinate descent on the volume of an enclosing ellipsoid, or on a potential function, or on both. The method can be enhanced by improving the lower bounds generated and by allowing the weights on inequalities to be decreased as well as increased, while still guaranteeing a decrease in volume. Three different initialization schemes are described, and preliminary computational results given. Despite the improvements discussed, these are not encouraging.

We investigate for viable models of inflation that can successfully produce dark matter (DM) from inflaton decay process, satisfying all the constraints from Cosmic Microwave Background (CMB) and from some other observations. In particular, we analyze near-inflection-point small field inflationary scenario with non-thermal production of fermionic DM from the decaying inflaton field during the reheating era. To this end, we propose two different models of inflation with polynomial potential. The potential of Model I contains terms proportional to linear, quadratic, and quartic in inflaton; whereas in Model II, the potential contains only even power of inflaton and the highest term is sextic in inflaton. For both the models, we find out possible constraints on the model parameters which lead to proper inflationary parameters from CMB data with a very small tensor-to-scalar ratio, as expected from a small-field model. With the allowed parameter space from CMB, we then search for satisfactory relic abundance for DM, that can be produced from inflaton via reheating, to match with the present-day cold dark matter (CDM) relic density for the parameter spaces of the DM $χ$ mass and Yukawa

This paper examines the current state of the science underlying cybersecurity research with an emphasis on the non-signature-based intrusion detection domain. First, the paper re-examines the base-rate fallacy originally published by Axelsson, putting the impact of false positives into context. Given the relative high numbers of false positives, the paper argues for deeper analysis of false positives, akin to the analysis that true positives are treated to. The second section of the paper examines the metrics being used to analyze non-signature intrusion detection techniques, the current status quo of employed metrics, and the impact of the status quo on scientific advancement. Finally, the paper analyzes the use of online attack graphs and their applicability, especially in scenarios of constrained environments, such as Internet of Things devices. The use of offline attack graphs in such constrained environments is also examined. In essence, a deep dive review identified multiple areas throughout the field in which the effectiveness and validity of the scientific method can be greatly improved, e.g., through removal of logical fallacies.

We present an analysis of 39 nuclei and their early-type hosts in the Virgo Cluster using ten broadband filters: F300W, F475W, F850LP, F160W, $u^*griz$, and $K_s$. We describe the Virgo Redux program, which provides high-resolution UV and NIR imaging. Combining this data with optical and NIR imaging from the ACS Virgo Cluster Survey and the Next Generation Virgo Cluster Survey, we estimate masses, metallicities and ages using simple stellar population (SSP) models. For 19 nuclei, we compare to SSP parameters derived from Keck and Gemini spectra and find reasonable agreement between the photometric and spectroscopic metallicity: the RMS scatter is 0.3 dex. We reproduce the nucleus-galaxy mass fraction of $0.33^{+0.09}_{-0.07}$ percent for galaxy stellar masses $10^{8.4}-10^{10.3} M_\odot$ with a typical precision of $\sim$35% for the nuclei masses. Based on available model predictions, there is no single preferred formation scenario for nuclei, suggesting that nuclei are formed stochastically through a mix of processes. Nuclei metallicities are statistically identical to those of their hosts, appearing $0.07 \pm 0.3$ dex more metal-rich on average -- although, omitting galaxies with

We present the results of a recent re-reduction of the data from the Very Large Array (VLA) Low-frequency Sky Survey (VLSS). We used the VLSS catalog as a sky model to correct the ionospheric distortions in the data and create a new set of sky maps and corresponding catalog at 73.8 MHz. The VLSS Redux (VLSSr) has a resolution of 75 arcsec, and an average map RMS noise level of $σ\sim0.1$ Jy beam$^{-1}$. The clean bias is $0.66\timesσ$, and the theoretical largest angular size is 36 arcmin. Six previously un-imaged fields are included in the VLSSr, which has an unbroken sky coverage over 9.3 sr above an irregular southern boundary. The final catalog includes 92,964 sources. The VLSSr improves upon the original VLSS in a number of areas including imaging of large sources, image sensitivity, and clean bias; however the most critical improvement is the replacement of an inaccurate primary beam correction which caused source flux errors which vary as a function of radius to nearest pointing center in the VLSS.

Recent results from Higgs boson and supersymmetry searches at the Large Hadron Collider provide strong new motivations for supersymmetric theories with heavy superpartners. We reconsider focus point supersymmetry (FP SUSY), in which all squarks and sleptons may have multi-TeV masses without introducing fine-tuning in the weak scale with respect to variations in the fundamental SUSY-breaking parameters. We examine both FP SUSY and its familiar special case, the FP region of mSUGRA/CMSSM, and show that they are beautifully consistent with all particle, astroparticle, and cosmological data, including Higgs boson mass limits, null results from SUSY searches, electric dipole moments, b -> s gamma, B_s -> mu^+ mu^-, the thermal relic density of neutralinos, and dark matter searches. The observed deviation of the muon's anomalous magnetic moment from its standard model value may also be explained in FP SUSY, although not in the FP region of mSUGRA/CMSSM. In light of recent data, we advocate refined searches for FP SUSY and related scenarios with heavy squarks and sleptons, and we present a simplified parameter space to aid such analyses.

A geometrical analysis of the bulk and anti-de Sitter boundary unitarity conditions of 3D "Minimal Massive Gravity" (MMG) (which evades the "bulk/boundary clash" of Topologically Massive Gravity) is used to extend and simplify previous results, showing that unitarity selects, up to equivalence, a connected region in parameter space. We also initiate the study of flat-space holography for MMG. Its relevant flat space limit is a deformation of 3D conformal gravity; the deformation is both non-linear and non-conformal, implying a linearisation instability.

We derive supernova (SN) bounds on muon-philic bosons, taking advantage of the recent emergence of muonic SN models. Our main innovations are to consider scalars $φ$ in addition to pseudoscalars $a$ and to include systematically the generic two-photon coupling $G_{γγ}$ implied by a muon triangle loop. This interaction allows for Primakoff scattering and radiative boson decays. The globular-cluster bound $G_{γγ}<0.67\times10^{-10}~{\rm GeV}^{-1}$ derived for axion-like particles carries over to the muonic Yukawa couplings as $g_a<3.1\times10^{-9}$ and $g_φ< 4.6\times10^{-9}$ for $m_{a,φ}\lesssim 100$ keV, so SN arguments become interesting mainly for larger masses. If bosons escape freely from the SN core the main constraints originate from SN1987A $γ$ rays and the diffuse cosmic $γ$-ray background. The latter allows at most $10^{-4}$ of a typical total SN energy of $E_{\rm SN}\simeq3\times10^{53}$erg to show up as $γ$ rays, for $m_{a,φ}\gtrsim 100$keV implying $g_a \lesssim 0.9\times10^{-10}$ and $g_φ\lesssim 0.4\times10^{-10}$. In the trapping regime the bosons emerge as quasi-thermal radiation from a region near the neutrino sphere and match $L_ν$ for $g_{a,φ}\simeq 10^{

For the past 25 years, BW Vulpeculae has been the topic of period analyses centered on a secular period change with a periodic variation superposed, presumed to be due to light time effects in a binary system. According to this paradigm, one would expect what seems like a period increase of about 0.5 s during or soon after the year 2001. I have continued photometric monitoring through the year 2012, adding 35 new timings of maximum and minimum light. This expected change in period did not occur, which rules out that interpretation of the period variation. As of 2012, the observed timings are about two hours early compared to those predicted by the quadratic ephemeris, but are very close to those predicted by the linear ephemeris. In fact, the period has remained constant for the last 32 years, indicating that the previous epochs of constant period are almost certainly the correct interpretation, though the cause of the period changes is still not clear. Continued photometric monitoring of BW Vul leads to the conclusion that the period changes are abrupt, followed by epochs of constant period lasting between 12 and at least 32 years.

We extend the investigation of special toroidal compactifications of heterotic string theory for which the half-BPS states provide representations of subgroups of the Conway group. We also explore dual descriptions of these theories and find that they are all linked to either F-theory or type IIA string theory on K3 surfaces with symplectic automorphism groups that are the same Conway subgroups as those of the heterotic dual. The matching with type IIA K3 dual theories includes both the matching of symmetry groups and a comparison between the Narain lattice on the heterotic side and the cohomology lattice on the type IIA side. We present twelve examples where we can identify a type IIA dual K3 orbifold theory as the dual description of the heterotic theory. In addition, we include a Mathematica package that performs most of the computations required for these comparisons.