SearcharxivSearch

arXiv subjects

Hyungwon Kim

Publications and source records attributed to Hyungwon Kim.

12 recordsLinked to original sources

Are you speaking my languages? On spoken language adherence in multimodal LLMs

While Large Language Model (LLM) based Automatic Speech Recognition (ASR) enables seamless multilingual use, models often misidentify the output language, compromising transcription fidelity and downstream application quality. To preserve flexibility and code-switching capabilities, we propose a soft prompting approach that hints at potential spoken languages without strictly constraining the output. We formally define this challenge as a lack of language adherence, introduce a novel metric to quantify violations, and evaluate three mitigation strategies: (1) zero-shot prompting for robust guidance under uncertainty, (2) supervised fine-tuning (SFT) to improve prompt adherence, and (3) Chain-of-Thought (CoT) reasoning to enforce adherence during decoding. We present a comparative analysis of these methods across multiple languages, evaluating effectiveness in reducing the language violation while maintaining overall ASR performance. Finally, we discuss trade-offs to guide strategy selection under various compute constraints.

cs.CL

Facade: High-Precision Insider Threat Detection Using Deep Contextual Anomaly Detection

Insiders with privileged access have the power to cause great harm to their organization. Even a single insider threat incident can be catastrophic, resulting in both financial losses and reputation damage. These threats are some of the most difficult to detect, as attack activity is interspersed in large volumes of legitimate activity. Although it is a serious threat, the literature is sparse aside from a few studies with various limitations, preventing their practical deployment in large-scale organizations. We present Facade: fast and accurate contextual anomaly detection, a high-precision, deep-learning system that has served as the last line of defense against insider threats at Google since 2018. Facade is an innovative self-supervised system that detects suspicious events by considering the context surrounding each event, including relevant facts about the user and resource involved. It is built around a new multi-modal model that is trained on corporate document access, SQL table access, and HTTP/RPC request logs. To overcome the scarcity of labeled incident data, Facade employs a novel contrastive learning strategy trained exclusively on benign activity. Facade detects insider attackers with an extremely low false-positive rate, lower than 0.01%. For single rogue events, such as the illegitimate access to a sensitive document, the false-positive rate is as low as 0.0003%. To the best of our knowledge, Facade is the only proposed insider risk anomaly detection system with a false positive rate low enough for use in large corporate environments.

cs.CR

Gaussian ensemble for quantum integrable dynamics

We propose a Gaussian ensemble as a description of the long-time dynamics of isolated quantum integrable systems. Our approach extends the Generalized Gibbs Ensemble (GGE) by incorporating fluctuations of integrals of motion. It is asymptotically exact in the classical limit irrespective of the system size and, under appropriate conditions, in the thermodynamic limit irrespective of the value of Planck's constant. Moreover, it captures quantum corrections near the classical limit, finite size corrections near the thermodynamic limit, and is valid in the presence of non-local interactions. The Gaussian ensemble bridges the gap between classical integrable systems, where a generalized microcanonical ensemble is exact even for few degrees of freedom and GGE, which requires thermodynamic limit. We illustrate our results with examples of increasing complexity.

cond-mat.stat-mech

Slowest local operators in quantum spin chains

We numerically construct slowly relaxing local operators in a nonintegrable spin-1/2 chain. Restricting the support of the operator to $M$ consecutive spins along the chain, we exhaustively search for the operator that minimizes the Frobenius norm of the commutator with the Hamiltonian. We first show that the Frobenius norm bounds the time scale of relaxation of the operator at high temperatures. We find operators with significantly slower relaxation than the slowest simple "hydrodynamic" mode due to energy diffusion. Then, we examine some properties of the nontrivial slow operators. Using both exhaustive search and tensor network techniques, we find similar slowly relaxing operators for a Floquet spin chain; this system is hydrodynamically "trivial", with no conservation laws restricting their dynamics. We argue that such slow relaxation may be a generic feature following from locality and unitarity.

cond-mat.stat-mech

Thermalization of entanglement

We explore the dynamics of the entanglement entropy near equilibrium in highly-entangled pure states of two quantum-chaotic spin chains undergoing unitary time evolution. We examine the relaxation to equilibrium from initial states with either less or more entanglement entropy than the equilibrium value, as well as the dynamics of the spontaneous fluctuations of the entanglement that occur in equilibrium. For the spin chain with a time-independent Hamiltonian and thus an extensive conserved energy, we find slow relaxation of the entanglement entropy near equilibration. Such slow relaxation is absent in a Floquet spin chain with a Hamiltonian that is periodic in time and thus has no local conservation law. Therefore, we argue that slow diffusive energy transport is responsible for the slow relaxation of the entanglement entropy in the Hamiltonian system.

cond-mat.stat-mech

Testing whether all eigenstates obey the Eigenstate Thermalization Hypothesis

We ask whether the Eigenstate Thermalization Hypothesis (ETH) is valid in a strong sense: in the limit of an infinite system, {\it every} eigenstate is thermal. We examine expectation values of few-body operators in highly-excited many-body eigenstates and search for `outliers', the eigenstates that deviate the most from ETH. We use exact diagonalization of two one-dimensional nonintegrable models: a quantum Ising chain with transverse and longitudinal fields, and hard-core bosons at half-filling with nearest- and next-nearest-neighbor hopping and interaction. We show that even the most extreme outliers appear to obey ETH as the system size increases, and thus provide numerical evidences that support ETH in this strong sense. Finally, periodically driving the Ising Hamiltonian, we show that the eigenstates of the corresponding Floquet operator obey ETH even more closely. We attribute this better thermalization to removing the constraint of conservation of the total energy.

cond-mat.stat-mech

Eigenstate Thermalization and Representative States on Subsystems

We consider a quantum system A U B made up of degrees of freedom that can be partitioned into spatially disjoint regions A and B. When the full system is in a pure state in which regions A and B are entangled, the quantum mechanics of region A described without reference to its complement is traditionally assumed to require a reduced density matrix on A. While this is certainly true as an exact matter, we argue that under many interesting circumstances expectation values of typical operators anywhere inside A can be computed from a suitable pure state on A alone, with a controlled error. We use insights from quantum statistical mechanics - specifically the eigenstate thermalization hypothesis (ETH) - to argue for the existence of such "representative states".

cond-mat.stat-mech

Entanglement spreading in a many-body localized system

Motivated by the findings of logarithmic spreading of entanglement in a many-body localized system, we more closely examine the spreading of entanglement in the fully many-body localized phase, where all many-body eigenstates are localized. Performing full diagonalizations of an XXZ spin model with random longitudinal fields, we identify two factors contributing to the spreading rate: the localization length ($ξ$), which depends on the disorder strength, and the final value of entanglement per spin ($s_\infty$), which primarily depends on the initial state. We find that the entanglement entropy grows with time as $\sim ξ\times s_\infty \log t$, providing support for the phenomenology of many-body localized systems recently proposed by Huse and Oganesyan [arXiv:1305.4915v1].

cond-mat.dis-nn

Ballistic spreading of entanglement in a diffusive nonintegrable system

We study the time evolution of the entanglement entropy of a one-dimensional nonintegrable spin chain, starting from random nonentangled initial pure states. We use exact diagonalization of a nonintegrable quantum Ising chain with transverse and longitudinal fields to obtain the exact quantum dynamics. We show that the entanglement entropy increases linearly with time before finite-size saturation begins, demonstrating a ballistic spreading of the entanglement, while the energy transport in the same system is diffusive. Thus we explicitly demonstrate that the spreading of entanglement is much faster than the energy diffusion in this nonintegrable system.

quant-ph

Heat and spin transport in a cold atomic Fermi gas

Motivated by recent experiments measuring the spin transport in ultracold unitary atomic Fermi gases (Sommer et al., 2011; Sommer et al., 2011), we explore the theory of spin and heat transport in a three-dimensional spin-polarized atomic Fermi gas. We develop estimates of spin and thermal diffusivities and discuss magnetocaloric effects, namely the the spin Seebeck and spin Peltier effects. We estimate these transport coefficients using a Boltzmann kinetic equation in the classical regime and present experimentally accessible signatures of the spin Seebeck effect. We study an exactly solvable model that illustrates the role of momentum-dependent scattering in the magnetocaloric effects.

cond-mat.quant-gas

Superdiffusive nonequilibrium motion of an impurity in a Fermi sea

We treat the nonequilibrium motion of a single impurity atom in a low-temperature single-species Fermi sea, interacting via a contact interaction. In the nonequilibrium regime, the impurity does a superdiffusive geometric random walk where the typical distance traveled grows with time as $\sim t^{d/(d+1)}$ for the $d$-dimensional system with $d\geq 2$. For nonzero temperature $T$, this crosses over to diffusive motion at long times with diffusivity $D\sim T^{-(d-1)/2}$. These results apply also to a nonzero concentration of impurity atoms as long as they remain dilute and nondegenerate.

cond-mat.quant-gas

Interfering directed paths and the sign phase transition

We revisit the question of the "sign phase transition" for interfering directed paths with real amplitudes in a random medium. The sign of the total amplitude of the paths to a given point may be viewed as an Ising order parameter, so we suggest that a coarse-grained theory for system is a dynamic Ising model coupled to a Kardar-Parisi-Zhang (KPZ) model. It appears that when the KPZ model is in its strong-coupling ("pinned") phase, the Ising model does not have a stable ferromagnetic phase, so there is no sign phase transition. We investigate this numerically for the case of ß1+1 dimensions, demonstrating the instability of the Ising ordered phase there.

cond-mat.dis-nn