arXiv subjects

Yeji Kim

Publications and source records attributed to Yeji Kim.

7 recordsLinked to original sources

arXiv2026

Does Role Specialization Matter for Explanation Faithfulness in Mixture-of-Experts?

Mixture-of-Experts (MoE) architectures have recently been extended with role-based mechanisms for interpretability. This is typically done by assigning semantic roles to individual expert components, for example roles like synergy, redundancy, and uniqueness in multimodal settings. However, whether such structural role decomposition preserves explanation faithfulness of the overall architecture remains largely underexplored. We hypothesize that inter-expert representation overlap weakens effective role separation and degrades attribution-based faithfulness, even when semantic roles are explicitly defined. To address this limitation, we introduce representation-level decorrelation regularization to explicitly reduce inter-expert similarity in latent space. Using representation decorrelation objectives, we encourage clearer specialization among experts by minimizing representation overlap. Our experiments show that across multiple multimodal benchmarks, this separation consistently improves explanation faithfulness, as measured by comprehensiveness, sufficiency, and their Area Over the Perturbation Curve (AOPC) summaries, while preserving task performance. We further show that these improvements are not limited to role-based architectures such as Interpretable Multimodal Interaction-aware MoE (I2MoE). Similar trends are observed in a standard sparse MoE baseline, suggesting that representation-level separation may provide a more general mechanism for enhancing explanation faithfulness in MoE systems. Overall, our findings suggest that structural role decomposition alone may be insufficient to guarantee faithful explanations and that representation-level separation helps improve explanation faithfulness. To support reproducibility, the source code and supplementary material are publicly available at https://github.com/dut0817/FL-I2MoE_Decor.

cs.LG↗

arXiv2026

Feature-level Interaction Explanations in Multimodal Transformers

Multimodal Transformers often produce predictions without clarifying how different modalities jointly support a decision. Most existing multimodal explainable AI (MXAI) methods extend unimodal saliency to multimodal backbones, highlighting important tokens or patches within each modality, but they rarely pinpoint which cross-modal feature pairs provide complementary evidence (synergy) or serve as reliable backups (redundancy). We present Feature-level I2MoE (FL-I2MoE), a structured Mixture-of-Experts layer that operates directly on token/patch sequences from frozen pretrained encoders and explicitly separates unique, synergistic, and redundant evidence at the feature level. We further develop an expert-wise explanation pipeline that combines attribution with top-K% masking to assess faithfulness, and we introduce Monte Carlo interaction probes to quantify pairwise behavior: the Shapley Interaction Index (SII) to score synergistic pairs and a redundancy-gap score to capture substitutable (redundant) pairs. Across three benchmarks (MMIMDb, ENRICO, and MMHS150K), FL-I2MoE yields more interactionspecific and concentrated importance patterns than a dense Transformer with the same encoders. Finally, pair-level masking shows that removing pairs ranked by SII or redundancy-gap degrades performance more than masking randomly chosen pairs under the same budget, supporting that the identified interactions are causally relevant. Code is available at https://github.com/dut0817/FL-I2MoE.

cs.LG↗

arXiv2025

Scale-robust Low Resistance Transport in Atomic Layer Deposited Topological Semimetal Wafers on Amorphous Substrate

As data-centric computing advances, energy-efficient interconnects are increasingly critical for AI-driven systems. Traditional metal conductors face severe limitations at nanoscale due to increased resistivity from surface scattering. In response, this study demonstrates the first wafer-scale realization of an amorphous topological semimetal, tantalum phosphide (TaP), grown directly on amorphous SiO2 substrates (without any seed layers) using low-temperature atomic layer deposition (ALD). The resulting TaP films exhibit unconventional resistivity scaling: decreasing resistivity with decreasing thickness, reaching 227 micro-ohm cm at ~2.3 nm film thickness. This behavior, observed without crystalline order or seed layers, indicates dominant surface conduction and establishes ALD-TaP as a promising candidate for back-end-of-line integration. The films also show excellent conformality, stoichiometry control, and thermal stability up to 600 degree C. A two-channel conduction model confirms surface-dominated transport in ultrathin regimes, further supported by enhanced conductivity in multi-stacked configurations. These findings highlight the potential of amorphous topological semimetals for future high-density, low-power electronic interconnects and expand the applicability of ALD for integrating novel quantum materials at scale.

cond-mat.mtrl-sci↗

arXiv2022

MolNet: A Chemically Intuitive Graph Neural Network for Prediction of Molecular Properties

The graph neural network (GNN) has been a powerful deep-learning tool in chemistry domain, due to its close connection with molecular graphs. Most GNN models collect and update atom and molecule features from the fed atom (and, in some cases, bond) features, which are basically based on the two-dimensional (2D) graph representation of 3D molecules. Correspondingly, the adjacency matrix, containing the information on covalent bonds, or equivalent data structures (e.g., lists) have been the main core in the feature-updating processes, such as graph convolution. However, the 2D-based models do not faithfully represent 3D molecules and their physicochemical properties, exemplified by the overlooked field effect that is a "through-space" effect, not a "through-bond" effect. The GNN model proposed herein, denoted as MolNet, is chemically intuitive, accommodating the 3D non-bond information in a molecule, with a noncovalent adjacency matrix $\bf{\bar A}$, and also bond-strength information from a weighted bond matrix $\bf{B}$. The noncovalent atoms, not directly bonded to a given atom in a molecule, are identified within 5 $\unicode{x212B}$ of cut-off range for the construction of $\bf{\bar A}$, and $\bf{B}$ has edge weights of 1, 1.5, 2, and 3 for single, aromatic, double, and triple bonds, respectively. Comparative studies show that MolNet outperforms various baseline GNN models and gives a state-of-the-art performance in the classification task of BACE dataset and regression task of ESOL dataset. This work suggests a future direction of deep-learning chemistry in the construction of deep-learning models that are chemically intuitive and comparable with the existing chemistry concepts and tools.

physics.chem-ph↗

arXiv2021

In-Kernel Control-Flow Integrity on Commodity OSes using ARM Pointer Authentication

This paper presents an in-kernel, hardware-based control-flow integrity (CFI) protection, called PAL, that utilizes ARM's Pointer Authentication (PA). It provides three important benefits over commercial, state-of-the-art PA-based CFIs like iOS's: 1) enhancing CFI precision via automated refinement techniques, 2) addressing hindsight problems of PA for in kernel uses such as preemptive hijacking and brute-forcing attacks, and 3) assuring the algorithmic or implementation correctness via post validation. PAL achieves these goals in an OS-agnostic manner, so could be applied to commodity OSes like Linux and FreeBSD. The precision of the CFI protection can be adjusted for better performance or improved for better security with minimal engineering efforts if a user opts in to. Our evaluation shows that PAL incurs negligible performance overhead: e.g., <1% overhead for Apache benchmark and 3~5% overhead for Linux perf benchmark on the latest Mac mini (M1). Our post-validation approach helps us ensure the security invariant required for the safe uses of PA inside the kernel, which also reveals new attack vectors on the iOS kernel. PAL as well as the CFI-protected kernels will be open sourced.

cs.CR↗

arXiv2017

Low-threshold optically pumped lasing in highly strained Ge nanowires

The integration of efficient, miniaturized group IV lasers into CMOS architecture holds the key to the realization of fully functional photonic-integrated circuits. Despite several years of progress, however, all group IV lasers reported to date exhibit impractically high thresholds owing to their unfavorable bandstructures. Highly strained germanium with its fundamentally altered bandstructure has emerged as a potential low-threshold gain medium, but there has yet to be any successful demonstration of lasing from this seemingly promising material system. Here, we demonstrate a low-threshold, compact group IV laser that employs germanium nanowire under a 1.6% uniaxial tensile strain as the gain medium. The amplified material gain in strained germanium can sufficiently surmount optical losses at 83 K, thus allowing the first observation of multimode lasing with an optical pumping threshold density of ~3.0 kW cm^-^2. Our demonstration opens up a new horizon of group IV lasers for photonic-integrated circuits.

physics.optics↗

arXiv2013

Warp products and (2+1) dimensional spacetimes

We investigate the geometrical aspects of the extended (2+1) dimensional Banados-Teitelboim-Zanelli spacetimes in the multiply warped product scheme. To do this, we analyze the interior physical properties by constructing the explicit warp functions in these regions.

gr-qc↗