arXiv subjects

SungGeun Kim

Publications and source records attributed to SungGeun Kim.

5 recordsLinked to original sources

arXiv2026

AURA: Agentic Diagnosis and Refinement for Production Recommender Systems at Scale

How and why does a recommender system fail the users it serves? Oftentimes, practitioners are left to improve their algorithms based on a combination of feedback from stakeholder teams, domain expertise, and insights from data analyses. Yet the nuances of how and where recommendations perform well or poorly for end users are difficult to discern from aggregate quantitative metrics. Whereas these metrics provide a high-level and incomplete picture, further granularity into the quality of recommendations and their patterns requires reasoning with domain understanding and objectivity, at scale. We contemplate this complex conundrum and describe a method and implementation that uses the latest AI agentic advances to provide actionable diagnoses and improvements for production recommender systems. We present AURA (Agentic Understanding and Refinement of recommender Algorithms), an end-to-end agentic system that performs qualitative evaluation at scale and can then generate improvements to our algorithms at the code level. Specialized agents read production engagement logs, from thousands of sessions to millions, and surface patterns and examples of how the recommender fails real users. The next step uses those diagnoses and context about the recommender's own code, data, and training pipeline to propose and implement refinements grounded in that codebase. We report the system design, initial tests on production data from two large consumer platforms at a major media-streaming company, safeguards, operational learnings, and early results toward a self-improving recommender system. Finally, the diagnostic gap AURA closes is not specific to streaming. The architecture is built to transfer: every domain-specific element enters through the configuration layer that already ported it between our two platforms. We map it concretely to e-commerce and online-retail recommendation.

cs.IR↗

arXiv2015

Optimum High-k Oxide for the Best Performance of Ultra-scaled Double-Gate MOSFETs

A widely used technique to mitigate the gate leakage in the ultra-scaled metal oxide semiconductor field effect transistors (MOSFETs) is the use of high-k dielectrics, which provide the same equivalent oxide thickness (EOT) as $\rm SiO_2$, but thicker physical layers. However, using a thicker physical dielectric for the same EOT has a negative effect on the device performance due to the degradation of 2D electrostatics. In this letter, the effects of high-k oxides on double-gate (DG) MOSFET with the gate length under 20 nm are studied. We find that there is an optimum physical oxide thickness ($\rm T_{OX}$) for each gate stack, including $\rm SiO_2$ interface layer and one high-k material. For the same EOT, $\rm Al_2O_3$ (k=9) over 3 $\rmÅ$ $\rm SiO_2$ provides the best performance, while for $\rm HfO_2$ (k=20) and $\rm La_2O_3$ (k=30), $\rm SiO_2$ thicknesses should be 5 $\rmÅ$ and 7 $\rmÅ$, respectively. The effects of using high-k oxides and gate stacks on the performance of ultra-scaled MOSFETs are analyzed. While thin oxide thickness increases the gate leakage, the thick oxide layer reduces the gate control on the channel. Therefore, the physical thicknesses of gate stack should be optimized to achieve the best performance.

cond-mat.mes-hall↗

arXiv2013

Engineering Nanowire n-MOSFETs at Lg < 8 nm

As metal-oxide-semiconductor field-effect transistors (MOSFET) channel lengths (Lg) are scaled to lengths shorter than Lg<8 nm source-drain tunneling starts to become a major performance limiting factor. In this scenario a heavier transport mass can be used to limit source-drain (S-D) tunneling. Taking InAs and Si as examples, it is shown that different heavier transport masses can be engineered using strain and crystal orientation engineering. Full-band extended device atomistic quantum transport simulations are performed for nanowire MOSFETs at Lg<8 nm in both ballistic and incoherent scattering regimes. In conclusion, a heavier transport mass can indeed be advantageous in improving ON state currents in ultra scaled nanowire MOSFETs.

cond-mat.mes-hall↗

arXiv2011

Effects of Interface Roughness Scattering on Radio Frequency Performance of Silicon Nanowire Transistors

The effects of an atomistic interface roughness in n-type silicon nanowire transistors (SiNWT) on the radio frequency performance are analyzed. Interface roughness scattering (IRS) is statistically investigated through a three dimensional full-band quantum transport simulation based on the sp3d5s?* tight-binding model. As the diameter of the SiNWT is scaled down below 3 nm, IRS causes a significant reduction of the cut-off frequency. The fluctuations of the conduction band edge due to the rough surface lead to a reflection of electrons through mode-mismatch. This effect reduces the velocity of electrons and hence the transconductance considerably causing a cut-off frequency reduction.

cond-mat.mes-hall↗

arXiv2011

Full 3D Quantum Transport Simulation of Atomistic Interface Roughness in Silicon Nanowire FETs

The influence of interface roughness scattering (IRS) on the performances of silicon nanowire field-effect transistors (NWFETs) is numerically investigated using a full 3D quantum transport simulator based on the atomistic sp3d5s* tight-binding model. The interface between the silicon and the silicon dioxide layers is generated in a real-space atomistic representation using an experimentally derived autocovariance function (ACVF). The oxide layer is modeled in the virtual crystal approximation (VCA) using fictitious SiO2 atoms. <110>-oriented nanowires with different diameters and randomly generated surface configurations are studied. The experimentally observed ON-current and the threshold voltage is quantitatively captured by the simulation model. The mobility reduction due to IRS is studied through a qualitative comparison of the simulation results with the experimental results.

cond-mat.mes-hall↗