arXiv subjects

Changdong Kim

Publications and source records attributed to Changdong Kim.

5 recordsLinked to original sources

arXiv2026

NanoQuant: Efficient Sub-1-Bit Quantization of Large Language Models

Weight-only quantization has become a standard approach for efficiently serving large language models (LLMs). However, existing methods fail to efficiently compress models to binary (1-bit) levels, as they either require large amounts of data and compute or incur additional storage. In this work, we propose NanoQuant, the first post-training quantization (PTQ) method to compress LLMs to both binary and sub-1-bit levels. NanoQuant formulates quantization as a low-rank binary factorization problem, and compresses full-precision weights to low-rank binary matrices and scales. Specifically, it utilizes an efficient alternating direction method of multipliers (ADMM) solver to precisely initialize latent binary matrices and scales, and then tunes the initialized parameters through a block and model reconstruction process. Consequently, NanoQuant establishes a new Pareto frontier in low-memory post-training quantization, and enables sub-1-bit compression. NanoQuant makes large-scale deployment feasible on consumer hardware. For example, it compresses Llama2-70B by 25.8$\times$ in just 13 hours on a single H100, enabling a 70B model to operate on a consumer 8 GB GPU. Code is available at https://github.com/SamsungLabs/NanoQuant.

cs.LG↗

arXiv2026

RaBiT: Residual-Aware Binarization Training for Accurate and Efficient LLMs

Efficient deployment of large language models (LLMs) requires extreme quantization, forcing a critical trade-off between low-bit efficiency and performance. Residual binarization enables hardware-friendly, matmul-free inference by stacking binary ($\pm$1) layers, but is plagued by pathological feature co-adaptation. We identify a key failure mode, which we term inter-path adaptation: during quantization-aware training (QAT), parallel residual binary paths learn redundant features, degrading the error-compensation structure and limiting the expressive capacity of the model. While prior work relies on heuristic workarounds (e.g., path freezing) that constrain the solution space, we propose RaBiT, a novel quantization framework that resolves co-adaptation by algorithmically enforcing a residual hierarchy. Its core mechanism sequentially derives each binary path from a single shared full-precision weight, which ensures that every path corrects the error of the preceding one. This process is stabilized by a robust initialization that prioritizes functional preservation over mere weight approximation. RaBiT redefines the 2-bit accuracy-efficiency frontier: it achieves state-of-the-art performance, rivals even hardware-intensive Vector Quantization (VQ) methods, and delivers a $4.49\times$ inference speed-up over full-precision models on an RTX 4090. Code is available at https://github.com/SamsungLabs/RaBiT.

cs.AI↗

arXiv2011

Highly Conductive 3D Nano-Carbon: Stacked Multilayer Graphene System with Interlayer Decoupling

We investigate electrical conduction and breakdown behavior of 3D nano-carbon-stacked multilayer graphene (s-MLG) system with complete interlayer decoupling. The s-MLG is prepared by transferring and stacking large-area CVD-grown graphene monolayers, followed by wire patterning and plasma etching. Raman spectroscopy was used to confirm the layer number. The D-band peak indicates low defect level in the samples. Electrical current stressing induced doping is performed to shift the charge-neutrality Dirac point and decrease the graphene/metal contact resistance, improving the overall electrical conduction. Breakdown experiments show the current-carrying capacity of s-MLG is largely enhanced as compared with that of monolayer graphene.

cond-mat.mtrl-sci↗

arXiv2011

Local Electrical Stress-Induced Doping and Formation of 2D Monolayer Graphene P-N Junction

We demonstrated doping in 2D monolayer graphene via local electrical stressing. The doping, confirmed by the resistance-voltage transfer characteristics of the graphene system, is observed to continuously tunable from N-type to P-type as the electrical stressing level (voltage) increases. Two major physical mechanisms are proposed to interpret the observed phenomena: modifications of surface chemistry for N-type doping (at low-level stressing) and thermally-activated charge transfer from graphene to SiO2 substrate for P-type doping (at high-level stressing). The formation of P-N junction on 2D graphene monolayer is demonstrated with complementary doping based on locally applied electrical stressing.

cond-mat.mtrl-sci↗

arXiv2010

Sub-optical resolution of single spins using magnetic resonance imaging at room temperature in diamond

There has been much recent interest in extending the technique of magnetic resonance imaging (MRI) down to the level of single spins with sub-optical wavelength resolution. However, the signal to noise ratio for images of individual spins is usually low and this necessitates long acquisition times and low temperatures to achieve high resolution. An exception to this is the nitrogen-vacancy (NV) color center in diamond whose spin state can be detected optically at room temperature. Here we apply MRI to magnetically equivalent NV spins in order to resolve them with resolution well below the optical wavelength of the readout light. In addition, using a microwave version of MRI we achieved a resolution that is 1/270 size of the coplanar striplines, which define the effective wavelength of the microwaves that were used to excite the transition. This technique can eventually be extended to imaging of large numbers of NVs in a confocal spot and possibly to image nearby dark spins via their mutual magnetic interaction with the NV spin.

quant-ph↗