Memory-aware efficient inference
On edge devices, model-weight movement and finite memory capacity often matter as much as arithmetic throughput. We study how quantization changes footprint, latency, and deployability.
memory · bandwidth · latencyKNU · Institute for Advanced Technology Convergence
LITE Lab
Lightweight Inference Technologies for Edge
Making compact language models deployable—and their quality claims auditable.
We study quantized language models, edge AI systems, and statistically reliable deployment from a hardware-aware perspective.
Led by Hyungkwang Lim, PhD · Principal Investigator, LITE Lab · Joint Appointment Professor at Kyungpook National University
Research stackConfiguration matters
01Language modelQwen · Llama
02QuantizationQ8 · Q4
03Inference systemengine · memory · kernels
04Edge hardwareJetson · heterogeneous devices
ANYTIME-VALID EVIDENCERelease only when degradation is bounded.
01 / Research question
Compression makes language models fit. Evidence determines whether they are fit to deploy.
A language model is not deployed as a weight file alone. It is deployed as a complete configuration—including the model, quantization method, inference engine, hardware, and generation settings.
LITE Lab develops methods that make the efficiency and reliability of that complete system measurable, auditable, and suitable for resource-constrained deployment.
02 / Research areas
From physical memory to trustworthy inference
Reliable quantization
We test whether a compressed model remains within a predeclared quality-loss tolerance, replacing informal score comparisons with controlled statistical evidence.
non-inferiority · paired evidenceConfiguration-specific evaluation
A reliability claim must bind the model, quantization, inference engine, hardware, and generation settings. A weight file alone is not the deployment object.
engine · hardware · generationNeuromorphic-to-LLM systems
Experience with resistive-switching memory and spiking neural networks informs a hardware-grounded approach to modern language-model systems.
devices · circuits · systems03 / Current research
Submitted manuscript · 2026
Anytime-Valid Non-Inferiority Certification for Quantized LLM Deployment Configurations
Can a quantized model be released before every audit item has been evaluated—without increasing the risk of accepting unacceptable degradation?
One-off benchmark comparisons do not control the risk of a bad release after repeated evaluation, optional stopping, or candidate selection.
Paired, item-level evidence combines non-inferiority testing, anytime-valid evaluation, finite audit frames, and a lifetime risk ledger.
Each certificate is cryptographically bound to the model, quantization, engine, hardware, generation settings, and audit record.
Qwen3 and Llama-3.2 models are evaluated on MMLU and GSM8K, with direct deployment measurements on an NVIDIA Jetson Orin Nano.
12model–task–quantization populations evaluated
≤ 4%observed false-certification rate in reported null simulations
167–423median items to certification in representative MMLU cases
11.8%cross-machine/build prediction disagreement for one Q4 artifact
These are findings from the reported case study, not general performance claims. The manuscript is submitted to Software Quality Journal; the anonymized PDF is not posted during editorial review.
KNU
Academic setting
Institute for Advanced Technology Convergence
Research at the intersection of AI and semiconductor systems
Kyungpook National University established the Institute for Advanced Technology Convergence in 2025 as one of its three university-level convergence institutes. Its research units span robotics, mobility, semiconductors, and AI, advancing core technologies through shared research platforms, global collaboration, and cultivation of emerging researchers.
경북대학교 첨단기술융합연구원은 연구중심대학 전환을 위해 2025년 출범한 3대 융합연구원 중 하나입니다. 로봇·모빌리티·반도체·AI 연구유닛을 기반으로 핵심기술 개발, 공동연구 플랫폼 구축, 글로벌 협력 및 연구후속세대 양성을 추진합니다.
Read KNU’s official institute announcement ↗04 / Research trajectory
One continuous systems question
Resistive memory
Nanoionic mechanisms and switching behavior
Neuromorphic systems
Artificial neurons, synapses, and unreliable hardware
Automation & orchestration
Web systems, AI-assisted workflows, and deployment practice
Efficient LLM assurance
Memory-aware inference and auditable release decisions
05 / People
A research group in formation
Principal Investigator
Hyungkwang Lim, PhD
Hardware-aware AI researcher and Joint Appointment Professor at KNU’s Institute for Advanced Technology Convergence.
Profile & full trajectory →Prospective students
Help build reliable, efficient AI.
We welcome inquiries from students interested in quantized LLMs, edge AI, hardware-aware machine learning, statistical model evaluation, or neuromorphic systems.
Recruiting information →Join LITE Lab
Build efficient AI with evidence.
Student researchers and research collaborators are currently being recruited.
hlim@knu.ac.kr ↗