Evaluating Large Language Models (LLMs) is one of the most resource-intensive steps in the AI pipeline. Conventional sequential execution often results in long runtimes, frequent out-of-memory (OOM) failures, and poor GPU utilization. Existing frameworks underuse modern GPU capabilities and provide limited failure prevention for large models. We present SpectraBench, an adaptive scheduling framework that optimizes LLM benchmarking through machine learning (ML)-driven resource orchestration. The framework follows a three-stage process: (1) Foundation Scheduling for cold-starts using heuristic rules, (2) Hybrid Scheduling that fuses heuristics with ML predictions via confidence-weighted decisions, and (3) Autonomous Scheduling using Random Forest ensembles for automated optimization. SpectraBench integrates continuous monitoring, predictive memory management, and dynamic parameter tuning to prevent failures and maximize efficiency. A two-phase optimization combining Convergence Analysis with Threshold Optimization establishes validated transition points between stages. Across 13 LLMs (0.5B–32B) and 10 benchmarks, SpectraBench achieves 10.3% shorter execution time, 11.8% higher throughput, and 5.5% lower thermal output. Hybrid Scheduling consistently outperforms both heuristic-only and ML-only baselines. SpectraBench is production-ready, modular, and compatible with single- and multi-GPU systems, enabling seamless integration into existing evaluation workflows. By addressing the memory pressure, scheduling, and failure-prevention challenges specific to LLM evaluation, SpectraBench delivers substantial gains in performance, resource efficiency, and reliability.

Deep learning-based speech enhancement methods have shown promising result in recent years. However, in practical applications, the model size and computational complexity are important factors that limit their use in end-products. Therefore, in products that require real-time speech enhancement with limited resources, such as TWS headsets, hearing aids, IoT devices, etc., ultra-lightweight models are necessary. In this paper, an ultra-lightweight network FSPEN is proposed for real-time speech enhancement task. We propose a full-band and sub-band network structure for extracting global and local features, and an inter-frame path extension method that can enhance network modeling capacity while preserving complexity. Experiments demonstrate that the proposed FSPEN achieves a performance of PESQ 2.97 on the VoiceBank+Demand dataset at 89M multiply-accumulate operation per second (MAC) and 79k parameters.

Indoor navigation is a representative application of an indoor positioning system that uses a variety of equipment, including smartphones with various sensors. Many indoor navigation systems utilize Wi-Fi signals, as well as a variety of inertial sensors, such as a 3D accelerometer, digital compass, gyroscope, and barometer, to improve the accuracy of user location tracking. The inertial sensors are vulnerable to changes in the surrounding environments and sensitive to users behavior, but little research has been conducted on sensor fusion under these conditions. In this paper, we propose a dynamic sensor fusion framework (DSFF) that provides accurate user tracking results by dynamically calibrating inertial sensor readings in a sensor fusion process. The proposed method continually learns the errors and biases of each sensor due to the changes in user behavior patterns and surrounding environments. The learned patterns are then dynamically applied to the user tracking process to yield accurate results. The results of experiments conducted in both a single-story and a multi-story building confirm that DSFF provides accurate tracking results. The scalability of the DSFF will enable it to provide more accurate tracking results with various sensors, both existing and under development.

A global indoor positioning system (GIPS) is a system that provides positioning services in most buildings in villages and cities globally. Among the various indoor positioning techniques, WLAN-based location fingerprinting has attracted considerable attention because of the wide availability of WLAN and relatively high resolution of the fingerprint-based positioning techniques. This paper introduces methods and tools to construct a GIPS by using WLAN fingerprinting. An unsupervised learning-based method is adopted to construct radio maps using fingerprints collected via crowdsourcing, and a probabilistic indoor positioning algorithm is developed for the radio maps constructed with the crowdsourced fingerprints. Along with these techniques, collecting indoor and radio maps of buildings in villages and cities is essential for a GIPS. This paper aims to collect indoor and radio maps from volunteers who are interested in deploying indoor positioning systems for their buildings. The methods and tools for the volunteers are also described in the process of developing an indoor positioning system within the larger GIPS. An experimental GIPS, named KAIST indoor locating system (KAILOS), was developed integrating the methods and tools. Then indoor navigation systems for a university campus and a large-scale indoor shopping mall were developed on KAILOS, revealing the effectiveness of KAILOS in developing indoor positioning systems. The more volunteers who participate in developing indoor positioning systems on KAILOS-like systems, the sooner GIPS will be realized.

The schedule of subway provides information on the arrival and departure of a train at stations. The subway passengers often refer to the subway schedule to make an appointment or to make their own schedule. Many apps on subway schedule are available for smartphone users. For example, in Korea, there are more than 10 apps that users can download to get the schedule of subway. However, most of the apps provide the schedule of subway based on fixed time table. As a result, there is no way to inform the differences between the time on the time table and the actual arrival time of a subway train. In this paper, we propose a method to provide correct information on the schedule of subway. Detecting the arrival of a train is required, and the difference between the time on time table and the actual arrival time of the train should be known. The method detects the arrival of a train based on the magnetometer sensing data. Then it identifies the arrived stop and computes the time differences between the schedule and the actual. The identified time differences are reflected to the time table to inform more correct schedule to the rest of the users. When we tested the detection accuracy of a subway train using the proposed technique at Seoul subway line, it showed over 90% accurate results. This indicates that the proposed method can be widely accepted to the subway schedule app developers.

Current testing practice in industry is often ineffective and slow to detect bugs, since most projects utilize manually generated test cases. Concolic testing alleviates this problem by automatically generating test cases that achieve high coverage. However, specialized execution platforms and resource constraints of embedded software hinder application of concolic testing to embedded software. To overcome these limitations, we have developed CONcrete and symBOLic (CONBOL) testing framework to unit test large size industrial embedded software automatically. To address the aforementioned limitations, CONBOL tests target units on a host PC platform by generating symbolic unit testing drivers/stubs automatically and applying heuristics to reduce false alarms caused by the imprecise drivers/stubs. We have applied CONBOL to four million lines long industrial embedded software and detected 24 new crash bugs. Furthermore, the development team of the target software adopted CONBOL to their development process to apply CONBOL to the revised target software regularly.