Computer vision
Semantic Prior-Guided Dual Decoder for Long-Tailed Human-Object Interaction Detection
International Conference on Pattern Recognition (ICPR) · 2026
A dual-decoder model guided by semantic priors for detecting rare human-object interactions in images.
Computer vision
AnyBald: Toward Realistic Diffusion-Based Hair Removal In-The-Wild
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) · 2026 · ModiFace
A diffusion model that removes hair from portraits taken in ordinary conditions while preserving face, identity and background. No adequate paired dataset existed, so the team synthesized their own.
Computer vision
FASTe: Framework With Application-Driven Spatio-Temporal Efficiency for Video Anomaly Detection
IEEE Access · 2026 · Lorex Technology
Finds unusual events in surveillance video from video-level labels only, and cuts the computing cost enough to run on the hardware inside a camera system.
Language models
EAGLIZE: Improving Stability of Knowledge Distillation
International Conference on Electronics, Information, and Communication (ICEIC) · 2026
A method for making knowledge distillation, where a small model learns from a larger one, more stable.
Health
Dynamic Mortality Prediction for Prostate Cancer Based on Irregularly Sampled Multimodal EHR and Musculoskeletal Image-Based Biomarkers
Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD) · 2026 · Sunnybrook Health Sciences Centre
A prognostic model that updates as the patient's health record grows, combining lab results, imaging, physician notes and weight. It outperformed models using any one type of data alone.
Language models
Beyond One-Step Distillation: Bridging the Capacity Gap in Small Language Models via Multi-Step Knowledge Transfer
EACL 2026 Student Research Workshop · 2026
Transfers knowledge from a large language model to a small one in several steps rather than one, to bridge the gap in capacity between them.
Language models
Bayesian Active Learning with Gaussian Processes Guided by LLM Relevance Scoring for Dense Passage Retrieval (BAGEL)
Findings of ACL 2026 · 2026
Uses Bayesian active learning, guided by language-model relevance scores, to improve passage retrieval.
Language models
Valid ≠ Necessary: Diagnosing Latent Inefficiency in Chain-of-Thought
Findings of ACL 2026 · 2026
Diagnoses reasoning steps in language-model chain-of-thought that are valid but not needed for the answer.
Language models
CascadeDebate: Multi-Agent Deliberation for Cost-Aware LLM Cascades
ACL 2026 Industry Track · 2026 · LG Electronics Toronto AI Lab
When a small language model is uncertain, a few lightweight agents deliberate before the query is passed to a larger model. It outperformed strong single-model cascades. Written by three students and their LG mentor after the placement had ended.
Robotics
SafeCasa: A Benchmark Dataset for Physical Safety in Household Robotic Manipulation
RSS 2026 Workshop on Trustworthy Embodied Foundation Models · 2026 · LG Electronics Toronto AI Lab
A robot can complete its task and still knock over a bottle on the way. SafeCasa separates task outcome from physical safety across more than a thousand recorded manipulations, with hazards marked frame by frame.
Robotics
Route Guided CBF-QP Repair for Safe Vision Language Action Models
RSS 2026 Workshop on Trustworthy Embodied Foundation Models · 2026 · LG Electronics Toronto AI Lab
Plans a short route around an obstacle and gives it to the robot's safety filter as guidance, improving both collision avoidance and task success.
Robotics
AUC-CoT: Action-Uncertainty-Conditioned Chain-of-Thought Reasoning for Vision-Language-Action Models
CoRL 2026 Workshop on Uncertainty in World Models (UncertainWM) · 2026 · LG Electronics Toronto AI Lab
Makes a vision-language-action model reason much faster at each step while keeping its task success, by tying its step-by-step reasoning to how uncertain it is about the next action.
Machine LearningUser BehaviorImbalanced Data
Predicting Interview Engagement in Real-Time Recruitment: Class Imbalance and Behavioral Feature Analysis
NML@ICDM 2025 (Workshop co-located with ICDM, Washington DC) · 2025 · Guhuza
Machine learning models for predicting candidate interview engagement on a real-time recruitment platform under extreme class imbalance, highlighting the importance of behavioral and temporal features alongside algorithmic match scores.
Computer vision
Enhancing Sidewalk Maintenance through Accurate Joint Deflection Measurement
IEEE International Conference on Consumer Electronics (ICCE) · 2025 · IRIS R&D Group
Stereo vision that measures sidewalk joint deflection more accurately than the monocular method.
Computer vision
Improving Human Pose-Conditioned Generation: Fine-tuning ControlNet Models with Reinforcement Learning
WACV 2025 Workshops · 2025 · LG Electronics Toronto AI Lab
Fine-tunes ControlNet with reinforcement learning so that generated figures match a requested pose.
Robotics
Simulating Mobile Robot Vision: An Analysis of RGB-D versus RGB-Based Distance Accuracy and CPU Optimization
International Conference on Artificial Intelligence in Information and Communication (ICAIIC) · 2025
Compares depth cameras and ordinary colour cameras for judging distance on a mobile robot, and the computing cost of each.
Language models
GEMMAS: Graph-based Evaluation Metrics for Multi Agent Systems
EMNLP 2025 Industry Track · 2025 · LG Electronics Toronto AI Lab
Models the exchange between language-model agents as a graph and scores how much new information each message adds, showing that two systems with almost the same accuracy can differ greatly in how much reasoning they waste.
Language models
Transparent Reference-free Automated Evaluation of Open-Ended User Survey Responses
EMNLP 2025 Industry Track · 2025 · Nexxt Intelligence
Scores open-ended survey responses automatically and explains its scores, without needing reference answers.
Language models
Quality-Aware Translation Tagging in Multilingual RAG System
5th Workshop on Multilingual Representation Learning · 2025 · LG Electronics Toronto AI Lab
Question answering over documents in several languages, tagging machine-translated passages by their translation quality so the system can weigh them.
Language models
MLM: Multi-linguistic LoRA Merging
NeurIPS 2025 Workshop on Efficient Reasoning · 2025 · LG Electronics Toronto AI Lab
A method for merging LoRA adapters across languages.
Speech
Voice Anonymization Using Emotion-Enriched Feature Integration with STT and TTS Models
4th Symposium on Security and Privacy in Speech Communication · 2024
Anonymizes speakers' voices with speech-to-text and text-to-speech models while carrying emotional features through.
InfrastructureData
A Construction Classification System Database for Understanding Resource Use in Building Construction
Scientific Data (Nature) · 2022
A comprehensive database enabling researchers to analyze resource consumption patterns in construction.
Healthcare AIClinical
Utilizing Machine Learning to Evaluate Sex-based Differences in Preoperative Hemoglobin Thresholds
Journal of Vascular Surgery · 2022
Applying ML algorithms to identify sex-based differences in surgical thresholds for improved patient outcomes.
Air QualityBuilding Science
Evaluation of Fixed and Adaptive Concentration Thresholds for Particle Filter Systems
Indoor Air · 2022
Research on optimizing air filtration systems through adaptive threshold algorithms for better indoor air quality.
Computer VisionSustainability
Note: Image-based Prediction of House Attributes with Deep Learning
ACM SIGCAS/SIGCHI COMPASS · 2022
A conference note demonstrating how deep learning models can infer key residential house attributes directly from imagery, laying the groundwork for scalable, data-driven housing and sustainability analysis.
