AI Systems Create Dependency Patterns Users Don't Recognize

A multimodal tutoring system built on fine-tuned Large Multimodal Models (LMMs) that provides step-by-step guidance combining text explanations with interactive visualizations. Generates accurate diagrams through code execution for math problems in geometry, calculus, and trigonometry. Enables natural multimodal student interaction with visual problem-solving aids.

An LLM-powered ride-hailing assistant with spatiotemporal-aware order planning and cost-effective dialogue system. Deployed in DiDi Chuxing app, achieving 93% order planning accuracy and 92% response generation accuracy. Includes MCP service API for ride-hailing research community integration and real-world interaction capabilities.

Real-time multi-microphone speech localization system for mobile devices using custom hardware with four integrated microphones on low-power microcontroller. Provides speaker diarization and directional guidance through visual arrows in UI for group conversations. Enables efficient audio localization algorithms for accessibility and meeting transcription use cases.

Comprehensive evaluation framework for benchmarking on-device LLMs on XR platforms. Measures performance consistency, processing speed, memory usage, and battery consumption across 68 model-device pairs. Uses 3D Pareto Optimality theory for optimal device-model selection. Tested on Magic Leap 2, Meta Quest 3, Vivo X100s Pro, and Apple Vision Pro.

A spreadsheet programming agent that scaffolds Excel development through structured plans based on professional workflows. Generates three potential next steps adaptively, uses pre-defined tools for component generation, and builds spreadsheets incrementally. Reduces cognitive load by 12.6% and produces higher-quality spreadsheets 2.3x more likely to be preferred.

A validated set of 17 design guidelines with recommendation levels for creating inclusive avatars in social VR. Covers disability expression through avatar appearance, body dynamics, assistive technology design, peripherals, and customization control. Derived from systematic literature review and interviews with 60 participants with disabilities, validated via heuristic evaluation with 10 VR practitioners.

A suite of minigames embedded in 3D virtual environment for deep cognitive assessment with continuous behavior logging. Four minigames explore processing speed, rule shifting, inhibitory control, and working memory. Provides process-informed metrics including gaze-based response times and movement trajectories, showing high test-retest reliability (ICC=0.52-0.83) for individual-specific behavior profiling.

An interactive system with temporal query language for collaborative specification of predictive modeling tasks between data scientists and domain experts. Enables rapid prototyping of model specifications with transparency about pre-processing choices. Allows domain experts to assess performance within data subgroups for temporal predictive models in healthcare and public services.

A method for discovering fine-grained human sub-activities from unlabeled ambient sensor data without pre-segmentation. Combines unsupervised feature extraction and clustering with visualization tool for efficient annotation. Enables domain experts to annotate only 0.05% of dataset by labeling representative cluster centroids for Human Activity Recognition in smart homes.

A framework that directly maps Time-series Foundation Model (TFM) outputs to Large Language Model representations without text conversion. Trained on synthetic data using periodicity prediction, validated on mental health classification from wearable sensing data (17,251 days, 256 participants). Maintains constant inference times regardless of input length while preserving time-series characteristics like auto-correlation.