Education

Master of Science in Computer Science, Georgia Institute of Technology, Aug. 2024 – May 2026
Master of Science in Computational Science and Engineering (Applied Mathematics), Georgia Institute of Technology, Aug. 2023 – May 2026
Bachelor of Engineering in Artificial Intelligence, Shandong University, Sep. 2019 – June 2023

Publications & Manuscripts

Y. Tang, M. Lin. Beyond Reversal Rates: A Mechanistic Audit of LLM Decision Revision Under Reputation Cues. ICML 2026 (International Conference on Machine Learning), under review.
Y. Tang, M. Lin. Retrieved but Ignored: Evaluating Evidence Use After Retrieval in Vision Language Models. NeurIPS 2026 (Conference on Neural Information Processing Systems), under review.
Y. Tang. UPD: Uncertainty-Preserving Distillation for Vision-Language Models. NeurIPS 2026 (Conference on Neural Information Processing Systems), under review.
Y. Tang. The Language of Surprise: LLM Compression as a Unifying Primitive for Knowledge Discovery. KDD 2026 (ACM SIGKDD Conference on Knowledge Discovery and Data Mining), Blue Sky Ideas Track, under review.
Y. Tang. Do Large Language Models Track Event Boundaries? Incremental Signals, Localization, and Robustness. In preparation for EMNLP 2026 (Conference on Empirical Methods in Natural Language Processing).
Y. Tang, Z. Y. Jiang, T. Changnawa, H.-S. Shih, I. Karadag, P. Kastner. High-Resolution Microclimate Forecasting with Morphology-Aware Spatiotemporal Models. In preparation for Building and Environment (SCI Q1, IF ≈ 7.6).
E. Cheng, R. Ma, R. Qi, Y. Tang. Image Colorization Using Generative Adversarial Networks. Proc. SPIE 12294, 7th Int. Symp. on Advances in Electrical, Electronics, and Computer Engineering, 122943Z, 2022. (Co-first author)

Research Experience

Research Assistant

Georgia Institute of Technology | Atlanta, GA | Sep. 2025 – Present

Advised by Prof. Mingfeng Lin.

First-author ICML 2026 submission: introduced a two-pass causal probe for LLM decision revision, distinguishing evidence-based belief updating from authority-driven compliance across 5 LLMs (31K+ trials on a 1,989-task benchmark).
First-author NeurIPS 2026 submission: identified a post-retrieval evidence-ignoring failure mode in multimodal RAG and introduced a retrieval-conditioned auditing framework, revealing that matched retrieval success can still mask sharply different evidence-use behavior across VLMs.
Built a cross-platform Electron desktop app (React/TypeScript/FastAPI) for side-by-side evaluation across 9+ LLM chat services and 100+ API models via OpenRouter; implemented a lightweight BrowserView automation engine (DOM injection, async polling, stability checks) cutting memory usage ~40% vs Playwright/Selenium.
Developed and productionized an LLM-as-a-Judge evaluation pipeline with structured JSON outputs, backed by a FastAPI service (15+ endpoints), SQLite persistence, and one-command cross-platform packaging (PyInstaller + electron-builder).

Graduate Researcher

Georgia Institute of Technology | Atlanta, GA | Jan. 2025 – Present

Advised by Prof. Patrick Kastner.

Architected a spatiotemporal modeling framework for high-frequency sensor data (947K samples), achieving large-scale training optimized on distributed HPC infrastructure.
Proposed a novel physics-informed sequence architecture (incorporating Gated Residual Networks, Variable Selection Networks, and multi-head attention) with structured inductive biases, achieving strong out-of-distribution generalization across unseen locations (Temp RMSE: 0.43 °C; RH RMSE: 1.3%).
Built a scalable sparse-to-dense inference pipeline for high-resolution spatial prediction (100K+ grid points), coupling Random Forest embeddings with Regression Kriging over high-dimensional geospatial covariates.
Evaluated zero-shot spatial transferability across rigorous deep learning baselines, resulting in a manuscript prepared for an SCI Q1 journal.

Undergraduate Researcher

Shandong University | Shandong, China | Feb. 2023 – June 2023

Developed an enhanced Bidirectional Rapidly-Exploring Random Tree (Bi-RRT) algorithm for autonomous vehicle path planning in complex parking lot environments.
Implemented adaptive probabilistic sampling and local trajectory smoothing modules, improving exploration efficiency and reducing curvature in dense obstacle fields.
Integrated real-time collision detection, dynamic obstacle avoidance, and kinematic feasibility validation for continuous, safe navigation under motion constraints.
Achieved 2× faster planning speed, ~35% smoother paths, and 15% shorter average trajectory length compared to baseline RRT and RRT*, validated across 100+ randomized test scenarios.

Research Intern

Shandong University | Shandong, China | Sep. 2021 – April 2022

Co-first & corresponding author of a peer-reviewed international conference paper (SPIE 2022) on automatic image colorization.
Designed a novel lightweight GAN pipeline (U-Net generator + ResNet18 discriminator) and introduced a YUV-channel separation technique, reducing training cost while boosting structural fidelity and perceptual sharpness.
Stabilized adversarial training with optimized objectives (re-weighted “realness” reliability term and tuned loss balance), improving color fidelity and transfer robustness under diverse textures and scenes.
Scaled experiments on 4.3K+ natural & animated images in PyTorch with extensive visual comparisons, consistently outperforming baselines in visual quality and detail preservation.

Work Experience

Machine Learning Engineer

GMI Cloud | Mountain View, California | May 2025 – August 2025

Optimized Flux-Schnell (12B DiT) multimodal inference on H100 clusters with GPU memory persistence & offload, reaching ~30 images/min/GPU and 1–2s latency (10–15× faster than baseline).
Scaled a distributed multi-GPU inference pipeline (NCCL all-reduce, ONNX → TensorRT) achieving linear throughput across nodes.
Built production-ready infrastructure (queuing, heartbeat monitoring, structured logging, GCS integration, content moderation) to ensure long-running stability & compliance.
Implemented a video super-resolution pipeline (Real-ESRGAN + FastAPI) with PSNR/SSIM evaluation, reducing 5s@24fps clip runtime by ~65% (284s → 100s) when integrated with Wan2.2 text-to-video.
Developed an AI-powered e-commerce try-on service (ComfyUI, Flux-Kontext + Segformer), delivering <5s per image outfit changing, background removal, and style transfer via secure RESTful APIs.

Projects

AI-Powered Product Recommendation System

Independent Venture | April 2024 – August 2024

Architected an AI-powered recommendation system that analyzes millions of Amazon product reviews to help users quickly discover the most relevant and high-quality items through semantic search and LLM-based understanding.
Developed a PySpark ETL pipeline to clean, tokenize, and embed reviews (768-dim via text-embedding-005), storing vectors and metadata efficiently in BigQuery for hybrid semantic retrieval.
Designed a hybrid retrieval engine (ScaNN + metadata filters) that improved nDCG@3 by +21% (0.85 vs 0.70) and achieved MRR = 0.88, using approximate nearest neighbors (TreeAH + AVQ) with reranking via FastAPI microservice.
Integrated Google Gemini with LangChain for RAG-based sentiment analysis and feature summarization, achieving 88% accuracy and 4.3 / 5 relevance for explainable recommendations.
Provisioned scalable infrastructure on GCP (Cloud Run, BigQuery, Cloud Storage) using Terraform, sustaining ~6 s query latency and 92% product-category coverage across 500 test queries.

Competition Experience

American Express Default Prediction

Kaggle Competition | May 2022 – August 2022 | Top 0.4% (20th/4,874 teams), Silver Medal

Trained a weighted ensemble of LightGBM (DART) and GPU-accelerated XGBoost on 16 GB of credit card transaction time-series covering spending, balances, delinquency, and repayment.
Ran 5-fold stratified cross-validation with a custom Gini-aligned training metric; iterated from a baseline through feature-enriched and feature-compressed model variants, checkpointing best models each round and tracking feature importance for selection.
Assembled the final submission as a weighted blend of four checkpoints (LightGBM at 30/25/25%, XGBoost at 20%), with weights assigned by per-model validation Gini; multi-seed training across variants kept variance low and generalization stable.

Teaching Experience

Teaching Assistant