Xueying Ding

Experience

My recent work combines academic research with industry collaborations across large-scale retrieval, anomaly detection, and financial time-series modeling.

Research Intern, UMAP, Snap Research
May 2025 - Present
Bellevue, WA
Studying decoder-only LLMs as general-purpose embedding models, with recent work on information flow, retrieval quality, long-context reasoning, and sparse attention mechanisms.
  • Proposed Hierarchical Token Prepending, a training-free method that improves decoder-based LLM embeddings on retrieval and general embedding benchmarks.
  • Investigated RLVR-based search agents and LLM-driven query rewriting for open-domain retrieval and recommendation settings.
  • Collaborated on Threshold Differential Attention, a sink-free sparse attention mechanism for long-context language models.
Quant Research Intern, Bloomberg
May 2024 - Aug 2024
New York City, NY
Worked on finance-oriented time-series foundation models and studied pre-training and adaptation strategies for noisy, multi-frequency financial signals.
  • Developed Delphyne, a finance-oriented pre-trained model for financial time series.
  • Investigated negative transfer in time-series pre-training and finance-aware adaptation strategies.
  • Evaluated downstream performance on return prediction, log-volume forecasting, volatility modeling, and broader forecasting benchmarks.
Research Intern, Capital One
Jun 2022 - Aug 2022
McLean, VA
Built human-interpretable anomaly detection tools for high-dimensional streaming financial data.
  • Developed explainable XStream for unsupervised fraud detection in streaming bank data.
  • Integrated feature importance, counterfactual reasoning, and group anomaly explanation tools.
  • Built a prototype human-in-the-loop interface for analyst investigation and feedback.