Zipeng Wu

Zipeng Wu

Temporal representation for time series and sequential data.

LinkedIn Email Google Scholar GitHub CV
Portrait of Zipeng Wu
PhD Researcher in Applied Mathematics
University of Birmingham

About me

I am a PhD researcher in Applied Mathematics at the University of Birmingham. My research focuses on temporal representation: how to represent temporal structure and use it to compare time series, predict future values, and classify sequences.

I study time series under non-stationarity, with work on decomposition, similarity, and retrieval-augmented forecasting. My current interests also include symbolic representation and temporal tokenization for foundation and world models, including action histories, VLA trajectories, and model rollouts.

I am also interested in sequence compression and symbolic representation: how compact representations retain information needed for a task. My recent work on TOMC Memory explores this question for long conversational histories, drawing inspiration from time-series representations.

Research directions

  • Representation and decomposition: temporal component recovery, decomposition evaluation, and signal feature extraction.
  • Similarity and retrieval: structural comparison and stationarity-aware retrieval for forecasting.
  • Prediction and regression: output dependencies, online learning, hierarchical forecasting, and interpretable grouped regression.
  • Current directions: time-series classification, sequence compression and symbolic representation, and temporal tokenization for foundation and world models. TOMC Memory is a recent exploration of compact representations for conversational histories.

Temporal Representation

The map places my published work under three research themes. My MRes and PhD work on representation theory provides the common basis. Current projects and accepted workshop work appear below the published work.

Zipeng Wu's research map. Representation and Decomposition includes the ICML 2026 decomposition benchmark and the MIIR 2025 battery ultrasound report. Similarity and Retrieval includes SARAF, KDD 2026. Prediction and Regression includes taxi demand, online multi-output regression, COVID-19 prediction, hierarchical load forecasting, and iTARGET. Research software, accepted workshop work, and current classification and symbolic representation projects are labeled separately.
Paper summaries and research connections ยท Editable PowerPoint

Recent work & news

TOMC Memory: Task-Oriented Memory Compilation. Recent work on compact representations of conversational histories, with CPU-based context preparation and explicitly saved task notes. Project overview.
Jul 2026 UKRI/AIRR Gateway Project Language-Action Time-Series Tokenization for Efficient VLA Policies allocated 10,000 GPUHR on Isambard-AI, with nominal compute-resource value GBP 45,000; compute resources only, not direct cash funding.
Jun 2026 UKRI/AIRR Gateway Project Time Series Language and Foundation Model allocated 10,000 GPUHR on Isambard-AI, with nominal compute-resource value GBP 45,000; compute resources only, not direct cash funding.
Jun 2026 Co-authored paper accepted to KDD 2026, CORE/ICORE A*.
May 2026 First-author paper accepted to ICML 2026, CORE/ICORE A*.