ICML 2026
Zipeng Wu
Temporal representation learning for foundation / world models and non-stationary time-series ML.
PhD Researcher in Applied Mathematics, University of Birmingham.
ICML Poster
Time-Series Decomposition as a Standalone Task: A Mechanism-Driven Diagnostic Benchmark.
First-author main-conference poster. The work treats decomposition as temporal representation extraction and evaluates whether methods preserve trend, oscillatory, residual, and mechanism-relevant structure under shape, phase, spectral-fidelity, and symbolic-regression diagnostics.
Poster Session 6, July 9, 10:30-12:15 KST, Hall A, #413.
What I Work On
- Temporal representation for foundation and world models: model rollouts, action histories, temporal memory, VLA trajectories, and sequence-level evaluation.
- Time-series representation under non-stationarity: decomposition, retrieval, forecasting diagnostics, similarity search, and robust evaluation.
- Research software for inspectable temporal components, compact JSON outputs, and shareable reports, including De-Time and EchoTime.