NTU's PACT Adapts Time-Series Foundation Models Without Forgetting Old Conditions
NanyangTechnologicalUniversity · hf · 2026-11-19
Researchers from NTU, ASTAR, and Khalifa University propose PACT (Piecewise Adaptation via Chart Transport) for adapting pretrained time-series foundation models to new operating conditions. Key insight: embedding spaces shift non-uniformly under condition drift, so global adaptation unnecessarily moves already-correct embeddings. PACT partitions the frozen embedding space into local charts and learns gated residual transports, achieving high macro-F1 with limited labeled calibration data and no condition identity at test time.
More from Research
- Tsinghua-led ALIVE makes streaming video models interactive with proactive monitoring and feedback-seeking — Tsinghua · 2026-11-19
- Stanford removes the decoder: direct pixel prediction from diffusion transformers cuts cost, lifts quality — StanfordUniversity · 2026-11-19
- AnisoFlow learns anisotropic prediction targets for flow models, boosting FID without touching architecture — StanfordUniversity · 2026-11-19
- Debiasing Without Annotations: Cross-View Consensus Boosts Worst-Group Accuracy — Weiwei Li · 2026-11-19
- KAIST Proposes PROLA: Low-Rank Adaptation for Spatiotemporal Weather Downscaling — MLVLab · 2026-11-19
- The Resolution Hypothesis: agent evolution stalls because noisy evals can't spot real improvements — ShanghaiTech · 2026-11-19