TimeSqueeze: Dynamic patching for efficient long-context time series forecasting

A dynamic patching mechanism that adaptively selects patch boundaries within each sequence based on local signal complexity.

NeurIPS

December 2, 2025

Topics:

Large time series forecasting models are limited by a trade-off between computational cost and preserving critical temporal details. TimeSqueeze is a new hybrid architecture that resolves this issue by using a lightweight encoder and adaptive patching to dynamically compress input data, focusing computation on information-rich regions. This approach significantly reduces computational requirements while improving forecasting accuracy.

Latest publications

TimeSqueeze: Dynamic patching

A mechanism that adaptively selects patch boundaries within each sequence based on local signal complexity. (NeurIPS)

BEDTime: A unified benchmark for automatically describing time series

The first benchmark dataset to assess models on each task, comprising four datasets reformatted for these tasks.

Foundation Models

Large Language Models (LLMs)

Reinforcement Learning

Retrieval-Augmented Generation (RAG)

Vision-Language Models (VLMs)

Reasoning & Chain-of-thought (CoT)

Foundation Models

Scaling-laws for large time-series models

Discovering power-law scaling relationships in large time-series transformer models, analogous to those found in language models.

Foundation Models

Large Language Models (LLMs)

Natural-Language Processing (NLP)

Foundation Models