OpenTSLM

lastmod 2025-10-06created 2025-10-062 edits

A language model that understands time-series data

About

• Existing foundation models handle text, images, audio, and video, but lack temporal reasoning over the time-series data that drives the real world (vitals, prices, telemetry, logs, etc.)
• Time Series Language Models(TSLMs) support time series as a native modality on par with text, so users can ask questions in natural language and receive explanations and recommendations
• OpenTSLM is a new language model designed to process medical time-series data together with clinical text
  • Adopts a cross-attention architecture, so it scales to long time series and, unlike existing approaches, is practical to operate at large scale
  • Two architectures are proposed: SoftPrompt (inserts time series as tokens) and Flamingo (cross-attention based), each with different trade-offs
  • Introduces three new Chain-of-Thought(CoT) datasets (HAR, Sleep, ECG-QA) to evaluate the model's reasoning performance
• With a model 200 times smaller than GPT-4o, achieves 4.4x accuracy in sleep stage classification, 6x in activity recognition, and 2x in ECG interpretation (efficiency 880x, 1,000x, and 400x respectively)
• The first in the medical field to process 12-lead ECG signals and text simultaneously, providing chain-of-thought reasoning validated by cardiologists
• Processes multiple variable-length time series simultaneously and integrates them with textual context to generate interpretable explanations verified by domain experts
• The study suggests potential expansion beyond medicine to various time-series application areas such as finance, supply chains, and industrial monitoring

See also

• Timeseries

Favorite site

• OpenTSLM - A language model that understands time-series data | GeekNews