Research map / Time series / Foundation models for time series
OpenTSLM: Time-Series Language Models for Reasoning over Multivariate Medical Text- and Time-Series Data
Patrick Langer, Thomas Kaar, Max Rosenblattl, Maxwell A Xu, Winnie Chow, Martin Maritsch, Robert Jakob, Ning Wang, Juncheng Liu, Aradhana Verma, Brian Han, Daniel Seung Kim and 10 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a pretrained Chronos-2 time-series encoder instead of a from-scratch encoder (ablation) | OpenTSLM-Flamingo · Sleep-CoT | F1: 49.33 -> 64.57 (Llama3.2-1B) | Improved |
| use a pretrained Chronos-2 time-series encoder (ablation) | OpenTSLM-Flamingo · TSQA | F1: 77.86 -> 99.69 (Gemma3-270M) | Improved |
| use a pretrained Chronos-2 time-series encoder (ablation) | OpenTSLM-Flamingo · HAR-CoT | F1: 62.93 -> 66.82 (Llama3.2-1B); 65.44 -> 62.72 (Gemma3-1B-pt) | Mixed |
| use a pretrained Chronos-2 time-series encoder (ablation) | OpenTSLM-Flamingo · ECG-QA-CoT | F1: 34.62 -> 41.92 (Llama3.2-1B) | Improved |
| train jointly on merged datasets instead of curriculum learning (ablation) | OpenTSLM-Flamingo Llama3.2-3B · HAR-CoT | F1: 62.77 -> 69.31 | Improved |
| train jointly on merged datasets instead of curriculum learning (ablation) | OpenTSLM-Flamingo Llama3.2-3B · Sleep-CoT | F1: 45.45 -> 52.13 | Improved |
| train jointly on merged datasets instead of curriculum learning (ablation) | OpenTSLM-Flamingo Llama3.2-3B · ECG-QA-CoT | F1: 40.25 -> 41.55 | Improved |
| add 50% Gaussian noise to the input series (ablation) | OpenTSLM-Flamingo Llama3.2-1B · Sleep-CoT | F1: 49.33 -> 12.76 | Hurt |
| add 50% Gaussian noise to the input series (ablation) | OpenTSLM-Flamingo Llama3.2-1B · ECG-QA-CoT | F1: 41.58 -> 39.92 | No clear effect |
| replace input with Gaussian noise (sensor failure) (ablation) | OpenTSLM-Flamingo Llama3.2-1B · HAR-CoT | F1: 62.93 -> 10.59 | Hurt |
| use cross-attention (Flamingo) instead of soft prompting for long or multiple series | OpenTSLM · training VRAM | peak VRAM | Improved |
| use soft prompting with small-context Gemma backbones | OpenTSLM-SoftPrompt · Sleep-CoT / ECG-QA-CoT | F1: Gemma3-270M 7.96 and 1.29 F1 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ITFormer — ITFormer: Bridging Time Series and Natural Language for Multi-Modal QA with Large-Scale Multitask Dataset (ICML 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- GEM: Empowering MLLM for Grounded ECG Understanding with Time Series and Images (NeurIPS 2025 · similarity 0.87)
- ITFormer: Bridging Time Series and Natural Language for Multi-Modal QA with Large-Scale Multitask Dataset (ICML 2025 · similarity 0.86)
- JoLT: Jointly Learned Representations of Language and Time-Series for Clinical Time-Series Interpretation (Student Abstract) (AAAI 2024 · similarity 0.86)
- ECG-Reasoning-Benchmark: A Benchmark for Evaluating Clinical Reasoning Capabilities in ECG Interpretation (NeurIPS 2026 · similarity 0.86)
- MAESTRO : Adaptive Sparse Attention and Robust Learning for Multimodal Dynamic Time Series (NeurIPS 2025 · similarity 0.86)
- Reading Your Heart: Learning ECG Words and Sentences via Pre-training ECG Language Model (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
