Paper Overview
- Field: NLP
- Authors: Hedi Zisling, Liran Nochumsohn, Ilan Naiman, Mihir Parmar, Omri Azencot
- Published: 2026-10-15
- arXiv: 2610.22327
- Line plots — preserve absolute measurements and the observed coordinate frame.
- Normalized delay embeddings — factor out positive amplitude rescaling and offsets while exposing temporal structure.
- arXiv: https://arxiv.org/abs/2610.22327
Summary
Time Series Language Models (TSLMs) enable language-based understanding and reasoning over temporal data, yet a fundamental question remains: what information should their time-series representations preserve, and what variation should they factor out? Existing TSLMs typically expose the language model to a single representation, imposing a fixed inductive bias across downstream tasks. The authors instead introduce complementary representations as a design principle for time-series reasoning.
Their model combines two visual views:
Dedicated visual encoders integrate both views into a shared language model built on Qwen3-VL-8B.
Results
On the TSRBench benchmark:
| Model | Overall Accuracy | |---|---| | Line plots only | 44.5% | | Strongest specialized TSLM baseline | 36.7% | | Dual-view (proposed) | 53.2% |
The dual-view model substantially outperforms specialized TSLM baselines, demonstrating that complementary representations provide an effective design principle for general-purpose TSLMs.