Paper 2604.17295
LLaTiSA: Towards Difficulty-Stratified Time Series Reasoning from Visual Perception to Semantics
- Published
- Apr 2026
- Research lab
- Independent
- Citations
- 1
- GitHub
- 78 stars
01 In brief
Summary
The paper introduces LLATISA, a Vision-Language Model (VLM) for time series reasoning (TSR), and HITSR, a hierarchical dataset of 83k samples.
The authors formalize a four-level TSR taxonomy (L1: Numerical Read-out, L2: Pattern Perception, L3: Semantic Reasoning, L4: Predictive Inference) to address fragmented task definitions and unreliable benchmarks.
HITSR covers L1-L3 with synthetic and real-world data, featuring verified Chain-of-Thought (CoT) annotations.
LLATISA uses a dual-view input (time series plot plus a numerical table image) to combine qualitative visual perception with precise numerical grounding.
It is trained via a three-stage curriculum aligned with the taxonomy.
Experiments on out-of-distribution benchmarks show LLATISA outperforms proprietary models (e.g., GPT-4o) and open-source baselines across L1-L3, with notable gains in numerical read-out (86.8% vs.
54.2% for GPT-4o) and pattern perception.
In ECG interpretation, LLATISA achieves higher lead assessment coverage and accuracy than domain-specific models, using only 2.5% of GEM's training data.
Ablations confirm the importance of CoT data and curriculum learning for generalization.
The paper also extends to L4 predictive inference, showing further improvements with a fourth curriculum stage.
02 From the paper
Abstract
Comprehensive understanding of time series remains a significant challenge for Large Language Models (LLMs). Current research is hindered by fragmented task definitions and benchmarks with inherent ambiguities, precluding rigorous evaluation and the development of unified Time Series Reasoning Models(TSRMs). To bridge this gap, we formalize Time Series Reasoning (TSR) via a four-level taxonomy of increasing cognitive complexity. We introduce HiTSR, a hierarchical time series reasoning dataset comprising 83k samples with diverse task combinations and verified Chain-of-Thought (CoT) trajectories. Leveraging HiTSR, we propose LLaTiSA, a strong TSRM that integrates visualized patterns with precision-calibrated numerical tables to enhance the temporal perception of Vision-Language Models (VLMs). Through a multi-stage curriculum fine-tuning strategy, LLaTiSA achieves superior performance and exhibits robust out-of-distribution generalization across diverse TSR tasks and real-world scenarios. Our code is available at https://github.com/RainingNovember/LLaTiSA.