Reproduce “LLaTiSA: Towards Difficulty-Stratified Time Series Reasoning from Visual Perception to Semantics” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yueyang Ding, HaoPeng Zhang, Rui Dai, Yi Wang, Tianyu Zong, Kaikui Liu, Xiangxiang Chu

Researchers from Amap, Alibaba Group, introduced a difficulty-stratified taxonomy for time series reasoning and developed HITSR, a large-scale, high-fidelity dataset for training and evaluating models. They also proposed LLATISA, a Vision-Language Model (VLM) that integrates visual patterns with numerical tables, achieving improved performance on diverse time series reasoning tasks, including real-world ECG interpretation.

Key insights

  • Integrating visual plots with numerical table images in a dual-view input framework effectively combines macroscopic pattern perception with point-accurate numerical verification, mitigating numerical hallucinations.
  • A multi-stage curriculum fine-tuning strategy, progressing from L1 (numerical read-out) to L3 (semantic reasoning), is crucial for robustly acquiring complex reasoning capabilities and improving generalization to out-of-distribution tasks.
  • Training with Chain-of-Thought (CoT) annotations is essential for enhancing model reasoning capabilities and preserving generative alignment, as demonstrated by significant performance deterioration when CoT is omitted.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.