Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

STReasoner: Empowering LLMs for Spatio-Temporal Reasoning in Time Series via Spatial-Aware Reinforcement Learning

Juntong Ni, Shiyu Wang, Ming Jin, Qi He, Wei Jin

TL;DR

This work introduces ST-Bench, a benchmark consisting of four core tasks, including etiological reasoning, entity identification, correlation reasoning, and in-context forecasting, developed via a network SDE-based multi-agent data synthesis pipeline, and proposes STReasoner, which empowers LLM to integrate time series, graph structure, and text for explicit reasoning.

Abstract

Spatio-temporal reasoning in time series involves the explicit synthesis of temporal dynamics, spatial dependencies, and textual context. This capability is vital for high-stakes decision-making in systems such as traffic networks, power grids, and disease propagation. However, the field remains underdeveloped because most existing works prioritize predictive accuracy over reasoning. To address the gap, we introduce ST-Bench, a benchmark consisting of four core tasks, including etiological reasoning, entity identification, correlation reasoning, and in-context forecasting, developed via a network SDE-based multi-agent data synthesis pipeline. We then propose STReasoner, which empowers LLM to integrate time series, graph structure, and text for explicit reasoning. To promote spatially grounded logic, we introduce S-GRPO, a reinforcement learning algorithm that rewards performance gains specifically attributable to spatial information. Experiments show that STReasoner achieves average accuracy gains between 17% and 135% at only 0.004X the cost of proprietary models and generalizes robustly to real-world data.

STReasoner: Empowering LLMs for Spatio-Temporal Reasoning in Time Series via Spatial-Aware Reinforcement Learning

TL;DR

This work introduces ST-Bench, a benchmark consisting of four core tasks, including etiological reasoning, entity identification, correlation reasoning, and in-context forecasting, developed via a network SDE-based multi-agent data synthesis pipeline, and proposes STReasoner, which empowers LLM to integrate time series, graph structure, and text for explicit reasoning.

Abstract

Spatio-temporal reasoning in time series involves the explicit synthesis of temporal dynamics, spatial dependencies, and textual context. This capability is vital for high-stakes decision-making in systems such as traffic networks, power grids, and disease propagation. However, the field remains underdeveloped because most existing works prioritize predictive accuracy over reasoning. To address the gap, we introduce ST-Bench, a benchmark consisting of four core tasks, including etiological reasoning, entity identification, correlation reasoning, and in-context forecasting, developed via a network SDE-based multi-agent data synthesis pipeline. We then propose STReasoner, which empowers LLM to integrate time series, graph structure, and text for explicit reasoning. To promote spatially grounded logic, we introduce S-GRPO, a reinforcement learning algorithm that rewards performance gains specifically attributable to spatial information. Experiments show that STReasoner achieves average accuracy gains between 17% and 135% at only 0.004X the cost of proprietary models and generalizes robustly to real-world data.
Paper Structure (42 sections, 7 equations, 11 figures, 9 tables, 1 algorithm)

This paper contains 42 sections, 7 equations, 11 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Problem and Dataset Construction
  4. Problem Definition
  5. Dataset Synthesis Pipeline
  6. Dataset Splits and Usage
  7. Our Solution: STReasoner
  8. Model Architecture
  9. Model Training
  10. Experiment
  11. Baselines.
  12. Main Results
  13. Zero-Shot Results on Real-World Data
  14. Ablation Study
  15. Effect of Spatial Reward Analysis
  16. ...and 27 more sections

Figures (11)

  • Figure 1: A traffic flow example of spatio-temporal reasoning. $Q$: query, $\mathcal{T}$: time series, $\mathcal{G}$: graph.
  • Figure 2: Overall framework of the Network SDEs-based multi-agent spatio-temporal data synthesis pipeline (upper) and spatio-temporal QA generation (lower). Detailed prompts for each agent are provided in Appendix \ref{['app:agent_prompt']}, and the pseudo algorithm is given in Algorithm \ref{['alg:data_synthesis']}.
  • Figure 3: An illustration of our proposed STReasoner with the S-GRPO algorithm.
  • Figure 4: RL training curves over steps.
  • Figure 5: S-GRPO Sensitivity Analysis.
  • ...and 6 more figures