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

1. Problem Definition

The paper addresses the critical gap in Spatio-Temporal Reasoning within time series analysis. While Large Language Models (LLMs) excel in textual and visual reasoning, and existing time series models focus on predictive accuracy (point forecasting), there is a lack of systems capable of explicit reasoning over complex spatio-temporal data.

• The Challenge: Real-world systems (traffic, power grids, disease spread) involve coupled spatial dependencies (graphs) and temporal dynamics. Current models fail to answer "why" and "how" questions (e.g., "Which source node caused congestion at Node 2 at 9:00?") because they lack the ability to integrate numerical time series, graph structures, and textual context for multi-step causal inference.
• Key Obstacles:
  1. Data Scarcity: No existing datasets pair spatio-temporal time series with aligned natural language descriptions explaining spatial entities and dependencies.
  2. Evaluation Gaps: No standardized benchmarks exist to decompose spatio-temporal reasoning into distinct tasks.
  3. Model Limitations: Existing RL methods rely on result-only rewards, failing to enforce spatial grounding. Models often exploit superficial temporal patterns rather than learning genuine spatial attribution.

2. Methodology

The authors propose a comprehensive framework comprising a data synthesis pipeline, a new benchmark, and a specialized reasoning model.

A. Data Synthesis: Network SDE-Based Multi-Agent Pipeline

To overcome data scarcity, the authors developed a pipeline using Network Stochastic Differential Equations (Network SDEs) to generate synthetic data with precise control over dynamics.

• Mechanism: The system uses a multi-agent framework (6 agents) including a Scenario Generation Agent, Parsing Agent, and Judge Agents.
• Core Logic: It models time series $X_i(t)$ for nodes in a graph $G$ using SDEs that include:
  • Drift/Diffusion: Intrinsic temporal dynamics.
  • Network Coupling: Spatial dependencies via time-varying adjacency weights $A_{ji}(t)$ and propagation delays $\tau_{ji}$.
• Alignment: The pipeline ensures strict alignment between the generated numerical data, the graph structure, and the textual scenario descriptions (e.g., distinguishing "demand source" nodes from "propagation" nodes).

B. Benchmark: ST-Bench

Based on the synthetic data, the authors constructed ST-Bench, a benchmark consisting of four core reasoning tasks:

1. Etiological Spatial Reasoning: Inferring global system-level semantics (e.g., identifying the root cause of a system-wide event).
2. Spatial Entity Identification: Recognizing the semantic roles of nodes (e.g., is this node a source or a relay?) based on time series patterns.
3. Spatial Correlation Reasoning: Evaluating causal relationships, including direct influence and multi-hop propagation.
4. In-Context Spatio-Temporal Forecasting: Predicting future values under spatial dependencies.

C. Model Architecture: STReasoner

STReasoner is a unified model designed to process multimodal inputs (Time Series, Graph, Text).

• Encoder: Uses a lightweight Time Series Encoder (5-layer MLP) to patchify and embed time series data.
• Integration: Time series embeddings are interleaved with text tokens (including the graph structure described in text) to form a single input sequence for the LLM (TS-LM).
• Training Strategy: A three-stage paradigm:
  1. Alignment Pretraining: Aligns time series embeddings with text using basic QA pairs (ST-Align).
  2. SFT (Supervised Fine-Tuning): Injects reasoning priors using Chain-of-Thought (CoT) data generated via rejection sampling.
  3. Spatial-Aware RL: Uses the proposed S-GRPO algorithm.

D. Algorithm: Spatial-Aware GRPO (S-GRPO)

Standard Group Relative Policy Optimization (GRPO) rewards only the final answer. S-GRPO introduces a contrastive reward mechanism to enforce spatial reasoning:

• Mechanism: For a given query, the model generates two groups of responses: one with the graph structure provided ($O_{sp}$) and one without ($O_{ns}$).
• Reward Function: The model receives an additional reward ($\alpha$) only if the performance with spatial information ($r_{sp}$) is significantly better than without ($r_{ns}$).
  $$R_i = \begin{cases} r_{sp}^i + \alpha & \text{if } r_{sp}^i > \beta r_{ns}^i \\ r_{sp}^i & \text{otherwise} \end{cases}$$
• Goal: This forces the model to learn that utilizing the spatial structure leads to better outcomes, preventing it from relying solely on temporal patterns.

3. Key Contributions

1. Formal Definition & Benchmark: First formal study of spatio-temporal reasoning in time series, introducing ST-Bench with four distinct reasoning tasks.
2. Data Synthesis Pipeline: A novel Network SDE-based multi-agent pipeline that generates high-quality, aligned spatio-temporal data and textual descriptions, solving the data scarcity problem.
3. STReasoner Model: A unified architecture integrating time series encoders with LLMs, trained via a three-stage process.
4. S-GRPO Algorithm: A novel reinforcement learning objective that explicitly incentivizes spatial grounding, ensuring models use graph structure for reasoning rather than superficial correlations.

4. Experimental Results

The authors evaluated STReasoner against proprietary models (GPT-5.2, Claude-4.5), open-source LLMs (Qwen3), and specialized time series models.

• Performance Gains: STReasoner achieved average accuracy gains of 17% to 135% over the best baselines across the four tasks.
  • Specifically, it outperformed proprietary models on Etiological (17% gain), Entity (135% gain), and Correlation (40% gain) tasks.
  • It achieved comparable forecasting performance (T4) with only a 2.6% difference.
• Cost Efficiency: STReasoner operates at 0.004× the cost of proprietary models (estimated based on API pricing).
• Zero-Shot Generalization: Trained entirely on synthetic data, STReasoner achieved 98.82% accuracy on a real-world causal QA dataset (CausalRivers), significantly outperforming GPT-5.2 (22.32%) and Claude-4.5 (83.18%).
• Ablation Studies:
  • S-GRPO vs. GRPO: Replacing standard GRPO with S-GRPO improved accuracy by 5.10% and significantly increased the model's reliance on spatial information (from ~69% to ~87% usage).
  • Modality: Explicitly encoding time series via a dedicated encoder was proven superior to prompting time series as text or images.

5. Significance

This work represents a paradigm shift in time series analysis by moving from pure prediction to explicit reasoning.

• Decision Support: It enables AI systems to provide interpretable, step-by-step explanations for complex system failures (e.g., tracing a power grid cascade or disease outbreak), which is vital for high-stakes domains like healthcare and infrastructure.
• Efficiency: It demonstrates that open-source models, when properly trained with spatial-aware RL and synthetic data, can outperform massive proprietary models at a fraction of the cost.
• Foundation for Future Research: The introduction of ST-Bench and the S-GRPO algorithm provides a new standard and methodology for developing reasoning-capable models for any domain involving structured, time-evolving data.