Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Visual Reasoning over Time Series via Multi-Agent System

Weilin Ruan, Yuxuan Liang

TL;DR

MAS4TS introduces a tool-driven multi-agent framework built on an Analyzer–Reasoner–Executor paradigm to enable explicit visual reasoning over time series plots and adaptive execution across forecasting, classification, imputation, and anomaly detection. It uses a Vision-Language Model to extract structured anchors from plots, constrains latent trajectory reconstruction with anchor-informed dynamics, and orchestrates a dynamic tool chain via a memory-guided router. The approach achieves state-of-the-art performance across multiple tasks and datasets, while maintaining a lightweight backbone and efficient inference through selective grounding and utility tools. Overall, MAS4TS demonstrates the practical value of combining visual reasoning with modular tool-based execution for robust, generalizable time-series analysis.

Abstract

Time series analysis underpins many real-world applications, yet existing time-series-specific methods and pretrained large-model-based approaches remain limited in integrating intuitive visual reasoning and generalizing across tasks with adaptive tool usage. To address these limitations, we propose MAS4TS, a tool-driven multi-agent system for general time series tasks, built upon an Analyzer-Reasoner-Executor paradigm that integrates agent communication, visual reasoning, and latent reconstruction within a unified framework. MAS4TS first performs visual reasoning over time series plots with structured priors using a Vision-Language Model to extract temporal structures, and subsequently reconstructs predictive trajectories in latent space. Three specialized agents coordinate via shared memory and gated communication, while a router selects task-specific tool chains for execution. Extensive experiments on multiple benchmarks demonstrate that MAS4TS achieves state-of-the-art performance across a wide range of time series tasks, while exhibiting strong generalization and efficient inference.

Visual Reasoning over Time Series via Multi-Agent System

TL;DR

MAS4TS introduces a tool-driven multi-agent framework built on an Analyzer–Reasoner–Executor paradigm to enable explicit visual reasoning over time series plots and adaptive execution across forecasting, classification, imputation, and anomaly detection. It uses a Vision-Language Model to extract structured anchors from plots, constrains latent trajectory reconstruction with anchor-informed dynamics, and orchestrates a dynamic tool chain via a memory-guided router. The approach achieves state-of-the-art performance across multiple tasks and datasets, while maintaining a lightweight backbone and efficient inference through selective grounding and utility tools. Overall, MAS4TS demonstrates the practical value of combining visual reasoning with modular tool-based execution for robust, generalizable time-series analysis.

Abstract

Time series analysis underpins many real-world applications, yet existing time-series-specific methods and pretrained large-model-based approaches remain limited in integrating intuitive visual reasoning and generalizing across tasks with adaptive tool usage. To address these limitations, we propose MAS4TS, a tool-driven multi-agent system for general time series tasks, built upon an Analyzer-Reasoner-Executor paradigm that integrates agent communication, visual reasoning, and latent reconstruction within a unified framework. MAS4TS first performs visual reasoning over time series plots with structured priors using a Vision-Language Model to extract temporal structures, and subsequently reconstructs predictive trajectories in latent space. Three specialized agents coordinate via shared memory and gated communication, while a router selects task-specific tool chains for execution. Extensive experiments on multiple benchmarks demonstrate that MAS4TS achieves state-of-the-art performance across a wide range of time series tasks, while exhibiting strong generalization and efficient inference.
Paper Structure (80 sections, 24 equations, 4 figures, 17 tables, 2 algorithms)

This paper contains 80 sections, 24 equations, 4 figures, 17 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Time Series Analysis
  4. Pretrained Large Models for Time Series
  5. Multi-Agent Systems
  6. Methodology
  7. Problem Formulation
  8. Multi-Agent Collaboration Framework
  9. Shared Memory Mechanism.
  10. Information Carrier Encoder.
  11. Latent Visual Reasoning
  12. VLM-based Anchoring.
  13. Latent Construction.
  14. Latent Fusion with Data Embedding.
  15. Tool-Integrated Execution and Optimization
  16. ...and 65 more sections

Figures (4)

  • Figure 1: The motivation of MAS4TS. Existing approaches (left) struggle to support intuitive reasoning and task-adaptive execution, while MAS4TS (right) organizes time series analysis into a unified paradigm that collaboratively analyzes, reasons, and executes.
  • Figure 2: Overview of the MAS4TS framework following an Analyzer-Reasoner-Executor paradigm. The Reasoner agent performs two sequential stages: Visual Reasoning (VLM-based anchor extraction) and Numeric Reasoning (latent trajectory reconstruction). A unified tool library supports the entire workflow, with primary integration at the Executor.
  • Figure 3: The ablation results on the Weather dataset with prediction horizon $H=720$. Each component contributes to the performance, while generative reasoner does not improve further.
  • Figure 4: Hyperparameter sensitivity analysis on ETTm1 ($H=720$). Each radar chart shows MSE (solid) and MAE (dashed) for: a) reasoning steps; b) memory dimension $D_m$; c) hidden dimension $D_h$; and d) feedforward dimension $D_r$.