A Survey of Reasoning and Agentic Systems in Time Series with Large Language Models

A comprehensive survey examines how large language models can reason about time series data through three structural topologies: direct reasoning, linear chain reasoning, and branch-structured reasoning. The research organizes methods across objectives including analysis, explanation, causal inference, and generation, emphasizing the need for evaluation practices that maintain evidence visibility and temporal alignment while balancing computational cost against reliability and reproducibility.

This survey addresses a critical gap in applying LLMs to time series analysis, moving beyond simple prediction toward genuine reasoning about temporal data. The work systematically categorizes approaches by reasoning topology—from single-step direct methods to complex branching structures that explore and revise—providing researchers and practitioners a mental model for selecting appropriate methods. The topology framework enables matching computational approaches to specific uncertainty levels and problem constraints, distinguishing between scenarios requiring explanation versus pure accuracy.

The research recognizes that time series reasoning in production systems faces distinct challenges absent from static data analysis: concept drift, streaming constraints, and long-horizon planning requirements. By emphasizing temporally-aligned evaluation and traceable evidence, the survey identifies a fundamental tension in deploying LLM-based systems at scale—achieving sufficient reasoning capacity for complex phenomena while maintaining reproducibility and cost efficiency. The inclusion of tool use, multimodality, and agent loops reflects how production systems combine LLMs with specialized components rather than treating language models as monolithic solvers.

For the AI industry, this systematization enables better architectural decisions when building time series applications in finance, infrastructure monitoring, and scientific domains where both accuracy and explainability matter. The emphasis on shift-aware evaluation and streaming settings acknowledges that real-world performance diverges from benchmark results, making reliability the central metric rather than point accuracy. The curated datasets, benchmarks, and open-source resources lower barriers for practitioners to implement these methods responsibly, potentially accelerating adoption of interpretable AI systems in domains where stakeholders demand understanding alongside predictions.

• →LLMs enable three distinct reasoning topologies for time series: direct, linear-chain, and branch-structured approaches, each suited to different problem constraints.
• →Evaluation practices must maintain visible evidence and temporal alignment rather than optimizing solely for accuracy metrics.
• →Production systems require cost and latency budgets as explicit design constraints, forcing tradeoffs between reasoning capacity and computational efficiency.
• →Streaming, concept drift, and long-horizon planning create distinct challenges that static benchmarks fail to capture in time series reasoning.
• →Future progress depends on shift-aware testbeds and benchmarks linking reasoning quality directly to downstream utility rather than isolated accuracy scores.