Survey of Reasoning and Agentic Systems for Time Series Analysis with Large Language Models

The survey, authored by a team from multiple institutions and accepted to Transactions on Machine Learning Research (TMLR), defines time series reasoning as a paradigm that treats time as a primary analytical axis and embeds intermediate evidence directly into model outputs. It organizes existing methods into three reasoning topologies—direct one-step reasoning, linear chain reasoning with explicit intermediate steps, and branch-structured reasoning that explores, revises, and aggregates—and crosses these with core objectives such as traditional time series analysis, causal inference, explanation, and generation. A compact tag set captures additional dimensions including decomposition, verification, ensembling, tool use, knowledge access, multimodality, and agent loops. The survey reviews methods across multiple application domains, identifying where each topology succeeds and where it breaks down in terms of faithfulness or robustness, and curates relevant datasets, benchmarks, and resources. Key guidance is offered on matching reasoning topology to uncertainty levels, grounding outputs in observable artifacts, and planning for distributional shift and streaming data. The authors stress that reasoning structures must balance grounding capacity and self-correction against computational cost and reproducibility. They conclude that closed-loop testbeds capable of trading off cost and risk under shift-aware, streaming, and long-horizon conditions will be essential for advancing the field.