LLM-Guided Planning for Time-Series Forecasting
A new research paper proposes using large language models (LLMs) as planning agents to guide time-series foundation models (TSFMs) in text-conditioned forecasting. The approach treats forecasting as a planning problem over TSFM-generated trajectories, where the frozen TSFM acts as a simulator proposing numerical continuations, and the LLM acts as a policy and value function to select candidates and evaluate trajectories against contextual text. This method avoids distorting temporal structure by not having the LLM directly generate forecast values. The paper is available on arXiv under ID 2607.24892.
Key facts
- Paper ID: arXiv:2607.24892
- Title: LLM as Forecasting Planner: Training-Free Text Conditioning for Time-Series Foundation Models
- Proposes using LLM as planning agent for TSFM
- TSFM acts as simulator, LLM as policy/value function
- Avoids direct LLM generation of forecast values
- Training-free approach
- Addresses text-conditioned time-series forecasting
- Published on arXiv
Entities
Institutions
- arXiv