FM-LLM: Frequency-Enhanced MoE Framework for LLM-Based Time Series Forecasting

A new study has been released on arXiv (ID: 2608.11623) introducing FM-LLM, a groundbreaking framework that improves large language models (LLMs) for forecasting time-series data using a frequency-based mixture-of-experts method. This approach addresses the limitations of existing cross-modal techniques that rely on text prompts for modality alignment, which can be heavy on computation and miss the unique spectral characteristics of time-series information. FM-LLM allows for a prompt-free adaptation of static LLMs through asymmetric coupling. It uses a Fourier Analysis Network (FAN) to align spectral tokens, integrating harmonic representations into the LLM. The asymmetric Mixture-of-Experts (MoE) decoder enables role specialization, with certain experts reconstructing periodic patterns while others capture residuals. You can check it out at https://arxiv.org/abs/2608.11623.