Distribution-Aware Diffusion-LLM for Robust Ultra-Long-Term Time Series Forecasting

The article introduces Diffusion-LLM, a novel framework that integrates a conditional diffusion model with a large language model (LLM) for improved ultra-long-term time series forecasting. Evaluated on six benchmarks, including ETT and Weather, Diffusion-LLM demonstrates significant performance improvements over existing LLM-based methods, particularly in few-shot scenarios, by enhancing probabilistic modeling and semantic alignment in a shared latent space. This advancement is crucial for practitioners seeking robust and generalizable solutions in multimodal time series forecasting using LLMs.