Time-LLM

Time-LLM (Jin et al., ICLR 2024) was one of the first serious attempts to use large language models for forecasting without fine-tuning them. Two mechanisms do the work: input time series are reprogrammed into text-prototype representations that look more natural to a language model, and a Prompt-as-Prefix mechanism injects declarative task context (domain knowledge, instructions) to steer reasoning. The backbone LLM, Llama-7B by default with GPT-2 and BERT options, stays completely frozen; only the reprogramming and projection layers train.

The authors report that it outperforms specialized forecasting models and excels at few-shot and zero-shot learning, and the paper has passed 1,000 citations. It was adopted industrially by XiMou Optimization for solar, wind, and weather forecasting, and folded into NeuralForecast and PyPOTS. The tradeoff is cost: running a 7B LLM for every forecast is heavy, and the authors state no major updates to the main codebase are planned, though PRs are welcome.

Author-reported (Jin et al., ICLR 2024): outperforms state-of-the-art specialized forecasting models on long-term benchmarks; strong few-shot and zero-shot transfer results. Independent replication is thin; several follow-up LLM-for-time-series papers report mixed results for frozen-LLM approaches.