Latent Instructions as Context Surrogates: Enhancing Frozen Time Series Forecasters with Instance-Adaptive Prompts

Time series foundation models (TSFMs) benefit from longer context windows, but with steeply diminishing returns. We propose instance-adaptive latent instructions, a set of learnable prompt tokens prepended to the input of a frozen TSFM that serve as compact context surrogates to encode distributional properties. A lightweight prompt generation module, consisting of an MLP and a learnable latent token basis, maps the statistics of input series to the data-specific prompts for each instance. Introducing only ${\sim}$0.4\% additional parameters and no modification to the pretrained model, our method with a short context consistently matches or outperforms the zero-shot baseline with doubled context length, while significantly reducing computational cost. Experiments on 62 long-context test sets from the GIFT-Eval benchmark with different backbones validate the effectiveness and efficiency of our approach.