source: arxiv artificial intelligence: gitco: gated inference-time context optimization in tsfms
level: technical
patch-based time series foundation models often suffer from context poisoning, where structurally unusual patches grab too much attention and quietly hurt zero-shot forecasts. the problem is not the model weights but the input context itself. gitco, a lightweight framework, fixes this at inference time without any parameter updates. it uses three parts: a gate to detect harmful patches, a router to decide which ones to suppress, and a critic to guide the process.
tested on timesfm 2.5 across 53 gift-eval datasets with k-fold cross-validation, gitco cut mean absolute scaled error by an average of 1.95%. it captured 89.9% of the maximum possible improvement, showing it gets close to the best achievable result. the method works by selectively identifying and removing patches that distort the model's attention, making forecasts more reliable without changing the underlying model.
the work also introduces context sensitivity profiles, a new way to describe how a model's accuracy changes based on input meta-features. these profiles map time series properties to expected gains from inference-time context tweaks. this helps practitioners understand when and why a model might fail and how much they can fix it by cleaning the input, all without expensive retraining or access to training data.
why it matters: it lets data scientists boost time series forecast accuracy instantly without retraining, saving compute and time.
source: arxiv artificial intelligence: gitco: gated inference-time context optimization in tsfms