Beyond Task-Specific Classifiers: In-Context Inference for Time Series Classification Foundation Models

Time series classification foundation models are commonly evaluated by freezing a pretrained backbone and fitting task-specific classifiers on extracted representations. While effective, this pipeline entangles representation quality with classifier choice, hyperparameter tuning, and per-dataset optimization, making it less suitable when deployment-time fitting is undesirable. We propose TIC-FM, a deployment-time training-free in-context inference framework that uses labeled support examples as context and predicts query labels without fitting task-specific classifiers. TIC-FM maps time series and labels into an ICL-compatible token space, consolidates long contexts with latent memory, and performs leakage-safe parallel inference via support-only label injection and split-masked reasoning. On the UCR archive, TIC-FM improves average accuracy over representative classifier-fitting baselines, especially in low-label regimes, suggesting that support-conditioned in-context inference is a practical alternative for time series classification.