Improving ICL Signal Fidelity

Background: Transformer-based models for time series forecasting can suffer from output oversmoothing when presented with noisy or heterogeneous context examples during in-context learning (ICL).

Question / Future Work: Further investigation is needed to develop and evaluate strategies that effectively balance denoising (resisting spurious signals) and selection (focusing on highly relevant support samples) to ensure high-capacity ICL models accurately capture abrupt fluctuations and fine-grained structure in time series, rather than producing overly smooth predictions.