Researchers introduce GITCO, a lightweight inference-time optimization framework that improves Time Series Foundation Models (TSFMs) by identifying and suppressing anomalous patches without modifying model weights. The method achieves a 1.95% average improvement in forecast accuracy on TimesFM 2.5, addressing the critical problem of context poisoning where structurally irregular data segments degrade zero-shot prediction quality.
Time Series Foundation Models represent a significant advancement in predictive analytics, yet they remain vulnerable to a subtle but pervasive degradation mechanism called context poisoning. Anomalous patches in time series data—irregular segments that deviate structurally from typical patterns—disproportionately capture model attention during inference, silently eroding forecast accuracy without obvious failure signals. GITCO addresses this challenge through an elegant inference-time intervention strategy that operates without retraining or modifying model parameters.
The framework's three-component architecture (Gate, Router, and Critic) works synergistically to detect harmful contextual elements and suppress their influence during prediction. This approach aligns with broader trends in machine learning toward post-hoc model optimization and dynamic input processing. Rather than investing computational resources in retraining large models, GITCO leverages the observation that inference-time context optimization can yield meaningful accuracy gains through selective filtering.
The introduction of context sensitivity profiles as a characterizable TSFM property opens new research directions. These profiles map time series meta-features to expected accuracy improvements under context intervention, revealing how model architecture and data statistics jointly determine robustness. The evaluation across 53 datasets using rigorous K-fold cross-validation demonstrates scalability and generalizability beyond isolated benchmarks.
For practitioners deploying time series models in production environments—financial forecasting, anomaly detection, demand prediction—this work suggests that inference-time optimization presents an underexplored efficiency frontier. Organizations can potentially improve existing model deployments without expensive retraining cycles, making this particularly valuable for resource-constrained applications.
- →GITCO achieves 1.95% average MASE improvement on TimesFM 2.5 through inference-time context optimization without parameter updates.
- →The framework identifies and suppresses anomalous patches that cause context poisoning in time series foundation models.
- →Context sensitivity profiles establish a new measurable property of TSFMs linking data characteristics to accuracy improvement potential.
- →The method captures 89.9% of the theoretical improvement upper bound across 53 diverse datasets.
- →Inference-time optimization offers an alternative to retraining for improving deployed time series models in production environments.