arXiv:2606.05332v1 Announce Type: new Abstract: Patch-based Time Series Foundation Models (TSFMs) suffer from context poisoning: structurally anomalous patches capture disproportionate attention and silently degrade zero-shot forecast quality. We propose improving TSFM accuracy at inference time by optimizing the input context rather than modifying model weights. We present GITCO (Gated Inference-Time Context Optimization), a lightweight three-component framework: Gate, Router, and Critic that selectively identifies and suppresses harmful patches without any parameter updates. Evaluated on Tim
The proliferation of Time Series Foundation Models (TSFMs) highlights the need for robust and efficient inference mechanisms as these models are increasingly deployed in real-world applications.
Improving the accuracy of TSFMs at inference time, particularly in zero-shot scenarios, expands their applicability and reliability across various industries without requiring extensive retraining or parameter updates.
The introduction of GITCO allows for more resilient TSFMs that can handle structurally anomalous data patches, leading to more trustworthy and immediately deployable forecasting solutions.
- · AI developers
- · Data scientists
- · Industries relying on time series forecasting
- · Open-source AI community
- · Companies relying on less efficient TSFM optimization methods
Improved performance and reliability of time series forecasting in critical applications.
Reduced operational costs due to less need for data cleaning or model fine-tuning for anomalous data.
Accelerated adoption of advanced TSFMs in novel, high-stakes environments where inference accuracy is paramount.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI