Attenuated in-context identification in time-series foundation models: diagnosis under counterfactual inputs and repair by synthetic forced-system fine-tuning

TL;DR

This paper analyzes the performance of time-series foundation models in identifying dynamics in engineering systems. The models exhibit attenuated in-context identification, but fine-tuning on synthetic forced systems improves their performance. The study highlights the importance of counterfactual inputs in diagnosing model limitations.

Why It Matters

The research has practical significance in applications where accurate what-if analysis is crucial, such as control systems and instrumented plants. Improved time-series foundation models can enhance decision-making and optimization in these domains. The findings also underscore the need for careful model evaluation and fine-tuning to ensure reliable performance.

Key Findings

• Attenuated in-context identification in time-series foundation models
• Fine-tuning on synthetic forced systems improves model performance
• Counterfactual inputs are essential for diagnosing model limitations