Time series foundation models (TSFMs) promise to be powerful tools for a wide range of applications. However, their internal representations and learned concepts are still not well understood. In this study, we investigate the structure and redundancy of representations across various TSFMs, examining the self-similarity of model layers within and across different model sizes. This analysis reveals block-like redundancy in the representations, which can be utilized for informed pruning to improve inference speed and efficiency. We also explore the concepts learned by these models, such as periodicity and trends. We demonstrate how conceptual priors can be derived from TSFM representations and leveraged to steer its outputs toward concept-informed predictions. Our work bridges representational analysis from language and vision models to TSFMs, offering new methods for building more computationally efficient and transparent TSFMs.

Modern data-driven systems often rely on complex pipelines to process and transform data for downstream machine learning (ML) tasks. Extracting these pipelines and understanding their structure is critical for ensuring transparency, performance optimization, and maintainability, especially in large-scale projects. In this work, we introduce a novel system, APEX-DAG ( A utomating P ipeline EX traction with D ataflow, Static Code A nalysis, and G raph Attention Networks), which automates the extraction of data pipelines from computational notebooks or scripts. Unlike execution-based methods, APEX-DAG leverages static code analysis to identify the dataflow, transformations, and dependencies within ML workflows without executing the code or the need to alter the code. Further, after an initial training phase, our system can identify pipelines that built with previously unseen libraries.