MarsTSC: VLM Agentic Reasoning Framework for Few-Shot Time Series Classification

MarsTSC, a novel framework, utilizes vision-language models (VLMs) for classifying multimodal time series with minimal examples. It features a self-improving knowledge repository that is progressively enhanced through reflective agentic reasoning. Three key roles are established: the Generator classifies through reasoning; the Reflector identifies reasoning mistakes to provide valuable insights; and the Modifier implements confirmed updates to the knowledge bank to avoid context collapse. A strategy for updating during testing allows for careful and ongoing refinement to address few-shot bias and shifts in distribution. The research can be accessed on arXiv (2605.09395).