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

Researchers have introduced MarsTSC, a novel Vision-Language Model (VLM) agentic reasoning framework designed for few-shot multimodal time series classification. Published on arXiv, this study addresses the challenges of limited data and distribution shifts in time series analysis. The framework features a self-evolving knowledge bank that is iteratively refined through reflective agentic reasoning involving three collaborative roles: a Generator for classification, a Reflector for diagnosing reasoning errors and identifying overlooked temporal features, and a Modifier for updating the knowledge bank to prevent context collapse. Additionally, a test-time update strategy is employed to mitigate few-shot bias. Extensive experiments across twelve mainstream benchmarks demonstrate that MarsTSC achieves substantial performance gains over classical and foundation model-based baselines using six different VLM backbones. The system not only improves accuracy but also provides interpretable rationales, grounding classification decisions in human-readable feature evidence. This advancement highlights the potential of agentic AI in enhancing complex data interpretation tasks with limited training samples.