TSRouter Dynamically Selects Models for Time Series Reasoning

Time series reasoning is a critical capability for many real-world applications, but it presents a challenge for AI models. Large Language Models (LLMs) excel at numerical precision when processing time series as text, yet struggle with global patterns. Conversely, Vision-Language Models (VLMs) effectively capture visual patterns but may lose fine-grained numerical details. Furthermore, different models have varying expertise and inference costs. To address this, researchers developed TSRouter, a dynamic routing framework that intelligently selects the optimal modality and model for each time series query. TSRouter constructs a heterogeneous graph that maps the complex interactions between tasks, query characteristics, modality attributes, and model capabilities. The framework formulates routing as a candidate scoring problem, evaluating each modality-model pair based on user-defined performance-cost preferences. TSRouter demonstrated substantial improvements (16% to 46% relative gains) over diverse baselines across four time series reasoning tasks. It also showed robust zero-shot generalization to unseen models and tasks, while reducing computational overhead through cost-aware optimization.