Emergent Temporal Reasoning: Distilling Chain-of-Thought for Zero-Shot Combinatorial Generalization

This work establishes a Reasoning-Rich Distillation framework extending prior research on time series analysis. We introduce a synthetic dataset of two time series plots with annotations and textual descriptions provided by a large pretrained foundation model, specifically Qwen3.5-35B-A3B (Vision), under an Answer-then-Explain strategy that justifies the ground-truth elements. The distilled smaller models are Qwen3.5-2B (Text) and Qwen3.5-2B (Vision), and we investigate whether leveraging vision (text-plus-image input) or appending a pseudo-chain-of-thought, corresponding to the learned description obtained in the synthetic dataset, improves their performance in multiple time series analysis. Our results highlight the improvement obtained by leveraging both modalities, showcasing more efficient pairwise compositional reasoning, namely small models' robustness on multi-series analysis ($N \ge 3$) while being trained only on two-time-series examples. We further push the evaluation boundary to $N=10$, demonstrating persistent pairwise compositional reasoning capabilities even at extreme out-of-distribution (OOD) complexity. This work demonstrates, for the first time, compositional generalization in time series analysis for 2B-parameter large language models.