Researchers propose TS-Reasoner, a method to align time series foundation models (TSFMs) with large language models (LLMs) for reasoning tasks. The goal is to enable TSFMs to understand and reason about time series data using LLMs' textual inputs. To achieve this, the authors develop a two-stage training recipe that first aligns the latent representations of TSFMs with LLMs through synthetic pairs of time series and textual captions, and then fine-tunes the model for instruction following. Experiments show that TS-Reasoner outperforms various open-source models on several benchmarks while using less than half the training data. --- Why it matters: This matters to researchers in AI because it addresses a key challenge in integrating time series analysis with large language understanding, enabling more effective decision-making in domains like finance and scientific discovery. The proposed method also demonstrates remarkable data efficiency, which is crucial for real-world applications where data is often limited. Source: https://arxiv.org/abs/2510.03519