Title: Enhancing Time Series Reasoning Through Adaptive Segment Selection

Abstract:

Reasoning tasks involving time series data typically begin with a natural language query, necessitating a focused analysis of the temporal data. Because relevant evidence might be distributed across the entire dataset or confined to specific, brief intervals, models must determine precisely which portions of the data to examine. Current methods generally encode the complete time series into a static representation prior to inference, irrespective of whether the full sequence is actually pertinent to the query.

To address this, we present ARTIST, a framework that redefines time-series reasoning as a sequential decision-making process. ARTIST integrates reasoning with adaptive temporal segment selection, employing a dual-component architecture comprising a controller and a reasoner. Through reinforcement learning, the controller is trained to identify and select informative data segments, while the reasoner generates reasoning traces and final answers conditioned on these selected segments. Unlike models that rely on static summaries of the entire sequence, ARTIST actively gathers task-specific information during inference.

For post-training optimization, we introduce a novel hierarchical policy optimization method that enables the model to excel in both identifying relevant segments and answering questions. We assessed ARTIST’s performance across six time-series reasoning benchmarks, comparing it against large language models, vision-language models, and previous time-series reasoning systems. The results demonstrate that ARTIST achieves an average accuracy improvement of 6.46 percentage points over the best-performing baseline, with the most significant advancements observed in tasks involving rare event localization and multi-segment reasoning. While supervised fine-tuning enhances performance, reinforcement learning yields further improvements by optimizing segment selection to be adaptive to the specific question. These findings highlight that selective data utilization is key to effective time-series reasoning.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC