Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adaptive Time Series Reasoning via Segment Selection

Shvat Messica, Jiawen Zhang, Kevin Li, Theodoros Tsiligkaridis, Marinka Zitnik

TL;DR

This work introduces ARTIST, which formulates time-series reasoning as a sequential decision problem, and uses a novel hierarchical policy optimization approach for post-training that allows the model to excel in both segment selection and question-answering behavior.

Abstract

Time series reasoning tasks often start with a natural language question and require targeted analysis of a time series. Evidence may span the full series or appear in a few short intervals, so the model must decide what to inspect. Most existing approaches encode the entire time series into a fixed representation before inference, regardless of whether or not the entire sequence is relevant. We introduce ARTIST, which formulates time-series reasoning as a sequential decision problem. ARTIST interleaves reasoning with adaptive temporal segment selection. It adopts a controller-reasoner architecture and uses reinforcement learning to train the controller role to select informative segments and the reasoner role to generate segment-conditioned reasoning traces and final answers. During inference, the model actively acquires task-relevant information instead of relying on a static summary of the full sequence. We use a novel hierarchical policy optimization approach for post-training that allows the model to excel in both segment selection and question-answering behavior. We evaluate ARTIST on six time-series reasoning benchmarks and compare it with large language models, vision-language models, and prior time-series reasoning systems. ARTIST improves average accuracy by 6.46 absolute percentage points over the strongest baseline. The largest gains appear on rare event localization and multi-segment reasoning tasks. Supervised fine-tuning improves performance, and reinforcement learning provides additional gains by optimizing question-adaptive segment selection. These results show that selective data use drives effective time-series reasoning.

Adaptive Time Series Reasoning via Segment Selection

TL;DR

This work introduces ARTIST, which formulates time-series reasoning as a sequential decision problem, and uses a novel hierarchical policy optimization approach for post-training that allows the model to excel in both segment selection and question-answering behavior.

Abstract

Time series reasoning tasks often start with a natural language question and require targeted analysis of a time series. Evidence may span the full series or appear in a few short intervals, so the model must decide what to inspect. Most existing approaches encode the entire time series into a fixed representation before inference, regardless of whether or not the entire sequence is relevant. We introduce ARTIST, which formulates time-series reasoning as a sequential decision problem. ARTIST interleaves reasoning with adaptive temporal segment selection. It adopts a controller-reasoner architecture and uses reinforcement learning to train the controller role to select informative segments and the reasoner role to generate segment-conditioned reasoning traces and final answers. During inference, the model actively acquires task-relevant information instead of relying on a static summary of the full sequence. We use a novel hierarchical policy optimization approach for post-training that allows the model to excel in both segment selection and question-answering behavior. We evaluate ARTIST on six time-series reasoning benchmarks and compare it with large language models, vision-language models, and prior time-series reasoning systems. ARTIST improves average accuracy by 6.46 absolute percentage points over the strongest baseline. The largest gains appear on rare event localization and multi-segment reasoning tasks. Supervised fine-tuning improves performance, and reinforcement learning provides additional gains by optimizing question-adaptive segment selection. These results show that selective data use drives effective time-series reasoning.
Paper Structure (29 sections, 20 equations, 4 figures, 4 tables, 1 algorithm)

This paper contains 29 sections, 20 equations, 4 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. ARTIST
  4. Problem Setup and Notation
  5. Controller-Reasoner Roles in ARTIST
  6. Interaction Trajectory Rollout
  7. Reward Functions
  8. Hierarchical Policy Optimization
  9. Experiments
  10. Benchmarking Results
  11. Analysis of Time Series Data Utilization
  12. Ablation Study
  13. Illustration of Time Series Reasoning in ARTIST
  14. Conclusion
  15. Controller and Reasoner Objectives and Rewards
  16. ...and 14 more sections

Figures (4)

  • Figure 1: (a) Time-series reasoning: answering a natural language question given a time series. (b) ARTIST alternates between reasoning and adaptive segment selection, choosing the next segment based on the question and intermediate outputs, and stopping once it can produce the final answer.
  • Figure 2: Overview of ARTIST. (a) Given a question and a time series, a high-level controller iteratively selects informative temporal segments, while a low-level reasoner produces intermediate reasoning traces and answers conditioned on the accumulated segment set. The controller decides whether to continue selecting or accept it as a final answer and segment list. (b) For each training example, multiple reasoning rollouts are generated per controller trajectory. Reasoner rewards are group-normalized across reasoning rollouts, while controller rewards are computed using the associated reasoning outcomes and normalized across controller trajectories. Both signals are used for a joint policy update.
  • Figure 3: Performance and question distribution across time-series usage levels. Bars (left axis) show the percentage of questions falling into each usage-percentage bin, and the line (right axis) shows accuracy for questions in that bin.
  • Figure 4: Examples of using ARTIST for reasoning with time series. Source: Etiological time series reasoning bechmark.