🤖 machine learning

Adaptive Time Series Reasoning via Segment Selection

This paper introduces ARTIST, a reinforcement learning-based framework that treats time series reasoning as a sequential decision problem to dynamically select informative temporal segments, thereby significantly outperforming existing models by avoiding static full-sequence encoding and focusing on task-relevant information.

Original authors: Shvat Messica, Jiawen Zhang, Kevin Li, Theodoros Tsiligkaridis, Marinka Zitnik

Published 2026-07-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shvat Messica, Jiawen Zhang, Kevin Li, Theodoros Tsiligkaridis, Marinka Zitnik

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of a crime scene, you are handed a massive, endless scroll of data that records every heartbeat, stock price, or weather temperature for a whole year. This is the world of time series reasoning, a field where computers try to answer questions like "Why did the patient's heart rate spike?" or "What caused the market to crash?" by looking at these long streams of numbers. Usually, when computers try to solve these puzzles, they are forced to read the entire scroll from start to finish, trying to memorize every single number before they can even think of an answer. It's like trying to find a specific typo in a library by reading every single book cover-to-cover, even if the typo is only on page 42 of one book. This approach is slow, wasteful, and often confuses the computer with too much irrelevant noise. The big question researchers are asking is: Can we teach computers to be smarter detectives, ones that know exactly which pages to flip to, skipping the boring parts to find the clues that actually matter?

Enter ARTIST, a new method that acts like a super-smart, adaptive detective for time series data. Instead of forcing the computer to stare at the whole timeline at once, ARTIST breaks the process into a fun, back-and-forth game between two roles: a Controller and a Reasoner. Think of the Controller as the detective's assistant who holds the map. When a question comes in (like "What happened in August?"), the Controller doesn't just guess; it decides, "Okay, I need to look at the first week to set a baseline," and then it grabs just that tiny slice of data. It hands this slice to the Reasoner, who is the detective doing the actual thinking. The Reasoner looks at that slice, says, "Hmm, that looks normal, but I need to check the next month to see if things changed," and asks the Controller for the next specific piece of the puzzle. They keep passing the baton—Controller grabs a segment, Reasoner thinks, Controller grabs another—until they have enough evidence to solve the case. This "interleaved" approach means the model only looks at the 30% to 70% of the data that is actually relevant, ignoring the rest.

The paper shows that this strategy works incredibly well. When tested on six different benchmarks involving everything from medical heart signals to financial trends, ARTIST beat the strongest existing computer models by a solid margin, improving average accuracy by 6.46 percentage points. The gains were even bigger for tricky tasks, like finding rare events or connecting clues across different time periods, where it jumped up to 12.5 percentage points better than the competition. The researchers found that this improvement comes from a clever training method called collaborative self-play reinforcement learning. In this setup, the two roles (Controller and Reasoner) practice together, learning from their own mistakes without needing a human teacher to tell them exactly which segment was "right" for every single question. The Controller learns to pick the best segments by checking if the Reasoner can consistently get the right answer using those segments, while the Reasoner learns to think clearly based on whatever segments it gets.

Crucially, the paper rules out the idea that the old way of doing things—feeding the computer the entire time series as a static block—is the best approach. The authors demonstrate that forcing the model to process the full sequence often dilutes the important clues with irrelevant noise, leading to worse performance. They also show that simply having a good reasoner isn't enough; if the Controller is frozen and can't adapt its choices based on the question, performance drops significantly. The results suggest that the key to solving these time series puzzles isn't just having a bigger brain, but having a brain that knows how to look efficiently. By using this selective, step-by-step strategy, ARTIST proves that sometimes, looking at less data actually helps you understand more.

Technical Summary: Adaptive Time Series Reasoning via Segment Selection (ARTIST)

Problem Definition

Time series reasoning tasks require models to answer natural language questions by identifying and analyzing specific, task-relevant parts of a time series. Unlike standard forecasting or classification tasks where the entire sequence is processed uniformly, reasoning tasks often depend on localized patterns (e.g., a specific spike, a regime shift, or a transient anomaly) that may span only a few intervals within a long sequence.

Existing approaches typically encode the entire time series into a fixed representation before inference, regardless of the question's specific requirements. This static approach has two primary limitations:

Information Dilution: Processing the full sequence mixes irrelevant context with task-relevant information, potentially obscuring salient local structures, especially in long sequences.
Lack of Adaptivity: The model cannot dynamically adjust its focus based on intermediate reasoning steps. For instance, a clinical query might first require checking for an abrupt spike, then shifting to a narrower window to test a hypothesis. Current methods lack the mechanism to interleave reasoning with the selection of specific temporal segments.

Furthermore, unlike vision-language tasks where bounding boxes or segmentation masks often provide explicit supervision for relevant regions, time series data rarely includes task-agnostic annotations. Relevant segments are inherently question-dependent, making it difficult to train models to select them without strong supervision.

Methodology: ARTIST

The authors introduce ARTIST (Adaptive Reasoning for TIme-Series via Temporal Selection), a framework that formulates time-series reasoning as a sequential decision problem. Instead of processing a static view, ARTIST interleaves reasoning with adaptive temporal segment selection.

Architecture: Controller-Reasoner Decomposition

ARTIST employs a single policy model $\pi_\theta$ that operates in two distinct roles via role-specific prompting:

Controller (High-Level): Receives the question, the full time series, the accumulated segment list, and the previous reasoning trace. It decides whether to continue the interaction or accept the current answer. If continuing, it selects a new informative temporal segment ( $s_i$ ) to inspect.
Reasoner (Low-Level): Receives the question and the accumulated list of selected segments. It generates intermediate reasoning steps and a final answer conditioned only on the retrieved segments.

This architecture explicitly decouples where to look (Controller) from how to reason (Reasoner).

Training: Hierarchical Collaborative Self-Play RL

The model is trained in two stages:

Supervised Fine-Tuning (SFT): The model is trained on curated traces that interleave natural language reasoning with explicit segment selection calls.
Collaborative Self-Play Reinforcement Learning (RL): The authors propose a novel hierarchical policy optimization approach to jointly optimize both roles without relying on external human annotations for segment relevance.
- Nested Rollouts: For each training example, the system generates $G$ interaction trajectories. For each trajectory, the Reasoner is resampled $N$ times conditioned on the final segment list selected by the Controller.
- Reward Design:
  - Reliability Reward (for Controller): Measures the consistency of the Reasoner's correctness across the $N$ resampled rollouts given the selected segments. This encourages the Controller to select segments that robustly support correct reasoning, rather than relying on stochastic "lucky" guesses.
  - Correctness Reward (for Reasoner): Based on the final answer's accuracy and format compliance.
- Hierarchical Advantage Calculation:
  - The Controller is optimized using trajectory-level advantages, propagating credit across the entire interaction sequence to learn a policy for building an informative set of segments over multiple steps.
  - The Reasoner is optimized using group-relative advantages on the final-round rollouts, disentangling its optimization from the variance in segment quality.
- Variance-Guided Sampling: To improve efficiency, the Reasoner update focuses on interaction trajectories that induce diverse outcomes (high variance in correctness), prioritizing groups that provide more informative learning signals.

Key Contributions

Adaptive Segment Selection: The paper presents a method where models iteratively choose temporal segments to inspect, updating reasoning based on retrieved information without assuming pre-defined segment labels.
Hierarchical Collaborative Self-Play RL: A post-training method that separates segment selection from answer generation, training each role with learning signals aligned to its specific function (reliability for selection, correctness for reasoning).
Empirical Superiority: ARTIST is evaluated on six diverse benchmarks (clinical, financial, environmental, and general domains) and outperforms strong baselines, including large language models (LLMs), vision-language models (VLMs), and prior time-series reasoning systems.

Results

Performance Gains: ARTIST improves average accuracy by 6.46 absolute percentage points over the strongest baseline across six benchmarks. The largest gains (up to 12.5 points) are observed in tasks requiring rare event localization and multi-segment reasoning.
Efficiency: The model typically uses only 30–70% of the input time series per query, demonstrating that selective temporal analysis improves accuracy without requiring full sequence consumption.
Ablation Studies:
- Removing the adaptive controller (processing the full series statically) results in a 9.30% drop in accuracy.
- Removing the Reliability Reward causes the most significant decline (21.44%), highlighting the necessity of optimizing for consistent reasoning rather than single-shot correctness.
- Replacing the hierarchical objective with a myopic (greedy) objective leads to a 12.28% performance drop, confirming the need for long-horizon planning in segment selection.
Modality Analysis: On the Sleep-QA dataset (EEG data), ARTIST's performance gap with VLMs was found to be primarily driven by the input modality (tokenized sequence vs. plotted image) rather than the reasoning mechanism itself. When ARTIST was instantiated with a VLM backbone, it surpassed VLM baselines.

Significance and Claims

The paper claims that selective data use drives effective time-series reasoning. By treating the time series as an active resource to be queried iteratively rather than a static context, ARTIST addresses the limitations of fixed-representation models. The work demonstrates that separating the acquisition of evidence (segment selection) from the generation of answers (reasoning) allows for more precise, interpretable, and accurate reasoning, particularly in scenarios involving long sequences and complex, multi-step queries. The authors note that while the approach incurs additional inference costs due to iterative interactions, the trade-off yields substantial accuracy improvements and better utilization of relevant temporal information.