acceptodds
Under review as a conference paper at ICLR 2027

AgenTS: Harnessing Runtime Tool Orchestration for Time Series Reasoning

Abstract

Time series reasoning involves understanding temporal patterns and reasoning over them to answer complex questions across domains such as healthcare, energy, and climate science. Large language models often struggle to abstract relevant numerical structure from long and noisy time series. Analytical tools can support this abstraction, but their heterogeneous capabilities and output structures require coordination according to the question and input data. To address these challenges, we propose AgenTS, a runtime harness that coordinates task analysis, tool configuration, execution, and evidence preparation for time series reasoning. The core idea is to disassemble inference into tool-based quantitative abstraction and evidence-based reasoning, making the abstraction process an explicit target of runtime control. Specifically, a Profiler constructs a per-instance runtime state that jointly characterizes task relevance and input series properties. An Orchestrator uses this state to construct and select ordered tool configurations under a utility criterion balancing task coverage, question relevance, task–data capability match, and computational cost. Procedural memory supports configuration reuse across similar instances, reducing repeated configuration generation. A Prober executes the selected configurations and aggregates their heterogeneous outputs into structured evidence for subsequent model reasoning, with final answers obtained through multi-sample voting. Experiments on two benchmarks covering 14 datasets and 10 reasoning tasks demonstrate strong performance, with an average improvement of more than 30% across tasks while using 33% fewer tools than the strongest baseline.

Primary area:
learning on time series and dynamical systems
Keywords:
time series reasoning, runtime harness, tool orchestration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_9c1pOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AgenTS, a runtime harness for time series reasoning that separates tool-based analysis from final model reasoning. A Profiler combines task relevance with temporal-property estimates; an Orchestrator selects cost-aware toolpaths and reuses paths through an online memory; a Synthesizer executes tools and reconstructs a structured evidence workspace. Experiments on TSR-Suite and MTBench compare several backbones and tool-using agents, with component ablations and tool-call/latency measurements. The reported results generally improve question-answering accuracy and forecasting MAE while using fewer tools.

Critical strengths

The harness addresses a concrete failure of tool-using models: accumulating noisy execution traces and repeatedly choosing tools can hurt reasoning. Workspace reconstruction is a simple, useful solution.

The evaluation combines multiple backbones and benchmarks

Critical weaknesses

are the gains coming mainly from better tool management, from powerful tools, or from both? The paper’s current comparisons do not separate these explanations clearly enough.

the online memory depends on ground-truth feedback.

Many fixed rules control the system but the paper dow not show whether its results remain strong when those rules are changed. (there are countless examples through out the paper: workspace confidences, path-ranking weights [Coverage 0.43, relevance 0.14, demand match 0.43, cost 0.05] and so on)

Questions for authors

do all agent baselines have the same set of tools or their tools differ?

are the examples used for memory warm-up randomly selected or carefully curated?

Reviewer_RZ6cOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AgenTS, a runtime tool orchestration framework designed for time-series reasoning with Large Language Models (LLMs). The framework consists of three core components. Profiler: Constructs a runtime context state by integrating query task types with statistical properties extracted from input time series. Orchestrator: Dynamically selects optimal toolpaths based on the Profiler's runtime state. Synthesizer: Executes selected tools, deduplicates shared execution calls, and aggregates tool outputs into structured evidence to guide final answer generation. Empirical evaluations show that AgenTS achieves state-of-the-art performance, achieving over 30% average improvement across diverse benchmark tasks while reducing tool invocation frequency by 33% compared to the strongest baseline.

Critical strengths
  1. Context-Aware Tool Orchestration: The framework introduces a principled orchestration mechanism that jointly factors in problem requirements and data-level statistical characteristics, significantly enhancing tool utilization efficiency for time-series agentic workflows.
  2. Trace Efficiency and Deduplication: AgenTS effectively minimizes redundant tool execution traces through shared-tool deduplication and structured evidence synthesis, streamlining inference efficiency.
Critical weaknesses
  1. Potential Data Leakage via Memory Updates: The paper states that memory starts empty at deployment and accumulates experience sequentially instance by instance. However, memory updates depend on ground-truth answers or future target values. It is unclear whether feedback during main experiments is drawn from an independent calibration split or sequentially from the evaluation stream itself. If test set ground truths are used for online updates without explicit delay, this raises concerns regarding potential data leakage prior to evaluation against baselines.
  2. Missing Comparison with Key Contemporary Baselines: AgenTS is compared against standard or earlier baselines (ReAct, TS-Reasoner, TS-Agent). However, the evaluation lacks empirical comparisons with state-of-the-art time-series agent architectures such as TimeART (cited in the paper but omitted from experimental evaluation) and TimeClaw (which similarly addresses runtime environment design and tool usage in time-series settings). Including these closely related baselines would render the empirical claims significantly more convincing.
  3. Underspecified Multivariate Processing Pipeline: Although the framework claims support for multivariate time-series data, the explicit mechanisms for handling multi-channel inputs are under-detailed. Specifically, the paper does not explain how Profiler features are aggregated across channels or whether the MTBench evaluations are conducted in a univariate or multivariate setting.
Questions for authors
  1. Memory Feedback Source & Evaluation Setup: What is the exact source of the correctness feedback used for memory updates in the primary experiments? If test set ground-truth labels are leveraged sequentially, when do these labels become available, and are baseline methods provided identical access to online feedback?
  2. Tool Library Parity Across Baselines: Do the baseline methods (ReAct, TS-Agent, TS-Reasoner) have access to the exact same tool set, implementations, and prompt descriptions as AgenTS? Clarifying this parity is critical to confirm that performance gains stem from orchestration strategy rather than asymmetric tool capabilities.
  3. Discrepancy in Tool Invocation Metrics: What accounts for the discrepancy in per-sample tool invocation counts reported in Figure 2 versus Table 4? Are these measurements based on distinct experimental settings, subsets, or evaluation protocols?
  4. Multivariate Implementation & MTBench Details: How does the Profiler aggregate statistical features across multiple series/channels in multivariate scenarios? Additionally, was the MTBench evaluation conducted using multivariate time-series inputs?
Reviewer_kYSiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AgenTS, a runtime harness for tool-augmented time series reasoning. The central idea is to decouple analytical tool execution from language model reasoning and provide the model with structured temporal evidence rather than accumulated tool interaction histories.

Critical strengths

The paper addresses an important challenge in agentic time series reasoning: coordinating statistical tools without accumulating excessive intermediate context or repeatedly performing unnecessary tool planning. Separating analytical computation from language model reasoning is a reasonable design choice that could improve the reliability of tool-augmented systems.

The paper also includes ablations of runtime state construction, toolpath planning, online memory, and workspace reconstruction. It also reports tool usage and inference time, which are relevant for evaluating practical agentic systems. These experiments provide useful evidence that the proposed components can affect downstream performance.

Critical weaknesses
  1. Potential evaluation leakage in the online learning memory. My most significant concern is the validity of the online learning protocol. In Section 3.3, Eq. (7)–(8) updates toolpath quality scores using outcomes obtained from final predictions. Appendix C.3 further specifies that the outcome for multiple-choice tasks depends on whether the predicted answer matches the ground-truth answer, while forecasting outcomes depend on prediction MAE.

  2. Limited methodological novelty relative to existing agentic tool orchestration and time series reasoning frameworks. Although the framework combines several useful components, the individual mechanisms appear relatively straightforward.

  3. Task-specific analytical tools may account for a substantial portion of the reported improvements.

Questions for authors
  1. How is ground-truth feedback obtained for online memory updates during benchmark evaluation?

  2. How much of the performance improvement comes from the specialized analytical tools rather than the proposed orchestration strategy?

  3. What is the advantage of the proposed online memory over simpler caching or retrieval strategies?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.