PaperlayerView exact revision on arXiv

Source-linked AI summary

Zero-Shot Time-Series Question Answering via Decoupled Perception and Reasoning

Jing Xie, Haochen Yuan, Yunbo Wang

arXiv:2610.04942v1cs.AI

TL;DR

TSQA methods do not reliably transfer across datasets where numerical inputs, question contexts, and answer requirements shift together. TSHARNESS separates numerical perception from contextual answering through a shared TPS and collaborative agents. Its workflow generalizes from one source dataset to five unseen target benchmarks, while target-side parameter training and reference-answer feedback are not required.

  • Problem

    TSQA methods gain temporal capabilities through numerical-text alignment and task-specific training, but performance gains on one dataset do not reliably transfer to another amid concurrent perceptual and contextual shifts.

    Source: §1 · pp. 1–2

  • Method

    TSHARNESS uses a TPS to separate numerical perception from contextual answering, with PERCEIVER and ANSWERER agents coordinating reusable memory, numerical tools, and active evidence requests.

    Source: §1 · pp. 1–2, §3.3 · p. 4

  • Results

    An analytical workflow constructed exclusively on one source dataset reliably generalizes across five unseen target benchmarks with diverse data distributions and answer specifications.

    Source: §5 · p. 10

  • Takeaways & Limitations

    Source-developed components, prompts, and model parameters remain frozen during transfer, with no target-side parameter training or reference-answer feedback required.

    Source: §3.1 · p. 3

    Perceptual capabilities are constrained by the pre-registered operations in the tool suite.

    Source: §5 · p. 10

Abstract

from arXiv · show

Time-series question answering (TSQA) requires grounding linguistic queries and diverse answer formats in complex numerical observations. However, existing methods heavily overfit to specific datasets and struggle to generalize when input series, question contexts, and answer requirements shift simultaneously. To address this challenge, we propose TSHarness, an agentic framework that establishes a decoupled workflow for cross-dataset zero-shot TSQA. At its core, TSHarness divides and conquers numerical perception and contextual reasoning via a structured Time-Series Perception State (TPS). Guided by a reusable memory of analytical knowledge, a learned Tool Selector adaptively invokes numerical tools to extract salient statistical and temporal features into the TPS. The answering agent then performs semantic reasoning over the TPS to generate target outputs, triggering iterative re-perception through the feedback loop when evidence is deemed insufficient. By separating numerical feature extraction from question-specific reasoning, TSHarness eliminates the need for target-side training or answer feedback, providing a generalizable, cost-efficient foundation for zero-shot TSQA.

1 INTRODUCTION

TSHARNESS addresses cross-dataset zero-shot TSQA by separating numerical perception from question-specific answering through a shared Time-Series Perception State (TPS). Its source-developed agents and tool-selection approach are evaluated across five heterogeneous target datasets without target-side adaptation.

1 INTRODUCTION

  • Existing TSQA methods gain temporal capabilities through numerical-text alignment and task-specific training, but gains on one dataset do not reliably transfer to another.The paper attributes the transfer gap to entangled perceptual and contextual shifts that also undermine parameter and analytical reuse.Source: §1 · p. 1
  • TSHARNESS uses a shared TPS to separate numerical perception from task-specific answer synthesis, with PERCEIVER and ANSWERER roles instantiated by a frozen backbone LLM.The PERCEIVER extracts quantitative temporal properties into the TPS; the ANSWERER uses that state to resolve the question.Source: §1 · p. 1
  • Figure 1 contrasts specialized models’ uneven cross-dataset performance with TSHARNESS’s decoupled perception-and-answering workflow through a unified TPS.The shifts include changes to time-series inputs, question context, and answer requirements.Source: §1 · p. 1, Figure 1 · p. 2
  • Reusable analytical memory guides numerical-tool use, while an active evidence loop lets the ANSWERER request targeted measurements when the TPS is insufficient.A lightweight RL-based TOOL SELECTOR balances answer correctness against execution cost by pruning redundant tool calls while maintaining competitive accuracy.Source: §1 · p. 2
  • TSHARNESS outperforms same-backbone baselines when transferred from one source dataset to five heterogeneous target datasets, without target-side adaptation.The framework is evaluated in a cross-dataset zero-shot setting.Source: §1 · p. 2

2 RELATED WORK

Related work spans language-based and pretrained forecasting, learned time-series-to-language methods, and tool-using agents. TSHARNESS distinguishes itself by structuring evidence passed from numerical perception to answering.

2 RELATED WORK

  • TSQA extends forecasting-related transfer work toward temporal reasoning and contextual sensor interpretation, while recent benchmarks identify cross-domain weaknesses in time-series-adapted LLMs.Learned approaches include ChatTime (Wang et al., 2025a), Time-MQA (Kong et al., 2025), ChatTS (Xie et al., 2025), and TimeOmni-1 (Guan et al., 2026).Source: §2 · p. 2
  • Prior agent frameworks develop tool execution, memory, feedback, and time-series workflows; TSHARNESS focuses on structuring evidence between perception and answering.TPS separates measured series properties into stable fields from task-specific calculations, while a complementary RL selector refines base-tool subsets with the backbone fixed.Source: §2 · p. 3

3 METHOD

TSHARNESS separates numerical perception from question-specific reasoning through a shared Time-Series Perception State (TPS), then uses an active evidence loop to answer cross-dataset questions. Its workflow combines tool-augmented perception with a learned selector that trades answer correctness against tool execution cost.

3.1 FRAMEWORK OVERVIEW

  • TSHARNESS represents each query as series and metadata, question context, and answer specifications, then iteratively updates the TPS before producing an answer.The source-developed components, prompts, and model parameters remain frozen during cross-dataset zero-shot transfer, with target-side training and answer feedback eliminated.Source: §3.1 · p. 3

3.2 PERCEPTION AND TOOL ORCHESTRATION

  • The PERCEIVER proposes eligible numerical tools using series characteristics and question context; validated execution yields structured records with field values, source tools, and evidence identifiers.Tools operate on original numerical arrays, preserving fine-grained temporal indices and analytical field meanings across domains.Source: §3.2 · p. 4

3.3 TIME-SERIES PERCEPTION STATE

  • The TPS combines a reusable static core with question-specific extensions and required evidence types, keeping measured properties distinct from task-dependent analyses.The core records metadata, statistical summaries, temporal primitives, and provenance; extensions retain analysis requests, returned evidence, supporting identifiers, and coverage status.Source: §3.3 · pp. 4–5

3.4 ANSWERING WITH THE HARNESS LOOP

  • The ANSWERER checks evidence coverage, requests targeted re-perception when needed, and can commit an answer using the available state when the request budget is exhausted.Unsupported, duplicate, or out-of-budget requests are rejected, and the output is parsed to match the specified answer requirements.Source: §3.4 · p. 5

3.5 RL-BASED TOOL SELECTION

  • A learned TOOL SELECTOR is trained to balance answer correctness against execution cost, then frozen for zero-shot target evaluations.It uses PPO with a reward that penalizes tool executions as a proxy for analysis overhead.Source: §3.5 · pp. 5–6

4 EXPERIMENTS

TSHARNESS transfers source-developed analytical capabilities to heterogeneous TSQA benchmarks without target-side training or answer feedback. Across experiments, numerical tools, evidence-triggered re-perception, and learned tool selection support accuracy and configurable execution costs.

4.1 EXPERIMENTAL SETUP

  • The study uses TSRBench to construct the harness and five target benchmarks for zero-shot transfer, sampling 412 QA examples per target for agentic comparisons without target-side training, tuning, or answer feedback.Comparisons include task-specialized models and agentic methods such as CoT, ReAct, and TimeClaw, using GPT-5.6-sol and Qwen3.8-27B backbones.Source: §4.1 · pp. 6–7

4.2 MAIN RESULTS

  • TSHARNESS exceeds the strongest Qwen3.8-27B baseline by 13.11 percentage points on TSRBench and by 3.88, 16.00, 7.28, and 0.49 points on Time-MQA, ChatTime, TSAQA, and TimeOmni, respectively.It reaches the highest accuracy on TSRBench and four of five target benchmarks; on CaTSBench, it scores 96.60% versus TimeClaw’s 96.84%.Source: Table 1 · p. 7, §4.2 · p. 7
  • Qwen3.8-27B TSHARNESS improves over zero-shot answering on all five targets, while GPT-5.6-sol improves on four and falls on TimeOmni from 75.73% to 74.76%.On ChatTime, Qwen3.8-27B accuracy rises from 74.13% to 96.10%.Source: §4.2 · p. 7

4.3 ABLATION STUDY OF TOOLS AND MEMORY

  • Removing numerical tools lowers average target accuracy from 74.61% to 63.06%, while removing Memory leaves it at 74.76%; the full harness gains 11.21 points over removing both.On TSRBench, disabling tools drops accuracy by 16.50 points, from 62.62% to 46.12%.Source: Table 2 · p. 8, §4.3 · p. 8

4.4 ANSWER REFINEMENT AND RE-PERCEPTION

  • After re-perception, error corrections exceed regressions on most benchmarks, with net gains of 9.22 points on TSRBench and 7.10 on ChatTime, while TSAQA and TimeOmni decline slightly.Evidence is fulfilled for 95.2% of requesting queries on TSRBench, 62.1% on TSAQA, and 25.5% on TimeOmni; ChatTime triggers no re-perception requests.Source: Figure 3 · p. 8, §4.4 · pp. 8–9

4.5 ANALYSES ON LEARNED TOOL SELECTION

  • Efficiency-oriented selection reduces tool executions by 56.39% with a 2.17% accuracy decrease to 72.54%; accuracy-oriented deployment reaches 74.71% while reducing executions by 1.76%.Efficiency-oriented selection also suppresses the heavy upper tail of source-benchmark tool invocations.Source: Figure 4 · p. 9, §4.5 · p. 9

5 CONCLUSION AND LIMITATIONS

TSHARNESS separates numerical perception from contextual reasoning and transfers a source-developed workflow to five unseen target benchmarks. Its generalization remains bounded by its tool suite, frozen selector policy, and the diversity and task overlap of source data.

5 CONCLUSION AND LIMITATIONS

  • A workflow constructed exclusively on one source dataset generalizes across five unseen target benchmarks with diverse data distributions and answer specifications.The framework coordinates PERCEIVER and ANSWERER roles, tool execution, procedural memory, active evidence requests, and a reinforcement-learning TOOL SELECTOR.Source: §5 · pp. 9–10
  • TSHARNESS’s perceptual capabilities are limited to pre-registered tool operations, and its frozen selector policy leaves test-time adaptation under severe domain shifts unexplored.Zero-shot performance also depends on source-benchmark diversity and task overlap, motivating larger open-world TSQA datasets.Source: §5 · p. 10

A DATASETS AND EVALUATION DETAILS

  • The appendix describes its dataset and evaluation details as supplements to the experimental setup, including QA tasks, input representation, evaluation subsets, and answer scoring.The system configuration is fixed before target evaluation, including model weights, prompts, tools, memory, adapters, and parsers.Source: A DATASETS AND EVALUATION DETAILS · p. 13, §A.2 · p. 13

A.1 BENCHMARK COVERAGE

  • The evaluation covers TSRBench for construction and in-domain evaluation, plus five targets testing transfer across series, contexts, and answer requirements.Table 3 organizes coverage by time-series inputs, QA capabilities, and answer forms.Source: §A.1 · p. 13, Table 3 · p. 13

A.2 INPUT REPRESENTATION AND TRANSFER PROTOCOL

  • Input representation separates observations from question context and answer requirements, while retaining series descriptions, candidate series, and temporal-ordering patch identities where relevant.The source test partition is reserved for in-domain evaluation; target benchmarks are excluded from development and tuning, and their answers are used only for scoring.Source: §A.2 · pp. 13–14

A.3 EVALUATION SUBSETS

The appendices describe evaluation subsets and scoring, then detail how source-developed tools and analytical memory populate the TPS and support evidence updates. The harness freezes its resources before deployment and uses bounded iterative analysis.

A.3 EVALUATION SUBSETS

  • Agentic comparisons use a fixed random sample of 412 questions per benchmark, while cross-evaluation uses the full test sets; CaTSBench image-dependent plot-retrieval questions are excluded from text evaluation.The samples cover different question types, with TimeOmni-1 questions drawn from its OOD split.Source: §A.3 · p. 14

A.4 ANSWER PARSING AND SCORING

  • Scoring parses answers according to each benchmark’s format, requires exact normalized agreement for true/false and temporal-ordering tasks, and averages target-benchmark accuracies without TSRBench.Invalid or unparseable completed responses score zero; percentage-point differences use unrounded accuracies.Source: §A.4 · p. 14

B HARNESS IMPLEMENTATION DETAILS

  • The harness implementation links tool interfaces, TPS construction, and evidence updates to separate numerical measurement from contextual answering.Two successful target-benchmark executions illustrate how numerical findings support contextual judgments.Source: B HARNESS IMPLEMENTATION DETAILS · p. 15

B.1 SOURCE CONSTRUCTION AND FROZEN DEPLOYMENT

  • Source development iteratively inspects batches of 16 questions to configure tools and analytical memory, then freezes those resources for deployment.Calculation-based tool–task pairings require at least two supporting observations and 95% agreement with source references.Source: §B.1 · p. 15

B.2 TOOLS AND ANALYTICAL GUIDANCE

  • The tool library covers statistical and temporal analyses, while eligibility checks and retrieved memory guidance constrain tool plans to applicable operations.Registered tools validate names, arguments, and ranges before execution.Source: §B.2 · p. 15, Table 4 · p. 16

B.3 FROM TOOL OUTPUTS TO TPS

  • TPS fields bind tool measurements to their sources and retain series identity, temporal locations, and question-specific extensions for the Answerer’s contextual judgment.The harness assigns numerical values from tool records, while the LLM selects fields and writes a summary.Source: §B.3 · pp. 16–17, Table 5 · p. 16

B.4 EVIDENCE UPDATES IN THE HARNESS LOOP

  • When evidence is insufficient, the Answerer can request further analysis; the harness validates and executes the request, attaches sourced results, and updates TPS coverage.Runtime limits bound tool calls, evidence requests, decision steps, and raw-series views.Source: §B.4 · pp. 17–18

B.5 SUCCESSFUL EXECUTIONS ON TARGET BENCHMARKS

Two target-benchmark examples show the harness using measured evidence for categorical and ordered answers. The accompanying Tool Selector details describe its policy, deployment constraints, and correctness–cost optimization.

B.5 SUCCESSFUL EXECUTIONS ON TARGET BENCHMARKS

  • In the ChatTime example, a −2.48 adjacent drop and a detected downward level shift at index 32 support the Answerer’s selection of “level shift.”The example separates locating and measuring a change from interpreting its type under the question’s definitions.Source: §B.5 · p. 18
  • A 0.642 boundary cost ranks patch order B,C,A,D ahead of the next candidate at 1.916, leading the Answerer to revise its preliminary order to the correct sequence.The extension sums adjacent-patch boundary costs, including the join from the given first patch.Source: §B.5 · p. 18

C.1 POLICY INPUTS AND NETWORK

  • The Tool Selector scores candidate operations using question, tool, plan, and series features, then adjusts a Perceiver-derived inclusion prior with a shared multilayer perceptron.The text features contribute 384 dimensions and 12 scalar features complete the policy input.Source: §C.1 · pp. 18–19

C.2 PERCEIVER PRIOR AND SUBSET SAMPLING

  • The policy samples nonempty tool subsets during training; efficiency-oriented deployment executes tools with inclusion probability at least 0.5, falling back to the highest-probability tool if needed.Retained tools preserve Perceiver-provided arguments and order.Source: §C.2 · p. 19

C.3 DEPLOYMENT CONFIGURATIONS

  • Accuracy-oriented deployment moderates learned adjustments and preserves minimum proposed evidence, with both deployment rules fixed before target evaluation and without target-side tuning or answer feedback.The matched comparison holds evidence presentation and answering procedures fixed; comparisons across configurations reflect deployment choices rather than one isolated selector parameter.Source: §C.3 · p. 19

C.4 ROLLOUT REWARD AND OPTIMIZATION

  • The selector is trained on rollout-level correctness–cost rewards with a tool-execution penalty, while KL and entropy regularizers constrain candidate inclusion distributions relative to the Perceiver prior.Each training question generates four complete harness rollouts.Source: §C.4 · p. 20, Table 7 · p. 20

C.5 CHECKPOINT SELECTION AND MATCHED EVALUATION

The section reports source-development checkpoint selection and a matched efficiency comparison, then describes target evaluations and supplementary breakdowns. Selector accuracy is slightly below the no-selector Controller in the reported TSRBench comparison.

C.5 CHECKPOINT SELECTION AND MATCHED EVALUATION

  • The final checkpoint is selected by source-development accuracy first and mean correctness–cost reward second, with earlier checkpoints preferred on ties.The reported deployment uses that checkpoint’s deterministic inclusion rule.Source: §C.5 · p. 20
  • 61.89% for the Controller versus 59.95% for the RL selector in the efficiency-oriented matched TSRBench comparison.The comparison shares the eligible tools, frozen backbone, Memory, extension rules, and evidence-request budget; the 62.62% harness result belongs to a separate main evaluation.Source: §C.5 · pp. 20–21

D SUPPLEMENTARY EXPERIMENTAL RESULTS

  • The efficiency-oriented target evaluation freezes the selected checkpoint and source-developed Tool and Memory resources, while supplementary results examine task categories, answer revisions, and tool-execution distributions.The target comparison covers five benchmarks; the accuracy-oriented comparison uses the same five with a matched no-selector control.Source: §C.5 · p. 21, D SUPPLEMENTARY EXPERIMENTAL RESULTS · p. 21

D.1 PERFORMANCE ACROSS CAPABILITIES AND TASK TYPES

The reported breakdowns show gains across source capabilities and several target question types, but not uniformly across categories or backbones. Improvements span temporal, relational, structured, and contextual judgments.

D.1 PERFORMANCE ACROSS CAPABILITIES AND TASK TYPES

  • On TSRBench, the Qwen-based harness improves all four capability groups, with reasoning rising from 36.26% to 65.50% and decision-making from 25.40% to 50.79%.The GPT-based harness improves reasoning, perception, and decision-making, while prediction decreases from 61.11% to 57.41%.Source: §D.1 · p. 21
  • On ChatTime, Qwen seasonality recognition rises from 60.70% to 91.20% and volatility recognition from 60.40% to 97.00%; both harnesses reach 100% on trend questions.GPT seasonality recognition increases from 66.70% to 91.10%.Source: §D.1 · p. 21
  • On TSAQA, Qwen temporal relationships rise from 13.24% to 45.59% and data transformation from 39.71% to 55.88%; on CaTSBench, peak timing rises from 70.59% to 91.18%.CaTSBench volatility comparison increases from 76.47% to 88.24%.Source: §D.1 · p. 21
  • On TimeOmni-1, Qwen scenario understanding rises from 74.72% to 83.71%, while GPT falls from 97.19% to 94.94%; improvements are not uniform across categories.The passage identifies the GPT result as illustrating the dependence of transfer gains on the backbone’s initial capability.Source: §D.1 · p. 22

D.2 ANSWER REVISIONS AFTER EVIDENCE INTEGRATION

The section compares initial hypotheses with final answers after TPS evidence integration, including evidence-request frequency. Qwen gains on every evaluated benchmark, while GPT’s revisions yield both gains and small losses.

D.2 ANSWER REVISIONS AFTER EVIDENCE INTEGRATION

  • GPT requests additional evidence on 19.90% of TSRBench questions and 48.54% of TimeOmni-1 questions, whereas Qwen mainly answers from the initial TPS.Revisit records questions with an accepted evidence request.Source: §D.2 · p. 22, Table 10 · p. 23
  • Qwen’s final accuracy rises from 44.17% to 62.62% on TSRBench, 74.30% to 96.10% on ChatTime, and 52.18% to 59.47% on TSAQA.The comparison scores each trajectory’s pre-TPS hypothesis and final prediction using the same benchmark parser.Source: §D.2 · p. 22
  • The initial hypothesis is a within-trajectory reference, distinct from the independently evaluated zero-shot baseline.Table 10 reports initial and final accuracy in percent, net change in percentage points, and revisit frequency.Source: §D.2 · p. 22, Table 10 · p. 23
  • GPT gains on TSRBench, Time-MQA, ChatTime, and CaTSBench, but revisions on TSAQA and TimeOmni-1 yield small net losses.These results characterize the complete evidence-integration process and do not isolate TPS representation or additional loop iterations.Source: §D.2 · p. 24

D.3 EFFICIENCY-ORIENTED EXECUTION ACROSS STAGES AND CAPABILITIES

The efficiency-oriented selector reduces tool executions, with savings concentrated in base perception and task-dependent accuracy effects across capabilities. The comparison concerns analytical tool use, not end-to-end runtime.

D.3 EFFICIENCY-ORIENTED EXECUTION ACROSS STAGES AND CAPABILITIES

  • Base-perception executions fall from 8.97 to 1.88, driving the reduction from 12.25 to 5.34 executions per question.Question-specific tools remain at 0.62 executions, while planned extensions rise from 2.67 to 2.84; later evidence acquisition remains active.Source: §D.3 · p. 24, Figure 5 · p. 24
  • Decision-making and reasoning retain accuracies of 47.62% and 66.67%, respectively, with more than 55% fewer executions, while perception and prediction accuracy decrease by 4.29 and 4.63 percentage points.Within prediction, numerical forecasting rises from 45.83% to 50.00%, while event forecasting falls from 69.44% to 47.22%.Source: §D.3 · p. 24, Figure 6 · p. 25
  • The reduction occurs on 411 questions and remains unchanged on one; the median execution count falls from six to three.Logical model-call totals are 3,393 and 3,392, and end-to-end runtime also depends on operation runtimes and LLM inference.Source: §D.3 · pp. 24–25
  • Tool executions fall by 56.39% while accuracy decreases from 61.89% to 59.95%; executions decrease in all four capability groups.The paired-bootstrap 95% interval for the execution reduction is 55.41–57.39%.Source: §D.3 · p. 24, Figure 6 · p. 25

D.4 SELECTOR OPERATING POINTS AND TRANSFER SCOPE

The section compares accuracy- and efficiency-oriented deployment settings, with target accuracy and tool cost measured separately. The settings show different accuracy–cost balances, but the comparison evaluates complete deployment configurations.

D.4 SELECTOR OPERATING POINTS AND TRANSFER SCOPE

  • The accuracy-oriented and efficiency-oriented configurations achieve mean target accuracies of 74.71% and 72.54%, respectively.The mean uses five target benchmarks and excludes source benchmark TSRBench.Source: Table 11 · p. 25, §D.4 · p. 25
  • Accuracy-oriented deployment retains 98.24% of matched Controller executions, while efficiency-oriented deployment retains 43.61%.Tool cost pools questions across the five target benchmarks; accuracy is the unweighted mean of their benchmark scores.Source: Table 12 · p. 25, §D.4 · p. 26
  • The five settings provide different accuracy–cost balances, but inference rules and evidence processing vary, so the comparison does not isolate a selector parameter.More conservative evidence acquisition sustains higher accuracy, while stronger filtering reduces tool execution.Source: §D.4 · p. 26
Loading 2610.04942v1…