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.
Figures & tables
Figure 1: Motivation and overview of TSHarness . (a) Specialized TSQA models exhibit uneven cross-dataset performance as time-series inputs, question context, and answer requirements change. (b) TSHarness decouples perception from answering through a unified TPS representation.
Figure 2: Workflow of TSHarness . Guided by procedural memory, the Perceiver populates the TPS with grounded measurements, refined by the Tool Selector . The Answerer interprets the accumulated evidence alongside the question, actively initiating an iterative feedback loop to request further evidence. Source-developed resources are completely frozen for zero-shot transfer.
Method
Base LLM
TSRBench
Target Benchmarks
Time-MQA
ChatTime
CaTSBench
TSAQA
TimeOmni
Specialized Time-Series Models
ChatTime
LLaMA-2-7B
19.95
28.79
88.77
9.63
11.36
19.66
Time-MQA
Qwen2.5-7B
30.47
54.98
37.07
26.05
45.07
32.04
Llama3-8B
20.95
54.58
33.34
51.27
40.30
33.25
Mistral-7B
36.27
55.91
35.47
47.79
42.61
31.55
Table 1: Cross-dataset zero-shot TSQA evaluation (Accuracy, %). TSHarness is developed exclusively on the source benchmark (TSRBench) and deployed for zero-shot evaluation on five unseen target benchmarks. TSHarness ⋆ incorporates the learned RL-based Tool Selector . Shaded orange cells indicate in-domain training for ChatTime and Time-MQA, directly contrasting with the strictly zero-shot deployment of TSHarness. The best results within each backbone group are highlighted in bold ; second-best results within the open-weight agentic group are underlined .
Variant
In-domain
Zero-shot
TSRBench
Time-MQA
ChatTime
CaTSBench
TSAQA
TimeOmni
Full TSHarness
62.62
65.78
96.10
96.60
59.47
55.10
w/o Tool
46.12
60.92
63.59
91.02
47.33
52.43
w/o Memory
60.44
68.45
96.12
96.12
58.74
54.37
w/o Tool + Memory
42.48
60.92
63.35
91.02
48.79
52.91
Table 2: Ablation study of core components across in-domain and zero-shot target datasets (Accuracy, %). We evaluate variants of TSHarness by disabling numerical tools and procedural memory using Qwen3.8-27B as the backbone LLM. Numerical tool orchestration proves indispensable for both in-domain and transfer accuracy. Best results are shown in bold .
Figure 3: Analysis of answer refinement and active re-perception dynamics. (a) Transitions from initial pre-TPS hypotheses to final verified answers across benchmarks. Diamond markers represent net accuracy changes in percentage points. (b) Evidence fulfillment rates among queries triggering the re-perception loop (%). Fulfilled and Incomplete indicate acquired evidence with full and partial request coverage, respectively. Gray indicates that no additional evidence was obtained through re-perception.
Figure 4: Deployment choices for learned tool selection. (a) Mean accuracy and total tool-execution ratios across all five target benchmarks (matched Controller = 100%). All five points are measured configurations; lines connect adjacent points. (b) The separate TSRBench cumulative distribution under efficiency-oriented deployment.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Benchmark
Time series and context
QA capabilities
Answer forms
TSRBench
Single or multiple series with domain context
Perception, reasoning, prediction, and decision-making
Multiple choice
ChatTime
Synthetic series of varying lengths
Trend, volatility, seasonality, and outlier recognition
Multiple choice
Time-MQA
Real-world series with contextual descriptions
Statistical properties, patterns, anomalies, and structural changes
True/false; multiple choice
CaTSBench
Series paired with textual descriptions
Caption and series retrieval, description consistency, and pairwise comparison
True/false; multiple choice
TSAQA
Cross-domain series and temporal patches
Classification, anomalies, characterization, comparison, transformations, and temporal relationships
True/false; multiple choice; ordered labels
TimeOmni-1
Series with event and domain context
Causality discovery, scenario understanding, and decision-making
Multiple choice
Appendix
Table 3: Time-series inputs, QA capabilities, and answer forms covered by the evaluation.
Tool family
Computation
Returned information
Statistics and value profile
Summary statistics, quantiles, extrema, and adjacent differences
Values, extrema locations, largest rise and drop
Trend
Linear fit on normalized time; slope scaled by series standard deviation
Direction, strength, normalized slope
Seasonality
Autocorrelation and spectral analysis
Period and periodicity strength
Volatility
Early/late standard deviations and their ratio
Variability level and change
Anomalies and change points
Median-deviation and local-residual scores; mean and variance changes across windows
Event positions, types, directions, and scores
Cross-series
Contemporaneous and lagged correlations
Channel pairs, correlations, and strongest lag
Appendix
Table 4: Representative tools used to construct TPS fields. Measurements and temporal locations are computed from the input observations.
Component
Contents and role
Statistics Sj
Distributions and value profiles, including means, standard deviations, extrema, and their locations.
Temporal features Fj
Trends, periodicity, volatility, segments, changes, anomalies, and cross-series relationships.
Task extensions Dr
Named analyses that add evidence required by the question beyond the common fields.
Appendix
Table 5: TPS field groups and their analytical roles.
LLM role
Inputs
Outputs
Perception planning
Bounded series view, question context, eligible tools, Memory summaries, and execution limits.
Tool names and arguments; a Memory query and retrieval choices.
TPS synthesis
Tool evidence, retrieved Memory entries, series identity, channels, length, and question context.
Measured fields to include and a factual summary hj describing the series.
Initial hypothesis
Raw numerical series, question context, answer choices, and output requirements.
A parsed candidate answer and its generated response, stored as an initial hypothesis in a named extension.
Initial extension planning
Core findings, question, evidence requirements, and available tools and Memory.
Analysis requests specifying operations, arguments, and required evidence.
Answerer
Question and answer requirements, TPS with the provisional hypothesis, evidence coverage, available operations, and request budget.
A task-formatted prediction with evidence references, or a request specifying the additional evidence needed.
Appendix
Table 6: Inputs and outputs of the LLM roles connecting Perception, TPS, and the Answerer. Tool execution and numerical field assignment are performed by the harness between calls.
Setting
Value
Source training / development questions
96 / 48
Training passes
1
Rollouts per question / questions per batch
4 / 4
Text-hash dimension / scalar features
128 per block / 12
Policy network
396→64→1 , tanh
Perceiver prior: proposed / other tools
0.8 / 0.2
Appendix
Table 7: Configuration of the learned Tool Selector in the matched base-tool experiment. The LLM and source-developed Tool and Memory resources remain fixed during policy optimization.
Qwen3.8-27B
GPT-5.6-sol
Capability / task
Zero-shot
TSHarness
Δ
Zero-shot
TSHarness
Δ
TSRBench
Decision-making
25.40
50.79
+25.40
44.44
53.97
+9.52
Perception
68.57
71.43
+2.86
82.86
87.14
+4.29
Prediction
50.00
59.26
+9.26
61.11
57.41
-3.70
Reasoning
36.26
65.50
+29.24
71.93
81.87
+9.94
Appendix
Table 8: Accuracy (%) across source capabilities and target question categories, using the matched same-backbone predictions underlying Table 1 . TSRBench is grouped by capability, ChatTime and TimeOmni-1 by task, and Time-MQA by answer format. Δ is TSHarness minus zero-shot in percentage points; bold indicates the higher score within each pair.
Qwen3.8-27B
GPT-5.6-sol
Capability / task
Zero-shot
TSHarness
Δ
Zero-shot
TSHarness
Δ
CaTSBench
Caption retrieval: cross-domain
100.00
100.00
0.00
100.00
100.00
0.00
Caption retrieval: perturbed
100.00
100.00
0.00
100.00
100.00
0.00
Caption retrieval: same-domain
100.00
100.00
0.00
100.00
100.00
0.00
Paraphrase consistency
100.00
100.00
0.00
100.00
100.00
0.00
Appendix
Table 9: Task-level accuracy (%) on CaTSBench and TSAQA. All evaluated task categories are included. Both conditions use the same questions within each benchmark. Bold and Δ follow Table 8 .
Benchmark
Initial Accuracy (%)
Final Accuracy (%)
Δ (pp)
Revisit (%)
Qwen3.8-27B
TSRBench
44.17
62.62
+18.45
0.00
Time-MQA
59.47
65.78
+6.31
0.24
ChatTime
74.30
96.10
+21.80
0.00
CaTSBench
93.45
96.60
+3.16
0.00
TSAQA
52.18
59.47
+7.28
0.24
Appendix
Table 10: Accuracy before and after TPS evidence integration within the same harness trajectory. Initial Accuracy scores the recorded pre-TPS hypothesis; Final Accuracy scores the final prediction. Both are reported in percent. Δ is the net accuracy change in percentage points (pp), computed before rounding. Revisit gives the percentage of questions with an accepted evidence request. The initial hypothesis is a within-trajectory reference, distinct from the independently evaluated zero-shot baseline. Time-MQA uses the same subset as Table 8 .
Figure 5: Tool execution breakdown for the matched Controller and RL conditions. (a) Mean executions at each stage. (b) Mean executions by tool across all stages, ordered by Controller usage; the seven question-specific analytical and verification tools are pooled as Question tools. Both panels use executions per question; requested extensions average zero for Controller and less than 0.01 for RL.
Figure 6: Execution–accuracy trade-off across TSRBench capabilities. (a) Mean tool executions per question across all stages. (b) Accuracy on the same questions. The selector reduces execution demand in every group while preserving accuracy in decision-making and reasoning. Colors identify the same policies in both panels.
Benchmark
Accuracy-oriented
Efficiency-oriented
Time-MQA
65.78
63.67
ChatTime
96.12
94.90
CaTSBench
96.84
93.69
TSAQA
59.71
57.04
TimeOmni
55.10
53.40
Five-target mean
74.71
72.54
Appendix
Table 11: Target accuracy (%) for the two deployment regimes. Accuracy-oriented values correspond to TSHarness ⋆ in Table 1 . Both regimes reuse the same learned checkpoint.
Configuration
Accuracy (%)
Relative cost (%)
Tool savings (%)
Efficiency-oriented
72.54
43.61
56.39
Intermediate 1
73.62
57.43
42.57
Intermediate 2
74.10
74.28
25.72
Intermediate 3
74.55
85.23
14.77
Accuracy-oriented
74.71
98.24
1.76
Appendix
Table 12: Measured accuracy–cost trade-offs across all five target benchmarks, plotted in Figure 4 (a). Tool savings equal 100% minus relative execution cost; each cost is normalized to its matched Controller.
Recent advances in large language models (LLMs) have given rise to time-series question answering (TSQA), which formulates time-series analysis as natural-language question answering. However, directly feeding raw numerical series into LLMs suffers from a tokenization bottleneck: Byte Pair Encoding fragments continuous values into unstable tokens whose embeddings lack meaningful metric structure, resulting in the loss of magnitude, scale, and trend information. Prior methods use patch-based encoders that split the series into fixed windows, locking in one granularity that breaks patterns and hides exact timesteps, through a separate module that rarely transfers across datasets with different lengths or sampling rates. To address this challenge, we propose CADE (Contrastive Alignment with Direct Embedding), a novel framework for TSQA built upon two key components: direct timestep embedding and semantic alignment. The proposed framework maps each timestep directly into the LLM embedding space through a point-wise linear encoder and MLP projector, preserving exact index-level access while eliminating the need for patching and padding. To further bridge the semantic gap between time-series and language representations, we introduce a novel one-directional supervised contrastive loss that aligns time-series embeddings with frozen class-name text anchors. Experimental results on the public Time-MQA benchmark demonstrate that our framework consistently improves performance across six TSQA tasks, outperforming both open-source and proprietary LLM baselines.
Large language models (LLMs) can struggle with time-series question answering (TS-QA), especially when numerical signals are serialized as text and require explicit computation. Tool-augmented approaches improve performance, but existing systems often intertwine language reasoning, computation, and perception, making it difficult to determine which components drive the gains. We present T-SMART, a neurosymbolic framework that separates these roles: a frozen LLM interprets questions and selects operations, deterministic tools perform numerical computation, and structured perception is invoked only when needed. Controlled paired ablations show that deterministic computation provides the dominant benefit, improving accuracy by 31.7 percentage points over direct LLM reasoning on serialized time series, while language understanding and perception offer smaller complementary gains. These results indicate that tool-augmented TS-QA benefits primarily from reliable numerical execution rather than additional language-model reasoning and provide a controlled framework for analyzing component contributions in neurosymbolic time-series systems.
Ivan Delgado, Himansi Gupta, Bishal Khatri +4
University of California, San Diego, CA, USA · West Virginia University, Morgantown, WV, USA
Time series reasoning demands both the perception of complex dynamics and logical depth. However, existing LLM-based approaches exhibit two limitations: they often treat time series merely as text or images, failing to capture the patterns like trends and seasonalities needed to answer specific questions; and when trained on a mix of simple and complex tasks, simpler objectives often dominate the learning process, hindering the development of deep reasoning capabilities. To address these limitations, we propose the Pattern-Aware Alignment and Balanced Reasoning model (PATRA), introducing a pattern-aware mechanism that extracts trend and seasonality patterns from time series to achieve deep alignment. Furthermore, we design a task-aware balanced reward to harmonize learning across tasks of varying difficulty, incentivizing the generation of coherent Chains of Thought. Extensive experiments show that PATRA outperforms strong baselines across diverse Time Series Question Answering (TSQA) tasks, demonstrating superior cross-modal understanding and reasoning capability.
Junkai Lu, Peng Chen, Xingjian Wu +4
East China Normal University, Shanghai, China · Aalborg University, Aalborg, Denmark