cs.LGNov 7, 2025
SaveQuAnTS: Question Answering on Time Series
Organizations: AI & ML Lab, TU Darmstadt · University of New South Wales · MBZUAI, Abu Dhabi · Hessian Center for AI (hessian.AI) ·
Abstract
Text offers intuitive access to information. This can, in particular, complement the density of numerical time series, thereby enabling improved interactions with time series models to enhance accessibility and decision-making. While the creation of question-answering datasets and models has recently seen remarkable growth, most research focuses on question answering (QA) on vision and text, with time series receiving minute attention. To bridge this gap, we propose a challenging novel time series QA (TSQA) dataset, QuAnTS, for Question Answering on Time Series data. Specifically, we pose a wide variety of questions and answers about human motion in the form of tracked skeleton trajectories. We verify that the large-scale QuAnTS dataset is well-formed and comprehensive through extensive experiments. Thoroughly evaluating existing and newly proposed baselines then lays the groundwork for a deeper exploration of TSQA using QuAnTS. Additionally, we provide human performances as a key reference for gauging the practical usability of such models. We hope to encourage future research on interacting with time series models through text, enabling better decision-making and more transparent systems.
Figures & tables
Figure 1: Text offers a highly intuitive medium of communication to interact with otherwise opaque multivariate time series.
| QA Pairs | Questions | Answ. | Time Series | ||||||||||
| Dataset | Scope | Level | #Total | #Unique | #Unique | Types | Dist. | b | m | o | #Unique | 0,00 T V | Human Eval. |
| OppQA-1800 1 | scene | high | 112,749 | 0 89,735 | 0 84,389 | 0 12 | balanced | ✓ | ✓ | ✗ | 0 1,362 | 1,800 77 | ✗ |
| ECG-QA | scene | high | 414,348 | 112,455 | 0 34,767 | 00 7 | balanced | ✓ | ✓ | ✗ | 16,052 | 5,000 12 | ✗ |
| TS Evol-Instr. | mixed | low | 0 44,802 | 0 44,516 | 0 43,973 | 0 10 | skewed | ✗ | ✗ | ✓ | 0 4,127 | 00 varying | ✗ |
| (Tune-)OpenSQA | local | low | 183,304 | 172,379 | 183,042 | n/a | skewed | ✗ | ✗ | ✓ | 35,683 | 00 varying | ✗ |
| QuAnTS (ours) | scene | high | 150,000 | 125,726 | 100,568 | 0 45 | balanced | ✓ | ✓ | ✓ | 30,000 | 0, 320 72 | ✓ |
Table 1: The QuAnTS dataset is large and varied. It guarantees diversity through the balanced distribution of 45 question and three answer types: b inary, m ultiple-choice, and o pen. The shape of the time series is specified in terms of the number of time steps and variates . QuAnTS provides human performances for reference.
Figure 2: QuAnTS is generated in several steps: An action sequence is sampled ➀, where for each we sample five question and answer types ➁. For diversity, each of them is then instantiated from a sampled template ➂. The time series from the human motion diffusion ➃ is then combined with the QA-pair and auxiliary data ➄. Example QA pairs are shown below. Dice ( ) indicate randomized operations to enhance dataset diversity.
Figure 3: The hierarchy of question and answer types we identified. The blue types are the most fundamental to time series QA and, therefore, included in QuAnTS. These tasks alone already make for a very challenging dataset, as we will see in Section 4 . The gray ones are left for future extensions.
Figure 4: QuAnTS’ question and open answer lengths are diverse.
| System | Accuracy (↑) | Precision (↑) | Recall (↑) | F1 (↑) | |
| Binary | Ablation: Only Question | 64.53% | 65.32% | 64.70% | 64.22% |
| Ablation: Only TS | 49.21% | 50.15% | 50.00% | 33.62% | |
| Humans ( ) | 77.50% | 81.08% | 74.66% | 77.74% | |
| Naive: TS + Question | 62.94% | 64.54% | 63.19% | 62.16% | |
| ChatTS | 51.69% | 55.38% | 25.36% | 34.79% | |
| xQA-Llama on GT | 79.44% | 84.91% | 72.42% | 78.17% |
Table 2: Dataset variants and ablations show that QuAnTS is well-formed. Furthermore, humans can successfully solve the binary and multi splits. The Naive Baseline (except for open) and ChatTS struggle to solve most splits of the QuAnTS dataset, whereas xQA successfully leverages the additional information. Human, ChatTS, and xQA rows provide reference scores under their respective evaluation setups; direct comparison is only possible within the sections marked by the gray dividers ( — ). indicates the total number of human responses. GT denotes ground-truth action labels and AE a supervised action encoder using pre-cut segments. More details on humans and xQA are provided in Section D.2 .
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: QuAnTS features a diverse and balanced question and answer types distribution. Some categories, such as comparison_counting_open, occur less often because of the particular circumstances that allow them to be posed without trivial responses.
| Data | Answer Type | # Samples | Eff. # Samples | Eff. Samples |
| Question + Answer | Binary | 61,804 | 45,253 | 73.220% |
| Multi | 43,599 | 41,021 | 94.087% | |
| Open | 44,597 | 33,194 | 74.431% | |
| Question + Answer + Action Labels | Binary | 61,804 | 61,791 | 99.979% |
| Multi | 43,599 | 43,598 | 99.998% | |
| Open | 44,597 | 44,540 | 99.872% |
Table 3: Effectively all combinations of question-answer pairs and action sequence labels are distinct. Furthermore, the probabilistic motion generation did not produce a single time series duplicate. All checks were performed using exact matching.
Figure 6: Effectively all generated numerical motion time series are unique. At , the effective dataset size under soft equality is 29,895 (99.65% of the full 30,000).
| Model | Temperature | Top- | Top- | Repetition Penalty |
| Qwen3 4B/8B (Reasoning) | 0.6 | 0.95 | 20 | 1.5 |
| Qwen3 4B | 0.7 | 0.8 | 20 | 1.5 |
| Llama 3.2 3B (Instruct) | 0.8 | 0.8 | 20 | 1.5 |
Table 4: Sampling parameters used across the four LLM judge candidates.
| Model | Avg. (↓) | Pearson (↑) | Spearman (↑) | Kendall (↑) |
| Qwen3 8B (AWQ, Reasoning) | 0.149 | 0.912 | 0.927 | 0.878 |
| Qwen3 8B (AWQ) | 0.254 | 0.843 | 0.857 | 0.794 |
| Qwen3 4B (AWQ) | 0.155 | 0.883 | 0.891 | 0.853 |
| Llama 3.2 3B (Instruct) | 0.430 | 0.640 | 0.654 | 0.603 |
Table 5: Qwen3 8B (AWQ, Reasoning) aligns best with human judgement. Highlighted bold is the best value per column, with the arrows (↓/↑) indicating which direction is desired. Average over 10 seeds.
| System | Answer Type | BF16 | NF4 | ||
| PPL (↓) | PPL/base | PPL (↓) | PPL/base | ||
| Base | — | 5.61 | 1.00 | 5.97 | 1.00 |
| Ablation: Only Question | Binary | ||||
| Multi | 9870.98 | 1758.13 | |||
| Open | 19.24 | 3.43 | 25.65 | 4.30 | |
| Ablation: Only TS | Binary | ||||
Table 6: Language modeling deteriorates drastically after fine-tuning for QuAnTS.
Figure 7: The study was performed on a diverse set of 105 participants , where each provided 20 answers spread over the three answer types.
Figure 8: Most actions are easily recognizable from the animated video clips.
Figure 9: Study participants were first presented with this introduction screen. The text for the binary and multi variants only differed in the description of the answer modalities in the last three paragraphs.
Figure 10: Study participants were presented with screens matching the respective answer modality. For the open variant, we also collected information on the identifiability of the action sequences.
Figure 11: The action encoder recognizes the 19 generated actions almost perfectly. Visualized are the accuracies on the QuAnTS test set.
Figure 12: The presence of different actions in the question or answer drives changes in QA performance. For open, the relation is less obvious, since in 2 out of 3 cases, a hard-to-detect action is in a distractor rather than the correct answer.
Figure 13: Overall, action recognizability is positively correlated with performance. We also report the respective Person’s correlation coefficient and Spearman’s rank correlation coefficient , along with their -values. The relationship is significant only for the open split.
Figure 14: xQA-Qwen on AE is ahead of humans on 35 of the 45 question types, especially on the multi and open splits. The clearest exception to that is descriptive_identification, where humans lead decisively across splits. The performance of humans ( responses) and xQA (all pairs) is shown with 95-percentile bootstrap intervals from resamples, with wider ones for humans due to fewer samples.
More in this topic
Time Series QA
Time-series question answering connects natural-language questions to evidence and reasoning over sequential data.
See all 32 papers