SABET-QA: Temporal Knowledge Graph Question Answering

Brahim Touayouch; Mirette Moawad; Dmitry Akulov

cs.CLAug 20, 2026

SABET-QA: Temporal Knowledge Graph Question Answering

Authors: Brahim Touayouch, Mirette Moawad, Dmitry Akulov

Organizations: QuickSort Research, Paris, France · ENS Paris-Saclay, École Polytechnique, France

Abstract

Question Answering over Temporal Knowledge Graphs (TKGQA) requires reasoning over time-sensitive facts, yet existing embedding-based methods struggle with multi-step queries due to single-pass reasoning pipelines. We propose SABET-QA, a framework that iteratively refines reasoning states across multiple hops via a bidirectional entity-temporal scoring mechanism and a slot-aware contextualization module that aligns question semantics with temporal KG embeddings. A differentiable working memory enables progressive hypothesis refinement, while auxiliary temporal boundaries serve as coarse supervision when available. Experiments on CronQuestions, Complex-CronQuestions, MultiTQ, and TimeQuestions demonstrate consistent improvements over strong baselines, particularly on complex multi-step temporal queries.

Figures & tables

Figure 1: Impact of the number of reasoning hops on Complex-CronQuestions. Left: Hits@1 for different question categories. Right: overall test Hits@1 and Hits@10.

Model	Entity Contextualization	Bidirectional Entity & Time Scoring	Hard Supervision	Complex-CronQuestions
				Hits@1	Hits@10
SABET-QA-Hard	✓	✓	✓	0.807	0.962
	✗	✓	✓	0.707	0.946
	✓	✗	✓	0.659	0.849
	✓	✓	✗	0.524	0.896

Table 1: Ablation study of SABET-QA-Hard on the Complex-CronQuestions test set. Removing any component degrades performance, with the largest drop observed when bidirectional entity and time scoring is disabled.

Model	Frozen LM	Frozen TKE	Complex-CronQuestions		CronQuestions
			Hits@1	Hits@10	Hits@1	Hits@10
SABET-QA	✓	✓	0.524	0.896	0.843	0.969
	✓	✗	0.446	0.871	0.773	0.950
	✗	✓	0.547	0.892	0.836	0.968
	✗	✗	0.476	0.882	0.784	0.946
SABET-QA-Hard	✓	✓	0.807	0.962	0.954	0.989

Table 2: Ablation study evaluating the impact of freezing the language model (LM) and temporal knowledge graph embedding (TKE) modules under both no supervision ( SABET-QA ) and hard supervision ( SABET-QA-Hard ). The comparison isolates the contribution of each component and demonstrates the effect of unfreezing the parameters on overall reasoning performance.

Model	Hits@1					Hits@10
Model	Overall	Complex	Simple	Entity	Time	Overall	Complex	Simple	Entity	Time
BERT	0.257	0.253	0.262	0.292	0.191	0.642	0.617	0.676	0.650	0.627
CronKGQA	0.646	0.391	0.987	0.698	0.550	0.886	0.806	0.993	0.901	0.858
EmbedKGQA	0.433	0.370	0.516	0.576	0.166	0.787	0.735	0.857	0.892	0.592
TempoQR	0.796	0.658	0.981	0.880	0.640	0.959	0.934	0.992	0.975	0.930
TempoQR-Hard	0.914	0.861	0.984	0.923	0.896	0.979	0.968	0.993	0.982	0.973

Table 3: Comparison against other methods on the CronQuestions test set. Metrics are reported for overall performance, question type (complex/simple), and answer type (entity/time).

Model	Hits@1			Hits@10
Model	Overall	Entity	Time	Overall	Entity	Time
BERT	0.087	0.097	0.068	0.421	0.351	0.567
CronKGQA	0.288	0.365	0.129	0.736	0.758	0.689
EmbedKGQA	0.260	0.361	0.050	0.618	0.742	0.360
TempoQR	0.438	0.585	0.132	0.853	0.906	0.743
TempoQR-Hard	0.632	0.721	0.448	0.933	0.942	0.914

Table 4: Comparison on the Complex-CronQuestions test set. Metrics are reported for overall performance and answer type (entity/time).

Model	MultiTQ						TimeQuestions
	Hits@1			Hits@10			Hits@1			Hits@10
	Overall	Entity	Time	Overall	Entity	Time	Overall	Entity	Time	Overall	Entity	Time
BERT	0.103	0.117	0.069	0.516	0.632	0.234	0.450	0.409	0.556	0.569	0.520	0.698
CronKGQA	0.278	0.387	0.011	0.527	0.724	0.046	0.326	0.296	0.405	0.454	0.409	0.569
EmbedKGQA	0.243	0.342	0.002	0.489	0.685	0.012	0.288	0.266	0.344	0.468	0.426	0.577
TempoQR	0.327	0.456	0.014	0.571	0.783	0.055	0.409	0.406	0.416	0.530	0.513	0.574

Table 5: Comparison on the MultiTQ and TimeQuestions test sets.

Panel of Figure 2 — Figure 2: Behavioral dynamics of the differentiable working memory ( $K=4$ ) across $30,000$ CronQuestions validation instances. The network dynamically routes computation depth, shifting attention mass to H3 for compositional temporal queries.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 3 — Figure 3: TComplEx pipeline used to initialize the temporal KG embeddings. The model learns complex-valued embeddings for entities, relations, and timestamps, and supports both entity ranking and timestamp ranking through the same compositional scoring function.

Panel of Figure 4 — Figure 4: Question encoding and slot contextualization in SABER-TQA. Question tokens are encoded by a pretrained language model and projected into the temporal KG embedding space. Head, tail, and temporal slot embeddings are retrieved from the shared TKG table and attend to the question representation, followed by gated residual fusion to produce contextualized role-specific vectors.

Panel of Figure 5 — Figure 5: Iterative hop-wise reasoning in SABET-QA. At each hop, the current latent reasoning state is combined with the global question summary and transformed into a hop-specific relation representation. Dedicated entity-oriented and temporal-oriented projections generate entity and timestamp scores, whose soft distributions are converted into a working-memory representation used to update the latent state for the next reasoning hop.

Dataset	Train	Valid	Test	Total	Split Ratio (Tr/Va/Te)
CronQuestions	350,000	30,000	30,522	410,522	85.3% / 7.3% / 7.4%
Complex-CronQuestions	35,795	5,020	5,528	46,343	77.2% / 10.8% / 11.9%
MultiTQ	386,787	57,979	54,584	499,350	77.5% / 11.6% / 10.9%
TimeQuestions	6,970	3,236	3,237	13,443	51.8% / 24.1% / 24.1%

Appendix

Table 6: Statistics of the temporal question answering datasets used in our experiments.

Dataset	Entities	Relations	Timestamps	Train Facts	Valid Facts	Test Facts	Total Facts
CronQuestions	125,726	406	9621	323,635	5000	5000	333,635
Complex-CronQuestions	125,726	406	9621	323,635	5000	5000	333,635
MultiTQ	10,488	502	4,017	322,958	69,224	69,147	461,329
TimeQuestions	118,010	884	1,636	227,564	4,997	4,998	240,597

Appendix

Table 7: Statistics of the temporal knowledge graphs associated with each benchmark.

Table 8: Answer type distributions across the benchmarks.

Setting	Value
Optimizer	Adam Kingma and Ba (2017)
Initial learning rate	$2\times 10^{-4}$ ( $6\times 10^{-4}$ for Timequestions Dataset )
Maximum epochs	20 (50 for Timequestions Dataset)
Training batch size	150
Validation batch size	150
Validation frequency	Every epoch

Appendix

Table 9: Training settings used for all experiments unless otherwise noted.

Explore similar work

Let the Agent Search: Autonomous Exploration Beats Rigid Workflows in Temporal Question Answering
Mar 2, 2026Xufei Lv, Jiahui Yang, Haoyuan Sun +5Temporal QA Knowledge Graph Question Answering
Tsinghua University · National University of Defense Technology
Temporal-Aware Heterogeneous Graph Reasoning with Multi-View Fusion for Temporal Question Answering
Feb 23, 2026Wuzhenghong Wen, Bowen Zhou, Jinwen Huang +5Knowledge Graph Question Answering Temporal QA
School of Internet of Things, Nanjing University of Posts and Telecommunications · Faculty of Medicine & Health, University of New South Wales, Sydney, Australia · Beijing Information Science and Technology University +2
SCoP: Structured Constraint Parsing for Evidence-Space Control in Temporal Knowledge Graph Question Answering
Sep 2, 2026Xiaokun Guo, Zhen Xu, Dongdong Huo +3Knowledge Graph Question Answering Temporal QA
Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China

Mar 2, 2026J/K move · Enter open · S save

Let the Agent Search: Autonomous Exploration Beats Rigid Workflows in Temporal Question Answering

Xufei Lv, Jiahui Yang, Haoyuan Sun, Xialin Su+4

Tsinghua University · National University of Defense Technology

Temporal QA Knowledge Graph Question Answering Temporal KGs LLM Reasoning with Graphs LLM Agents

Open paper AskPDF arXiv

Temporal Knowledge Graph Question Answering (TKGQA) is challenging because it requires multi-hop reasoning under complex temporal constraints. Recent LLM-based approaches have improved semantic modeling for this task, but many still rely on fixed reasoning workflows or costly post-training, which can limit adaptability and make error recovery difficult. We show that enabling an off-the-shelf Large Language Model (LLM) to determine its next action is already effective in a zero-shot setting. Based on this insight, we propose AT2QA, an Autonomous and Training-free Agent for TKG Question Answering. AT2QA empowers the LLM to iteratively interact with the TKG via a generic search tool, inherently enabling autonomous exploration and dynamic self-correction during reasoning. To further elicit the LLM's potential for complex temporal reasoning, we introduce a training-free experience mining mechanism that distills a compact few-shot demonstration library from successful self-generated trajectories. AT2QA also yields a transparent audit trail for every prediction. Experiments on three challenging benchmarks -- MultiTQ, Timeline-CronQuestion, and Timeline-ICEWS-Actor -- show that AT2QA achieves new state-of-the-art performance, surpassing the strongest baselines by 10.7, 4.9, and 11.2 absolute points, respectively. Our code is available at https://github.com/AT2QA-Official-Code/AT2QA-Official-Code

Abstract

Explore similar work

Let the Agent Search: Autonomous Exploration Beats Rigid Workflows in Temporal Question Answering

Temporal-Aware Heterogeneous Graph Reasoning with Multi-View Fusion for Temporal Question Answering

SCoP: Structured Constraint Parsing for Evidence-Space Control in Temporal Knowledge Graph Question Answering