MultiTQ: Unified Multi-Target Evaluation

• MultiTQ is a domain-specific term representing frameworks that unify multiple targets, tasks, or temporal reasoning challenges across disciplines.
• In multilingual NLP, it quantifies text quality across 115 languages via pairwise preference learning on coherence, fluency, simplicity, and acceptability.
• In temporal KGQA and related fields, MultiTQ benchmarks stress-test alternative decompositions for joint modeling, enabling improved inference and system comparisons.

Searching arXiv for papers and usages of “MultiTQ” to ground the article. arXiv search query: "MultiTQ temporal knowledge graph question answering" MultiTQ is used in several distinct senses across recent research. In the materials considered here, it denotes or closely tracks multilingual text-quality evaluation for LLMs, a multi-granularity temporal knowledge graph question answering benchmark and its associated methods, multi-task question answering over time series, joint multi-tissue eQTL analysis, and several quantum formulations centered on multi-target or multitask processing. This breadth indicates that MultiTQ is best understood as a domain-specific label whose meaning is fixed by the surrounding methodology, data model, and evaluation protocol rather than by a single universal definition (Pokharel et al., 12 Nov 2025, Lv et al., 2 Mar 2026, Kong et al., 26 Feb 2025, Flutre et al., 2012, Artag et al., 10 Mar 2026).

1. Scope and terminological usage

In the supplied research record, MultiTQ functions as a compact label for different kinds of “multi-target,” “multilingual text quality,” or “multi-task question answering” problems. Some papers use the term directly; others state that they do not explicitly define “MultiTQ” but nonetheless occupy the conceptual space the label typically denotes. The result is a family resemblance rather than a single canonical object (Pokharel et al., 12 Nov 2025, Kong et al., 26 Feb 2025, Hai et al., 2024).

Area	Referent	Defining emphasis
Multilingual NLP	MTQ-Eval	Multilingual text-quality preference learning
Temporal KGQA	MultiTQ benchmark	Multi-granularity temporal question answering
Time-series reasoning	Time-MQA / TSQA	Multi-task QA over temporal data
Multivariate forecasting	TQ / TQNet	Temporal Query mechanism for multivariate correlations
Statistical genomics	joint multi-tissue eQTL framework	Shared activity across tissues
Quantum computing	MTQA / MILQ / multi-target compilation / MTQCL	Parallel or multi-target quantum processing

The multilingual NLP usage is explicit in the MTQ-Eval paper’s explanation that, although the paper “does not use or define the term ‘MultiTQ’ explicitly,” its formulation directly addresses the multilingual text-quality problem that “MultiTQ” typically denotes. The time-series QA literature makes a parallel statement: Time-MQA and TSQA “fill that conceptual space” for temporal data. In multi-target quantum compilation, the supplied interpretation states that “MultiTQ” may be taken as shorthand for “multi-target quantum compilation” (Pokharel et al., 12 Nov 2025, Kong et al., 26 Feb 2025, Hai et al., 2024).

A plausible implication is that MultiTQ is not a stable acronym in the way that benchmark names such as MultiTQ in temporal KGQA are stable. Instead, the term recurrently names frameworks that unify multiple targets, tasks, languages, tissues, or temporal conditions under one learning or inference mechanism.

2. MultiTQ as multilingual text-quality evaluation

In multilingual NLP, MultiTQ corresponds to the problem instantiated by MTQ-Eval: learning a multilingual evaluator that prefers high-quality over low-quality text across 115 languages. The framework defines text quality using four dimensions—coherence, fluency, simplicity, and linguistic acceptability—and casts evaluation as a binary decision, with pairwise preference learning over high-quality versus low-quality versions of the same passage. High-quality passages come from Belebele’s human-translated parallel data; low-quality passages are created by controlled word shuffling that randomly swaps three to six words per passage. This yields 4,600 passage-level training instances across 115 languages, with 20 samples per language, split into 2,300 normal and 2,300 degraded passages (Pokharel et al., 12 Nov 2025).

MTQ-Eval trains open-source base LLMs through Direct Preference Optimization rather than prompt-only judging. The paper evaluates Llama 3.1 8B Instruct and Aya Expanse 8B, using LoRA with rank 64, alpha 128, dropout 0.05, 1 epoch, batch size 2, and learning rate $5\mathrm{e}{-7}$. Prompts standardize the task by instructing the model to rate the four quality dimensions and return only [[0]] or [[1]]. The underlying claim is that a learned evaluator aligned to quality preferences is more scalable than few-shot prompt judging in low-resource settings (Pokharel et al., 12 Nov 2025).

The reported results are consistently positive. On MELA, MTQ-Eval improves MCC over supervised fine-tuning in both model families; for Aya, the average MCC is 0.39 versus 0.21 for SFT. On the Belebele-derived 115-language test set, Llama improves from MCC 0.18 to 0.24, KL divergence from 0.06 to 0.12, and F1 from 0.56 to 0.59; Aya improves from MCC 0.14 to 0.26, KL divergence from 0.04 to 0.15, and F1 from 0.44 to 0.60. Resource-level analysis shows gains in both high-resource and low-resource languages, though the gains are stronger in high-resource settings, especially for Aya. The same quality-alignment training also transfers modestly to downstream tasks: on MMS sentiment, average F1 rises from 0.49 to 0.51 for Llama and from 0.48 to 0.52 for Aya; on XL-Sum summarization, MTQ-Eval matches or exceeds SFT on G-Eval coherence, consistency, and fluency, with higher low-resource averages of +0.03 in Llama and +0.01 in Aya (Pokharel et al., 12 Nov 2025).

The limitations are equally explicit. The method excludes languages where space tokenization or word order makes shuffling ineffective, such as Japanese and Basque; it captures only some degradation modes; and it does not analyze fairness across underrepresented families and scripts. The paper also notes that performance is strongest in high-resource Indo-European languages and weaker in many African and South Asian languages, which raises cross-lingual parity concerns (Pokharel et al., 12 Nov 2025).

3. MultiTQ as a temporal KGQA benchmark

In temporal knowledge graph question answering, MultiTQ is a benchmark rather than a general concept. It is described as a large, multi-granularity benchmark for Temporal Knowledge Graph Question Answering constructed from ICEWS05-15, targeting complex temporal reasoning over knowledge graphs with multi-hop chains, strict temporal operators, and both single-target and multi-target queries. The benchmark covers temporal constraints at year, month, and day granularities over a temporal span exceeding 3,600 days, and it includes operators such as Before/After, First/Last, Equal, plus implicit temporal constraints (Lv et al., 2 Mar 2026).

The dataset statistics reported for this benchmark are substantial. AT2QA describes MultiTQ as containing more than 461K temporal quadruples and approximately 500K unique question–answer pairs, with train/dev/test counts of 386,787 / 57,979 / 54,584. It further reports that approximately 73% of the questions are single-target and approximately 27% are multiple-target. Temp-R1 uses the same split totals and characterizes “Single” as simple factual lookup or single-hop reasoning and “Multiple” as complex multi-hop reasoning with temporal ordering and constraints (Lv et al., 2 Mar 2026, Gong et al., 26 Jan 2026).

The task formalization is standard for temporal KGQA. A temporal knowledge graph is a set of quadruples

$$G = \{(e_h, r, e_t, \tau)\} \subset E \times R \times E \times T,$$

and a valid answer must satisfy the semantics of the question together with its temporal conditions. Different papers evaluate the benchmark with different protocols. AT2QA reports Hits@1 exact match on the full benchmark; Temp-R1 and RTQA report Hits@1 and Hits@10 with breakdowns by question type and answer type; RoMem evaluates a 500-question stratified sample with retrieval metrics such as MRR, Hit@3, and Hit@10 and answer metrics such as Acc@5 and Acc@10; STAR-RAG samples 1,000 test questions and reports Hit@1, Hit@5, and Hit@10 (Lv et al., 2 Mar 2026, Gong et al., 26 Jan 2026, Li et al., 13 Apr 2026, Zhu et al., 19 Oct 2025).

This variation in protocol is central to interpreting the literature. The benchmark identity is stable, but the reported numbers are not directly interchangeable unless the sampling regime, metric family, and retrieval setting are matched.

4. Systems developed for MultiTQ

MultiTQ has become a focal evaluation bed for several methodological families in temporal reasoning: program induction, recursive decomposition, reinforcement-learned autonomous agents, training-free search agents, temporal GraphRAG, and structured temporal memory. The shared problem is temporal consistency under multi-hop and multi-constraint reasoning, but the operational strategies differ sharply (Chen et al., 2024, Gong et al., 4 Sep 2025, Gong et al., 26 Jan 2026, Lv et al., 2 Mar 2026, Zhu et al., 19 Oct 2025, Li et al., 13 Apr 2026).

System	Setting	Reported result
Prog-TQA	MultiTQ test set	Hits@1 0.797, Hits@10 0.934
Temp-R1	MultiTQ test set	Hits@1 0.780 overall, 0.550 Multiple
RTQA	MultiTQ test set	Hits@1 0.765 overall, 0.424 Multiple
AT2QA	MultiTQ full benchmark	88.7% Hits@1, 75.1% on multiple-target
STAR-RAG	1,000 sampled test questions	Hit@1 30.5%, Hit@5 41.5%, Hit@10 47.7%
RoMem	500 sampled questions	MRR 0.337, Acc@5 0.366 in one setup

Prog-TQA approaches MultiTQ through symbolic programs with temporal operators such as FilterBefore, FilterAfter, FilterFirst, FilterLast, GetYear, GetMonth, and FilterRange. On MultiTQ it reports Hits@1 of 0.797 overall, 0.750 on multiple-constraint questions, 0.817 on single-constraint questions, 0.790 for entity answers, and 0.815 for time answers. Its ablations are large: removing self-improvement reduces overall Hits@1 from 0.797 to 0.583 and multiple-question Hits@1 from 0.750 to 0.379; replacing the linker also causes major losses (Chen et al., 2024).

RTQA is training-free and frames complex questions through recursive decomposition, a bottom-up solver, and an answer aggregator. On MultiTQ it reports Hits@1 of 0.765 overall, 0.424 on Multiple, 0.902 on Single, 0.692 for entity answers, and 0.942 for time answers. Removing the decomposer reduces Multiple Hits@1 from 0.424 to 0.214, while removing multi-answer aggregation reduces it to 0.341, which directly quantifies the contribution of decomposition and fault-tolerant aggregation (Gong et al., 4 Sep 2025).

Temp-R1 is an autonomous end-to-end agent trained with reinforcement learning. Its action space separates internal actions—plan, think, filter, rank—from the external search action and terminal answer action, and it uses reverse curriculum learning so that difficult Multiple questions are seen first. On MultiTQ it reports Hits@1 of 0.780 overall, 0.550 on Multiple, 0.888 on Single, 0.714 for entity answers, and 0.969 for time answers. The reverse curriculum is decisive in the ablation table: without it, Multiple Hits@1 falls from 0.550 to 0.143 (Gong et al., 26 Jan 2026).

AT2QA pushes the agentic interpretation further by dispensing with training and granting an off-the-shelf LLM autonomy through iterative interaction with a general Search tool. On MultiTQ it reports 88.7% Hits@1, an absolute improvement of +10.7% over prior SOTA, and 75.1% Hits@1 on multiple-target queries, a +20.1% gain over Temp-R1. Its ablations show how heavily the method depends on temporal tooling: removing time-window constraints drops accuracy from 88.7% to 59.2%, while removing temporal sorting or structural filtering yields 79.6% and 78.2%, respectively (Lv et al., 2 Mar 2026).

Retrieval-centric methods attack a narrower subproblem: time-consistent evidence selection. STAR-RAG builds a time-aligned rule graph and runs seeded personalized PageRank over that graph. On MultiTQ it reports Hit@1 of 30.5%, Hit@5 of 41.5%, and Hit@10 of 47.7%, with especially strong multiple-event performance of 44.4% Hit@1. RoMem instead augments agentic memory with continuous phase rotation and a Semantic Speed Gate over relation volatility. On a sampled 500-question MultiTQ setting, it reports MRR 0.337, Hit@3 0.384, Hit@10 0.502, Acc@5 0.366, and Acc@10 0.392 in a closed-source setup, with similarly strong gains in an open-source setup (Zhu et al., 19 Oct 2025, Li et al., 13 Apr 2026).

Taken together, these systems show that MultiTQ has become a benchmark for contrasting explicit temporal operators, recursive subproblem formation, autonomous search, temporal retrieval, and temporally structured memory. This suggests that the benchmark’s main scientific role is not merely ranking systems, but stress-testing alternative decompositions of temporal reasoning itself.

5. Time-series interpretations of MultiTQ

In time-series research, MultiTQ is associated with unified natural-language interaction over temporal data. Time-MQA defines a framework in which a model $f_\theta$ maps a time series $X$, context $C$, and question $Q$ to an answer $A$, where the answer may be a number, vector, class label, set of timestamps, or free-form text. The accompanying TSQA dataset contains approximately 200k question–answer pairs, reported more precisely as 192,843 total, spanning 12 domains and five task families: forecasting, imputation, anomaly detection, classification, and open-ended reasoning QA (Kong et al., 26 Feb 2025).

The technical significance of Time-MQA is the collapse of heterogeneous time-series workloads into a single QA interface. Forecasting uses input lengths uniformly sampled in $[64,256]$ and horizons in $[8,32]$; anomaly detection uses lengths in $[8,256]$; classification uses lengths in $$G = \{(e_h, r, e_t, \tau)\} \subset E \times R \times E \times T,$$0; and all numeric series are serialized as comma-separated text lists. Continual pretraining is performed on Mistral 7B, Llama-3 8B, and Qwen-2.5 7B with LoRA of rank 16 and $$G = \{(e_h, r, e_t, \tau)\} \subset E \times R \times E \times T,$$1. On the reported evaluation, Mistral 7B achieves forecasting MSE 1.35 versus GPT-4o 1.79, and Qwen achieves anomaly detection accuracy 0.68 and judgment accuracy 0.82. The paper explicitly states that Time-MQA plus TSQA “fills that conceptual space” for MultiTQ in time series (Kong et al., 26 Feb 2025).

A second time-series usage is multivariate forecasting through the Temporal Query mechanism. TQNet treats channels as tokens and uses periodically shifted learnable query vectors to encode global inter-variable patterns, while keys and values come from the raw input to capture local sample-level correlations. The mechanism is parameterized by $$G = \{(e_h, r, e_t, \tau)\} \subset E \times R \times E \times T,$$2, where $$G = \{(e_h, r, e_t, \tau)\} \subset E \times R \times E \times T,$$3 is a period length hyperparameter. TQNet is deliberately lightweight—a single TQ-MHA layer plus a shallow MLP—and it reports Top-2 performance in 22/24 metrics across 12 datasets. Representative averages include Electricity at 0.164/0.259 MSE/MAE, PEMS03 at 0.097/0.203, and PEMS07 at 0.075/0.171 (Lin et al., 19 May 2025).

The two time-series usages are methodologically different. Time-MQA is a language-model formulation of heterogeneous temporal tasks, whereas TQNet is a forecasting architecture centered on multivariate correlations. The commonality is that both reinterpret temporal analysis as a unified multi-task or multi-target problem.

6. Statistical genomics and quantum formulations

In statistical genomics, the supplied material uses MultiTQ to describe a joint multi-tissue eQTL framework. For a given gene–SNP pair, the goal is to determine whether the SNP is an eQTL in any tissue and, if so, in which subset of tissues. The framework introduces a latent activity vector $$G = \{(e_h, r, e_t, \tau)\} \subset E \times R \times E \times T,$$4, models cross-tissue effect sharing through a multivariate normal prior, computes configuration-specific approximate Bayes factors, and estimates sharing patterns by empirical Bayes. Reanalysis of transformed B cells, T cells, and fibroblasts identifies 63% more genes with eQTLs at FDR $$G = \{(e_h, r, e_t, \tau)\} \subset E \times R \times E \times T,$$5 than tissue-by-tissue analysis, and the hierarchical model estimates approximately 88% $$G = \{(e_h, r, e_t, \tau)\} \subset E \times R \times E \times T,$$6 of detectable eQTLs shared across all three tissues and approximately 8% tissue-specific (Flutre et al., 2012).

In quantum research, several closely related formulations appear. MTQA, or multi-tasking quantum annealing, embeds multiple independent QUBOs into spatially disjoint regions of a single annealer and solves them simultaneously. On D-Wave Advantage 6.4, it maintains solution quality comparable to single-problem quantum annealing and simulated annealing while reducing time-to-solution; for small dense graphs with $$G = \{(e_h, r, e_t, \tau)\} \subset E \times R \times E \times T,$$7, the reported parallel capacity reaches up to 130 MVCP or 128 GPP instances in non-isolated mode (Artag et al., 10 Mar 2026). MILQ, a scheduler and cutter for heterogeneous clusters of QPUs, frames the problem as unrelated parallel machines and reports up to 26% total improvement over a baseline approach (Seitz et al., 2023). Multi-target quantum compilation jointly compiles several target unitaries into a shared circuit architecture with per-target parameters, minimizing the average infidelity

$$G = \{(e_h, r, e_t, \tau)\} \subset E \times R \times E \times T,$$8

and reports substantial depth reductions relative to baselines in noiseless benchmarks (Hai et al., 2024). MTQCL, finally, is a logical formalism extending standard Quantum Computational Logic by allowing an arbitrary number and placement of target qubits and defining truth through a circuit-dependent projector over the target set (Sergioli, 2018).

These genomics and quantum usages do not share application domains, but they share a structural motif: a single inferential or computational formalism is designed to handle multiple targets, tissues, problems, or target qubits simultaneously. This suggests that the persistence of the MultiTQ label across fields is tied less to a fixed acronym expansion than to a recurring design principle of joint modeling across heterogeneous but related units.