ChronoQA: Temporal Question Answering

• ChronoQA is a label for temporally grounded QA that requires models to reason over event order and time-scoped facts across various domains.
• It encompasses multiple benchmarks and systems—including news, narrative, dialogue, and structured data—that assess retrieval-augmented generation with temporal, causal, and character consistency.
• Temporal reasoning in ChronoQA demands explicit time representation, preservation of chronology during retrieval, and separation of factual correctness from temporal alignment.

ChronoQA is a label used in recent arXiv literature for temporally grounded question answering in which the correct answer depends on when events happened, how they are ordered, how time-scoped facts evolve, or how evidence from different dates must be aligned. In current usage, the term does not denote a single canonical benchmark. It names a Chinese benchmark for temporal-sensitive Retrieval-Augmented Generation (RAG) over news (Chen et al., 17 Aug 2025), appears as the target setting for chronology-aware narrative RAG in ChronoRAG (Kim et al., 26 Aug 2025), and is also used for a narrative benchmark centered on temporal, causal, and character consistency under RAG (Zhang et al., 6 Jun 2025). A separate paper explicitly states that, in the historical-newspaper setting, “ChronoQA is not the paper’s formal dataset title”; the official name there is ChroniclingAmericaQA (Piryani et al., 2024).

1. Terminology and scope

The recent literature uses “ChronoQA” in several non-identical ways. Taken together, these papers suggest that the label functions as a recurring shorthand for temporal question answering across multiple substrates—news corpora, narrative texts, dialogue histories, knowledge bases, tables, and multimodal archives—rather than as a single standardized task definition (Chen et al., 17 Aug 2025).

Label in literature	Formal resource or system	Reported scope
ChronoQA	Chinese temporal-sensitive RAG benchmark	5,176 QA pairs from over 300,000 news articles, 2019–2024
ChronoQA / ChronoRAG setting	“Chronological Passage Assembling in RAG framework for Temporal Question Answering”	NarrativeQA evaluation with a 1,111-question “Time Questions” subset
ChronoQA	Narrative RAG benchmark in “Respecting Temporal-Causal Consistency”	497 QA pairs from 9 public-domain narrative works
ChronoQA as informal reference	ChroniclingAmericaQA	487K question-answer pairs from historical newspapers, 1800–1920

The ambiguity is not merely terminological. In the Chinese RAG benchmark, ChronoQA is explicitly a dataset for “temporal-sensitive Retrieval-Augmented Generation” with annotated temporal type, time-expression type, answer type, temporal scope, and reference-document count (Chen et al., 17 Aug 2025). In the ChronoRAG paper, ChronoQA names the temporal question answering regime in which narrative answers depend on “the temporal relationship among multiple events” rather than isolated snippets (Kim et al., 26 Aug 2025). In the E$^2$RAG paper, ChronoQA is a benchmark for narrative documents that tests “temporal, causal, and character consistency” under retrieval (Zhang et al., 6 Jun 2025). In contrast, ChroniclingAmericaQA is a formally distinct dataset whose paper states that ChronoQA is only an informal reference, not the official title (Piryani et al., 2024).

A common misconception is therefore that ChronoQA refers to one benchmark analogous to a single shared leaderboard. The published record does not support that reading. It instead presents several resources and systems that converge on a shared research question: how to answer queries whose semantics are inseparable from temporal structure.

2. Core problem formulations

Across these papers, ChronoQA is defined less by domain than by the form of reasoning required. The Chinese ChronoQA benchmark organizes questions into absolute, aggregate, and relative temporal types; distinguishes explicit from implicit time expressions; and includes both single-document and multiple-document settings (Chen et al., 17 Aug 2025). This formulation makes temporal resolution itself part of the task: a model must not only retrieve relevant evidence, but also anchor references such as “last year,” compare temporally separated events, and maintain logical consistency when evidence is distributed across multiple documents.

Narrative ChronoQA emphasizes a different failure mode. The ChronoRAG paper argues that long-context narrative questions are often not answerable from “a single isolated sentence,” because the answer lies in the temporal relation among events and in the surrounding story flow. Questions such as “When did X happen?” and “What happened after Y?” require the model to identify what happened, when it happened, and how one event led to another (Kim et al., 26 Aug 2025). The E$^2$RAG benchmark extends this by treating temporal reasoning as inseparable from causal consistency and character consistency. Its eight reasoning categories include Causal Consistency, Character Consistency, Setting / Environment / Atmosphere, Symbolism / Imagery / Motifs, Thematic / Philosophical / Moral, Narrative / Plot Structure, Social / Cultural / Political, and Emotional / Psychological (Zhang et al., 6 Jun 2025).

Dialogue-based ChronoQA-style work pushes the formulation further from static QA toward persistent temporal state. Chronos, a temporal-aware conversational memory framework, targets “long-term, multi-session question answering over dialogue histories that span weeks or months.” Its motivating queries require knowing not only what happened, but whether it is recent or outdated, how facts evolve across sessions, and how multi-hop reasoning should aggregate temporally grounded events (Sen et al., 17 Mar 2026). ChronoScope isolates yet another aspect—temporal scope stability—defined as the ability to preserve, override, or transfer time-scoped assumptions across dialogue turns, especially when follow-up questions omit explicit dates (Atri et al., 24 Apr 2026).

Structured QA work shows that the same conceptual problem appears over symbolic stores. TempQA-WD targets temporal KBQA with reasoning over time points, intervals, before/after relations, overlap, and derived temporal spans (Neelam et al., 2022). CRONQUESTIONS and TimeQuestions frame temporal KGQA as answering over temporally scoped facts, with answer types that can be entities or timestamps and with question classes such as Before/After, First/Last, Time Join, Explicit, Implicit, Temporal Answer, and Ordinal (Saxena et al., 2021, Jia et al., 2021). Temporal tabular QA, in turn, converts evolving infobox snapshots into SQL over a normalized relational schema so that temporal logic becomes explicit in joins, aggregations, and date comparisons (Thanga et al., 29 Nov 2025).

3. Benchmark resources and data regimes

The benchmark ecosystem around ChronoQA is heterogeneous in corpus type, annotation granularity, and evaluation target.

Resource	Domain and representation	Reported scale
ChronoQA (Chen et al., 17 Aug 2025)	Chinese news RAG benchmark	5,176 QA pairs; over 300,000 news articles
ChroniclingAmericaQA (Piryani et al., 2024)	Historical newspapers; OCR text, corrected text, scanned pages	487K QA pairs; 1800–1920
TempQA-WD (Neelam et al., 2022)	Wikidata/Freebase temporal KBQA with SPARQL	839 questions
CRONQUESTIONS (Saxena et al., 2021)	Temporal KGQA over Wikidata-derived temporal KG	410k unique QA pairs
TimeQuestions (Jia et al., 2021)	Large temporal KGQA benchmark	16,181 temporal questions
ChronoScope (Atri et al., 24 Apr 2026)	Multi-turn temporal scope benchmark from Wikidata	1,469,628 interaction chains

The Chinese ChronoQA dataset is built from over 300,000 news articles published from January 1, 2019 to August 30, 2024, and contains 5,176 question-answer pairs. Its reported composition includes 2,529 Absolute, 1,911 Aggregate, and 736 Relative questions; 2,000 Explicit and 3,176 Implicit time-expression cases; 3,261 Single-document and 1,915 Multiple-document examples; and temporal scopes split into 1,946 Long-term, 2,736 Mid-term, and 494 Short-term instances (Chen et al., 17 Aug 2025). The paper also reports structural annotations including question_date, temporal_expression_type, temporal_scope, temporal_granularity, temporal_type, answer_type, reference_document_count, and golden_chunks, making it suitable for retrieval diagnostics as well as answer evaluation.

ChroniclingAmericaQA targets historical, diachronic QA under archival conditions rather than contemporary RAG. It is built from a subset of Chronicling America covering 1800–1920 and reports 487K question-answer pairs, with a more exact split of 439,302 train, 24,111 dev, and 24,084 test pairs (Piryani et al., 2024). Its distinctive contribution is that the same benchmark supports three settings: raw noisy OCR text, GPT-corrected OCR text, and scanned page images. This materially changes the failure model: temporal QA must coexist with OCR noise, archaic language, and complex newspaper layout.

Narrative ChronoQA resources are smaller but structurally specialized. The E$^2$RAG ChronoQA benchmark is built from nine public-domain narrative works and contains 497 question-answer pairs with passage-level supervision, including start and end sentences and byte offsets for the supporting excerpt (Zhang et al., 6 Jun 2025). The ChronoRAG paper does not introduce a new dataset, but evaluates on NarrativeQA and derives a 1,111-question “Time Questions” subset from temporal keywords “When,” “While,” “During,” “After,” and “Before,” from a base collection of 355 stories and scripts and 10,557 question-answer pairs (Kim et al., 26 Aug 2025).

In the structured-QA lineage, TempQA-WD contains 839 questions with Wikidata SPARQL for all questions and a fine-grained annotated subset of 175 examples (Neelam et al., 2022). CRONQUESTIONS is the largest temporal KGQA dataset in that line, with 410k unique question-answer pairs over a temporal KG derived from Wikidata (Saxena et al., 2021). TimeQuestions contains 16,181 temporal questions compiled from eight KGQA datasets and manually categorized into Explicit, Implicit, Temporal Answer, and Ordinal classes (Jia et al., 2021). ChronoScope shifts the unit of analysis from isolated questions to controlled multi-turn chains, reporting 1,469,628 interaction chains spanning 2 to 11 turns and organized into 11 chain families such as Carryover, Scope Switch, Temporal Narrative, and Bridged Multi-PID (Atri et al., 24 Apr 2026).

4. Retrieval-augmented and memory-based architectures

ChronoQA research has made chronology a first-class retrieval signal rather than a downstream afterthought. ChronoRAG exemplifies this move for narrative documents. Its pipeline has an offline graph-construction stage and an online retrieval-and-generation stage. The document is first divided into fixed-length chunks, exemplified as 100 tokens for Layer 0; groups of 10 chunks are summarized; relation descriptions are extracted from those summaries; and only relation descriptions, not full entity descriptions, are used for indexing in Layer 1. At retrieval time, ChronoRAG first retrieves high-precision relation descriptions from Layer 1, then uses their child indices to fetch Layer 0 chunks, and finally assembles each retrieved relation with neighboring Layer 1 passages in index order so that the model sees a coherent local event sequence. The paper specifies top-$k=20$ retrieval, a 1,500-token context limit, embedding-similarity scoring without BM25, arctic-Snowflake-embed-l for embeddings, meta-llama-3-8B-Instruct for summarization and extraction, and unifiedqa-v2-t5-3b-1363200 for final answering (Kim et al., 26 Aug 2025).

E$^2$RAG addresses a related narrative failure mode by arguing that KG-RAG collapses all mentions of an entity into one node and thereby erases evolving contextual state. Its remedy is a dual-graph design with an entity subgraph $\mathcal{G}_{\text{ent}}$, an event subgraph $\mathcal{G}_{\text{evt}}$, and a bipartite mapping $\mathcal{B}$ that links entity mentions to events when the entity name appears in the event description. Query-time processing extracts entity and event cues, retrieves seed nodes from vector stores, performs one-hop expansion across the bipartite graph, ranks only the passages attached to the expanded set, and assembles the final prompt from raw passages plus a linearized subgraph (Zhang et al., 6 Jun 2025). This architecture is explicitly designed to preserve time-specific entity states and event grounding.

The Chinese ChronoQA benchmark paper evaluates retrieval baselines rather than proposing a single new architecture, but its baseline findings are methodologically important. It compares Native RAG, Temporal Filter, Query Rewrite, and Query Decomposition, and reports that Query Decomposition performs best overall, especially for multi-document questions (Chen et al., 17 Aug 2025). This result suggests that ChronoQA queries often benefit from explicit subproblem factorization before retrieval, particularly when the answer requires aggregation or chaining over time-linked documents.

Chronos generalizes chronology-aware retrieval from documents to long-term conversational memory. Its dual-calendar architecture stores extracted $$\langle \text{subject}, \text{verb}, \text{object} \rangle$$ event tuples with start_datetime and end_datetime in an event calendar, while preserving raw dialogue turns in a turn calendar. At query time it generates retrieval guidance with a small LLM-powered template generator, performs initial retrieval over the turn calendar via dense search, reranking, and context expansion, and then runs a ReAct-style tool loop with search_turns, search_events, grep_turns, and grep_events. The turn retrieval stage takes the top 100 dense candidates, reranks them with Cohere Rerank v3, keeps the top 15, and expands each with one preceding and one following turn from the same session (Sen et al., 17 Mar 2026). This design operationalizes temporal reasoning as structured range matching over event intervals plus iterative LLM synthesis.

Multimodal timestamped QA extends the same intuition to video. CourseTimeQA defines a lecture-video setting in which the system retrieves top-$k$ timestamped segments and generates a grounded answer under a median end-to-end latency budget below 2.5 seconds on one GPU. CrossFusion-RAG precomputes query-agnostic multimodal segment embeddings from ASR and sampled frames, retrieves from FAISS, reranks with a small cross-attentive reranker, and applies MMR diversification before generation (Kovalev et al., 29 Nov 2025). While this work is not named ChronoQA, it shows that temporal retrieval design principles transfer to timestamped cross-modal QA.

5. Structured temporal reasoning over knowledge bases, tables, and claims

A second major lineage treats ChronoQA as explicit symbolic reasoning over temporally scoped structures. TEQUILA is an early modular formulation for temporal KBQA. It detects temporal intent, decomposes the question into non-temporal sub-questions and temporal constraints, retrieves answers from an underlying KB-QA engine, and then applies interval reasoning with relations such as BEFORE, AFTER, and OVERLAP (Jia et al., 2019). Its central claim is that temporal semantics can be layered on top of standard KB-QA backends rather than requiring a monolithic new parser.

TempQA-WD formalizes temporal KBQA as semantic parsing to SPARQL over Wikidata while retaining parallel Freebase answers from TempQuestions. It supplies SPARQL for all questions and a fine-grained subset with AMR, $^2$0-expressions, entity linking, and relation linking annotations, making interpretability a benchmark feature rather than an afterthought (Neelam et al., 2022). Exaqt moves toward end-to-end neural temporal KGQA with a two-stage pipeline: high-recall answer-graph construction using Group Steiner Trees plus BERT relevance modeling, followed by high-precision answer prediction with an R-GCN enhanced by time-aware entity embeddings and attention over temporal relations (Jia et al., 2021).

CRONQUESTIONS and CRONKGQA shift the field toward large-scale temporal KGs. CRONQUESTIONS is built over a temporal KG with 328k facts and roughly 5k event-facts and supports question families including Simple Entity, Simple Time, Before/After, First/Last, and Time Join (Saxena et al., 2021). CRONKGQA couples a BERT question encoder with temporal KG embeddings such as TComplEx, producing separate entity and time score vectors before joint softmax prediction (Saxena et al., 2021). QC-MHM further argues that PLM question encoders over-focus on surface entities and under-model entity transfer induced by temporal constraints. Its remedy is question calibration with time-constrained KG concepts, timestamp-order-aware embeddings, and multi-hop GNN reasoning over a temporal subgraph (Xue et al., 2024).

Temporal reasoning over evolving tables has recently adopted a database-centric formulation. “Evidence-Guided Schema Normalization for Temporal Tabular Reasoning” generates a 3NF schema from Wikipedia infobox snapshots, populates SQLite, and answers questions by text-to-SQL execution. The paper identifies three principles—normalization that preserves context, semantic naming that reduces ambiguity, and consistent temporal anchoring—and argues that schema quality can matter more than model capacity (Thanga et al., 29 Nov 2025). This reframes part of ChronoQA as a representation-design problem: poorly normalized or weakly anchored time structure can make correct temporal reasoning inaccessible even to stronger generators.

ChronoFact shows that temporal QA and temporal verification share a core dependence on timeline structure. It verifies claims by extracting events from the claim and evidence, building claim and evidence timelines, comparing them at token, event, and time levels, classifying each claim event, classifying chronological consistency, and combining those decisions with logical rules (Barik et al., 2024). This suggests a close relationship between ChronoQA and timeline-based fact verification: in both, event truth and event order must be jointly modeled.

6. Evaluation patterns, empirical findings, and open issues

Empirical results across these works converge on a common point: temporal alignment is often the dominant bottleneck. In ChronoRAG, chronology-aware passage assembling yields the best reported NarrativeQA performance, with ROUGE-L 0.308 on the full set and 0.268 on the Time Questions subset, ahead of RAPTOR-CT at 0.297 and 0.261 and RAPTOR-TT at 0.295 and 0.259. Ablations show that removing passage assembling drops performance to 0.295 overall and 0.252 on Time Questions; removing chunk summarization lowers it to 0.272 and 0.233; and removing relation extraction reduces it to the NaiveRAG baseline of 0.255 and 0.227 (Kim et al., 26 Aug 2025). These numbers support the paper’s claim that preserving local chronology during retrieval is not a marginal refinement but a core ingredient.

The Chinese ChronoQA benchmark reports a different pattern: state-of-the-art LLMs perform moderately on single-document questions but drop significantly on multi-document questions, and among retrieval baselines Query Decomposition performs best overall, especially for multiple-document instances (Chen et al., 17 Aug 2025). This indicates that time-sensitive RAG failure is not only about missing date filters; it is also about decomposing temporally compositional queries into retrievable subproblems.

Chronos reports unusually high accuracy on long-term dialogue memory: 92.60% for Chronos Low and 95.60% for Chronos High on LongMemEvalS, with the event calendar identified as the largest contributor in ablations and a 58.9% gain on the baseline (Sen et al., 17 Mar 2026). At the same time, the paper explicitly notes benchmark issues and judgment variability, including inconsistent references and unstable LLM-as-judge behavior. This is a methodological caution: ChronoQA evaluation over free-form answers can inherit substantial variance from the judging procedure itself.

Structured benchmarks likewise show that temporal QA is far from saturated. TempQA-WD reports low baseline performance for SYGMA on Wikidata, with overall F1 0.32 and especially weak medium-complexity performance (Neelam et al., 2022). Exaqt improves over prior KGQA systems on TimeQuestions with P@1 0.565, MRR 0.599, and Hit@5 0.664, while identifying Ordinal questions as particularly difficult (Jia et al., 2021). CRONKGQA reaches Hits@1 0.647 on CRONQUESTIONS and, according to the paper, improves accuracy by 120% over the next best method (Saxena et al., 2021). QC-MHM pushes further, reporting Hits@1 0.971 overall and 0.946 on complex questions in CronQuestions, with the largest ablation drop caused by removing question calibration (Xue et al., 2024). For evolving tables, the SQL-based normalization pipeline reaches 80.39 EM and 82.11 F1, exceeding a 68.89 EM baseline and supporting the claim that schema quality is a first-class model component (Thanga et al., 29 Nov 2025).

Several diagnostic benchmarks show that high single-turn factual competence does not imply robust temporal reasoning. ChronoSense finds modest performance on the full Allen interval algebra and temporal arithmetic, with strong asymmetries between converse relations and evidence that models may rely on memorization of named events rather than interval reasoning (Islakoglu et al., 6 Jan 2025). ChronoScope shows that temporal scope stability is frequently violated in multi-turn settings, with models often drifting toward present-day assumptions even under Gold Context and with chain accuracy collapsing under Self-Conditioned evaluation (Atri et al., 24 Apr 2026). ChronoFact attributes a large portion of improvement to explicit chronological-order classification, and its ablation removing the chronological order classifier produces the largest drop on ChronoClaims (Barik et al., 2024). In multimodal settings, CHRONOSIGHT finds a large human–model gap that it terms chronological blindness, with the best open model scoring 0.40 under direct prompting against a human average of 0.89, and with parse failures and vocabulary-mapping issues confounding pure perceptual assessment (Goswami et al., 15 Jun 2026).

The limitations reported across the literature are consistent. ChronoRAG notes linear offline construction cost, fixed-chunk boundary issues, and specialization to narrative text (Kim et al., 26 Aug 2025). Chronos points to storage overhead from dual indexes, added inference latency from dual-calendar reasoning, and fabrication on abstention-style questions (Sen et al., 17 Mar 2026). ChroniclingAmericaQA emphasizes OCR noise, archaic language, and page-layout complexity as persistent confounders (Piryani et al., 2024). ChronoScope is intentionally template-generated and entity-centric, which sharpens diagnosis but narrows ecological validity (Atri et al., 24 Apr 2026). CHRONOSIGHT is synthetic and still-image-based rather than photographic or video-grounded (Goswami et al., 15 Jun 2026).

A plausible implication is that ChronoQA research is converging on three non-interchangeable requirements. First, systems must represent time explicitly, whether as ordered passages, datetime intervals, temporal KG qualifiers, snapshot identifiers, or SQL time anchors. Second, retrieval must preserve chronology rather than merely rank semantically similar fragments. Third, evaluation must separate factual correctness from temporal correctness, because a fact can be individually true yet temporally misaligned with the question’s scope. Across the surveyed papers, those three requirements recur even when the data modality and modeling stack differ substantially.