TableVQA: Visual Table Question Answering

• TableVQA is the task of answering natural-language questions by reasoning over table images using computer vision, NLP, and symbolic methods.
• Recent approaches include end-to-end vision–language models and hybrid structured pipelines that enhance transparency and multi-step reasoning.
• Benchmarks reveal challenges like visual noise, multilingual discrepancies, and efficiency–accuracy trade-offs in processing real-world table images.

TableVQA, or Table Visual Question Answering, is the task of answering natural-language questions by reasoning over images of tables. The field sits at the intersection of computer vision, natural-language processing, and symbolic reasoning, and demands the reliable extraction, representation, and multi-step reasoning over structured tabular layouts rendered as images. Key research advances address challenges in table understanding, pipeline transparency, multi-lingual and noisy domains, and alignment between vision-LLMs and explicit reasoning toolchains.

1. Formal Problem Definition and Benchmark Datasets

TableVQA requires a system to map a pair $(T, Q)$—where $T \in \mathcal{I}$ is a table image and $Q \in \mathcal{Q}$ a natural-language question—to an answer $A \in \mathcal{A}$ (Yutong et al., 8 Oct 2025). Typical objectives minimize the negative log-likelihood $-\frac{1}{N}\sum_{i=1}^N \log P(A_i | T_i, Q_i)$ over annotated datasets. Exact-match accuracy is the prevailing metric, often supplemented with F1 or STS-based overlap for open-ended or list-style answers (Gautam et al., 13 Apr 2026).

Major benchmarks for TableVQA include:

Dataset	#Images	QA Pairs	Features
TableVQA-Bench	894	1,500	4 domains, real & synthetic
DenTab	2,000	2,208	Real-world estimates, HTML
ComTQA	1,591	9,070	Arithmetic, logical inference
INDOTABVQA	1,593	6,372	4 languages, 3 table styles
MirageTVQA	~30,000	58,480	24 languages, visual noise
ReTabVQA	60	120	Multi-step quantitative

TableVQA-Bench covers real and synthetic tables (WikiTableQuestions, TabFact, FinTabNet), while DenTab focuses on noisy, administrative tables with detailed role annotations. ComTQA and MirageTVQA stress-test reasoning in multilingual and imperfect settings (Kim et al., 2024, Hamdi et al., 17 Apr 2026, Zhao et al., 2024, Singh et al., 21 Nov 2025).

2. Methodological Paradigms and Model Architectures

Two methodological avenues dominate TableVQA:

• End-to-end vision–LLMs (VLMs): Directly predict answers from images and questions using joint vision–language architectures. Notables are GPT-4V, Gemini-ProV, Qwen-VL, TabPedia, and InternVL3 (Kim et al., 2024, Zhao et al., 2024). End-to-end models rely on encoder–decoder or transformer-based cross-attention between visual (e.g., ViT, Swin) features and language tokens.
• Hybrid structured pipelines: These decompose TableVQA into perception (table detection/recognition), serialization (CSV/HTML), explicit reasoning (either chain-of-thought or program synthesis), and programmatic execution. ExpliCIT-QA exemplifies this approach by chaining multimodal table understanding (with VLLM and CoT prompting), natural-language reasoning, code generation (Python/Pandas), deterministic execution, and interpretable natural-language explanation; all intermediate artifacts are exposed for inspection (Lagos et al., 15 Jul 2025). TALENT reframes the task as perception–narration, where a small VLM emits both natural-language and symbolic (Markdown/HTML) representations, and a downstream LLM is tasked with holistic reasoning (Yutong et al., 8 Oct 2025).

TabPedia introduces the “concept synergy” mechanism, fusing multi-grained visual embeddings (high-/low-res) and tasks (detection, parsing, QA) within a single LLM-centric transformer by using meditative tokens for flexible cross-attention and feature re-weighting (Zhao et al., 2024).

3. Reasoning, Transparency, and Tool Augmentation

Auditability and the traceability of reasoning are increasingly foregrounded in TableVQA, particularly for high-stakes domains:

• Chain-of-Thought (CoT) and Visual Chain-of-Thought: ReFocus introduces explicit visual reasoning by prompting models to iteratively edit table images—masking columns, highlighting rows—thus concretizing multihop selective attention. Each “thought” consists of a visual action and accompanying rationale, which guides VLMs to the answer with improved robustness, yielding up to 11% absolute gain on tabular tasks (Fu et al., 9 Jan 2025).
• Code-based Reasoning: ExpliCIT-QA translates CoT steps into transparent code (Python/Pandas), which is executed and explained. All intermediate steps—image-to-CSV extraction, stepwise CoT rationale, code, executed answer, and justification—enable external auditing and error localization, especially beneficial for domains needing audit trails such as finance or healthcare (Lagos et al., 15 Jul 2025).
• Arithmetic and Logic Routing: DenTab demonstrates that even when structure recognition (measured by S-TEDS) is strong (>90%), direct TableVQA accuracy on arithmetic and logic tasks lags. Its Table Router Pipeline categorizes question types, generates deterministic programs in a custom DSL for complex arithmetic/consistency cases, and executes them over the parsed table. This approach can boost arithmetic-related accuracy by up to +19.8 percentage points (Hamdi et al., 17 Apr 2026).
• LLM-Centric Design: TALENT empirically confirms that scaling the LLM in an OCR/description-to-LLM pipeline yields superior returns compared to scaling the VLM, supporting the architectural focus on language-based reasoning engines (Yutong et al., 8 Oct 2025).

4. Multilingual, Domain-Specific, and Robust TableVQA

The robustness and generalization of TableVQA systems are evaluated increasingly under adversarial, multilingual, and document-realistic conditions:

• Multilinguality: INDOTABVQA and MirageTVQA reveal severe cross-lingual drops, with accuracy reductions of 30–50 percentage points on non-English scripts, especially Hindi and Arabic (Gautam et al., 13 Apr 2026, Singh et al., 21 Nov 2025). Fine-tuning on small target-language datasets and leveraging spatial priors (table bounding boxes) incrementally increase accuracy (up to +17.8 point improvements).
• Structural and Visual Noise: MirageTVQA explicitly models realistic noise (blur, skew, scan artifacts) and finds modern VLMs can lose 5–35% absolute accuracy under noisy conditions, with English-first bias persisting (Singh et al., 21 Nov 2025).
• Table Layout Diversification: Datasets such as INDOTABVQA sample bordered, borderless, and colorful styles, finding that borderless tables present the greatest challenge, but spatial priors mitigate some of the difficulty (Gautam et al., 13 Apr 2026).

5. Evaluation, Limitations, and Open Challenges

Evaluation in TableVQA centers on exact-match accuracy, though F1 and STS metrics are increasingly used for more flexible answer sets (Kim et al., 2024, Gautam et al., 13 Apr 2026). Key observations highlight persistent bottlenecks:

• Perception vs. Reasoning Disentanglement: SOTA models may recover structure but fail at multi-step arithmetic and consistency even when given oracle HTML (perception-free) (Hamdi et al., 17 Apr 2026).
• Vision-Text Performance Gaps: LLMs over text-formatted tables outperform vision input MLLMs by 20–30 percentage points, with vision query (token) count and input resolution further influencing performance (Kim et al., 2024).
• Efficiency–Accuracy Trade-offs: Modular/hybrid pipelines (e.g., TALENT, Table Router) achieve comparable or superior accuracy to monolithic VLMs at far lower computational cost, critically enabling deployment in resource-constrained environments (Yutong et al., 8 Oct 2025).
• Transparency Over Raw Performance: Explicit, auditable pipelines (e.g., ExpliCIT-QA) may trail raw accuracy of giant end-to-end VLMs, but provide full justification for each answer, a requisite in sensitive or regulated domains (Lagos et al., 15 Jul 2025).

Remaining challenges include robust cross-lingual generalization, compositional logic over imperfect perceptions, efficient structure recognition under domain noise, and the need for finer-grained supervision and interpretability tools (Singh et al., 21 Nov 2025, Gautam et al., 13 Apr 2026).

6. Future Directions

Emerging trends in TableVQA research focus on:

• Hybrid neural-symbolic approaches for improved header and footnote handling (Lagos et al., 15 Jul 2025).
• Graph-based and richer intermediate representations (cell-wise concepts, table-grid graphs) to disentangle structure from content (Zhao et al., 2024).
• Scalable and multi-turn document reasoning to enable cross-table or page-level QA (Zhao et al., 2024).
• Robust augmentation during pretraining, including noise models and multilingual, script-diverse corpora (Singh et al., 21 Nov 2025).
• Constraint-based answer validation and adaptive pipeline routing for resource-efficient inference (Hamdi et al., 17 Apr 2026).
• Open, domain-driven benchmarks reflecting the spectrum of real-world administrative and scientific layouts (Hamdi et al., 17 Apr 2026, Zhao et al., 2024).

The field is moving swiftly toward end-to-end yet auditable architectures capable of fluent, transparent reasoning over visually diverse, multilingual, and noisy table images.