DD-VQA+: Enhanced Visual Question Answering

• DD-VQA+ is an umbrella term for VQA tasks that require multi-sensor fusion, discrete reasoning, and structural simplification beyond traditional image recognition.
• It integrates diverse methods including adverse-driving scenarios, rich document processing, long-document evidence aggregation, and degradation-driven prompting.
• These approaches collectively enhance robustness and accuracy, addressing challenges like sim-to-real transfer, multi-scale reasoning, and modality fusion.

Searching arXiv for the cited works to ground the article in current paper metadata. arxiv_search(query="(Tao et al., 11 Mar 2026) DriveXQA Cross-modal Visual Question Answering adverse driving", max_results=5) arxiv_search(query="DriveXQA Cross-modal Visual Question Answering for Adverse Driving Scene Understanding", max_results=10) DD-VQA+ is best understood as an umbrella designation for enriched visual question answering settings in which answering requires more than single-image recognition or shallow language grounding. In the literature summarized here, the label is instantiated in at least four ways: as cross-modal adverse-driving VQA with multi-sensor fusion, as document VQA requiring discrete reasoning over visually rich pages, as agentic evidence aggregation over long or multi-document corpora, and as degradation-driven prompting that deliberately reduces image fidelity to emphasize structural cues (Tao et al., 11 Mar 2026, Zhu et al., 2022, Zheng et al., 15 Apr 2026, Suri et al., 2024, Han et al., 6 Apr 2026). This suggests that DD-VQA+ is not a single canonical benchmark name, but a family of VQA formulations that extend conventional VQA with stronger robustness, grounding, retrieval, or reasoning requirements.

1. Conceptual scope and terminological usage

Across these works, the common denominator is not a fixed dataset or architecture, but an expanded task definition. The “+” consistently denotes additional structure beyond ordinary VQA: multiple sensing channels in driving, explicit extraction-versus-reasoning decisions in documents, sequential tool use and working memory in long-document settings, or deliberate perceptual simplification through image degradation. In all cases, the answer remains natural-language or answer-string prediction, but the path to that answer is constrained by nontrivial evidence selection or multimodal alignment.

Strand	Added capability	Representative work
Adverse-scene VQA	Multi-sensor fusion under weather and sensor failure	DriveXQA / MVX-LLM
Document VQA	Extraction plus discrete reasoning	TAT-DQA / MHST
Long-document QA	Agentic evidence aggregation	Doc-$V^*$
Multi-document QA	Parallel visual and textual RAG with modality fusion	VisDoMRAG
Perceptual VQA	Degradation-driven prompting with tools and ICL	DDP

A common misconception is that DD-VQA+ necessarily refers to document QA alone. The adverse-driving formulation contradicts that reading, because it uses the same “enhanced VQA” logic in a sensor-fusion setting rather than a page-understanding setting. Another misconception is that the “+” always means supplying more information. The degradation-driven line instead argues that performance can improve when high-frequency detail is intentionally suppressed.

2. Cross-modal adverse-driving VQA

In autonomous driving, DD-VQA+ is concretely realized by DriveXQA, a multimodal dataset and benchmark for adverse driving scene understanding, together with the MVX-LLM architecture for token-efficient fusion (Tao et al., 11 Mar 2026). DriveXQA contains 7,885 frames and 102,505 QA pairs, with an 80% train / 10% validation / 10% test split. Each scene has exactly 13 QA pairs, stratified into 2 global scene questions, 8 allocentric questions, and 3 ego-vehicle centric questions. Manual quality control corrects misidentified weather, vehicle counts, and ego-centric questions such as lane occupancy by cross-referencing LiDAR point clouds and multi-view RGB imagery.

The dataset explicitly targets degraded perception. It includes four complementary sensor streams: RGB cameras from four viewpoints, depth cameras from four viewpoints, event cameras from four viewpoints, and LiDAR point clouds. It also covers five weather conditions—rainy, night, foggy, cloudy, and sunny—and five sensor failure cases: Motion Blur, Overexposure, Underexposure, LiDAR Jitter, and Event Low-resolution. The hierarchical question design separates holistic scene reasoning from allocentric spatial reasoning and ego-vehicle safety reasoning. Answers are free-form natural language rather than multiple-choice labels.

MVX-LLM operationalizes the fusion problem through a Dual Cross-Attention projector. RGB, depth, and event images are processed by a shared CLIP-based ViT, while LiDAR is processed by PointNet++. Each visual modality begins with a $7 \times 7$ spatial grid, or 49 tokens, and is aligned to 48 tokens before fusion:

$$F_{\text{rgb}}' = F_{\text{rgb}} W_{\text{align}}^{\text{rgb}} \in \mathbb{R}^{B \times 48 \times 512}$$

with analogous mappings for depth and event features. Spatial cross-attention uses RGB as queries and concatenated depth and event tokens as keys and values:

$$F_s = \text{MultiHeadAttn}_s(Q_{\text{rgb}}, K_{\text{multi}}, V_{\text{multi}})$$

while channel cross-attention operates on transposed features:

$$F_c = \text{MultiHeadAttn}_c(Q_{\text{rgb}}^\top, K_{\text{multi}}', V_{\text{multi}}')$$

and the two paths are fused as

$F_{\text{fused}} = 0.5 \times (F_s + (F_c)^\top).$

LiDAR is fused downstream via token aggregation. Among aggregation heads, QAttn is the implementation default and achieves the best GPTScore of 55.5.

The benchmark emphasizes robustness rather than nominal-scene performance. Overall, MVX-LLM obtains 55.5 GPTScore and 58.1 human score, compared with 52.8 for the best token fusion baseline and 48.2 for Honeybee projector with GAP. Under fog, a highlighted difficult condition, MVX-LLM reaches 53.5 versus 25.1 for the “prepend” baseline. The attention ablation shows 24.0 for sAttn alone, 20.4 for cAttn alone, 50.5 for sAttn + cAttn, and 55.5 for sAttn + cAttn + QAttn. The modality ablation likewise indicates best synergy with all four modalities: RGB + Depth + Event + LiDAR yields 55.5, whereas RGB + Depth gives 23.4 and Depth + Event + LiDAR gives 39.7. Training uses cross-entropy loss, Adam with learning rate $1 \mathrm{e}{-4}$, batch size 16, LoRA rank 16, and distributed mixed-precision training on 4 $\times$ NVIDIA A100 40GB GPUs.

3. Discrete reasoning over visually rich documents

In document-centered DD-VQA+, the defining extension is discrete reasoning over pages that contain both semi-structured tables and unstructured text. TAT-DQA, introduced as an extension of TAT-QA, consists of 3,067 document pages and 16,558 question-answer pairs sampled from real-world financial reports, and it is explicitly designed so that arithmetic, comparison, counting, sorting, and related operations are necessary rather than incidental (Zhu et al., 2022).

The task differs from ordinary DocVQA because answerability depends on combining layout, document image, and textual evidence with numerical operations. Financial reports distribute relevant information across tables, footnotes, and prose. A question may therefore require multi-hop evidence alignment, unit normalization, aggregation across rows or columns, or comparisons across years. The answer space includes strings, numbers, and dates, but the key distinction is not answer type; it is whether the system can identify when extraction suffices and when explicit reasoning is required.

MHST is described as a multimodal model that ingests text, layout, and visual image, and chooses between two strategies. The extraction strategy directly finds answer spans or strings such as names, dates, and categorical facts. The reasoning strategy performs numerical or logical operations over grounded evidence, such as difference, ratio, minimum, maximum, comparison, counting, or sorting. The paper reports that MHST significantly outperforms baseline methods, but also states that performance still lags far behind that of expert humans. That gap is important: it indicates that DD-VQA+ in the document domain is not merely an OCR problem, but a joint grounding-and-reasoning problem in which the document’s visual organization remains semantically active.

A plausible implication is that TAT-DQA formalizes a boundary case for VQA: the visual signal is neither a natural scene nor a pure text rendering, but a structured artifact whose semantics are distributed across typography, layout, and numerical content. This is why extraction-only formulations are insufficient.

4. Agentic retrieval and evidence aggregation in long-document QA

A second major branch of DD-VQA+ moves from single-page reasoning to interactive evidence acquisition. Doc-$V^*$ recasts multi-page Document VQA as a sequential decision process in which an OCR-free agent alternates between semantic retrieval, targeted page fetching, and answer generation while maintaining a structured working memory (Zheng et al., 15 Apr 2026). The action space contains <retrieval_page>, <fetch_page>, and <answer>, and the policy is instantiated by Qwen-2.5-VL-7B-Instruct in a ReAct-style “think–act” protocol. The system begins from a thumbnail overview in which pages are resized to $256 \times 256$, grouped up to $7 \times 7$0 per overview image, and arranged into near-square grids. High-resolution page tokens are fetched only when needed.

The training objective combines imitation learning and reinforcement learning. The trajectory reward is

$7 \times 7$1

with $7 \times 7$2, $7 \times 7$3, and $7 \times 7$4. Stage I uses 9,019 expert trajectories, and Stage II applies GRPO on 2,048 stratified examples. Inference uses greedy decoding, a fixed horizon $7 \times 7$5, and adaptive retrieval budget $7 \times 7$6. Empirically, Doc-$7 \times 7$7 (GRPO) reaches 64.5 ANLS on DUDE, 86.2 ANLS on MP-DocVQA, 77.2 F1 on SlideVQA, 42.1 accuracy on MMLongBench-Doc, and 56.3 accuracy on LongDocURL. Relative to RAG Top-5 at the same 7B scale, gains include +12.3 on DUDE, +8.8 on MP-DocVQA, and +18.5 on LongDocURL. In the $7 \times 7$8-page regime, it scores 40.7 versus 30.9 for RAG and 21.9 for All Pages.

VisDoMRAG addresses a related but distinct problem: multi-document QA over visually rich corpora containing tables, charts, and slides (Suri et al., 2024). VisDoMBench comprises 2,271 queries and 1,277 documents, with an average of 8.36 documents per query and 128.69 pages per query. The method runs textual and visual RAG in parallel, then performs consistency-constrained modality fusion:

$7 \times 7$9

Textual retrieval uses OCR via PyTesseract, 3000-character chunks with 10% overlap, and BGE-1.5 embeddings; visual retrieval uses page-level indexing with ColQwen2 or ColPali. Evidence curation, chain-of-thought reasoning, and answer generation are performed separately per modality before fusion.

On end-to-end QA, the representative average Word Overlap F1 scores are 12.09 for Qwen2-VL long-context, 31.65 for text RAG with Qwen2-VL, 39.06 for visual RAG with Qwen2-VL, and 39.94 for VisDoMRAG with Qwen2-VL. With GPT-4o, visual RAG scores 49.02 and VisDoMRAG scores 50.01. The paper summarizes the overall advantage as 12–20% over unimodal and long-context baselines. The methodological significance is that DD-VQA+ here becomes a retrieval-and-fusion problem across both modality and document boundaries.

5. Degradation-driven prompting as a VQA augmentation

A different interpretation of DD-VQA+ appears in Degradation-Driven Prompting, where the central claim is that reducing visual detail can improve VQA reasoning (Han et al., 6 Apr 2026). The framework applies degradation twice within a three-stage agentic pipeline: task classification, tool management, and target prompting with a critic. The first degradation step downsamples the image to approximately 150 pixels in maximum dimension for tool application; the second imposes an 80-pixel structural bottleneck for final reasoning.

For physical attributes, “80p” is an absolute resolution cap rather than 80% scaling. If the original image size is $$F_{\text{rgb}}' = F_{\text{rgb}} W_{\text{align}}^{\text{rgb}} \in \mathbb{R}^{B \times 48 \times 512}$$0, the downsampling factor is

$$F_{\text{rgb}}' = F_{\text{rgb}} W_{\text{align}}^{\text{rgb}} \in \mathbb{R}^{B \times 48 \times 512}$$1

and the resized image is $$F_{\text{rgb}}' = F_{\text{rgb}} W_{\text{align}}^{\text{rgb}} \in \mathbb{R}^{B \times 48 \times 512}$$2. Additional visual tools include white background masks,

$$F_{\text{rgb}}' = F_{\text{rgb}} W_{\text{align}}^{\text{rgb}} \in \mathbb{R}^{B \times 48 \times 512}$$3

orthometric lines, red boxes, crops, heavy Gaussian blur, and contrast enhancement. For perceptual phenomena, the framework distinguishes VA, CI, MI, GI, GSI, and VI, then routes the input through specialized tool combinations. In-context learning exemplars are used especially for physical-attribute tasks.

The empirical claim is that “less is more.” DDP with Gemini-3-Pro reaches 92.1 on MMBench, 94.5 on SEED-Bench, 99.1 on ScienceQA, and 89.4 on VQAv2, compared with 88.4, 87.2, 98.1, and 85.6 respectively for Gemini-3-Pro alone. On V*Bench, DDP obtains 89.3 overall, with 89.2 on Attribute and 89.5 on Spatial. The ablation is particularly revealing: removing degradation drops V*Bench overall performance from 89.3 to 80.6, a larger reduction than removing visual tools or prompt engineering. On ColorBlind, Gemini-3-Pro scores 15.33% Pass@1, whereas DDP with Gemini-3-Pro scores 29.33%.

This branch of DD-VQA+ reverses an otherwise common assumption in VQA research. More pixels are not always more evidence. When the target property is geometric, contextual, or illusory, high-resolution texture can act as confounding signal rather than useful information.

6. Evaluation regimes, limitations, and open problems

The evaluation regimes across DD-VQA+ variants are heterogeneous because the failure modes are heterogeneous. DriveXQA uses BLEU-4, ROUGE-L, METEOR, CIDEr, sentence embedding-based similarity, and GPTScore normalized to a 0–100 scale with

$$F_{\text{rgb}}' = F_{\text{rgb}} W_{\text{align}}^{\text{rgb}} \in \mathbb{R}^{B \times 48 \times 512}$$4

and reports strong correlation between GPT-4o-mini scores and four human evaluators on 360 samples, with Spearman $$F_{\text{rgb}}' = F_{\text{rgb}} W_{\text{align}}^{\text{rgb}} \in \mathbb{R}^{B \times 48 \times 512}$$5 (Tao et al., 11 Mar 2026). TAT-DQA emphasizes benchmark construction for discrete reasoning and reports that model performance remains substantially below expert humans (Zhu et al., 2022). Doc-$$F_{\text{rgb}}' = F_{\text{rgb}} W_{\text{align}}^{\text{rgb}} \in \mathbb{R}^{B \times 48 \times 512}$$6 uses ANLS, F1, and accuracy depending on benchmark (Zheng et al., 15 Apr 2026). VisDoMBench uses Word Overlap F1 for end-to-end QA and ANLCS for retrieval and evidence extraction (Suri et al., 2024). DDP primarily reports accuracy and Pass@1 (Han et al., 6 Apr 2026).

The limitations are similarly domain-specific. DriveXQA is simulation-based and single-frame, so sim-to-real transfer and temporal reasoning remain open. TAT-DQA exposes persistent problems in OCR robustness, table parsing, and numerical grounding. Doc-$$F_{\text{rgb}}' = F_{\text{rgb}} W_{\text{align}}^{\text{rgb}} \in \mathbb{R}^{B \times 48 \times 512}$$7 has been evaluated on one backbone, leaves multi-document scenarios unexplored, and still admits hallucinated outputs. VisDoMRAG depends on multiple retrieval and generation stages, which raises latency and composes errors from OCR, retrieval, and fusion. DDP can fail when fine-grained high-frequency detail is itself the answer, and its behavior is sensitive to tool choice and degradation parameters.

Taken together, these works suggest that DD-VQA+ is converging on a broader research program rather than a single benchmark lineage. Its core question is how to make VQA systems reliable when evidence is distributed across sensors, pages, documents, or visual scales. The answers currently explored include token-efficient cross-modal fusion, explicit extraction-versus-reasoning decomposition, agentic evidence gathering with working memory, modality-consistent retrieval-augmented generation, and deliberate visual simplification. The field’s open problems therefore include sim-to-real robustness, temporal and multi-page compositionality, evidence attribution under long-context constraints, and principled control over when more detail helps and when less detail yields better answers.