DocExplainerV0: Modular QA Grounding

• DocExplainerV0 is a plug-and-play spatial grounding module for document QA that separates answer generation from evidence localization for improved interpretability.
• It employs a dual-branch design using a frozen SigLIP2 backbone and a lightweight bounding-box regressor, ensuring compatibility with diverse VLMs.
• The system significantly boosts spatial localization metrics compared to prompting-only methods, emphasizing the need for explicit grounding in reliable document QA.

DocExplainerV0 is a plug-and-play spatial grounding module for document question answering with vision-LLMs. It is designed to address a specific interpretability gap in document QA: modern VLMs can often return a correct answer string while failing to indicate where on the page the supporting evidence is located. The system therefore decouples answer generation from spatial localization, using the base VLM for textual answering and a separate bounding-box predictor for grounding. In the formulation introduced in "Towards Reliable and Interpretable Document Question Answering via VLMs" (Chen et al., 12 Sep 2025), the central claim is that document QA is not fully interpretable unless it can both answer correctly and localize the supporting evidence.

1. Problem formulation and interpretability target

DocExplainerV0 is motivated by the observation that current generative VLMs are optimized mainly for autoregressive next-token prediction rather than for explicit spatial reasoning or bounding-box prediction. This produces a mismatch between textual correctness and spatial grounding: a model may answer a question such as “What is the total amount due?” correctly, yet fail to identify where that amount appears in the document image. In document workflows, that failure matters because users typically require not only extracted content but also source highlighting and page-level traceability (Chen et al., 12 Sep 2025).

Within this framing, spatial localization is treated as a first-class requirement of interpretability. A bounding box provides traceability, allows inspection of the source evidence, supports robustness checks when the answer is correct but the box is wrong, and matches enterprise document-processing requirements in which provenance is operationally necessary. This suggests that answer quality alone is an insufficient proxy for reliability in document QA.

The module is explicitly scoped to document QA with VLMs rather than to general document classification or structured extraction. Its task is to recover the answer region corresponding to a question-answer pair, not to explain the entire document representation. The paper therefore positions DocExplainerV0 as both a model component and an evaluation framework for the gap between textual answering and grounded evidence.

2. Architectural design and localization model

The architecture of DocExplainerV0 is deliberately decoupled. A VLM generates the answer text, and a separate localization module predicts the answer bounding box. This separation is the core design choice: instead of requiring a single model to jointly learn answer generation and grounding, DocExplainerV0 treats localization as an independent module that can be attached to existing systems, including proprietary models where fine-tuning is not feasible (Chen et al., 12 Sep 2025).

The bounding-box regressor is built on top of a frozen SigLIP2 vision-language encoder. Its inputs are the document image and the question text. Visual and textual embeddings are projected into a shared latent space, after which fully connected layers and a regression head predict the box coordinates. The predicted box is parameterized as

$[x_1, y_1, x_2, y_2]$

with normalized coordinates in $[0,1]$ during training and page-coordinate mapping at evaluation time.

The dual-branch structure shown in the paper consists of a visual branch, a text branch, latent-space fusion, and coordinate regression. Because the SigLIP2 backbone is frozen and only the regressor is trained, the localization module is lightweight relative to full end-to-end VLM retraining. A plausible implication is that the module is intended to preserve compatibility with heterogeneous upstream answer generators rather than to replace them.

This design distinguishes DocExplainerV0 from systems that integrate explanation into the answering backbone itself. Rather than forcing the VLM to emit both answer and spatial justification directly, it introduces a post-answer grounding step that remains model-agnostic at the level of the answering system.

3. Training objective and inference pipeline

DocExplainerV0 is trained only for localization. The paper uses BoundingDocs v2.0 for supervision, trains for 20 epochs on a single NVIDIA L40S-1-48G GPU, and selects the checkpoint with the highest mean IoU on validation (Chen et al., 12 Sep 2025).

The regression loss is Smooth L1 loss, or Huber loss, applied to normalized box coordinates: $$L(x,y) = \begin{cases} 0.5 (x-y)^2 & \text{if } |x-y| < 1, \ |x-y| - 0.5 & \text{otherwise}. \end{cases}$$ The paper presents this as a localization-only training objective; the base VLM is not retrained as part of this stage.

Inference is modular. First, the VLM is prompted with the document image and question. Second, it returns the answer text. Third, the answer and/or question-document representation is passed to DocExplainerV0. Fourth, the module predicts the bounding box. The system then returns a JSON object with "content" for the answer and "position" for the coordinates. The prompting format used in the experiments asks VLMs to respond in JSON with both answer and position, using coordinates in the range $[0,1000]$ to match the dataset annotation scale.

The paper evaluates three prompting strategies for the base VLMs: zero-shot prompting, chain-of-thought prompting with QA examples, and anchor-based prompting with OCR-extracted words and positions. Anchor-based prompting is the best prompting strategy for localization, while chain-of-thought tends to help answer accuracy most. The limited localization gains from prompting are a central empirical argument for the decoupled module: prompting alone does not reliably induce spatial grounding.

4. Dataset, metrics, and empirical results

The benchmark used for DocExplainerV0 is BoundingDocs v2.0, described as containing 48,151 documents, 237,437 pages, and 249,016 QA pairs across 8 languages and multiple document types including invoices, contracts, forms, receipts, and multilingual corpora. The paper notes that every entry includes a rephrased_question field, that XFUND questions were rephrased with Claude 3.7 Haiku for multilingual consistency, and that image–OCR alignment issues in some MP-DocVQA entries were fixed. Evaluation is explicitly restricted to questions whose answers can be localized to a single bounding box, excluding cases that require multi-region reasoning (Chen et al., 12 Sep 2025).

Two metric families are used. Textual correctness is evaluated with Average Normalized Levenshtein Similarity, or ANLS, in the range $[0,1]$. Spatial grounding is evaluated with IoU and MeanIoU: $$\text{IoU} = \frac{|B_\text{pred} \cap B_\text{gt}|}{|B_\text{pred} \cup B_\text{gt}|}.$$ This pairing is central to the paper’s thesis: document QA systems should be judged jointly on correctness and grounding.

The reported results show a large gap between textual performance and localization quality for prompting-only VLMs. For SmolVLM-2.2B, zero-shot yields ANLS 0.527 and MeanIoU 0.011, anchors yield 0.543 and 0.026, and CoT yields 0.561 and 0.011. For Qwen2.5-VL-7B, zero-shot yields 0.691 and 0.048, anchors 0.694 and 0.051, and CoT 0.720 and 0.038. Claude Sonnet 4 under zero-shot prompting reaches ANLS 0.737 and MeanIoU 0.031. These figures support the paper’s claim that correct answers frequently lack reliable localization.

Adding DocExplainerV0 improves localization substantially. SmolVLM with DocExplainerV0 reaches ANLS 0.572 and MeanIoU 0.175. Qwen with DocExplainerV0 reaches ANLS 0.689 and MeanIoU 0.188. The central numerical claim is therefore the large MeanIoU increase relative to prompting-only grounding, especially from 0.011 to 0.175 for Smol and from 0.048 to 0.188 for Qwen.

The paper also includes a naive OCR-based baseline: ask the VLM for the answer, search that string in OCR text, use the first word if no exact match is found, and return the matched OCR span’s box. On this benchmark, the OCR baseline remains stronger in MeanIoU, reaching 0.405 for Smol + Naive OCR and 0.494 for Qwen + Naive OCR. The authors attribute this to the fact that BoundingDocs often contains exact answer strings in OCR text. They nevertheless argue that OCR search is brittle and does not generalize to inferred answers that are not literally present in the document.

5. Position within explainable DocVQA research

DocExplainerV0 belongs to a broader shift in explainable document question answering from black-box answering toward explicit grounding. Its specific contribution is the decoupling of answer generation from localization. In that respect it differs from self-explainable DocVQA systems that make explanation part of the prediction pipeline itself (Souibgui et al., 12 May 2025).

DocVXQA, for example, learns a relevance mask $M$ over the document image, constructs the bottleneck representation $T = X \odot M$, and answers from the masked image. Its explanations are trained to be contextually sufficient and representation-efficient, using a multi-objective loss with cross-entropy, sparsity, and alignment to a ColPali prior (Souibgui et al., 12 May 2025). CoExVQA goes further by separating two explanation roles: a question-evidence heatmap, an answer-region box, and answer decoding exclusively from the grounded region. Its chain-of-explanation design makes the decoder see only the predicted answer region, creating a hard information bottleneck, and on PFL-DocVQA its crop variant reaches 0.78 ANLS versus 0.66 for DocVXQA (Indrehus et al., 7 May 2026).

Against that background, DocExplainerV0 occupies a different point in the design space. It does not claim self-explainability by construction in the same sense as systems that restrict downstream decoding to grounded evidence. Instead, it offers compatibility with existing VLMs, including proprietary systems such as Claude Sonnet 4, and can be attached after the fact. This suggests a division between two explainability regimes in document QA: integrated explainability, where explanation is architecturally coupled to answer prediction, and modular explainability, where grounding is added as an external capability. DocExplainerV0 is a clear instance of the latter.

The paper also contributes a standardized evaluation perspective. By benchmarking open and closed VLMs on both ANLS and MeanIoU, it quantifies the gap between answer quality and grounding quality rather than assuming that the former implies the latter. That diagnostic framing is one of its main research contributions.

6. Limitations, failure modes, and significance

The paper is explicit that DocExplainerV0 is not a complete solution. The most immediate limitation is that it still underperforms the naive OCR baseline on the current benchmark. Another limitation is the evaluation restriction to single-box answer cases, which excludes multi-region reasoning even though many real document questions require evidence from several regions. The authors also note that BoundingDocs v2.0 favors OCR-based methods because answers often have exact textual matches, and that localization remains hard even when the textual answer is correct (Chen et al., 12 Sep 2025).

The dominant failure pattern identified in the experiments is a correct textual answer paired with a near-random or clearly wrong box. DocExplainerV0 reduces that problem but does not eliminate it. This suggests that spatial grounding in document QA is not a trivial byproduct of good multimodal generation. Rather, it is a separate competence that must be modeled and evaluated directly.

The broader significance of DocExplainerV0 lies in its reframing of document QA reliability. A system that produces only the correct answer string is not necessarily trustworthy if it cannot show where that answer came from. By separating answer generation from localization, freezing a SigLIP2 backbone for box regression, and evaluating both answer correctness and grounding quality, DocExplainerV0 establishes a benchmarked modular approach to interpretable document QA. Its main contribution is therefore not only the localization regressor itself, but also the argument that reliable document question answering requires both textual accuracy and verifiable spatial evidence.