EaGERS: Explanation-Guided Region Selection for DocVQA

Updated 6 July 2026

EaGERS is a training-free, model-agnostic pipeline for Document Visual Question Answering that uses natural language explanations to guide region selection.
It decomposes DocVQA into explanation generation, region grounding via multimodal similarity and majority voting, and masked answer generation to enhance interpretability.
The approach improves transparency and reproducibility while reducing deployment costs, though it may incur increased inference time.

Searching arXiv for the specific term and closely related DocVQA grounding work. arXiv.search(query="EaGERS Document Visual Question Answering", max_results=5, sort_by="relevance") arXiv.search(query="\"Explanation-Guided Region Selection\" DocVQA", max_results=10, sort_by="relevance") arXiv.search(query="DocVQA vision LLM grounding explanations masked image", max_results=10, sort_by="relevance") EaGERS, short for Explanation-Guided Region Selection, is a fully training-free and model-agnostic pipeline for Document Visual Question Answering (DocVQA) that seeks to make answer generation more transparent and reproducible by forcing the model to answer only from document regions that are explicitly justified by a generated natural-language explanation (Lagos et al., 15 Jul 2025). Its central problem formulation is to enforce that the answer can be reconstructed solely from document regions that are explicitly grounded and verbalised, without any additional training of the VLM. The method therefore decomposes DocVQA into explanation generation, explanation-to-region grounding, and masked answer generation, using pretrained multimodal components rather than fine-tuning.

1. Concept and problem formulation

EaGERS addresses two practical problems identified for DocVQA systems: lack of interpretability / grounding, in which models output an answer without showing which part of the document justified it, and the need for task-specific training or fine-tuning, which can make deployment costly and brittle (Lagos et al., 15 Jul 2025). The method is designed around the question of how to constrain answer generation so that it depends only on visually localized evidence that has first been described in natural language.

The pipeline’s defining idea is to ask a vision-LLM not for an answer in the first pass, but for a natural language rationale describing where the answer should come from in the image. That rationale is then grounded to spatial sub-regions of the document, and the model is queried again on a version of the image in which only the selected regions remain visible. This design makes the explanation a semantic bridge between the question and the final masked-answer stage.

A plausible implication is that EaGERS treats explanation not as post hoc justification but as an operational constraint on inference. In the reported formulation, transparency arises because the answer is generated only after irrelevant regions have been removed, rather than because the model merely emits an accompanying explanation.

2. Three-stage pipeline

EaGERS consists of three stages: Explanation generation, Region selection by multimodal similarity + majority voting, and Masked answer generation (Lagos et al., 15 Jul 2025). The experiments use Qwen2.5VL-3B as the base VLM, although the method is described as model-agnostic.

In the first stage, the input is a document image $I$ and a question $q$ , and the output is an explanation or rationale $e$ . The explanation is intended to be spatially informative, as in descriptions such as “the invoice number is near the top right corner” or “the date appears under the heading.” The paper emphasizes that this explanation is not the final answer. Its role is to indicate where relevant evidence is likely to reside. A limitation already identified at this stage is that the explanation may be inaccurate and can therefore direct later stages toward irrelevant regions.

In the second stage, the document image is partitioned into a configurable uniform grid of $m \times n$ cells, yielding $m \cdot n$ sub-regions. The reported experiments use $5 \times 5$ and $5 \times 10$ grids, corresponding to 25 and 50 cells. The rationale is embedded and compared against each region crop using multiple pretrained multimodal encoders, and the final set of retained regions is obtained by ranking and voting.

In the third stage, EaGERS creates a masked version of the original image in which all grid cells outside the selected set are filled with black. The same VLM is then queried again with the question $q$ and the masked image $I_{\text{mask}}$ . The explanation is not explicitly provided in this second pass; its influence is mediated entirely through the spatial mask.

3. Grounding mechanism and evidence selection

The grounding stage uses three pretrained multimodal models: BLIP, CLIP, and ALIGN (Lagos et al., 15 Jul 2025). Each grid cell is represented as an image crop and embedded by each model, while the rationale $e$ is also embedded into the corresponding shared latent space. For each embedder $q$ 0, if $q$ 1 is the embedding of the explanation and $q$ 2 is the embedding of region $q$ 3, the similarity is computed using cosine similarity:

$q$ 4

For each embedder, regions are ranked by cosine similarity, and the top $q$ 5 regions are selected, with $q$ 6 of the total number of grid cells, rounded up. The paper describes this setting as a trade-off between spatial granularity and the inclusion of irrelevant context. The final selected set $q$ 7 is obtained by majority voting across the three embedders. If there is a tie, it is resolved using the average cosine similarity across all embedders, favoring regions with more consistent relevance.

The authors motivate this ensemble construction as a robustness mechanism: each pretrained vision-LLM may carry different biases, and combining them helps mitigate model-specific failures across heterogeneous document layouts. They also note that the use of a fixed grid improves reproducibility, although an adaptive grid might better fit irregular layouts. This suggests that EaGERS prioritizes deterministic and portable behavior over layout-adaptive spatial precision.

4. Training-free and model-agnostic properties

EaGERS is described as training-free because it does not require parameter updates, fine-tuning, or extra supervision (Lagos et al., 15 Jul 2025). It relies instead on pretrained VLMs for explanation and answer generation, pretrained multimodal encoders for region grounding, and simple similarity computations with voting. No additional training objective is introduced.

It is described as model-agnostic because the pipeline does not depend on a special architecture or task-specific head. In principle, any VLM that can produce a rationale and accept an image-plus-question input for answering can be used within the framework. Although the reported experiments use Qwen2.5-VL-3B, the method is intended for general-purpose multimodal models.

The stated benefits are lower deployment cost than fine-tuning, better reproducibility, improved transparency, easier integration into enterprise pipelines, and the absence of any need for dedicated DocVQA-specific training data. A plausible implication is that EaGERS is positioned as an inference-time control layer rather than a new end-to-end trained architecture.

5. Experimental setup and reported results

Experiments were run on the validation split of the DocVQA Single Page dataset (Lagos et al., 15 Jul 2025). Images were resized while preserving aspect ratio to reduce inference cost. Two grid configurations were tested—25 cells $q$ 8 and 50 cells $q$ 9—and a margin expansion of 15\% around the unmasked sub-region was also evaluated. The paper reports Exact Match (EM) and Average Normalized Levenshtein Similarity (ANLS).

The main reported results are as follows:

Model	EM (%)	ANLS
$e$ 0	64.20	75.25
$e$ 1	66.67	77.94
$e$ 2	72.72	82.52
$e$ 3	74.50	83.31
Qwen2.5-VL-3B	71.17	82.90

The best-performing configuration is $e$ 4, with EM = 74.50\% and ANLS = 83.31. Compared with direct Qwen2.5-VL-3B inference, this corresponds to an Exact Match gain of 3.33 points and an ANLS gain of 0.41 points.

The configuration study reported in the paper yields three main findings. First, adding a 15\% masking margin consistently improves performance across both grid sizes. Second, finer granularity helps, since the 50-cell grid outperforms the 25-cell grid in both margin settings. Third, the best trade-off is 50 cells + 15\% margin, which the authors interpret as evidence that small overlaps mitigate boundary issues when relevant text lies near cell borders.

6. Transparency, reproducibility, limitations, and disambiguation

The paper attributes four main strengths to EaGERS: transparency, because the answer is derived from identified regions; reproducibility, because the pipeline uses a fixed grid and deterministic selection; traceability, because explanations are linked to visible regions; and performance, because the best setup yields modest gains over the base model (Lagos et al., 15 Jul 2025). At the same time, several limitations are stated explicitly.

The first limitation is that a fixed grid may not fit irregular document layouts, and the paper notes that adaptive segmentation or object detection might work better. The second is dependence on explanation quality: if the VLM rationale is wrong, region selection may also be wrong. The third is increased inference cost, since the pipeline requires multiple model passes and is therefore slower than direct inference. The fourth is the need for stronger fidelity evaluation: the current experiments focus on EM and ANLS rather than direct localization fidelity.

The future directions named in the paper are adaptive grid partitioning, more robust fusion methods such as Reciprocal Rank Fusion (RRF), agreement analysis with Krippendorff’s alpha $e$ 5, evaluation on datasets with answer-localization annotations, quantitative spatial metrics like IoU, and user studies on explanation quality.

A recurring source of confusion is nomenclature. EaGERS in (Lagos et al., 15 Jul 2025) denotes Explanation-Guided Region Selection for DocVQA. It is distinct from EAGLE, an evidence-aligned multi-agent VQA method mentioned in the abstract of (Wang et al., 29 May 2026), although the details provided for that arXiv entry indicate that the supplied document was actually an ACL template rather than a substantive method paper. It is also distinct from EAGER, which refers to Embedding-Assisted Entity Resolution for Knowledge Graphs (Obraczka et al., 2021), and from EGS, also described as EaGERS in one summary, which concerns Edge Generation Scheduling for real-time DAG tasks (Sun et al., 2023). This suggests that “EaGERS” is not yet a uniquely stabilized acronym across arXiv-indexed work, and precise citation by title or domain is necessary to avoid conflation.