Saverio Vadacchino

—

Unclaimed · AI-structured profileOpenAlex

Document AI researcher: cost-efficient page stream segmentation and multimodal input representations for document QA

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-13) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

Publication-volume metrics are not yet available for this profile. This axis would measure the number of outputs over time, venue types, and authorship roles; the two 2026 arXiv preprints in the Evolve catalog are the only counted items so far.

Impact—

Citation and influence data are insufficient for this axis. It would capture citation counts, downstream adoption, and referencing work; no such statistics are provided in the source bibliometric snapshot, so no score is estimated.

Focus0/100

The focus score is 0, computed from an even field distribution: 0.5 in document AI / natural language processing (page stream segmentation) and 0.5 in natural language processing (multimodal document AI). Work currently spans two threads rather than concentrating in one. Notably, both threads sit inside document AI and share vision-language models, so the spread is narrow in substance even though the distribution itself is perfectly balanced.

Collaboration36/100

The collaboration score is 36, based on 5 distinct coauthors in the source data. Both cataloged papers carry a middle authorship position with no corresponding-author role. Frequent coauthors include Nikhil Reddy Pottanigari, Sepideh Kharaghani, Alejandro Posada, and Ying Zhang, suggesting a compact, stable working group producing arXiv preprints together.

Activity—

Publishing-cadence data are not yet available. This axis would measure output frequency and consistency over time, such as papers per year and gaps between releases; with only two 2026 preprints on record, the timeline is too short to characterize activity.

Reproducibility55/100

The reproducibility score is a mean of 55 across 2 scored papers, both rated medium on reproducibility readiness. Evolve's verdicts flag the same limitation in each: the input-representation comparison and the MSBD batching method both rest on single-run evidence without significance testing, which caps confidence even as the experimental setups are described as controlled and honest.

Research summary

Saverio Vadacchino is a researcher working at the intersection of document AI and natural language processing. This Evolve profile is built from two arXiv preprints released in 2026, and the field distribution splits evenly at 0.5/0.5 between two adjacent threads: document AI for page stream segmentation and multimodal document AI. The first cataloged paper, 'Text, Pixels, or Both?', evaluates text, image, and hybrid input representations for document question answering within a 50-page cap, finding that page images improve accuracy at growing latency cost; Evolve characterizes the work as controlled and honest, but lacking significance testing. The second introduces MSBD, which batches page-boundary decisions for page stream segmentation into a single LLM call, holding accuracy near a model-specific knee, again with single-run evidence. Both papers carry medium reproducibility readiness, giving a reproducibility mean of 55 over 2 scored papers. Vadacchino holds a middle authorship position on both papers without a corresponding-author role, and the collaboration footprint is compact: 5 distinct coauthors, including Nikhil Reddy Pottanigari, Sepideh Kharaghani, Alejandro Posada, and Ying Zhang. Output, impact, and activity metrics are not yet available for this profile.

Collaboration pattern

The recorded pattern is a small, stable team: 5 distinct coauthors in the source data, with Nikhil Reddy Pottanigari, Sepideh Kharaghani, Alejandro Posada, and Ying Zhang recurring across the catalog. Vadacchino takes a middle authorship position on both 2026 arXiv preprints and is not the corresponding author on either, consistent with contributing within a compact group rather than leading a large lab network; all cataloged outputs so far appear on arXiv.

Research keywords

document question answeringinput representationvision-language modelsOCRlatency-cost trade-offLLM-as-a-judgepage stream segmentationdocument boundary detectionzero-shot LLM evaluationmulti-instance promptinginference cost efficiency

Highlights

Input representations for multimodal document QA
2026 arXiv study 'Text, Pixels, or Both?' compares text, pixel, and hybrid inputs for document question answering within a 50-page cap, finding page images lead on accuracy at growing latency cost; Evolve calls it controlled, honest work, but with no significance testing.
MSBD: batching page-boundary decisions into one LLM call
'Splitting Documents at Lower Cost' (2026) introduces Multi-Split Boundary Decisions, batching page-boundary decisions for LLM-based page stream segmentation into a single LLM call while holding accuracy to a model-specific knee; Evolve finds it plausible, but the single-run evidence lacks significance checks.
A recurring evaluation-rigor gap
Both 2026 preprints share the same flagged limitation: evidence without significance testing, from the document QA comparison to MSBD's single-run results. Both are rated medium on reproducibility readiness, yielding a reproducibility mean of 55 across 2 papers.

Timeline

2026Released 'Text, Pixels, or Both? Evaluating Input Representations for Multimodal Document QA' on arXiv, in a middle authorship position.
2026Released 'Splitting Documents at Lower Cost' on arXiv, introducing MSBD for LLM-based page stream segmentation, again in a middle authorship position.
2026Both preprints entered the Evolve catalog with medium reproducibility readiness, producing a reproducibility mean of 55 across 2 papers.

Papers in the Evolve catalog

Frequent co-authors

Nikhil Reddy Pottanigari×2 Sepideh Kharaghani×2 Alejandro Posada×2 Ying Zhang×2 Ying Zhang×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometric scores and raw metrics are computed from public bibliographic data only; no institution, country, or ORCID was provided, so this profile is name-resolved.
Only the two 2026 arXiv preprints in the Evolve catalog are counted; output, impact, and activity axes are missing because no volume or citation data is available.
The reproducibility mean of 55 covers only 2 scored papers, both rated medium, and Evolve's verdicts note absent significance testing in both.
Both papers list a middle authorship position with no corresponding role, so research-leadership status should not be inferred from this data.