Sepideh Kharaghani

—

Unclaimed · AI-structured profileOpenAlex

NLP researcher in multimodal document AI: input representations, page stream segmentation, and LLM-based evaluation

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—

Not yet scoreable: no publication-count or venue metrics are in the current data. This axis would measure total and per-year research output; the record here is three 2026 arXiv preprints reviewed by Evolve.

Impact—

No citation or uptake data is available yet, so impact is unscored. This axis would measure scholarly influence — citations, downstream use, and adoption. All three catalog entries are recent 2026 preprints without accumulated citation evidence.

Focus0/100

Focus scores 0 on this radar, derived from a field distribution split evenly into thirds: natural language processing, document AI/page stream segmentation, and multimodal document AI. Because all three labels sit within one NLP-document AI cluster and the three reviewed papers share evaluation-method questions, topical concentration is likely higher than the raw zero score alone implies.

Collaboration41/100

Collaboration scores 39, computed from 6 distinct coauthors across the three reviewed papers. Every catalog entry places this researcher in a middle authorship position without corresponding-author role, consistent with contributing inside a small stable team — Nikhil Reddy Pottanigari, Ying Zhang, Saverio Vadacchino, Alejandro Posada, Ramin Fahimi, and Noah Bolger — rather than leading an independent group or an extended external network.

Activity—

Insufficient data to score: the axis measures publishing cadence and year-over-year research rhythm, but only a single publication year (2026) appears in the catalog, so no temporal trend or ongoing-activity signal can be computed.

Reproducibility55/100

Reproducibility scores 55, the mean across 3 papers carrying reproducibility assessments; all three are rated medium readiness. Evolve's verdicts converge on the same gaps — no significance testing in the document-QA comparison, single-run evidence for the MSBD segmentation method, and untested validity against human preference for the scaffold metrics — so the axis reflects careful setup held back by statistical rigor.

Research summary

Sepideh Kharaghani is a natural language processing researcher whose reviewed work centers on document AI: how multimodal inputs, document structure, and LLM-based evaluation behave in practice. Three 2026 arXiv preprints, all reviewed in the Evolve catalog, define the record to date. 'Text, Pixels, or Both?' compares text, page-image, and hybrid inputs for multimodal document question answering within a 50-page cap, finding page images lead on accuracy at growing latency cost. 'Splitting Documents at Lower Cost' introduces MSBD, which batches page-boundary decisions into a single LLM call for page stream segmentation, holding accuracy to a model-specific knee while reducing inference cost. 'Preserving What Matters' documents LLM-as-a-judge saturation in summarization evaluation and proposes semantic scaffold metrics (FPS, QPS, EPS) that retain scoring spread. A consistent thread is empirical care: controlled comparisons, explicit cost accounting, and honest reporting of limitations. Evolve's verdicts likewise flag a shared weakness — absent significance testing, single-run evidence, and untested validity against human preference — reflected in a reproducibility score of 55 (mean across 3 papers, all rated medium readiness). The researcher appears in a middle authorship position on all three papers, working with 6 distinct coauthors (collaboration score 39). Output, impact, and activity metrics are not yet available for this profile.

Collaboration pattern

All three reviewed papers are arXiv preprints with this researcher in a middle author position and no corresponding role, alongside 6 distinct coauthors. The recurring collaborator list — Nikhil Reddy Pottanigari, Ying Zhang, Saverio Vadacchino, Alejandro Posada, Ramin Fahimi, Noah Bolger — matches the 6-person coauthor pool, pointing to one small, stable working group producing evaluation- and efficiency-focused document AI studies rather than a wide external network.

Research keywords

document question answeringpage stream segmentationdocument boundary detectionsummarization evaluationLLM-as-a-judgemetric saturationsemantic scaffoldvision-language modelsOCRinput representationinference cost efficiencyzero-shot LLM evaluation

Highlights

Page images outperform text for document QA within a 50-page cap
In 'Text, Pixels, or Both?', a controlled comparison of text, page-image, and hybrid inputs shows page images lead on document QA accuracy, with latency cost growing alongside the page budget — honest, controlled work that stops short of significance testing.
MSBD cuts segmentation cost by batching boundary decisions into one LLM call
The 'Splitting Documents at Lower Cost' paper proposes Multi-Split Boundary Decisions, holding page stream segmentation accuracy to a model-specific knee while reducing inference cost; Evolve judges it plausible but notes the evidence comes from single runs without significance checks.
Scaffold metrics keep evaluation signal after LLM-judge saturation
'Preserving What Matters' shows LLM-as-a-judge saturates in summarization evaluation and proposes semantic scaffold metrics (FPS, QPS, EPS) that preserve scoring spread; Evolve views them as plausible diagnostics whose validity against human preference remains untested.

Timeline

2026All three Evolve-reviewed papers appear as arXiv preprints in a single year, 2026, forming the researcher's entire catalog record to date.
2026'Text, Pixels, or Both?' compares text, pixel, and hybrid input representations for multimodal document question answering within a 50-page cap.
2026'Splitting Documents at Lower Cost' introduces MSBD, batching page-boundary decisions into a single LLM call for page stream segmentation.
2026'Preserving What Matters' reports LLM-judge saturation in summarization evaluation and proposes the FPS/QPS/EPS semantic scaffold metrics.

Papers in the Evolve catalog

Frequent co-authors

Nikhil Reddy Pottanigari×3 Ying Zhang×3 Ying Zhang×2 Saverio Vadacchino×2 Alejandro Posada×2 Ramin Fahimi×1 Noah Bolger×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.

Identity is name-resolved from public bibliographic data; no ORCID, institution, or country is on record, so attribution to a specific 'Sepideh Kharaghani' carries disambiguation risk.
Bibliometric scores are computed from public bibliometric data only; no citation counts, h-index, or venue-level metrics were provided.
The profile rests on 3 Evolve-reviewed 2026 arXiv preprints; no peer-reviewed journal or conference record is included.
Output, impact, and activity dimensions are unscored because the underlying metrics were not available in the input data.