Razvan-Gabriel Dumitru

University of Arizona · United States

Unclaimed · AI-structured profileOpenAlex

University of Arizona researcher bridging medical image segmentation and LLM agent reliability evaluation.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-15) 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—

This axis measures total publication volume and its growth over a researcher's career. The underlying full output data is not yet available in this profile — only the recent three-year count of 15 works is recorded — so the dimension is left unscored rather than estimated.

Impact—

This axis would quantify citation-based influence — citations, h-index trajectory and uptake by later work. The citation data needed to score it is not available in this snapshot beyond h-index 2 and i10-index 1, so it is left unscored rather than estimated.

Focus0/100

Focus scores 0/100 because the field distribution is an even 0.5/0.5 split between medical image analysis and LLM agent reliability and evaluation — two communities with different venues and methods. Topic records reinforce this breadth: topic modeling (12), advanced neural network applications (9), NLP techniques (8), domain adaptation and few-shot learning (3) and medical image segmentation (2). The spread reads as an exploratory early-career phase rather than a settled specialty.

Collaboration54/100

Collaboration scores 54/100, grounded in 14 distinct coauthors over the recent three-year window. The frequent-coauthor list divides into a recurring core (Darius Peteleaza, Bogdan Neamtu, Arpad Gellert, Mariana Sandu, Claudiu Matei) and a second circle (Harsh Raj, David Lee, Anas Mahmoud, Renxiong Wang, Chenguang Wang, Tong Zhao, Yunzhong He) that lines up with the two-field split. Both cataloged papers are middle-authored and not corresponding.

Activity81/100

Activity is the strongest axis at 81/100. OpenAlex counts rise monotonically across the four recorded years: 2 works in 2023, 4 in 2024, 5 in 2025 and 6 in 2026, cumulating to 15 works in the recent three-year window. Both cataloged 2026 arXiv preprints fall within this acceleration, one in medical image segmentation and one in LLM agent evaluation, indicating parallel output in both fields rather than a single-track sprint.

Reproducibility55/100

Reproducibility scores 55/100, a mean over the 2 Evolve-reviewed papers, both rated medium readiness. Evolve's verdicts flag concrete gaps on each: NeuroTS-Net reports top BraTS-PEDS Dice but no significance tests and a possible recipe confound; Continual Search shows F1 gains from 0.349 to 0.498 but notes prompt asymmetry and weak statistics. Results look promising, yet statistical rigor is the recurring reservation.

Research summary

Razvan-Gabriel Dumitru is a researcher at the University of Arizona whose recorded work splits evenly between two lines: deep-learning medical image analysis and the reliability and evaluation of LLM-based agents — OpenAlex field weights are 0.5 and 0.5. The bibliometric snapshot (identity: openalex; no ORCID) records an h-index of 2, an i10-index of 1, and 15 works over the recent three years, with yearly output rising from 2 in 2023 to 4 in 2024, 5 in 2025 and 6 in 2026 — an early-career ramp scored 81/100 on activity. Topic records show topic modeling (12) as the deepest thread, ahead of advanced neural network applications (9), NLP techniques (8), domain adaptation and few-shot learning (3) and medical image segmentation (2). Evolve has reviewed two 2026 arXiv preprints: NeuroTS-Net, a compact 3D encoder-decoder for multi-class pediatric brain tumor segmentation in multi-parametric MRI reporting top BraTS-PEDS Dice against nnU-Net and MedNeXt; and Continual Search, an iterative prompting scheme lifting LLM-judge root-cause F1 on long agent traces from 0.349 to 0.498. Both carry medium reproducibility readiness (mean 55/100 across 2 papers), with Evolve flagging absent significance testing in each. Collaboration scores 54/100 on 14 distinct coauthors in the recent window, and focus scores 0/100 on the even two-field split, consistent with an exploratory early-career profile.

Collaboration pattern

With 14 distinct coauthors in the recent three-year window, collaboration sits mid-range (54/100). The frequent-coauthor list splits into a recurring core — Darius Peteleaza, Bogdan Neamtu, Arpad Gellert, Mariana Sandu, Claudiu Matei — and a second circle including Harsh Raj, David Lee, Anas Mahmoud, Renxiong Wang, Chenguang Wang, Tong Zhao and Yunzhong He, roughly matching the two research lines. In both cataloged papers this researcher holds a middle author position and is not the corresponding author.

Research keywords

pediatric brain tumor segmentationmulti-class semantic segmentation3D convolutional neural networkmultiparametric MRILLM-as-judgeagent failure analysislong-horizon agentsroot-cause attributioniterative promptingagentic evaluationtopic modelingdomain adaptation and few-shot learning

Highlights

Top reported pediatric BraTS 2026 Dice with the smallest, fastest model
NeuroTS-Net, a 3D encoder-decoder for multi-class semantic segmentation of pediatric brain tumors in multi-modal MRI, reports the top published pediatric BraTS 2026 Dice against nnU-Net and MedNeXt while being the smallest and fastest model compared. Evolve calls the results strong but flags missing significance tests and a possible recipe confound; reproducibility is rated medium.
Turning LLM-judge root-cause attribution into an iterative search
Continual Search iteratively prompts LLM judges to keep searching for failure causes in long-horizon agent traces, lifting F1 from 0.349 to 0.498 on long traces. Evolve calls the gains solid but tempers them with prompt-asymmetry and weak-statistics concerns; reproducibility is rated medium.
Two-track 2026 output across segmentation and agent evaluation
The 2026 count of 6 works in the OpenAlex year-by-year series coincides with two Evolve-reviewed arXiv preprints in different fields — pediatric tumor segmentation and LLM agent failure attribution — reflecting the even 0.5/0.5 field split and the 81/100 activity score.

Timeline

2023First recorded output in the OpenAlex year-by-year series: 2 works.
2024Output doubles to 4 works in the OpenAlex year-by-year series.
20255 works recorded in the OpenAlex year-by-year series, continuing the upward trend.
20266 works so far — the highest recorded year — including both Evolve-reviewed arXiv preprints (NeuroTS-Net and Continual Search); the recent three-year window totals 15 works with 14 distinct coauthors.

Papers in the Evolve catalog

Frequent co-authors

Darius Peteleaza×1 Bogdan Neamtu×1 Arpad Gellert×1 Mariana Sandu×1 Claudiu Matei×1 Harsh Raj×1 David Lee×1 Anas Mahmoud×1 Renxiong Wang×1 Chenguang Wang×1 Tong Zhao×1 Yunzhong He×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.

Bibliometrics are computed from OpenAlex only; no ORCID is linked, so this profile is name-resolved from public bibliographic data and homonyms are possible.
Output and impact dimensions are unscored because the underlying data is missing from the snapshot; citation data beyond h-index 2 and i10-index 1 is not available.
Reproducibility (55/100) averages only 2 Evolve-reviewed papers, both arXiv preprints and not necessarily peer-reviewed.
Yearly counts reflect OpenAlex records, which may lag or omit works in non-indexed venues.