Martin Takáč

Mohamed bin Zayed University of Artificial Intelligence · United States

Unclaimed · AI-structured profileOpenAlex

Optimization researcher at MBZUAI building reproducible benchmarks for diffusion optimizers and LLM defenses

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) 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 scored: this identity's snapshot lacks usable publication-volume data — indexed per-year counts exist only for 2022 and 2026. This axis would measure total works, annual output, and venue spread across the full record.

Impact—

Not yet scored: citation coverage is too sparse to evaluate — the record shows h-index 1 and i10-index 0. This axis would measure total citations, h-index trajectory, and downstream uptake of the work.

Focus100/100

Focus is the record's strongest signal (score 100). Topic weights concentrate on Topic Modeling (7) and Stochastic Gradient Optimization Techniques (6), followed by Advanced Optimization Algorithms Research (4), with smaller clusters in NLP Techniques (3) and Adversarial Robustness in Machine Learning (3). The field distribution and both catalog papers — diffusion-optimizer benchmarking and LLM copyright-defense evaluation — stay within this optimization-plus-evaluation lane.

Collaboration53/100

Collaboration scores 53: the recent three-year window lists 12 works with 13 distinct coauthors, plus 12 recurring collaborators including Salem Lahlou, Sebastian U. Stich, Eric Moulines, Maxim Panov, Nils Lukas, and Arman Bolatov. The pattern is team-based benchmark building: a last-author (senior) placement on the diffusion-optimizer benchmark and a middle-author placement on CopyShield suggest supervision combined with hands-on contribution.

Activity75/100

Activity scores 75, grounded in recent3yWorks = 12: twelve works within the recent three-year window. Per-year counts are lopsided — a single indexed work in 2022 versus 12 in 2026 — indicating a currently active, benchmark-producing record rather than a dormant one; both catalog papers appeared as 2026 arXiv preprints.

Reproducibility100/100

Reproducibility is a top-scoring axis (100): reproMean is 100 across 1 assessed paper (reproPapers = 1), and both Evolve-reviewed papers carry 'high' reproducibility readiness. The diffusion-optimizer benchmark is described as 'well-run and reproducible', and CopyShield is noted for a rigorous shared protocol and honest caveats. The sample is small but consistent — quality on few papers, not volume.

Research summary

Martin Takáč is profiled here as a machine-learning researcher affiliated with the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), with an identity record name-resolved from OpenAlex public bibliographic data. The indexed snapshot shows a sharply concentrated topic profile: Topic Modeling (7) and Stochastic Gradient Optimization Techniques (6) lead, supported by Advanced Optimization Algorithms Research (4), NLP Techniques (3), and Adversarial Robustness in Machine Learning (3). Activity is current and dense — 12 works in the recent three-year window with 13 distinct coauthors — earning a maximum focus score (100) and a maximum reproducibility score (100). The Evolve catalog adds two 2026 arXiv benchmark papers. In 'Optimizers for Diffusion Models: A Controlled Benchmark' Takáč is the last author on what Evolve assesses as the first controlled optimizer benchmark for discrete diffusion: well-run and reproducible, though text-task budgets are only 3% of published training length. In 'CopyShield' he is a middle author on a cross-level benchmark of three LLM copyright defenses with a rigorous shared protocol and honest caveats. Citation-impact data is too sparse to score on this identity (h-index 1, i10-index 0), so the profile reads as a benchmark-building, reproducibility-first record rather than a citation-ranked one.

Collaboration pattern

Collaboration scores 53: the recent three-year window lists 12 works with 13 distinct coauthors, plus 12 recurring collaborators including Salem Lahlou, Sebastian U. Stich, Eric Moulines, Maxim Panov, Nils Lukas, and Arman Bolatov. The pattern is team-based benchmark building — last-author (senior) placement on the diffusion-optimizer benchmark and middle-author placement on CopyShield suggest supervision combined with hands-on contribution rather than solo or honorary roles.

Research keywords

Diffusion modelsDiscrete diffusionOptimizer benchmarkingStochastic gradient optimizationTopic modelingLLM memorizationCopyright defenseContrastive decodingAdversarial robustnessReproducible machine learning

Highlights

First controlled optimizer benchmark for discrete diffusion
As last author on 'Optimizers for Diffusion Models: A Controlled Benchmark' (arXiv 2026), Takáč anchors a study Evolve assesses as the first controlled optimizer benchmark for discrete diffusion — well-run and reproducible (high readiness), with the stated limitation that text-task budgets are only 3% of published training length.
CopyShield: cross-level benchmark of LLM copyright defenses
A middle-author contribution to CopyShield (arXiv 2026), a controlled cross-level benchmark of three LLM copyright defenses. Evolve notes a rigorous shared protocol and honest caveats, with the limitation that a single method per level constrains generality; reproducibility readiness is high.
A reproducibility-first benchmark record
Both 2026 catalog papers are benchmark studies rated 'high' for reproducibility readiness, and the snapshot's reproducibility mean is 100 (on 1 assessed paper). The consistent pattern across both papers: controlled comparisons, shared protocols, and explicit limitation reporting.

Timeline

2022Single indexed work in the OpenAlex snapshot's per-year counts — the earliest recorded activity in this identity record.
2026Publication of 'Optimizers for Diffusion Models: A Controlled Benchmark' (arXiv), with Takáč in the last-author position.
2026Publication of 'CopyShield: A Cross-Level Benchmark of Copyright Defenses in LLMs' (arXiv), middle-author position.
2026Peak of the record: 12 works in the recent three-year window with 13 distinct coauthors; focus and reproducibility both score 100.

Papers in the Evolve catalog

2026-09-19
Optimizers for Diffusion Models: A Controlled Benchmark
2026-09-01
CopyShield: A Cross-Level Benchmark of Copyright Defenses in LLMs

Frequent co-authors

Maryam Alshehyari×1 Dushyant Singh Chauhan×1 Samuele Poppi×1 Salem Lahlou×1 Nils Lukas×1 Arman Bolatov×1 Egor Shulgin×1 David Li×1 Abduragim Shtanchaev×1 Sebastian U. Stich×1 Maxim Panov×1 Eric Moulines×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 come from a single OpenAlex snapshot (identity: 'openalex'); indexed per-year counts cover only 2022 and 2026, with h-index 1 and i10-index 0 — the record likely captures only a fraction of full career output.
No ORCID is on file; the profile is name-resolved from public bibliographic data and should be identity-verified before reuse.
Output and impact dimensions are not scored because citation coverage in the snapshot is insufficient.
The reproducibility score (100) rests on a small sample: 1 assessed paper (reproPapers = 1) plus 2 Evolve-reviewed catalog papers.