Machine learning researcher on emergent communication and compositional generalization in multi-agent language models
2
papers here
2
reviewed by Evolve
2
first-author
2
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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 scored: the OpenAlex record shows only four works (2020: 1; 2026: 3) with an h-index of 0, too thin to judge productivity or venue quality. This axis would measure publication volume, venues, and output consistency.
Impact—
Not scored: the h-index and i10-index are both 0 and no citation aggregates are available, so influence cannot yet be assessed. This axis would measure citations, uptake, and downstream reuse of the research.
Focus0/100
The recorded field distribution splits exactly 50/50 between two tightly allied subfields — multi-agent language models and compositional generalization, and emergent communication in multi-agent LLM systems — yielding a concentration score of 0, i.e. a deliberately narrow agenda around how language models transmit and reuse information. OpenAlex topics widen slightly outward (topic modeling, 3; domain adaptation and few-shot learning, 2; evolutionary algorithms, 2), hinting at earlier breadth beyond the current focus.
Collaboration0/100
The collaboration score is 0 because OpenAlex records zero distinct coauthors and the frequent-coauthor list is empty. Both Evolve-reviewed 2026 papers carry Marincat as first and corresponding author with no other recorded authors, a pattern consistent with solo scholarship — though thin metadata can also hide uncredited collaborators. With no co-authorship network on record, this axis says more about data coverage than about an aversion to teamwork.
Activity40/100
Activity scores 40, reflecting a sparse but re-igniting record: OpenAlex lists one work in 2020 and three in 2026, with three works within the last three years. The visible pattern is a long gap followed by a concentrated 2026 burst in which both Evolve-reviewed preprints appeared. Whether this marks a return to research or simply lagging metadata ingestion cannot be determined from the available counts.
Reproducibility100/100
This is the profile's strongest axis: reproducibility scores a perfect 100, the mean across 2 assessed papers, both rated high for reproducibility readiness. The methodology is explicit — one study is preregistered across sixty societies, and the other is described as rigorously sealed despite small scale. For an early-career record, a 100/100 on this axis is distinctive: the work prioritizes verifiability and transparent design over volume.
Research summary
Narcis Marincat is an early-stage machine-learning researcher whose reviewed work examines how language models communicate and generalize. Both Evolve-cataloged papers are 2026 arXiv preprints with Marincat as first and corresponding author: a preregistered sixty-society confirmation that evidence masking drives compositional generalization in a small language-model relay, and a study showing that latent communication interfaces between language-model cells harden into initialization-anchored dialects that act as severe negative-transfer priors. Both are rated high for reproducibility readiness, and the profile's perfect 100 reproducibility score (mean across 2 assessed papers) reflects that rigor-first, preregistration-driven methodology. The underlying OpenAlex record is thin: h-index 0, i10-index 0, four works split across 2020 (1) and 2026 (3), and zero distinct coauthors recorded, consistent with solo authorship or incomplete attribution. OpenAlex topics reach beyond the catalog — topic modeling (3), domain adaptation and few-shot learning (2), evolutionary algorithms (2), plus single counts in NLP techniques and neural networks — hinting at a broader or earlier footprint than the two 2026 papers alone. Output and impact remain unscored for insufficient data; this profile should be read as an early-stage snapshot anchored by two rigorously documented preprints.
Collaboration pattern
Every recorded signal points to solo work: OpenAlex lists zero distinct coauthors, the frequent-coauthor list is empty, and both cataloged 2026 papers carry Marincat as first and corresponding author with no other named authors. The sixty-society preregistered design implies broader infrastructure, yet no individual collaborators are recorded. The score of 0 should be read as a data-coverage limitation as much as a collaboration pattern.
Research keywords
Highlights
Preregistered sixty-society test: evidence masking drives compositional generalization
Latent dialects between language-model cells become negative-transfer priors
A reproducibility-first methodology at the start of a career
Both reviewed papers earn a perfect 100 mean reproducibility score across 2 assessed papers, with preregistration and sealed experimental design as the signature; verifiability is prioritized over output volume at this early stage of the record.
Timeline
- 2020Earliest recorded output in the OpenAlex profile: one publication, outside the Evolve catalog and predating the reviewed work.
- 2026Three works recorded in a single year, including both Evolve-reviewed preprints, each with Marincat as first and corresponding author.
- 2026Publishes the preregistered sixty-society study on evidence masking and compositional generalization on arXiv; reproducibility readiness rated high.
- 2026Publishes 'Portable Semantics, Private Dialects' on latent communication and negative transfer between language-model cells.
Papers in the Evolve catalog
- first authorcorresponding
What You Can't See Is Still What You Learn: A Preregistered Sixty-Society Confirmation That Evidence Masking Drives Compositional Generalization
- first authorcorresponding
Portable Semantics, Private Dialects: Reuse and Negative Transfer in Latent Communication Between Language-Model Cells
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 OpenAlex only; the h-index (0), i10-index (0), and coauthor counts reflect that record's coverage, not a verified full publication history.
- Identity is resolved from public bibliographic data; the 2020 work and the broader topic mix (topic modeling, evolutionary algorithms) sit outside the two reviewed papers and could reflect a namesake — no ORCID or affiliation was available to confirm.
- Output and impact dimensions are unscored due to insufficient citation data; only two papers carry full Evolve reviews.
- Both reviewed works are arXiv preprints; peer-reviewed venue status is not recorded in the available data.