FPT University researcher in educational AI: sub-3B essay scoring and bias audits of knowledge tracing.
2
papers here
0
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-28) 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—
No output score yet: the OpenAlex record lists h-index 0, i10-index 0 and just 3 recent works, too thin to compute this axis. It would measure publication volume and consistency across the researcher's career.
Impact—
No impact score yet: with h-index 0 and i10-index 0 on OpenAlex, citation uptake cannot be assessed. This axis would measure citation-based influence — how widely the essay-scoring and knowledge-tracing audits are taken up by later work.
Focus—
No focus score yet: OpenAlex logs five topics — topic modeling, NLP techniques, graph algorithms, AI ethics, and stochastic gradient optimization — each with a count of 1, so thematic concentration cannot be computed. This axis would measure consistency of research themes.
Collaboration32/100
Collaboration scores 32 on a 0-100 scale built from real bibliometric data: the OpenAlex record lists 4 distinct coauthors, matching Evolve's frequent-coauthor set — Dang Quang Minh, Nguyen Huu Loi, Truong Viet Vu, and Nguyen Thai Anh. Both catalog papers are co-authored, with first authorship on the essay-scoring benchmark and a middle position on the knowledge-tracing audit, suggesting a small, stable, likely FPT University-based group rather than a large distributed network.
Activity40/100
Activity scores 40: the recent 3-year window contains 3 works, and countsByYear logs all 3 in 2026 alone, with no earlier years recorded. Output is concentrated in a single recent burst — two arXiv preprints in the Evolve catalog plus one further work — indicating an emerging but currently active publishing rhythm.
Reproducibility—
No reproducibility score yet: Evolve rates both catalog papers' reproduction readiness as medium, but the axis itself needs code releases and replication evidence to compute. It would measure how independently others can rerun the reported essay-scoring and knowledge-tracing experiments.
Research summary
Nguyen Dung Son is an early-career researcher at FPT University (Vietnam) working at the intersection of natural language processing and educational artificial intelligence. The Evolve catalog contains two 2026 arXiv papers. As first author, Son led a benchmark asking how far sub-3B open language models can go in zero-shot essay scoring on an 8 GB consumer GPU, testing whether rubric-decomposed prompting rescues small local models on the ASAP-AES benchmark; the best quadratic weighted kappa reached 0.388 — statistically careful work, but still behind a simple length-only baseline, a negative result reported candidly. The second paper, with a middle author position, delivers the first cross-architecture audit of demographic bias in deep knowledge tracing, using ABROCA and careful statistics while openly noting that headline p-values sit at the permutation floor. Together they suggest a research style centered on honest, resource-aware evaluation of educational AI rather than leaderboard chasing. The OpenAlex record — h-index 0, i10-index 0, 3 works in the recent 3-year window, 4 distinct coauthors — confirms an emerging profile: collaboration and activity are scored (32 and 40), while output, impact, focus, and reproducibility lack sufficient data. Logged topics span topic modeling, NLP techniques, graph algorithms, AI ethics, and stochastic gradient optimization, indicating breadth during a formative phase.
Collaboration pattern
With 4 distinct coauthors on OpenAlex, Nguyen Dung Son works in a compact team matching Evolve's frequent-coauthor list: Dang Quang Minh, Nguyen Huu Loi, Truong Viet Vu, and Nguyen Thai Anh. Authorship roles vary — first author on the essay-scoring benchmark, middle position on the knowledge-tracing bias audit — consistent with a small, FPT University-anchored group where roles rotate. The collaboration score of 32 reflects limited network breadth so far, not absent collaboration.
Research keywords
Highlights
Sub-3B open models on one consumer GPU: a candid zero-shot essay scoring benchmark
First cross-architecture audit of demographic bias in deep knowledge tracing
Honest negative results as an early signature
Both Evolve-reviewed papers are described as candid and statistically careful, and both report limits openly — a best QWK of 0.388 losing to a length-only baseline, and headline p-values at the permutation floor. This transparent evaluation style is a distinctive marker of an emerging profile.
Timeline
- 2026First-author arXiv benchmark on zero-shot essay scoring with sub-3B open language models on an 8 GB consumer GPU (best QWK 0.388).
- 2026Co-authored the first cross-architecture audit of demographic bias in deep knowledge tracing, using ABROCA.
- 2026OpenAlex logs 3 works in the recent 3-year window with 4 distinct coauthors; h-index and i10-index both at 0, reflecting an emerging record.
Papers in the Evolve catalog
Frequent co-authors
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 metrics are drawn from OpenAlex only; no ORCID is linked, so identity resolution is name-based on public bibliographic data.
- The record is early-stage: h-index 0 and i10-index 0 mean the output, impact, focus and reproducibility axes are unscored due to insufficient data, not evidence of low quality.
- Qualitative assessment rests on 2 Evolve-reviewed papers, both 2026 arXiv preprints; peer-reviewed venue information is absent.
- countsByYear contains only 2026 (3 works), so longer-term productivity trends cannot be characterized.