First-author NLP researcher in mechanistic interpretability, LLM code generation, and low-resource evaluation rigor
3
papers here
3
reviewed by Evolve
3
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-20) 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—
Output volume cannot be scored yet: the profile lacks verified publication counts, venue records, and yearly productivity totals, so this axis would measure how many peer-reviewed works this researcher produces per year once such indexing data is added.
Impact—
Impact is unscored because citation counts, h-index, and downstream uptake are absent from the current snapshot; this axis would measure how often the three 2026 preprints are cited and adopted once citation indexing becomes available.
Focus0/100
Focus scores 0 because the field distribution is perfectly flat: three fields — parameter-efficient fine-tuning with mechanistic interpretability, NLP for code / LLM code generation, and general NLP — each hold about 33% of the reviewed output. The portfolio is broad rather than concentrated in one niche, so the zero reflects even spread across NLP sub-areas, not inactivity.
Collaboration32/100
Collaboration scores 32 with 4 distinct coauthors recorded. All three catalogued papers are first-authored and none is corresponding, suggesting a small, stable working cluster rather than a wide network; recurring names include Phat T. Tran-Truong, Xuan-Bach Le, and Nghia Duong-Trung. The modest score marks a compact co-author pool, which keeps the 2026 output consistent but limits external exposure.
Activity—
Activity over time cannot be charted yet: the profile has no annual publication counts or rolling productivity window beyond the three 2026 preprints, so this axis would measure publishing cadence and recent-output trends once multi-year indexing data exists.
Reproducibility85/100
Reproducibility is the strongest axis, averaging 85 across the 3 reviewed papers, two rated high readiness and one medium. MechSparse's negative result and the audit-quality counterfactual study carry high readiness, while LangSelect is tempered by GPT-5-only evidence and a verbose Python baseline — a pattern of documented, re-runnable evaluation setups.
Research summary
Son Ha Xuan is an early-stage NLP researcher whose public footprint, as resolved from bibliographic data, rests on three 2026 arXiv preprints — each with first-author position, none yet marked corresponding. The papers cut across parameter-efficient fine-tuning with mechanistic interpretability (MechSparse), cost-aware LLM code generation (LangSelect), and synthetic-data selection for low-resource African NLP, with the field distribution an even three-way split (~33% each). A unifying signature is evaluation rigor: MechSparse is framed as a well-instrumented negative result in which mechanism-guided sparse PEFT selection never beats cheap heuristics such as activation-norm; the audit-quality study shows selector rankings decoupling from downstream F1 across 8 African-language cells; LangSelect reports a 50.3% cost headline whose evidence base (GPT-5-only, verbose Python baseline) the review flags. Reproducibility is the strongest measured dimension — a mean of 85 across 3 reviewed papers, two rated high readiness. Collaboration is comparatively thin (score 32, 4 distinct coauthors), with recurring partners including Phat T. Tran-Truong, Xuan-Bach Le, and Nghia Duong-Trung. Output, impact, and activity metrics are not yet available, so citation-based standing cannot be assessed.
Collaboration pattern
The recorded pattern is a compact, stable cluster: 4 distinct coauthors, with Phat T. Tran-Truong, Xuan-Bach Le, and Nghia Duong-Trung recurring. Son Ha Xuan takes first-author position on all three 2026 papers, none yet corresponding, consistent with an early-career researcher driving each project within a small working group.
Research keywords
Highlights
MechSparse: a well-instrumented negative result on sparse PEFT selection
LangSelect: target-language routing as a real code-gen cost lever
Audit rankings decouple from downstream utility in African-language NLP
Timeline
- 2026Publishes MechSparse, a first-author arXiv study showing that mechanism-guided sparse PEFT selection never beats cheap heuristics such as activation-norm.
- 2026Publishes LangSelect, a first-author arXiv study on cost-aware target-language routing for LLM code generation with a 50.3% headline figure.
- 2026Publishes a first-author counterfactual study of synthetic-data selectors for low-resource African NLP, spanning 8 language cells.
Papers in the Evolve catalog
- first author
MechSparse: Mechanism-Guided Sparse PEFT Selection Is Task-Shaped
- first author
LangSelect: Cost-Aware Target-Language Routing for LLM Code Generation
- first author
When Audit Quality Fails to Predict Downstream Utility: A Counterfactual Study of Synthetic-Data Selectors for Low-Resource African NLP
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 scores are computed from the platform's OpenAlex-based snapshot only; output, impact, and activity axes are missing because no citation or productivity index is attached to this profile.
- No institution, country, or ORCID is provided; the profile is name-resolved from public bibliographic data, and the name Son Ha Xuan may match multiple researchers.
- All catalogued works are 2026 arXiv preprints; peer-reviewed venues, awards, and career facts are outside the current data scope.
- Reproducibility (85) is computed over only 3 reviewed papers; field and collaboration metrics cover the same small sample.