Xuan-Bach Le

—

Unclaimed · AI-structured profileOpenAlex

NLP researcher in mechanism-guided PEFT, cost-aware LLM code generation, and low-resource data selection

papers here

reviewed by Evolve

first-author

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 cannot be scored yet: the underlying bibliometric feed offers no verified publication-count or productivity series beyond the three arXiv papers in the Evolve catalog. This axis would measure total and per-year publication volume once the author record is fully linked.

Impact—

Impact is unscored because citation counts, h-index, and downstream uptake are absent from the current snapshot. This axis would measure citation accumulation and influence of the three cataloged papers once citation data are linked to this author record.

Focus0/100

Focus scores 0, reflecting an almost perfectly even topical spread: the field distribution assigns about one-third each to mechanism-guided parameter-efficient fine-tuning, NLP/machine learning for code, and general natural language processing. Across the three cataloged papers no single area dominates, so the profile reads as deliberately multi-front rather than narrowly specialized, spanning interpretability methods, code-generation systems, and African-language data selection.

Collaboration32/100

Collaboration scores 32 on a compact footprint: four distinct coauthors appear across the three cataloged papers, with repeated collaboration involving Son Ha Xuan, Phat T. Tran-Truong, and Nghia Duong-Trung. Authorship positions alternate between last (twice) and middle (once), and no paper flags this researcher as corresponding author, suggesting contribution within a small, stable team rather than an extended network.

Activity—

Activity cannot be scored: the snapshot lacks a verified yearly publication series, so a recent output rate cannot be computed. All three cataloged papers are dated 2026, which hints at current momentum, but a single-year window is too narrow to measure sustained research activity.

Reproducibility85/100

Reproducibility is the strongest measured axis, with a mean score of 85 across three assessed papers. MechSparse and the African-language synthetic-data study carry high reproducibility readiness, while LangSelect is rated medium, reflecting its GPT-5-only evidence. The pattern — instrumented negative results and transparent counterfactual accounting — points to work built for verification rather than headline claims.

Research summary

Xuan-Bach Le is a natural language processing researcher whose reviewed work spans three near-equally weighted areas — parameter-efficient fine-tuning with mechanistic interpretability, NLP/machine learning for code, and broader low-resource NLP — each accounting for about one-third of the cataloged output. The Evolve catalog contains three arXiv papers, all dated 2026: MechSparse (last author), LangSelect (middle author), and a counterfactual study of synthetic-data selectors for low-resource African NLP (last author); none carries a corresponding-author flag. A unifying theme is rigorous negative and counterfactual results: mechanism-guided sparse PEFT selection never beat activation-norm heuristics on primary metrics, audit rankings decoupled from downstream F1 across eight African-language cells, and target-language routing was framed as a real but carefully hedged cost lever given GPT-5-only evidence. Reproducibility is the strongest measured dimension, with a mean readiness score of 85 across three assessed papers (two rated high, one medium). The collaboration footprint is compact — four distinct coauthors, with repeated work alongside Son Ha Xuan, Phat T. Tran-Truong, and Nghia Duong-Trung — for a collaboration score of 32, while the focus score of 0 reflects an even three-way split across fields. Output, impact, and activity axes lack sufficient bibliometric data and are left unscored.

Collaboration pattern

The collaboration footprint is deliberately small: four distinct coauthors across three papers, yielding a collaboration score of 32. Recurring partners are Son Ha Xuan, Phat T. Tran-Truong, and Nghia Duong-Trung. Authorship alternates between last position (MechSparse, the African-NLP study) and middle position (LangSelect), with no corresponding-author flag in the catalog — a pattern consistent with steady contribution inside a compact, stable working group rather than network expansion.

Research keywords

parameter-efficient fine-tuningLoRA / QLoRAactivation patchingmechanistic interpretabilitysparse mask selectionLLM code generationtarget-language routingcost-aware inferencesynthetic data selectionLLM-as-judgelow-resource NLPAfrican languages NLP

Highlights

Mechanism-guided sparse PEFT selection fails to beat cheap heuristics
Evolve's review of MechSparse (last-author position, 2026) calls it a well-instrumented negative result: activation-patching-guided sparse mask selection never beats cheap heuristics such as activation-norm on primary metrics, and the paper carries high reproducibility readiness.
Target-language routing emerges as a real code-generation cost lever
LangSelect (middle author, 2026) frames cost-aware target-language routing for LLM code generation, with careful accounting, LinUCB-based routing, and verification-gated fallback; Evolve tempers the 50.3% headline given GPT-5-only evidence and a verbose Python baseline, rating reproducibility readiness medium.
Synthetic-data audit quality decouples from downstream utility in African NLP
A counterfactual study (last author, 2026) finds synthetic-data audit rankings fail to predict downstream F1 across eight African-language cells; Evolve deems it a transparent, credible case study, though underpowered at a single seed, with high reproducibility readiness.

Timeline

2026Publishes MechSparse, showing that mechanism-guided sparse PEFT selection is task-shaped and never beats activation-norm heuristics on primary metrics.
2026Contributes to LangSelect, introducing cost-aware target-language routing with LinUCB and verification-gated fallback for LLM code generation.
2026Publishes a counterfactual study showing synthetic-data audit rankings decouple from downstream F1 across eight African-language cells.

Papers in the Evolve catalog

Frequent co-authors

Son Ha Xuan×3 Phat T. Tran-Truong×3 Nghia Duong-Trung×1 Phat T. Tran-Truong×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.

Radar scores are derived from public bibliometric data covering only the three Evolve-reviewed arXiv papers; no external publication record is linked.
Institution, country, and ORCID are unavailable, so this profile is name-resolved from public bibliographic data.
Output, impact, and activity axes are unscored because the underlying bibliometric data is insufficient.
All cataloged papers are 2026 arXiv preprints; no peer-reviewed venue information is present in the input data.