NLP researcher spanning parameter-efficient fine-tuning, mechanistic interpretability, and LLM code generation.
3
papers here
3
reviewed by Evolve
0
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—
Insufficient data: no aggregate publication counts or venue history are available yet. This axis would measure total research output — papers per year and their venues — across the researcher's full record, beyond the three cataloged arXiv preprints.
Impact—
Insufficient data: no citation counts, h-index, or venue-level influence metrics are available. This axis would measure scholarly impact — citations, referencing works, and downstream use of the three 2026 preprints — once such data exists.
Focus0/100
Focus scores 0 because the raw field distribution is split exactly evenly — about 0.33 each — across parameter-efficient fine-tuning with mechanistic interpretability, LLM code generation, and general NLP. The three papers share an evaluation-and-measurement methodology rather than one topical lane, so current topical concentration is minimal despite a consistent empirical style.
Collaboration32/100
Collaboration scores 32 from a raw count of 4 distinct coauthors across 3 papers — a compact network. All cataloged papers place the researcher in a middle author position with no corresponding role. Son Ha Xuan, Xuan-Bach Le, and Nghia Duong-Trung recur as frequent collaborators, indicating a stable small team rather than a broad external network.
Activity—
Insufficient data: all three cataloged papers are dated 2026, leaving no multi-year record. This axis would measure publication cadence — output frequency, submission consistency, and research continuity — and cannot be scored from a single year.
Reproducibility85/100
Reproducibility is the strongest measured axis, with a mean of 85 across 3 papers: MechSparse and the African-language counterfactual study rate high, while LangSelect rates medium, consistent with its GPT-5-only evidence and verbose-baseline caveats. The overall pattern points to careful instrumentation and transparent reporting of experimental settings.
Research summary
Phat T. Tran-Truong is an early-stage natural language processing researcher whose Evolve-reviewed record to date comprises three arXiv preprints, all dated 2026, in each of which they appear as a middle author rather than first or corresponding author. The cataloged work splits evenly across three areas — parameter-efficient fine-tuning with mechanistic interpretability, LLM code generation, and low-resource African NLP — each accounting for roughly a third of the field distribution. A consistent thread is rigorous, measurement-first evaluation: MechSparse is a well-instrumented negative result showing that activation-patching-guided sparse PEFT selection never beats cheap activation-norm heuristics; LangSelect frames target-language routing as a real cost lever for LLM code generation, though Evolve notes that GPT-5-only evidence and a verbose Python baseline temper the 50.3% headline; and the African NLP study finds synthetic-data audit rankings decoupling from downstream F1 across eight language cells. Reproducibility is the strongest measured axis, averaging 85 across three papers (two high, one medium readiness). With output, impact, and activity unscored, this profile should be read as provisional.
Collaboration pattern
Across the three cataloged papers, Phat T. Tran-Truong appears as a middle author within a compact team of 4 distinct coauthors (collaboration score 32). Son Ha Xuan, Xuan-Bach Le, and Nghia Duong-Trung recur as frequent collaborators, suggesting a stable core group that rotates topics across PEFT interpretability, code generation, and low-resource NLP rather than a wide external network. No corresponding-author role appears in the cataloged work.
Research keywords
Highlights
A well-instrumented negative result in mechanism-guided PEFT
Target-language routing as a cost lever for LLM code generation
Synthetic-data audits fail to predict downstream utility
Timeline
- 2026Co-authored MechSparse, reporting that activation-patching-guided sparse PEFT selection never beats activation-norm heuristics on primary metrics.
- 2026Co-authored LangSelect, introducing cost-aware target-language routing with contextual-bandit (LinUCB) selection and verification-gated fallback for LLM code generation.
- 2026Co-authored a counterfactual study of synthetic-data selectors for low-resource African NLP, finding that audit quality decouples from downstream F1.
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 scores are computed from a small OpenAlex-derived snapshot covering only 3 Evolve-reviewed arXiv preprints; they may not reflect the full publication record.
- No institution, country, ORCID, or citation data was provided; the profile is name-resolved from public bibliographic data, and 'Phat T. Tran-Truong' may be ambiguous.
- Output, impact, and activity axes are unscored because the underlying data is missing, not because performance is poor.
- All cataloged papers are 2026 arXiv preprints with middle authorship; career-stage and author-contribution claims are not supported by the data.