Early-career researcher linking mechanistic interpretability of multilingual LLMs with generative recommender systems
3
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
1
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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 yet scored: the snapshot lacks reliable publication totals for this OpenAlex identity, so aggregate output cannot be computed. This axis would measure total works, per-year production, and venue spread across the full career.
Impact—
Not yet scored: citation coverage is too thin to interpret (h-index 1, i10-index 0), so influence is left unquantified. This axis would measure total citations, citation trajectory, and highly cited works once coverage matures.
Focus0/100
Scores 0/100 on concentration: the field split is exactly 50/50 between mechanistic interpretability of multilingual LLMs and recommender systems (IR/ML), and indexed topics spread across five clusters — topic modeling (4), NLP techniques (4), speech recognition and synthesis (4), recommender systems (3), and IR/search behavior (3). This reads as a deliberate dual-track profile rather than drift, but it dilutes any single-field signal.
Collaboration54/100
Scores 54/100 on a moderate network: 14 distinct coauthors across 19 works in the recent three-year window, with recurring partners including Nhu Vo, Phuong Nguyen, Nu Uyen Phuong Le, Wray Buntine, Massimo Piccardi, and Inigo Jauregi Unanue. Coauthorship spans both research tracks, and catalog positions (last, middle, middle) point to collaborative contribution rather than solo leadership so far.
Activity87/100
Scores 87/100 on a steep ramp: indexed works rose from 1 in 2024 to 4 in 2025 and 14 in 2026, totalling 19 works in the recent three-year window — a high tempo for an early-career profile. All three Evolve-catalog papers appeared in 2026 on arXiv, consistent with the acceleration shown in the by-year counts.
Reproducibility100/100
Scores 100/100: the mean reproducibility rating across the 2 papers Evolve reviewed is maximal, and both carry high repro-readiness. The Gemma 2/3 SAE study is framed as controlled causal validation, and EPIC ships thorough ablations; Evolve's caveats concern scope (6 languages, 16k SAEs; small catalogs) rather than reproducibility.
Research summary
Dung D. Le is an early-career researcher at Johns Hopkins University whose indexed work splits evenly between two tracks: mechanistic interpretability of multilingual large language models and recommender systems at the intersection of information retrieval and machine learning (a 50/50 field distribution in the OpenAlex snapshot; identity resolved from public bibliographic data). Activity is the clearest signal: indexed works rose from 1 in 2024 to 4 in 2025 and 14 in 2026 — 19 works across the recent three-year window with 14 distinct coauthors, scored 87/100. Reproducibility matches it: the 2 papers Evolve reviewed both rate high repro-readiness, mean 100/100. In the interpretability track, 'Recurrence Is Not Enough' causally validates multilingual SAE translation features in Gemma 2 and 3, showing that feature recurrence overstates causal transfer in a well-controlled but narrow design (6 languages, 16k SAEs). On the recommendation side, EPIC adds explicit posterior item conditioning to semantic-ID diffusion recommenders, reporting solid Amazon gains with thorough ablations. A third 2026 preprint, En-ViMedNER, contributes an English-Vietnamese parallel biomedical corpus with UMLS semantic type annotations, extending the multilingual NLP thread into domain corpora. Citation-based impact is not yet scored (h-index 1, i10-index 0), and focus scores 0/100 because the portfolio spreads exactly across five topic clusters, from topic modeling to speech recognition.
Collaboration pattern
Collaboration is moderate (54/100): 14 distinct coauthors across 19 recent-3y works, so the network is real but not yet dense. Recurring names include Nhu Vo, Phuong Nguyen, Nu Uyen Phuong Le, Wray Buntine, Massimo Piccardi, Inigo Jauregi Unanue, Thanh Tam Nguyen, and Quoc Viet Hung Nguyen, spanning both the interpretability and recommender-system tracks. On catalog papers the researcher appears in last- and middle-author positions, never as corresponding author — consistent with collaborative contribution plus one senior-role authorship.
Research keywords
Highlights
Causal check on multilingual SAE translation features
Posterior item conditioning for semantic-ID diffusion recommendation
English-Vietnamese biomedical corpus with UMLS annotations
A 2026 preprint releasing a parallel English-Vietnamese biomedical corpus annotated with UMLS semantic types; it is not yet Evolve-reviewed, so no verdict or repro-readiness rating is available.
Timeline
- 2024Earliest indexed output in the snapshot: OpenAlex's by-year counts record 1 work.
- 2025Output grows to 4 indexed works as the dual research tracks take shape.
- 2026Acceleration year: 14 indexed works, including all three Evolve-catalog preprints — the Gemma 2/3 SAE causal-validation study, EPIC for diffusion recommendation, and the En-ViMedNER corpus.
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 snapshot is OpenAlex-only; identity is name-resolved from public bibliographic data, and 'Dung D. Le' is a common Vietnamese name, so some indexed works may belong to different people.
- Output and impact axes are unscored: citation coverage is thin (h-index 1, i10-index 0), so influence cannot yet be reliably assessed.
- Only 2 of the 3 catalog papers carry an Evolve verdict; En-ViMedNER is listed without review, keywords, or repro-readiness rating.
- By-year counts cover 2024-2026 only; earlier career output and venue-level detail are absent from the snapshot.