Trung Le

Monash University

Unclaimed · AI-structured profileOpenAlex

ML researcher at Monash University spanning LLM safety, Bayesian parameter-efficient fine-tuning, and multi-hop QA.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) 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 underlying publication-volume data for this profile is unavailable, so this axis cannot be estimated. It would measure cumulative paper counts and productivity trends across the researcher's career.

Impact—

Not yet scored: citation counts, h-index and related influence data are not available in this snapshot. This axis would measure how widely the three reviewed papers and other works have been cited by the community.

Focus0/100

Scores 0, reflecting an evenly split field distribution: LLM safety / adversarial machine learning, machine learning (Bayesian deep learning / parameter-efficient fine-tuning), and natural language processing each account for roughly 0.33 of the catalog. Across the three reviewed 2026 papers, no single field dominates, so the radar treats the portfolio as maximally dispersed. This breadth spans safety detection, geometric Bayesian fine-tuning, and retrieval-based QA rather than one concentrated line.

Collaboration51/100

Scores 51, based on 12 distinct coauthors across the reviewed catalog. The frequent-coauthor list is anchored by Quoc Viet Vo, Damith C. Ranasinghe, Ehsan Abbasnejad, Quang-Duy Tran, Bao Duong, Phuoc Nguyen, Thin Nguyen, and Thien Huu Nguyen, among others. The moderate score reflects a real but not yet extensive network: all three papers are middle-author, non-corresponding contributions, consistent with team-based rather than lead collaborative roles.

Activity—

Not yet scored: recent publication cadence data is insufficient in this snapshot. This axis would measure the timing and consistency of output — for example, papers per year — across the researcher's recent career.

Reproducibility55/100

Scores 55 (reproMean) across 3 scored papers — a middle-of-the-road readiness signal, with all three reviewed papers rated medium repro-readiness. Evolve's verdicts note a shared weakness: HyperProve's benchmark gains and StePS's fine-tuning improvements both lack significance tests, and StePS evaluates only one base model, which tempers an otherwise method-transparent profile.

Research summary

Trung Le is a researcher at Monash University whose Evolve profile is built from a reviewed catalog of three 2026 arXiv preprints, in all of which Le is a middle-position, non-corresponding author. The catalog spans three fields in an even split (each roughly 0.33): LLM safety / adversarial machine learning, machine learning (Bayesian deep learning / parameter-efficient fine-tuning), and natural language processing — a breadth that yields a focus score of 0. SAFEGuard is a training-free two-stage jailbreak detector combining a perplexity–Jensen-Shannon-divergence hybrid fluency measure with gradient matching, reported strong on Llama-2/Mistral but weaker on Llama-3.1. StePS applies Riemannian Stein variational gradient descent to Stiefel-manifold LoRA for calibrated Bayesian parameter-efficient fine-tuning, with promising but small, untested gains on one base model. HyperProve tackles multi-hop question answering with answer-guided hypergraph expansion over atomic facts, reaching top mean scores on three benchmarks without significance tests. Collaboration scores 51 with 12 distinct coauthors; reproducibility scores 55 (reproMean) across 3 scored papers, all rated medium repro-readiness. Output, impact, and activity metrics are not yet available.

Collaboration pattern

The reviewed catalog connects Trung Le to 12 distinct coauthors, with recurring names including Quoc Viet Vo, Damith C. Ranasinghe, Ehsan Abbasnejad, Quang-Duy Tran, Bao Duong, Phuoc Nguyen, Thin Nguyen, and Thien Huu Nguyen. On all three 2026 papers Le is a middle-position, non-corresponding author, suggesting contributions within larger team efforts — a pattern consistent with the moderate collaboration score of 51 rather than a solo or lead-author mode.

Research keywords

jailbreak detectionadversarial machine learninglarge language modelsparameter-efficient fine-tuningLoRABayesian deep learninguncertainty quantificationmulti-hop question answeringretrieval-augmented generationStein variational gradient descentperplexityhypergraph retrieval

Highlights

SAFEGuard: training-free two-stage jailbreak detection for LLMs
Evolve's review describes a training-free two-stage detector for optimization-based jailbreak attacks that combines a perplexity–Jensen-Shannon-divergence hybrid fluency measure with gradient matching, performing strongly on Llama-2 and Mistral while weakening on Llama-3.1.
StePS: Riemannian SVGD brings Bayesian calibration to Stiefel-manifold LoRA
The reviewed StePS paper applies Stein variational gradient descent on the Stiefel manifold for geometry-aware Bayesian parameter-efficient fine-tuning; Evolve notes promising gains, but on one base model with small margins and no significance tests.
HyperProve: answer-guided hypergraph expansion for multi-hop QA
HyperProve uses atomic facts and question decomposition to guide iterative hypergraph retrieval, achieving top mean scores on 3 benchmarks in a controlled setup — though Evolve flags that the gains lack significance tests.

Timeline

2026Publishes SAFEGuard on arXiv: a training-free two-stage jailbreak detector combining perplexity–JSD fluency analysis with gradient matching for optimization-based attacks on LLMs.
2026Publishes StePS on arXiv: geometry-aware Bayesian parameter-efficient fine-tuning on the Stiefel manifold via Stein variational gradient descent.
2026Publishes HyperProve on arXiv: answer-guided hypergraph expansion for multi-hop question answering, with top mean scores on 3 benchmarks.

Papers in the Evolve catalog

Frequent co-authors

Quoc Viet Vo×1 Damith C. Ranasinghe×1 Ehsan Abbasnejad×1 Quang-Duy Tran×1 Bao Duong×1 Phuoc Nguyen×1 Thin Nguyen×1 An Nguyen Phu×1 Dung Nguyen Quang×1 Luu Hieu An×1 Linh Ngo Van×1 Thien Huu Nguyen×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.

Bibliometric scores are derived from OpenAlex-based data; the output, impact, and activity axes lack underlying data and are therefore not scored.
"Trung Le" is a common Vietnamese name; this profile is name-resolved from public bibliographic data and may merge distinct individuals.
All reviewed papers are 2026 arXiv preprints with middle-author positions; no peer-reviewed venue information and no first- or corresponding-author work is within scope.
Collaboration and reproducibility scores cover only the three papers in Evolve's catalog, not the researcher's full publication record.