NUS researcher spanning LLM evaluation methodology, on-policy distillation, and audio deepfake detection.
3
papers here
3
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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 data is insufficient: this snapshot indexes no paper counts, venue breakdown, or productivity totals beyond the three Evolve-reviewed 2026 preprints. This axis would measure publication volume and venue quality once a fuller bibliographic record exists.
Impact—
Impact cannot be scored: no citation counts, h-index, or download statistics are available for this profile. This axis would measure scholarly influence — citations, uptake, and reuse of the three preprints — once citation data is ingested.
Focus0/100
Focus scores near the floor reflect the raw field distribution: the record splits evenly, one-third each in LLM evaluation methodology, LLM post-training and distillation, and audio deepfake detection. With three equally weighted directions and no dominant field yet, the 0 score measures concentration, not productivity — the portfolio is deliberately broad, and a clearer specialty may emerge as more output is indexed.
Collaboration53/100
A collaboration score of 53 sits mid-scale, backed by 13 distinct coauthors across the three catalogued papers. Each paper assembles a largely different team — names recurring in the record include Qianli Shen, Ruomeng Ding, Yanxi Chen, Daoyuan Chen, Yaliang Li, Pin-Yu Chen, and Wenqi Wei — pointing to project-based collaboration across institutions rather than a single fixed group. Note that 'Xiang Li' itself appears among coauthor strings, a likely name-collision artifact.
Activity—
Activity data is insufficient: with only three 2026 preprints and no full publication history, publication cadence cannot be computed. This axis would measure research pace and consistency over recent years once a longer record is available.
Reproducibility85/100
Reproducibility is the strongest scored axis: a mean of 85 across three reviewed papers. Evolve rates the LLM-evaluation and distillation work high readiness — the evaluation paper is called rigorous and the distillation work earns an honest-evaluation note despite single-run experiments — while the ROGUE deepfake paper is rated medium. The pattern suggests careful, well-documented methodology across both machine-learning and speech-security projects.
Research summary
Xiang Li is a researcher at the National University of Singapore whose Evolve-reviewed record to date consists of three 2026 arXiv preprints spread evenly — roughly one-third each — across LLM evaluation methodology, LLM post-training and distillation, and speech-and-audio security (audio deepfake detection). The evenly split field distribution yields a focus score of 0, signalling breadth over concentration. The three papers also show varied authorship roles: first author of ROGUE, an adversarially trained LLM-agent workflow generator for deepfake detection; a middle author on a hierarchical-Bayes Neyman-allocation scheme that cuts stochastic LLM evaluation variance by 13–34% versus uniform sampling; and last author of Direct-OPD, where masking low-JSD token states beats log-ratio rewards in 7 of 8 settings. Evolve's reproducibility assessment averages 85 across the three papers, with high readiness for the evaluation and distillation work and medium readiness for ROGUE. Thirteen distinct coauthors appear in the record, and a collaboration score of 53 suggests moderate, project-based teamwork rather than one fixed group. Citation, output-volume, and activity metrics are not yet available, so impact claims should be withheld; the profile is name-resolved from public bibliographic data since 'Xiang Li' is a common name.
Collaboration pattern
Collaboration is moderate (score 53), with 13 distinct coauthors across three papers, and each paper draws a largely different team. Recurring names include Qianli Shen, Ruomeng Ding, Yanxi Chen, Daoyuan Chen, Yaliang Li, Pin-Yu Chen, Wenqi Wei, Yunshi Lan, Yibo Zhao, and Zixuan Yang, suggesting per-project partnerships across institutions rather than one long-term group. The coauthor list also contains 'Xiang Li' itself, a likely name-collision artifact in bibliographic records.
Research keywords
Highlights
ROGUE: adversarially trained LLM-agent workflows for audio deepfake detection
Speculative Evaluation cuts stochastic LLM evaluation variance by 13–34%
Selective supervision for on-policy distillation: not every token is worth distilling
Timeline
- 2026First-authored ROGUE paper appears on arXiv, using adversarial learning to generate robust workflows for audio deepfake detection.
- 2026Co-authors (middle position) Speculative Evaluation of Stochastic LLMs, a hierarchical-Bayes variance-reduction method for LLM evaluation.
- 2026Last-authored Direct-OPD paper proposes selective supervision for on-policy distillation.
- 2026All three catalogued preprints land in a single year, evenly spanning LLM evaluation, distillation, and audio deepfake detection.
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.
- 'Xiang Li' is a common name; this profile is name-resolved from public bibliographic data, and some coauthor records may belong to other researchers with the same name.
- Bibliometric scores are computed from public bibliometric data only; citation counts, h-index, and total publication volume are absent from this snapshot.
- All three catalogued papers are 2026 arXiv preprints; peer-reviewed venue acceptance is not reflected in this profile.
- ORCID and country are not provided, so identity verification relies solely on the stated institution and bibliographic matching.