ML researcher across LLM evaluation, graph learning, speech models, and audio deepfake defense; values reproducibility
5
papers here
5
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-01) 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: the catalog holds only five 2026 arXiv entries, with no total publication count, venue history, or yearly productivity series. This axis would measure cumulative papers, publication venues, and output tempo.
Impact—
Impact is not yet measurable: no citation counts, h-index, or usage statistics exist in this snapshot. The axis would track citations to the five cataloged preprints and uptake of methods like ROGUE and QUARTET.
Focus0/100
Focus scores 0 because activity is maximally dispersed: the field distribution assigns an even 0.2 to each of five areas — LLM evaluation methodology, speech and NLP, audio deepfake detection, medical image analysis, and graph representation learning. No dominant field has emerged across the five reviewed 2026 papers, which reads either as deliberate breadth or as a research identity still crystallizing.
Collaboration70/100
Collaboration scores 69 on the strength of 31 distinct coauthors across only five cataloged papers — a broad network for so small a sample. Recurring partners include Qianli Shen, Ruomeng Ding, Yanxi Chen, Daoyuan Chen, Yaliang Li, and a speech cluster of Junyi Ao, Kainan Peng, Mingbo Ma, and Xutai Ma, suggesting parallel collaboration tracks in speech/audio and in general machine learning.
Activity—
Activity cannot be scored: every entry in the snapshot is a 2026 arXiv preprint, so no multi-year publishing cadence exists yet. This axis would track publication tempo, intervals between releases, and sustained engagement once a longer history is linked.
Reproducibility66/100
Reproducibility scores 66, the mean readiness across all five reviewed papers: two rated high (Speculative Evaluation of Stochastic LLMs; QUARTET), two medium (ROGUE; M3D-Net), and one low (Spoken Language Models that Think Aloud, limited by proprietary data). The pattern shows solid methodological description with recurring gaps in baselines, ablations, and data release.
Research summary
Xiang Li is an early-stage machine-learning researcher whose 2026 arXiv record, as cataloged by Evolve, spans five adjacent areas in near-equal measure: LLM evaluation methodology, speech and natural language processing, audio deepfake detection, medical image analysis, and graph representation learning — the field distribution assigns 0.2 to each. Across five reviewed papers, Li appears once as first author (ROGUE, adversarial workflow generation for audio deepfake detection) and once as last author (M3D-Net for mammography classification), with the other three at middle authorship. The methodological throughline is measurement-aware, rigorous machine learning: hierarchical-Bayes Neyman allocation with async speculation cuts LLM evaluation variance 13–34%; a quad-branch graph transformer beats RelGT on 8 of 12 RelBench tasks; asynchronous think-aloud decoding cuts spoken-reasoning silence from 12.8s to 0.36s at near-chain-of-thought accuracy; and adversarially trained LLM-agent workflows harden deepfake detectors against corruption. Evolve's reproducibility screen averages 66 across the five papers (two high, two medium, one low readiness). Collaboration is broad: 31 distinct coauthors, including recurring speech and general-ML partners, giving a collaboration score of 69. Citation-level output and impact metrics are not yet available, so this profile should be read as a name-resolved, early-stage snapshot built from public bibliographic data.
Collaboration pattern
Li's collaboration footprint is wide for a five-paper sample: 31 distinct coauthors, yielding a collaboration score of 69. Two clusters stand out — a speech/audio group (Junyi Ao, Kainan Peng, Mingbo Ma, Xutai Ma, Zhenyu Tang, Yinghao Li) behind the think-aloud and deepfake-detection work, and a general-ML group (Qianli Shen, Ruomeng Ding, Yanxi Chen, Daoyuan Chen, Yaliang Li) around evaluation, relational learning, and medical imaging. Li shifts among middle, first, and last authorship rather than anchoring a single lab line.
Research keywords
Highlights
First-author ROGUE: adversarially trained LLM-agent workflows harden audio deepfake detection
Speculative Evaluation cuts LLM eval variance 13–34% with hierarchical-Bayes Neyman allocation
QUARTET graph transformer wins 8 of 12 RelBench tasks against RelGT
Asynchronous think-aloud decoding cuts spoken-reasoning silence from 12.8s to 0.36s
Last-author M3D-Net reaches 97.78% mammography validation accuracy
Timeline
- 2026Posts first-author ROGUE, an adversarial-learning approach to LLM-agent workflow generation for audio deepfake detection, on arXiv.
- 2026Contributes as middle author to Spoken Language Models that Think Aloud, cutting reasoning silence from 12.8s to 0.36s.
- 2026Contributes as middle author to QUARTET, a quad-branch graph transformer beating RelGT on 8 of 12 RelBench tasks.
- 2026Appears as last author of M3D-Net, reporting 97.78% validation accuracy for mammography classification.
- 2026Contributes as middle author to Speculative Evaluation of Stochastic LLMs, reducing evaluation variance 13–34%.
Papers in the Evolve catalog
Speculative Evaluation of Stochastic LLMs
M3D-Net: Hierarchical Coordination of Spatial Context, Feature Reuse, and Differential Attention for Mammography Classification
QUARTET: Quad-branch cross-Attention and Random-walk Traces for Enhancing Transformers on Relational Graphs
Spoken Language Models that Think Aloud
- first author
Robust Workflow Generation via Adversarial Learning for Audio Deepfake Detection
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 figures derive from a public OpenAlex-based snapshot; this record carries no institution, country, or ORCID.
- "Xiang Li" is a common name; the profile is name-resolved from public bibliographic data and co-author overlap, so identity attribution is probabilistic.
- The catalog covers only five 2026 arXiv preprints; citation counts, h-index, venue history, and multi-year activity are unavailable, leaving output, impact, and activity unscored.
- Reproducibility scores and verdicts reflect Evolve's review of these five papers only and should not be generalized to the researcher's full body of work.