Audio ML researcher spanning audio deepfake detection and recursive self-improvement for audio language models
2
papers here
2
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 data is not yet available; this axis would measure total publication volume, venue types, and authorship roles over time. With only two 2026 arXiv preprints reviewed, no reliable score can be computed or estimated.
Impact—
Impact data is not yet available; this axis would measure citations and downstream uptake of the reviewed work. Since the two papers are recent 2026 preprints with no citation record in the input data, no score is computed here.
Focus0/100
The radar focus score is 0, and the raw field distribution splits exactly 50/50 between speech and audio signal processing (audio deepfake detection) and audio machine learning. In practice this reflects a two-track portfolio: detector-side work on synthetic speech and model-side work on audio language understanding. Both tracks share self-supervised audio representations and refinement-style training, so the even split reads as breadth within one domain rather than scattered output.
Collaboration53/100
The collaboration score is 50, based on 11 distinct coauthors across the two reviewed 2026 papers. The recurring pool includes Zhizheng Wu, Kunyu Feng, Li Wang, Wan Lin, Shengbo Cai, Yingda Shen, Ming-Hao Hsu, Qinke Ni, Liqiang Zhang, Teddy Sun, and Steve Yevs. Wang took first authorship on one paper and a middle position on the other, consistent with contributing core ideas within a larger team rather than working solo.
Activity—
Activity data is not yet available; this axis would measure annual publishing cadence and research continuity across years. The input contains only the single publication year 2026, so no multi-year trend can be established, and this axis is left unscored.
Reproducibility38/100
The reproducibility score is 38, an average computed over 2 reviewed papers. The two releases sit at different levels: EvoAudio is rated low because it reports gains of up to 6.3 points across 13 training rounds but ships no error bars and no code, while CoRELoop is rated medium thanks to strong ablations on its pooled EER drop from 4.85% to 3.74%, though it lacks external state-of-the-art comparisons. Releasing code and variance estimates would lift this axis fastest.
Research summary
Yuxiang Wang is an audio machine learning researcher whose reviewed work concentrates on two tightly linked problems: improving audio language models and detecting synthetic speech. The Evolve catalog holds two 2026 arXiv papers. As first author, Wang's EvoAudio builds a closed loop that constructs verifiable audio QA data with synthesis tools and applies GRPO-style reinforcement learning over 13 rounds, reporting gains of up to 6.3 points on five large audio language models. As a middle author, Wang contributed to CoRELoop, which applies LoRA-based controlled recurrent refinement to a frozen self-supervised deepfake detector, cutting pooled EER from 4.85% to 3.74% with strong ablations. The field distribution splits evenly, 50/50, between audio deepfake detection and audio machine learning, and the radar focus score is 0. The collaboration score is 50, drawing on 11 distinct coauthors including Zhizheng Wu and Kunyu Feng. The reproducibility score of 38, averaged over 2 papers, reflects mixed openness: no error bars or code on one paper, solid ablations but no external SOTA comparison on the other. This is a name-resolved profile built from public bibliographic data.
Collaboration pattern
Wang's reviewed work is team-based: 11 distinct coauthors appear across the two 2026 papers, with Zhizheng Wu, Kunyu Feng, and Li Wang among the recurring names. Moving from first author on EvoAudio to a middle authorship position on CoRELoop suggests flexibility between leading and supporting roles. The collaboration score of 50 sits mid-range; no affiliation data is available, so cross-institution patterns cannot yet be confirmed.
Research keywords
Highlights
Recursive self-improvement lifts audio language models by up to 6.3 points
LoRA-based recurrent refinement cuts deepfake detection EER from 4.85% to 3.74%
A shared refinement recipe across detection and understanding
Across both 2026 papers Wang applies iterative refinement — recurrent LoRA-based refinement for deepfake detection, recursive self-improvement with verifiable rewards for audio QA — one methodological thread spanning the two halves of the recorded field distribution.
Timeline
- 2026First-author arXiv preprint EvoAudio introduces recursive self-improvement for audio understanding, reporting gains of up to 6.3 points on five large audio language models over 13 rounds.
- 2026Co-authors CoRELoop, applying LoRA-based controlled recurrent refinement to audio deepfake detection and cutting pooled EER from 4.85% to 3.74%.
- 2026Both preprints appear in the same year, establishing a two-track profile that splits evenly between audio deepfake detection and audio machine learning.
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.
- Bibliometrics are derived from OpenAlex-sourced public bibliographic data; no ORCID, institution, or country linkage was available for this profile.
- 'Yuxiang Wang' is a common name: this profile is name-resolved from public bibliographic data and may merge or split individuals with the same name.
- The radar is computed from only 2 reviewed arXiv preprints; output, impact, and activity axes lack sufficient data and are left unscored.
- Both papers are 2026 arXiv preprints, so citation impact and peer-reviewed venue status are not yet reflected anywhere in this profile.