Audio and spoken-language ML researcher at CUHK Shenzhen; RL-based self-improvement and TTS prosody evaluation
3
papers here
3
reviewed by Evolve
0
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—
Not yet scored: the snapshot contains no publication-volume or productivity data for this author, so the output axis cannot be computed. It would measure the count and trajectory of peer-reviewed and preprint outputs over time.
Impact—
Not yet scored: citation counts and downstream influence indicators are absent from the available data, so no impact value can be reported. The axis would measure how often and how widely this author's papers are cited and built upon.
Focus0/100
The focus axis scores 0/100, reflecting an evenly spread field distribution: the reviewed papers sit one-third each in speech processing/audio AI (TTS evaluation), audio machine learning (speech, sound, and music understanding), and speech and natural language processing. Since all three converge on spoken-audio foundation models and GRPO-based reinforcement learning, the spread reads as one coherent audio-ML program rather than scattered interests.
Collaboration68/100
Scoring 68/100, the collaboration axis rests on 30 distinct co-authors in the bibliometric snapshot, with a recurring core including Zifan Guan, Longyu Lu, Junan Zhang, Meiguang Jin, and Junfeng Ma. Wu takes the last-author position on EvoAudio and co-author positions on the other two 2026 papers, consistent with a supervisory role in a sizeable, stable audio-ML group rather than isolated bilateral work.
Activity—
Not yet scored: only three 2026 papers are visible in this snapshot, with no year-by-year publication record, so activity cannot be computed. The axis would measure publishing cadence and the recent output trend across the full career timeline.
Reproducibility20/100
The reproducibility axis is low at 20/100, computed from 3 scored papers (reproMean 20). All three reviewed 2026 preprints are rated low: EvoAudio reports gains to 6.3 points on five LALMs but ships no code or error bars; the think-aloud model relies on proprietary data without an interleaving baseline; the prosody judge is a single-seed result on private data. Careful evaluation design, weak release practice.
Research summary
Zhizheng Wu, affiliated with the Chinese University of Hong Kong, Shenzhen, works across spoken-language and audio machine learning, with recent focus on building and evaluating audio and spoken language models. Evolve has reviewed three of the group's 2026 arXiv preprints: EvoAudio, which wraps audio language models in a closed-loop self-improvement pipeline that synthesizes verifiable audio QA and reports gains of up to 6.3 points on five LALMs over 13 rounds; Spoken Language Models that Think Aloud, which cuts reasoning silence from 12.8 to 0.36 seconds at near chain-of-thought accuracy; and a multi-dimensional prosody judge for live-streaming text-to-speech that distills Gemini into an open reward model trained with span-local GRPO. A recurring methodological signature is reinforcement learning with verifiable or model-based rewards — GRPO appears in two of the three papers — paired with deliberate evaluation design. The bibliometric snapshot is narrow but coherent: three fields split evenly at one third each, 30 distinct co-authors, a collaboration score of 68, focus 0, reproducibility 20, and all three reviewed papers rated low for reproducibility given private data, missing code, and absent error bars. Output, impact, and activity scores are not yet available.
Collaboration pattern
The snapshot records 30 distinct co-authors, and the recurring list — Zifan Guan, Longyu Lu, Junan Zhang, Meiguang Jin, Junfeng Ma, Yuxiang Wang, Shengbo Cai, Yingda Shen, Ming-Hao Hsu, Qinke Ni, Liqiang Zhang — suggests a stable, sizeable audio-ML group at CUHK Shenzhen. Wu holds the last-author position on EvoAudio and co-author positions on the other two 2026 papers, consistent with supervising several parallel efforts rather than hands-on first-author work.
Research keywords
Highlights
Closed-loop self-improvement for audio language models
Cutting reasoning silence in spoken language models
An open prosody judge for live-streaming TTS
Timeline
- 2026Last author on EvoAudio, a recursive self-improvement framework reporting gains of up to 6.3 points on five audio language models over 13 rounds.
- 2026Co-authors Spoken Language Models that Think Aloud, reducing spoken reasoning silence (Lsil) from 12.8 to 0.36 seconds.
- 2026Contributes to multi-dimensional prosody judgment for live-streaming speech synthesis, distilling Gemini into an open reward model trained with span-local GRPO.
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 scores (collaboration 68, focus 0, reproducibility 20, 30 distinct co-authors) come from a public bibliometric snapshot only; no citation or venue-level data was available.
- Output, impact, and activity axes are unscored because the underlying productivity and citation data is missing from this snapshot.
- All reviewed papers are 2026 arXiv preprints rated low for reproducibility (private or proprietary data, no code releases, no error bars); findings are single-source and unverified by Evolve.
- No ORCID or disambiguating identifiers were provided; this profile is name-resolved from public bibliographic data, and 'Zhizheng Wu' is a shared name — verify the CUHK Shenzhen affiliation before reuse.