Alibaba researcher bridging LLM agent self-evolution and long-form audio/spoken-language benchmark evaluation.
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) 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 catalog holds only three 2026 arXiv preprints with no citation, volume, or venue data attached, so this axis — paper output volume, publication venues, and authorship role over time — cannot be measured yet.
Impact—
Not yet scored: no citation counts, h-index, or downstream-usage signals accompany these preprints in the provided data, so this axis — how often and how widely the work is cited and adopted — remains unmeasured.
Focus0/100
Scores 0, reflecting a spread profile: the bibliometric field distribution splits evenly at about 33.3% across three labeled areas — LLM agent self-evolution, audio-language model evaluation, and speech and natural language processing. All three 2026 papers share an evaluation-and-benchmarking orientation, so the dispersion is partly an artifact of field labeling, but no single field yet dominates the record.
Collaboration62/100
Scores 62, computed from 21 distinct coauthors across the reviewed papers. The frequent-coauthor names — Xize Cheng, Ruimin Pei, Yongkang Wu, Shangyi Zheng, Yaqing Zhang, Deyang Li, Jianjun Tao, Xiang Zhang, Zeyu Yang, Zibo Bi, and Pei Zhang — point to a stable team around this researcher at Alibaba Group, with mid-list authorship rather than first or corresponding positions on all three papers.
Activity—
Not yet scored: all three catalog entries are 2026 arXiv preprints with no year-by-year publication counts available, so this axis — publication cadence and sustained research activity over time — cannot yet be computed.
Reproducibility55/100
Scores 55, the mean readiness across 3 reviewed papers, each rated medium. A consistent pattern emerges: strong documentation — MuLA-Bench's multi-tier auditing, Vox-Infinity's provenance annotation, TTSE's rule lifecycle design — is offset by incomplete openness, including partially released data in Vox-Infinity, LLM-constructed gold answers and LLM judging in both audio benchmarks, and a pivotal TTSE ablation lacking significance testing.
Research summary
Xinyu Zhang is a researcher at Alibaba Group (China) whose Evolve profile, resolved from public bibliographic data, currently rests on three 2026 arXiv preprints in two adjacent areas: LLM agent self-evolution and speech/audio evaluation. TTSE: A Two-Track Online Self-Evolution Framework proposes FACT/TIP dual-track agent memory separating declarative from procedural knowledge with rule lifecycle management; Evolve's review credits its clean risk theory while flagging a pivotal ablation without significance testing and a warehouse component that drives most SOPBench gain. The other two papers are benchmarks: MuLA-Bench, a well-documented multilingual long-form audio understanding benchmark covering 16 languages via multi-tier auditing, and Vox-Infinity, claimed as the first provenance-annotated long-context speech benchmark, where the recency-effect finding is solid but LLM-built QA, LLM judging, and partial data release limit rigor. All three papers place this researcher in a middle author position, never as corresponding author. Bibliometrics show 21 distinct coauthors (collaboration 62), an even split across three labeled fields (focus 0), and mean reproducibility readiness of 55 across 3 reviewed papers, each rated medium. Output, impact, and activity scores are not yet available.
Collaboration pattern
The record shows 21 distinct coauthors and a compact recurring team — Xize Cheng, Ruimin Pei, Yongkang Wu, Shangyi Zheng, Yaqing Zhang, Deyang Li, Jianjun Tao, Xiang Zhang, Zeyu Yang, Zibo Bi, and Pei Zhang — consistent with group-based work at Alibaba Group. Across all three 2026 papers this researcher sits in a middle author position and is never the corresponding author, suggesting a contributor role within a larger benchmark-building effort rather than sole leadership.
Research keywords
Highlights
Dual-track agent memory: FACT/TIP framework for online self-evolution
16-language long-form audio benchmark with multi-tier auditing
Provenance-annotated long-context speech benchmark claim
Timeline
- 2026Vox-Infinity posted to arXiv, benchmarking long-context spoken language models with claimed answer-provenance annotation (middle authorship, non-corresponding).
- 2026MuLA-Bench posted to arXiv, a multilingual long-form audio understanding benchmark covering 16 languages via multi-tier auditing (middle authorship, non-corresponding).
- 2026TTSE: A Two-Track Online Self-Evolution Framework posted to arXiv, introducing FACT/TIP dual-track agent memory (middle authorship, non-corresponding).
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.
- Xinyu Zhang is a common name; this profile is name-resolved from public bibliographic data and is not confirmed by an ORCID (none provided).
- Radar scores (collaboration 62, focus 0, reproducibility 55) come from Evolve's bibliometric snapshot only; output, impact, and activity metrics were unavailable.
- All three catalog entries are 2026 arXiv preprints with medium reproducibility readiness; no peer-reviewed publication or citation data is included.
- The frequent-coauthor list is name-based; with common Chinese names it may merge or split distinct individuals.