Speech and multimodal AI researcher building long-context spoken benchmarks and audio-visual dialogue pipelines
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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 scored yet: output would measure total publication volume, venue spread, and pace of new papers, but Evolve's reviewed record for this researcher holds only two 2026 arXiv preprints, too few to compute.
Impact—
Not scored yet: impact would measure citations, citation velocity, and downstream adoption of the two 2026 benchmark papers; they are too recent and coverage too thin for any estimate.
Focus0/100
Focus scores 0/100 because Evolve's field distribution splits exactly evenly: 0.5 speech and natural language processing versus 0.5 multimodal machine learning (speech and vision). In practice the two reviewed 2026 papers sit on adjacent sides of one speech-centric agenda — a first-authored spoken-language long-context benchmark and a middle-authored audio-visual dialogue effort combining data synthesis and RL training. The even split signals breadth across benchmarking and training rather than scattered unrelated topics.
Collaboration66/100
Collaboration scores 66/100 across 26 distinct coauthors. Twelve recur as frequent collaborators, including Haolin He, Yunfei Chu, Qi Chen, Wen Huang, Yuan Feng, Muzhi Zhu, Zheqi Dai, Haoning Xu, Dongchao Yang, Chunyat Wu, Zining Liang, and Zhengxi Liu — a large core team typical of data-engine and benchmark-heavy projects. The authorship pattern mixes roles: first author on Vox-Infinity and a middle position on OmniVChat, indicating both led and team-based contributions.
Activity—
Not scored yet: activity would measure publication cadence and sustained yearly research output over time; with only two 2026 preprints in scope, no multi-year activity trend can be established.
Reproducibility55/100
Reproducibility averages 55/100 over the two papers Evolve could assess (reproPapers: 2), and both are rated medium readiness. Evolve's verdicts flag concrete limitations: Vox-Infinity uses LLM-built QA, LLM judging, and only partial data release; OmniVChat relies on an LLM judge whose bias remains even though gains transfer to real recordings. The 55 mean reflects benchmarks whose artifacts are incompletely released or judged by models rather than verified ground truth.
Research summary
Xize Cheng works at the intersection of speech processing and multimodal machine learning, asking how far current spoken-language and audio-visual models can genuinely understand long, natural inputs. Evolve's reviewed catalog holds two 2026 arXiv preprints. The first, Vox-Infinity, is a first-authored benchmark claiming to be the first provenance-annotated long-context speech benchmark; Evolve judges its recency-effect finding solid while noting that LLM-built QA, LLM judging, and partial data release limit rigor. The second, OmniVChat, is a middle-author collaboration that pairs a multi-agent synthetic data engine with a rubric benchmark and reinforcement learning for native audio-visual dialogue; Evolve notes the gains transfer to real recordings, though judge bias and trade-offs remain. Evolve's field distribution splits evenly, 0.5 speech and natural language processing and 0.5 multimodal machine learning (speech and vision), which drives a focus score of 0/100. Collaboration scores 66/100 across 26 distinct coauthors, twelve of them frequent collaborators. Reproducibility averages 55/100 over the two reviewed papers, both rated medium readiness. Output, impact, and activity are not yet scored, so this reads as an early-stage, benchmark-centric profile.
Collaboration pattern
Across Evolve's reviewed record, Xize Cheng works with 26 distinct coauthors, and 12 recur as frequent collaborators — including Haolin He, Yunfei Chu, Qi Chen, Wen Huang, Yuan Feng, Muzhi Zhu, Zheqi Dai, Haoning Xu, Dongchao Yang, Chunyat Wu, Zining Liang, and Zhengxi Liu. The pattern is large-team, data- and benchmark-heavy research: first-author leadership on Vox-Infinity combined with a middle-author team role on OmniVChat. Evolve's collaboration score is 66/100.
Research keywords
Highlights
First-authored provenance-annotated long-context speech benchmark
Recency effect stands out as the robust finding
Synthetic data engine plus RL for native audio-visual dialogue
Timeline
- 2026Posted first-authored Vox-Infinity on arXiv, a provenance-annotated benchmark probing long-context understanding in spoken language models.
- 2026Co-authored OmniVChat on arXiv, spanning synthetic data generation, rubric benchmarking, and reinforcement-learning training for native audio-visual dialogue.
- 2026Both preprints appeared in the same year, splitting evenly between speech/NLP and multimodal speech-and-vision work.
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.
- This profile is name-resolved from public bibliographic data (ORCID 0000-0001-9708-3225); no institution or country is on record.
- Only two reviewed papers, both 2026 arXiv preprints, are in scope; output, impact, and activity axes lack sufficient data and are unscored.
- The reproducibility mean of 55 averages only 2 papers, both rated medium readiness, so it rests on a very small sample.
- Frequent-coauthor names come from Evolve's co-occurrence data and have not been disambiguated against other researchers with similar names.