First-author work on native audio-visual dialogue at CUHK: synthetic data engines, rubric benchmarks, and RL training
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—
Output is not scored yet: the bibliometric snapshot contains no reliable publication count or venue record for this researcher, so this axis, which would measure total output and share of first or corresponding-author papers, is left blank rather than estimated.
Impact—
Impact is not scored yet: no citation counts or citation-based indicators are available in the current data, so this axis, which would measure citations, uptake, and influence on follow-up work, is reported as missing instead of guessed.
Focus0/100
Focus computes to 0 on the 0-100 scale. The underlying field distribution splits evenly, 0.5 to speech-and-audio-visual-interaction work and 0.5 to speech-and-vision work, across the two Evolve-reviewed papers. Read literally, the even split yields minimal concentration, though the two subfields are adjacent branches of the same multimodal machine learning area, so the low score reflects breadth within one theme rather than scattered research interests.
Collaboration70/100
Collaboration scores 70 out of 100, with 32 distinct coauthors recorded. The reviewed papers list frequent collaborators including Qi Chen, Yunfei Chu, Qiuqiang Kong, Jin Xu, Xie Chen, Ziyang Ma, Yuxuan Wang, and Yifan Yang, a roster typical of large audio and speech language-model teams. Working across a synthetic-data engine and a benchmark effort implies multi-role teams spanning data engineering, modeling, and evaluation rather than small two-person authorship.
Activity—
Activity is not scored yet: the snapshot provides no publication-frequency or year-by-year volume data, so this axis, which would measure publishing cadence and sustained research activity over time, is left unscored pending more bibliometric coverage.
Reproducibility78/100
Reproducibility averages 78 out of 100 across the 2 reviewed papers. Paper-level readiness differs: Omni Demand Understanding (ODU-Bench) is rated high, helped by an explicit benchmark pipeline, while OmniVChat is rated medium, consistent with its heavier synthesis-and-RL training stack. Evolve's verdicts flag shared risks in both: an LLM-judged metric carries 60% of ODU-Bench's score, and OmniVChat retains judge-bias concerns despite gains transferring to real recordings.
Research summary
Haolin He is a multimodal machine learning researcher at the Chinese University of Hong Kong (ORCID 0009-0008-3493-3550) whose Evolve-visible work centers on speech-grounded interaction: half of the reviewed field distribution sits in speech and audio-visual interaction and half in speech and vision. The catalog contains two 2026 arXiv papers. As first author of OmniVChat, He led a full-stack effort — a synthetic-data engine, a rubric-based benchmark, and reinforcement learning for native audio-visual dialogue — with gains that transfer to real recordings, though Evolve notes residual judge bias and capability trade-offs. The second paper, Omni Demand Understanding (ODU-Bench), is a middle-authorship contribution that turns demand detection and contextual user-intent inference into an explicit benchmark; its pipeline is solid, but an LLM-judged metric carries 60% of the score. Bibliometrically, the profile scores 70/100 on collaboration with 32 distinct coauthors, including frequent collaborators such as Qi Chen, Yunfei Chu, Qiuqiang Kong, Jin Xu, Xie Chen, and Ziyang Ma. Focus computes to 0 on the 0-100 scale from an even 0.5/0.5 field split, and reproducibility averages 78/100 across the two papers (ODU-Bench high, OmniVChat medium). Output, impact, and activity axes lack underlying data and are intentionally unscored; this remains an early-stage profile built from a small, preprint-heavy snapshot.
Collaboration pattern
His catalogued work is team-produced: 32 distinct coauthors and a collaboration score of 70/100. Frequent collaborators named in the reviewed snapshot include Qi Chen, Yunfei Chu, Wen Huang, Qiuqiang Kong, Jin Xu, Xie Chen, Yifan Yang, Zihan Liu, Yuxuan Wang, Ziyang Ma, and Ruiyang Xu. Combining a synthetic-data engine, RL training, and benchmark construction points to role-specialized teams typical of contemporary audio-LLM groups rather than small bilateral collaborations.
Research keywords
Highlights
First-author full stack for native audio-visual dialogue
Turning demand and intent inference into an explicit benchmark
Benchmark-and-evaluation-first research signature
Across both catalogued 2026 papers, the recurring pattern is to build the measurement layer — rubric benchmark and reward for dialogue, demand and intent benchmark for interaction — alongside model training; reviewed reproducibility averages 78/100 over the two papers, with readiness spanning medium to high.
Timeline
- 2026Posts OmniVChat to arXiv as first author: synthesis, benchmarking, and RL training for native audio-visual dialogue.
- 2026Contributes to Omni Demand Understanding (ODU-Bench), a benchmark for contextual user-intent inference in multimodal interaction.
- 2026Both papers enter the Evolve catalog; reviewed reproducibility averages 78 across the two, with ODU-Bench rated high and OmniVChat medium readiness.
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 signals come from OpenAlex-style public bibliographic data only; citation counts, h-index, and total publication volume are absent, so output, impact, and activity remain unscored.
- Qualitative judgments rest on only 2 Evolve-reviewed papers, both 2026 arXiv preprints; no peer-reviewed venue record is present in this snapshot.
- The profile is name-resolved from public bibliographic data tied to ORCID 0009-0008-3493-3550; common-name collisions cannot be fully excluded.
- Affiliation is listed as the Chinese University of Hong Kong with no country field; currency of the affiliation is not independently verified.