Conversational AI researcher at Peng Cheng Laboratory: full-duplex speech models, intent benchmarks, and agent memory
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-14) 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—
No verified publication count or venue list is available for this profile yet, so output is unscored. This axis would measure total scholarly production, the split between preprints and peer-reviewed papers, and the share of first or corresponding-author work.
Impact—
Citation counts, h-index-style indicators, and usage signals are absent from the current snapshot, so impact is unscored. This axis would measure how often this researcher's papers are cited and taken up by follow-up work and open-source projects.
Focus0/100
The focus score is 0 because the raw field distribution splits evenly three ways: multimodal machine learning for speech and audio-visual interaction, speech processing and conversational AI, and NLP for conversational AI and agent memory each account for 33.3% of the record. Read substantively, the spread is narrow in theme — all three lines centre on spoken and multi-turn dialogue — but under an even-split metric this record reads as broad rather than specialised.
Collaboration86/100
The collaboration score of 86 rests on a large co-author network: 73 distinct collaborators appear in the bibliometric record, with 12 recurring partners including Qi Chen, Yunfei Chu, Zihan Liu, and Ziyang Ma. All three catalogued papers place Wang in a middle author position on sizeable teams, consistent with a contributor who embeds in group efforts across speech, multimodal, and NLP lines rather than a solo or strictly first-author workflow.
Activity—
Year-by-year publication records are not present in the provided data, so activity is unscored. This axis would measure recent cadence — papers per year, gaps, and whether the 2026 burst of three preprints reflects a sustained pace or a momentary peak.
Reproducibility58/100
The reproducibility score of 58 is the mean over three assessed papers, matching reproMean 58 and reproPapers 3. The spread is wide: ODU-Bench rates high for reproducibility readiness, the VoiceChat full-duplex model medium, and the memory-RL work low, reflecting Evolve's concerns about LLM-judge circularity and thin release detail. The average makes reproducibility the middle-strength axis in this profile.
Research summary
Hui Wang is a researcher at Peng Cheng Laboratory whose published work concentrates on conversational AI across three linked layers: speech interaction, multimodal intent understanding, and memory for long-term dialogue. The Evolve catalog contains three 2026 arXiv papers, each with Wang in a middle author position. NemotronLabs VoiceChat presents an open full-duplex speech-to-speech model with native tool calling, praised for turn-taking and barge-in behavior but showing weak argument accuracy (42.2%, Pass@1 33.0%). ODU-Bench formalizes demand detection and contextual intent inference in multimodal interaction as a benchmark with a solid pipeline, though an LLM-judged metric carries 60% of the score. The third applies reinforcement learning to memory management in long-term conversations via a two-agent setup with delayed rewards, reporting strong gains that Evolve tempers for LLM-judge circularity and thin reproducibility. Bibliometric indicators point to a highly collaborative contributor: 73 distinct coauthors and a collaboration score of 86, with reproducibility scoring 58 as the mean over three assessed papers. The field distribution splits evenly across speech/audio-visual multimodal learning, speech processing and conversational AI, and NLP for agent memory, producing a focus score of 0 — thematically narrow, numerically broad. Citation, output-volume, and per-year activity data are not available in the current snapshot, so those axes remain unscored in this early-stage profile.
Collaboration pattern
Co-authorship data shows 73 distinct collaborators and a collaboration score of 86, with 12 recurring partners — including Qi Chen, Yunfei Chu, Haolin He, Zihan Liu, Yuxuan Wang, and Ziyang Ma — suggesting stable small clusters inside a large rotating network. On all three catalogued papers Wang sits in a middle author position (none corresponding), consistent with team-based research spanning speech, multimodal, and NLP groups.
Research keywords
Highlights
First open full-duplex speech-to-speech model with native tool calling
Making contextual intent inference a measurable benchmark
Reinforcement learning for memory in long-term conversations
Timeline
- 2026Releases NemotronLabs VoiceChat, an open full-duplex speech-to-speech model with native tool calling capabilities.
- 2026Publishes ODU-Bench, formalizing demand detection and contextual intent inference in multimodal interaction as an explicit benchmark.
- 2026Proposes interactive memory learning, applying two-agent reinforcement learning with delayed rewards to long-term conversation memory.
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 are computed from a public bibliometric snapshot; citation, output-volume, and per-year activity data were unavailable, so those radar axes are left unscored.
- 'Hui Wang' is a common name; this profile is name-resolved from public bibliographic data and should be verified against ORCID or an institutional page before reuse.
- All three catalogued papers are 2026 arXiv preprints; no peer-reviewed venue information is within this data's scope.
- The reproducibility score of 58 averages three Evolve-reviewed papers whose per-paper readiness ranges from high to low, so the single number should not be read as uniform.