Xize Cheng

Output—

Output cannot be scored yet: reliable publication-volume records for this OpenAlex identity are insufficient, so total and per-year output remain unknown. This axis would measure cumulative and recent publication counts once the author record is disambiguated.

Impact—

Impact is not scored yet: citation-linked metrics such as h-index and citation counts are not reliably available for this identity, so influence cannot be estimated. This axis would measure citation accumulation and uptake of the benchmarks once citation records resolve.

Focus0/100

Focus scores 0/100 because recorded output spreads evenly across three adjacent fields — audio-visual machine learning for embodied AI, audio-language model evaluation, and speech NLP each take about 33% — so no single field yet dominates. Topic counts still cluster on speech and audio (Speech Recognition and Synthesis: 3; Music and Audio Processing: 3), suggesting concentration within one modality rather than scattered interests.

Collaboration65/100

Collaboration scores 65/100. The OpenAlex record lists 25 distinct coauthors, and 12 recur frequently — including Jin Xu, Xinyu Zhang, Yunfei Chu, Zhifang Guo, and Xiong-Hui Chen — pointing to a stable core team plus a wide per-project network. All three cataloged 2026 papers are multi-author team efforts in which Cheng holds a middle position except Vox-Infinity, where he is first author; none lists him as corresponding author.

Activity52/100

Activity scores 52/100. Recent output is thin but concentrated: 5 works indexed in the last three years, and the counts-by-year series records 5 works in 2026 alone — essentially all visible activity clusters in a single year. The catalog confirms this burst, with three 2026 arXiv benchmark papers spanning multilingual audio, long-context speech, and embodied audio-visual settings; earlier-year counts are absent from the record.

Reproducibility43/100

Reproducibility scores 43/100, the mean across 3 assessed papers. MuLA-Bench and Vox-Infinity rate medium: their conditional-difficulty and recency-effect findings are considered credible, but LLM-constructed gold standards, LLM judging, and partial data release limit rigor. OmniEcho rates low: its 2,972-question FOA benchmark shows real but modest gains, and SOTA claims rest on its own benchmark rather than external validation.