Evaluation scientist at Alibaba Group building benchmarks for computer-use, audio-language, and embodied AI agents
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-09) 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 metrics such as publication volume and yearly counts are not yet available for this profile. The axis would measure total papers, venue distribution, and the arXiv-versus-peer-reviewed split; only three reviewed papers currently anchor it.
Impact—
Impact data — citations, h-index, or downstream adoption of the 2026 benchmarks — is absent from the current snapshot, so this axis stays unscored. It would measure how often these evaluation resources are cited or reused by others.
Focus0/100
Focus scores 0, reflecting an unusually even three-way split of reviewed output: agent evaluation and benchmarking, audio-language model evaluation, and audio-visual embodied AI each account for about 33.3% of the catalog. The unifying thread is evaluation methodology — benchmarks, judging, and evidence grounding — rather than one application domain, so the score reads breadth over depth.
Collaboration68/100
Collaboration scores 68, computed from 30 distinct coauthors across just three reviewed papers — roughly ten new collaborators per paper. All three list Jin Xu in a middle author position rather than first or corresponding. Recurring partners include Xize Cheng, Ruixun Liu, Yuxuan Wang, Zhifang Guo, and Yunfei Chu, consistent with a group-internal contributor pattern spanning three sub-fields.
Activity—
Activity metrics such as publication cadence and year-over-year velocity are missing from the snapshot. With all three reviewed papers dated 2026, no multi-year trend exists yet; this axis would track sustained research engagement over time.
Reproducibility43/100
Reproducibility scores 43, the mean readiness across three assessed papers. OSWorld-Pro and MuLA-Bench rate medium — both document protocols thoroughly (2814 subgoals with 67k human labels; 16-language multi-tier auditing) — while OmniEcho rates low, its SOTA claims resting on its own benchmark. Benchmark releases help, but single-judge, single-run judging tempers trust.
Research summary
Jin Xu is an AI researcher at Alibaba Group whose recent work centers on how advanced models should be evaluated rather than on any single model architecture. All three reviewed papers are 2026 arXiv releases and share one methodological agenda: rigorous, auditable benchmarking. OSWorld-Pro adds process-based evaluation for computer-use agents, with 305 tasks, 2814 subgoals, and 67k human labels; judge validation is solid, though a single-judge, single-run design tempers trust. MuLA-Bench contributes a 16-language long-form audio understanding benchmark built through multi-tier auditing, with credible findings on conditional difficulty but LLM-constructed gold standards that complicate cross-model comparison. OmniEcho introduces a real-world first-order ambisonics benchmark — 2,972 QA pairs and 900 navigation tasks — plus an FOA-augmented omni-modal model whose gains are real yet modest. Radar scores reflect this breadth: focus is 0, with the catalog split evenly (~33.3% each) across agent evaluation, audio-language model evaluation, and audio-visual embodied AI; collaboration is 68, drawn from 30 distinct coauthors across three papers; reproducibility averages 43. On every reviewed paper Jin Xu holds a middle author position without corresponding authorship, indicating a contributor role within larger teams. Output, impact, and activity metrics are not yet available for scoring.
Collaboration pattern
Across three 2026 papers, Jin Xu worked with 30 distinct coauthors — roughly ten new collaborators per paper — and holds a middle author position on each rather than first or corresponding. The frequent-coauthor list (Xize Cheng, Ruixun Liu, Yuxuan Wang, Jiacheng Xie, Yuhuan You, Donghua Cai, Junming Lin, Xiong-Hui Chen, Zhifang Guo, Yunfei Chu, Qize Yang, Yiwu Zhong) suggests parallel team efforts spanning agent evaluation, audio-language research, and audio-visual embodied AI.
Research keywords
Highlights
Process-based evaluation for computer-use agents
A 16-language long-form audio understanding benchmark
First-order ambisonics benchmark for omni-modal embodied agents
Timeline
- 2026Co-authored OSWorld-Pro, adding process-based evaluation for computer-use agents (305 tasks, 2814 subgoals, 67k human labels) to the arXiv catalog.
- 2026Co-authored MuLA-Bench, a 16-language long-form audio understanding benchmark built via multi-tier auditing.
- 2026Co-authored OmniEcho, introducing a real-world FOA benchmark (2,972 QA + 900 navigation tasks) for omni-modal embodied agents.
- 2026All three reviewed papers carry 2026 dates, establishing an evaluation-centred agenda across agents, audio-language models, and embodied AI.
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.
- Radar scores and counts derive only from the provided bibliometric snapshot and reviewed catalog; output, impact, and activity axes lack sufficient data to score.
- All three reviewed papers are 2026 arXiv preprints; no peer-reviewed venue, citation, or award information falls within the current data scope.
- "Jin Xu" is a common name; this profile is name-resolved from public bibliographic data and Evolve's reviewed catalog only.
- All reviewed papers list a middle author position with no corresponding authorship, so per-paper individual contribution cannot be attributed from this data.