Machine learning researcher (NUDT) across language-model interpretability, multimodal embeddings, and 3D point clouds
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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 yet scored: the catalog's three arXiv preprints lack the publication-volume and venue data this axis needs. It would measure total and per-year research output across preprints and peer-reviewed venues.
Impact—
Not yet scored: no citation counts, h-index, or download statistics appear in the input data. This axis would measure citations to the three preprints and uptake of their methods by other research teams.
Focus0/100
Scores 0 because output is fully spread: the raw field distribution splits evenly, with machine-learning interpretability, multimodal representation learning, and 3D point cloud understanding each at about one-third of the record. Zero here reflects maximal breadth across three distinct subfields rather than missing data; no single field exceeds a third of the reviewed output.
Collaboration62/100
Scores 62 from bibliometric collaboration data: 21 distinct coauthors appear in the record, with 12 recurring partners including Ke Cheng, Xin Xu, Yixiao Chen, and Lei Xin. All three reviewed papers list Zhang in a middle author position, consistent with work inside larger teams rather than solo or corresponding-author projects. The network spans all three research fields, indicating broad but team-embedded collaboration.
Activity—
Not yet scored: only the 2026 preprint year is available, with no submission cadence or venue timeline. This axis would measure recent publication frequency and ongoing research tempo.
Reproducibility55/100
Scores 55, the mean across three assessed papers in Evolve's reproducibility review. Each of the three preprints is individually rated medium on reproducibility readiness: the induction-circuit study's patching-probe evidence is called unusually rigorous, but PACE reports no significance tests and MSSP provides no variance data for its thin ScanNet margin.
Research summary
Jun Zhang is a machine learning researcher at the National University of Defense Technology whose reviewed work to date spans three areas in roughly equal measure: mechanistic interpretability of language models, multimodal representation learning, and 3D point cloud scene understanding. The Evolve catalog lists three 2026 arXiv preprints, on each of which Zhang is a middle-position author rather than first or corresponding author. The interpretability paper proposes that hybrid attention-recurrence architectures organize induction circuits around the preceding token, with patching-probe evidence Evolve describes as unusually rigorous; PACE introduces progressive angular-to-norm contrastive training to stabilize dot-product multimodal embeddings, reporting MMEB gains without significance tests; and MSSP tackles unsupervised semantic segmentation of 3D point clouds through multi-scale, spatially constrained partitioning, reporting the best unsupervised mIoU on S3DIS and ScanNet. Bibliometric signals are early-stage: 21 distinct coauthors give a collaboration score of 62; an even one-third split across the three fields yields a focus score of 0, indicating breadth rather than concentration; and a reproducibility score of 55 averages three assessed papers, each rated medium on reproducibility readiness. Output, impact, and activity scores are not yet computable from the available data. Because Jun Zhang is a common name and no ORCID is provided, this profile is name-resolved from public bibliographic data.
Collaboration pattern
Bibliometric data records 21 distinct coauthors, with 12 recurring collaborators including Ke Cheng, Xin Xu, Yixiao Chen, Lei Xin, and Jianbo Zhao. On all three reviewed 2026 preprints Zhang appears as a middle author rather than first or corresponding author, suggesting contributions embedded in larger project teams. Co-authorship spans all three fields — interpretability, multimodal learning, and point cloud understanding — indicating several parallel collaboration lines rather than one stable group.
Research keywords
Highlights
How hybrid architectures organize induction circuits
PACE: stabilizing dot-product multimodal embeddings
MSSP: unsupervised segmentation of 3D point clouds
Timeline
- 2026Releases the MSSP preprint on arXiv, claiming the best unsupervised mIoU for 3D point cloud semantic segmentation on S3DIS and ScanNet via multi-scale spatially-constrained partitioning.
- 2026Releases PACE on arXiv, proposing progressive angular-to-norm contrastive embedding training for multimodal models with reported MMEB gains.
- 2026Releases a mechanistic-interpretability study of induction circuits in hybrid attention-GDN architectures, keyed to the preceding token.
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 OpenAlex-derived public bibliometric data only.
- All three catalog papers are 2026 arXiv preprints; no peer-reviewed venue information is available.
- Output, impact, and activity axes lack underlying data and are unscored; no citation counts are provided.
- No ORCID or country is on record; Jun Zhang is a common name, so this profile is name-resolved from public bibliographic data and may not uniquely identify the individual.