Tsinghua researcher spanning LLM agent harnesses for science, quantization mechanisms, and real-time video generation
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: this snapshot lacks publication-volume data such as total output counts, venue types, and yearly productivity, so this axis — which measures research output volume — cannot be assessed.
Impact—
Not yet scored: citation counts, h-index, and downstream adoption metrics are absent from the input data, so this axis — which measures scholarly influence and uptake of the work — cannot be estimated.
Focus0/100
Scored 0: the bibliometric field distribution splits evenly three ways — 33.3% AI-for-science agent harnesses, 33.3% video generation, and 33.3% LLM quantization — across the three 2026 papers in the catalog. This indicates a deliberately broad portfolio rather than a single concentrated line; the score reflects even dispersion across fields, not inactivity.
Collaboration80/100
Scored 80: the co-author network includes 53 distinct collaborators, with 12 recurring partners such as Jianfei Chen, Yimi Wang, and Liuzhenghao Lv. Across the three reviewed 2026 papers Zhu appears in middle and last author positions without a corresponding-author role, consistent with contributions embedded in larger team-based deep-learning projects rather than solo-led studies.
Activity—
Not yet scored: yearly submission cadence and recent publication timing beyond the three 2026 papers are not provided, so this axis — which measures ongoing research pace — has insufficient underlying data to assess.
Reproducibility43/100
Scored 43, the mean readiness across the three assessed papers. It varies by project: the OpenAI4S agent harness and the quantization mechanistic study rate medium, while Vidu S2 rates low — Evolve notes it lacks ablations, significance tests, and released code despite broad benchmark claims.
Research summary
Jun Zhu is a Tsinghua University researcher whose reviewed output concentrates in three parallel areas: LLM agent harnesses for scientific discovery, mechanistic analysis of post-training quantization for large language models, and real-time video generation. Three 2026 arXiv papers anchor the profile. OpenAI4S: Code as Action, Science as Sessions introduces an agent harness built on persistent kernels and provenance tracking, reporting 7.83 versus Claude Code's 5.79–6.36 across 36 science scenarios; Evolve cautions that part of the gain is built into the harness's own evaluator. Why Does Post-Training Quantization Work? argues that pretrained blocks cancel inherited quantization error while LM-head geometry protects top tokens — an account Evolve judges well-supported. Vidu S2, where Zhu is the last-listed author, claims real-time 720p interactive avatars and streaming video editing with broad benchmark wins, though without ablations, significance tests, or code. Bibliometric signals show strong collaboration (score 80; 53 distinct coauthors; 12 frequent partners) and a focus score of 0, reflecting an even three-way split across fields. Reproducibility readiness averages 43 over the three assessed papers, ranging from medium to low. Because 'Jun Zhu' is a common name, this profile is name-resolved from public bibliographic data.
Collaboration pattern
The collaboration footprint is broad: 53 distinct coauthors and a stable core of 12 frequent collaborators, including Jianfei Chen, Gongbo Zhang, Hao Li, Yu Wang, Mujie Lin, Liuzhenghao Lv, Yicheng Mao, Yimi Wang, Minhan Tang, Zhengxiang Jiang, Yusong Wang, and Jiayu Yao. On the three reviewed 2026 papers Zhu holds middle positions on two and the last position on Vidu S2, with no corresponding-author role — a pattern of contributing within larger team-based projects.
Research keywords
Highlights
Agent harness for code-as-action science sessions
Explaining why post-training quantization works
Real-time interactive and editable video generation
Timeline
- 2026Publishes OpenAI4S, an agent harness for code-as-action science sessions benchmarked against Claude Code on 36 scenarios.
- 2026Co-authors the mechanistic study on why post-training quantization works for large language models.
- 2026Appears as last-listed author on Vidu S2, claiming real-time 720p interactive video generation and streaming editing.
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 public bibliographic records only; citation counts, venue tiers, and funding data are not part of this snapshot.
- 'Jun Zhu' is a common name; the profile is name-resolved from public bibliographic data, and identity details (ORCID, country) are unverified.
- Coverage is limited to the three 2026 arXiv papers reviewed by Evolve; earlier or non-arXiv work is not represented.
- Output, impact, and activity dimensions are unscored because the underlying metrics (citations, yearly volume) were not provided.