Zhen Zhang

Research profile

Research summary

Zhen Zhang, affiliated with Baidu in China, is an early-stage, dual-track researcher whose reviewed work spans two quite different areas: scientific machine learning for turbulence modelling, and LLM post-training with reinforcement learning. On the first track, a 2026 first-author arXiv paper, 'Well-posedness of neural turbulence closures and tangent dissipation', develops a well-posedness theorem for coupled neural turbulence closures — touching monotone-operator structure, eddy viscosity and tangent dissipation — and supports it with channel-flow experiments; Evolve rates its reproducibility readiness high. On the second track, Zhang contributes as a middle author to 'RLVR²: Reinforcement Learning with Verifiable Rubric-based Ranking' (2026), which converts rubric scores into within-group Bradley-Terry rankings used as reinforcement-learning rewards, reporting broad gains at three model scales, though Evolve notes single-run results, no significance tests and closed data. OpenAlex resolves this identity to 8 distinct co-authors, an exact 50/50 field split between the two tracks, 1 work in 2026 and 1 work in the trailing three years; h-index and i10-index are recorded as 0 and output and impact scores are unavailable. Given a very common name and name-resolution noise, this profile should be read as a grounded snapshot of the reviewed papers rather than a complete career record.

Research keywords

Highlights

• Turning rubric scores into verifiable ranking rewards (RLVR²)

  This 2026 arXiv paper converts rubric scores into within-group Bradley-Terry rankings that serve as reinforcement-learning reward signals, reporting broad gains across three model scales. Evolve's verdict: a promising reward-aggregation idea, but evidence rests on single runs, without significance tests, and training data is closed — reproducibility readiness rated low.

• A well-posedness theorem for neural turbulence closures

  First-authored arXiv work (2026) proves a well-posedness result for coupled neural turbulence closures, engaging monotone-operator structure, eddy viscosity and tangent dissipation, and backs the theory with channel-flow experiments. Evolve calls the theorem sound with an unverified global margin and rates reproducibility readiness high.

• One profile, two reproducibility cultures

  The two reviewed papers pull in opposite directions: the turbulence-closure theory is rated high for reproducibility, while the RLVR² pipeline is rated low. Together with an exact 50/50 field split, this makes the profile a compact snapshot of an early, industrially embedded dual-track research agenda.

Timeline

1. 2026First-author arXiv paper 'Well-posedness of neural turbulence closures and tangent dissipation' appears — the turbulence-modelling track, rated high for reproducibility.
2. 2026Contributes as a middle author to 'RLVR²: Reinforcement Learning with Verifiable Rubric-based Ranking' — the LLM post-training track, rated low for reproducibility.
3. 2026OpenAlex attributes 1 work to this identity in 2026, the only year with indexed output in the current snapshot, placing both tracks in the same burst.

Papers in the Evolve catalog

• 2026-09-20

  RLVR$^{2}$: Reinforcement Learning with Verifiable Rubric-based Ranking

• 2026-09-17first author

  Well-posedness of neural turbulence closures and tangent dissipation

Frequent co-authors