Zeyang Li

Massachusetts Institute of Technology · United States

Unclaimed · AI-structured profileOpenAlex

Early-stage MIT researcher uniting safe reinforcement learning and generative modeling through rigorous theory

papers here

reviewed by Evolve

first-author

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 underlying output record is insufficient. This axis would measure publication volume, authorship position, and venue tier over time; only 2 first-author 2026 arXiv preprints are verified here.

Impact—

Not yet scored. Impact would measure citation-based influence; this OpenAlex record shows h-index 0 and i10-index 0, so there is no citation evidence to assess — expected for 2026 preprints, not a negative judgment.

Focus0/100

Focus scores 0/100, the lowest signal in the radar, reflecting a split portfolio rather than a single line: fieldDist assigns 0.5 to machine learning (generative modeling) and 0.5 to safe reinforcement learning and control. Indexed topics likewise scatter — Reinforcement Learning in Robotics (4), Advanced Control Systems Optimization (2), Adaptive Dynamic Programming Control (2), plus single records in scientometrics and recommender systems.

Collaboration32/100

Collaboration scores 32/100, at the low end. The recent window lists 4 distinct coauthors, and the recurring set comprises Navid Azizan, Yunan Wang, Paolo Giaretta, and Sunbochen Tang — a small, stable circle. Both Evolve-cataloged papers are first-authored without corresponding authorship, consistent with contribution inside a compact supervised team rather than a large consortium or multi-institution network.

Activity52/100

Activity scores 52/100, mid-range. The OpenAlex snapshot records 5 works within the last three years, with countsByYear concentrated entirely in 2026 — no earlier years appear in the window. Both catalog papers carry 2026 arXiv timestamps, so current output is recent and clustered; the score reflects a short observable burst rather than a sustained multi-year publication cadence.

Reproducibility55/100

Reproducibility scores 55/100 (reproMean), computed from 2 assessed papers (reproPapers), both rated medium reproReadiness. Evolve's review notes concrete but narrow artifacts: the safe meta-RL paper runs two HalfCheetah tasks with 3 seeds and thin baselines, while Newton Matching provides rigorous KL-descent and quadratic-convergence theory with no empirical results shown — theory-forward work whose experimental replication base is still thin.

Research summary

Zeyang Li is an early-stage researcher at the Massachusetts Institute of Technology (US) whose recorded work spans two machine-learning lines: safe reinforcement learning and control, and generative modeling. This profile is name-resolved from OpenAlex bibliographic data. The snapshot records 5 works in the last three years, all counted in 2026, alongside 4 distinct coauthors, h-index 0, and i10-index 0 — a preprint-heavy, early-career footprint. Field distribution splits evenly (0.5 generative modeling, 0.5 safe RL and control), and indexed topics are led by Reinforcement Learning in Robotics (4), Advanced Control Systems Optimization (2), and Adaptive Dynamic Programming Control (2), with single records in scientometrics and recommender systems. The Evolve catalog holds two first-author 2026 arXiv preprints: Safe Meta-Reinforcement Learning via Information Space Reachability, praised for clean contraction theory in belief space but tested only on two HalfCheetah tasks with 3 seeds and thin baselines; and Newton Matching for Generative Modeling, a unified fine-tuning and sampling framework with KL-descent and quadratic-convergence guarantees but no empirical results shown. Reproducibility scores 55/100 across these 2 assessed papers, both medium readiness.

Collaboration pattern

Zeyang Li's recorded collaboration footprint is small and stable: 4 distinct coauthors in the recent OpenAlex window, with the recurring set comprising Navid Azizan, Yunan Wang, Paolo Giaretta, and Sunbochen Tang. Both catalog papers are first-authored but not corresponding, and the collaboration radar score is 32/100 — a compact, supervised-team pattern around MIT rather than a broad multi-institution network. Collaboration claims here rest on bibliographic co-authorship only.

Research keywords

safe reinforcement learningmeta-reinforcement learningHamilton-Jacobi reachabilityBayes-adaptive MDPbelief statestate-wise safety constraintsdiffusion modelsflow matchingreward fine-tuningsampling from unnormalized densitiesreverse KL divergenceinformation geometry

Highlights

Safe meta-RL via belief-space reachability, with clean contraction theory
First-author 2026 arXiv preprint formulates safe meta-reinforcement learning over Bayes-adaptive MDPs using Hamilton-Jacobi reachability in information space, enforcing state-wise safety constraints on belief states. Evolve's verdict: clean contraction theory and sound framing, but evidence limited to two HalfCheetah tasks, 3 seeds, and thin baselines.
Newton Matching: one framework for reward fine-tuning and sampling
Second first-author 2026 preprint proposes Newton Matching, unifying reward fine-tuning and sampling from unnormalized densities for diffusion models and flow matching. Evolve highlights KL-descent with quadratic-convergence guarantees grounded in information geometry and reverse KL divergence, but notes no empirical results are shown yet.

Timeline

2026Posts first-author arXiv preprint 'Safe Meta-Reinforcement Learning via Information Space Reachability', studying safe meta-RL with belief-space reachability.
2026Posts first-author arXiv preprint 'Newton Matching for Generative Modeling', unifying reward fine-tuning and sampling for diffusion and flow models.
2026OpenAlex snapshot records 5 works in the recent window and 4 distinct coauthors; both 2026 preprints enter the Evolve catalog rated medium for reproducibility readiness.

Papers in the Evolve catalog

2026-09-14first author
Safe Meta-Reinforcement Learning via Information Space Reachability
2026-09-04first author
Newton Matching for Generative Modeling: A Unified Framework for Fine-Tuning and Sampling

Frequent co-authors

Navid Azizan×2 Yunan Wang×1 Paolo Giaretta×1 Sunbochen Tang×1

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.

Bibliometrics come from OpenAlex only; 'Zeyang Li' is a common name, so this profile is name-resolved from public bibliographic data and may merge or split distinct people.
No ORCID is linked, and h-index 0 / i10-index 0 reflect an early-stage record that likely undercounts actual output and citations.
Output and impact axes are unscored (dimMissing) because citation-side data is insufficient; no values are estimated.
Qualitative evidence covers only 2 Evolve-reviewed arXiv preprints from 2026; neither is a peer-reviewed publication yet.