Leqi Zheng

Tsinghua University · China

Unclaimed · AI-structured profileORCID OpenAlex

Early-career Tsinghua researcher in LLM post-training with RL and benchmark design for code, education, and legal NLP.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) 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 cannot be scored yet: the snapshot provides per-year work counts but no verified full publication list and no venue-normalized production measure, so this axis — which would rate total research output weighted by venue quality and role — is left unscored.

Impact—

Impact is left unscored: the record shows an h-index of 2 and an i10-index of 0 across roughly 29 recorded works — too early-stage for a stable citation-impact rating — and the snapshot lacks per-paper citation data that this axis would require.

Focus0/100

The focus score is 0, reflecting maximal dispersion: the field distribution splits evenly three ways — LLM post-training/reinforcement learning, NLP in education, and database systems/LLM code generation (33.3% each) — while the five largest topic clusters (topic modeling 10, recommender systems 8, graph neural networks 7, information retrieval 5, machine learning in healthcare 5) reach into further areas.

Collaboration75/100

Collaboration scores 75, drawn from 43 distinct coauthors — a wide, Tsinghua-centered network with twelve recurring partners including Zhaolu Kang, Fang Niu, Chaokun Wang, Yantao Liu, Tailong Luo, and Lei Wei. In the Evolve catalog the researcher holds middle-author positions on three of four papers and first authorship on one, consistent with team-based work rather than solo leadership.

Activity98/100

Activity scores 98, the strongest signal in the record: yearly output climbed from 1 work in 2020 and 1 in 2024 to 8 in 2025 and 19 in 2026, with 28 works in the most recent three-year window. The cadence is dominated by 2026 arXiv output, including all four Evolve-cataloged papers.

Reproducibility55/100

Reproducibility scores 55, averaged over three assessed papers (reproMean 55), each rated medium readiness: ProcArena, SLATE, and the gradient-aligned-rewards RLVR paper are checkable in principle, but Evolve reviews flag LLM-constructed gold tasks, a small system count, and tiny benchmarks as limits on replication confidence.

Research summary

Leqi Zheng is an early-career researcher at Tsinghua University whose recorded profile is defined by an unusually steep activity curve: output rose from one work in 2020 and one in 2024 to eight in 2025 and nineteen in 2026, giving 28 works in the last three years against an h-index of 2 and an i10-index of 0. The Evolve catalog holds four 2026 arXiv papers that illustrate this range: a first-author reinforcement-learning method (gradient-aligned rewards for RLVR math reasoning), the ProcArena benchmark for direct and interactive NL-to-PL/SQL development, the SLATE benchmark asking whether AI-generated slides actually teach, and JPO for structured legal reasoning in criminal judgment prediction. The spread is structural, not incidental: the field distribution splits evenly three ways across LLM post-training/reinforcement learning, NLP in education, and database systems/LLM code generation, while the top topics — topic modeling (10), recommender systems (8), graph neural networks (7) — reach further. The radar's focus score of 0 alongside an activity score of 98 captures exactly this: very high throughput, low specialization. Collaboration is broad (43 distinct coauthors, 12 recurring partners) at 75, and reproducibility is middling at 55 across three assessed papers, all rated medium readiness. Output and impact cannot yet be scored from this snapshot.

Collaboration pattern

Research is team-based and Tsinghua-anchored: 43 distinct coauthors with a recurring core of 12 — Zhaolu Kang, Fang Niu, Chaokun Wang, Yantao Liu, Tailong Luo, Lei Wei, Chenghua Zhu, Junhao Gong, Jiachen Qian, Eric Hanchen Jiang, Jiaxin Liu, and Yuan Wang — spanning the reinforcement-learning, education-NLP, and database/code-generation lines. In the Evolve catalog the researcher takes middle-author roles on three papers and first authorship on the gradient-aligned-rewards RLVR paper.

Research keywords

RLVR (reinforcement learning with verifiable rewards)GRPOreward shapinggradient alignmentLLM benchmark designexecution-based evaluationNL-to-PL/SQLstored proceduresinteractive code generationAI slide generationlanguage teaching qualitylearning gain

Highlights

First-author gradient-aligned rewards for RLVR math reasoning
As first author, proposed expert-CoT gradient-cosine reward shaping that reportedly lifts RLVR math pass@k at under 9% overhead, with NTK-based motivation; Evolve credits the theory but notes the gains are small on tiny benchmarks (medium reproducibility).
ProcArena: first multi-scenario NL-to-PL/SQL benchmark
Middle-author benchmark covering direct and interactive PL/SQL development from natural language with execution-based evaluation; Evolve credits the design but flags LLM-built gold tasks and missing significance tests (medium reproducibility).
SLATE: do AI-generated slides actually teach?
Middle-author benchmark for language-teaching quality and learner knowledge acquisition, using a VLM-as-learner proxy and near/far transfer measures; Evolve calls the design rigorous but notes the headline D1/D2 dissociation rests on only 10 systems.
JPO: Juris Policy Optimization for legal reasoning
Middle-author 2026 paper on structured legal reasoning in criminal judgment prediction; it is listed in the catalog but not yet Evolve-reviewed, so no verdict, keywords, or reproducibility rating exist for it.

Timeline

2020First recorded work appears in the OpenAlex snapshot: 1 work in 2020, the earliest data point in the record.
2024A single recorded work precedes the current ramp-up: the counts show 1 work in 2024 before output accelerates.
2025Output accelerates to 8 works, marking the start of the current productive phase.
2026Peak year: 19 recorded works, including all four Evolve-catalog papers — one first-author RLVR method, two benchmarks (ProcArena, SLATE), and the JPO legal-reasoning paper.

Papers in the Evolve catalog

Frequent co-authors

Zhaolu Kang×2 Fang Niu×2 Chaokun Wang×2 Yantao Liu×1 Tailong Luo×1 Lei Wei×1 Chenghua Zhu×1 Junhao Gong×1 Jiachen Qian×1 Eric Hanchen Jiang×1 Jiaxin Liu×1 Yuan Wang×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 an OpenAlex-resolved profile only; 'Leqi Zheng' is a common Chinese name, so identity rests on public bibliographic matching plus the listed ORCID, not independent verification.
Only 4 of roughly 29 recorded works have Evolve-reviewed detail; the remaining works are counted in yearly totals but not individually assessed.
Output and impact axes are unscored because the snapshot lacks a verified publication list and per-paper citation data.
Reproducibility (55) averages 3 assessed papers; the fourth catalog paper (JPO) has no review, keywords, or readiness rating yet.