Ge Zhang

—

Unclaimed · AI-structured profileOpenAlex

LLM self-evolution and agent-evaluation benchmark builder; last author on Aspire and S3Gym

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—

Publication volume is not scored yet: OpenAlex coverage for this name-resolved profile is incomplete, with only three cataloged 2026 works fully reviewed. This axis would measure total output, venue spread, and year-over-year publishing consistency.

Impact—

Citation impact cannot be scored: the snapshot records h-index and i10-index as zero, reflecting missing or unlinked citation data rather than measured influence. This axis would track citations, citation velocity, and uptake of the self-evolution benchmarks.

Focus—

Research focus is not scored: OpenAlex topic tags exist (Topic Modeling, Software Engineering Research, NLP Techniques), but the verifiable corpus is too small to rate concentration. This axis would measure how tightly output clusters around core themes.

Collaboration72/100

The collaboration score of 72 rests on 36 distinct coauthors in the recent OpenAlex record — a broad network for an early-stage profile. Twelve frequent collaborators recur, including Jiajun Shi, Jiaheng Liu, Wenhao Huang, Shen Yan, Yuhao Wu, and Yuxuan Zhang. On the two cataloged 2026 benchmarks, Aspire and S3Gym, Ge Zhang appears in the last-author position, consistent with a senior or supervisory role, while REER-PT lists them among middle authors.

Activity67/100

The activity score of 67 reflects a recent, accelerating cadence: OpenAlex logs 9 works over the last three years, split as 3 works in 2025 and 6 in 2026 — a doubling within the observed window. All three Evolve-cataloged papers are 2026 arXiv releases spanning agent self-evolution benchmarks and pre-training data augmentation, indicating sustained submission activity concentrated in the current year.

Reproducibility—

Reproducibility is not scored: no artifact or dataset availability data exists yet. The two benchmarks were rated low-to-medium for reproducibility readiness; a profile-level score needs more reviewed outputs. This axis would measure released code, data, and evaluation harnesses.

Research summary

Ge Zhang works on large language model self-evolution, agent evaluation, and pre-training data. This profile is resolved from public OpenAlex bibliographic records; no verified institutional affiliation is attached, so biography beyond the publication record is not asserted. The cataloged record concentrates on 2025–2026: OpenAlex attributes 9 works over the last three years, 3 in 2025 and 6 in 2026 — an accelerating cadence consistent with the activity score of 67. The snapshot lists 36 distinct coauthors, supporting a collaboration score of 72; frequent collaborators include Jiajun Shi, Jiaheng Liu, Wenhao Huang, and Shen Yan. Topic attribution is led by Topic Modeling and Software Engineering Research (4 works each), followed by Natural Language Processing Techniques (3), Information Retrieval and Search Behavior (2), and AI-based Problem Solving and Planning (2). Three 2026 arXiv papers sit in the Evolve catalog, two with Ge Zhang as last author: Aspire, a benchmark for vague-goal LLM self-evolution with hidden expert evaluation, and S3Gym, which asks whether self-testing and self-judging become self-improvement. REER-PT, a middle-author paper, applies reverse-engineered reasoning annotations to perplexity-guided pre-training data augmentation. Evolve's reviews consistently note protocol-sound, honest negative results paired with statistically thin evaluation. Citation indicators (h-index, i10-index) read zero in this snapshot and are treated as missing.

Collaboration pattern

Ge Zhang's OpenAlex record lists 36 distinct coauthors, and twelve recur frequently, including Jiajun Shi, Jiaheng Liu, Wenhao Huang, Shen Yan, Yuhao Wu, Yuxuan Zhang, Xinping Lei, Zexuan Wang, Jinkai Liu, Haoran Que, Ting Huang, and Renming Pang. The pattern suggests team-based benchmark building: large co-author pools, last-author placement on the two 2026 agent benchmarks, and a middle-author slot on REER-PT. Collaboration scores 72 on this profile's radar.

Research keywords

LLM self-evolutionLLM agentsbenchmark designself-improvementself-judgingin-context learningexperience reusepre-training data augmentationreverse-engineered reasoningchain-of-thoughtvague goalssynthetic data

Highlights

Aspire: testing whether models can self-evolve from vague goals
A 2026 arXiv benchmark for vague-goal LLM self-evolution using a hidden expert evaluation set. Evolve's review credits honest reporting of negative results — gains are rare — while flagging statistically thin evaluation and a harness that sits below the reference implementation.
S3Gym: can self-testing and self-judging become self-improvement?
A 2026 arXiv benchmark, with Ge Zhang as last author, finding that LLM self-improvement is task-dependent. Evolve rates the protocol sound and the negative results honest, but notes that tiny 3-episode evaluations lack statistical rigor; reproducibility readiness is medium.
REER-PT: perplexity-guided pre-training data augmentation
A middle-author 2026 arXiv paper augmenting pre-training data with reverse-engineered reasoning annotations. Evolve notes perplexity drops and gains on a 680M-parameter model that look strong, but single-run results come without significance tests; reproducibility readiness is low.

Timeline

2025OpenAlex records 3 works in 2025, the earliest year covered by this profile's counts-by-year data.
2026Output doubles to 6 works; Aspire appears on arXiv, a benchmark for vague-goal LLM self-evolution with hidden expert evaluation, with Ge Zhang as last author.
2026S3Gym is released on arXiv, again with last-author placement, showing that self-testing and self-judging yield only task-dependent self-improvement.
2026REER-PT appears on arXiv as a middle-author contribution, using reverse-engineered reasoning annotations for perplexity-guided pre-training data augmentation.

Papers in the Evolve catalog

Frequent co-authors

Jiajun Shi×3 Jiaheng Liu×3 Wenhao Huang×3 Shen Yan×3 Yuhao Wu×2 Yuxuan Zhang×2 Xinping Lei×2 Zexuan Wang×2 Jinkai Liu×2 Haoran Que×1 Ting Huang×1 Renming Pang×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.

Identity is resolved by name from OpenAlex bibliographic data; 'Ge Zhang' is a common name and no ORCID or affiliation is attached, so attribution risk exists.
All bibliometrics come from the OpenAlex snapshot; citation indicators (h-index, i10-index) read zero and are treated as missing, not as evidence of no impact.
Qualitative judgments rest on only three Evolve-reviewed papers, all 2026 arXiv preprints; no peer-reviewed venue record is included.
Institution, country, and career history are absent from the input data and are therefore not described.