Early-career LLM self-evolution researcher; first author of the S3Gym benchmark and co-author on Aspire and REER-PT
3
papers here
0
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
0
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—
Not yet scorable: the OpenAlex record lists only 3 works in the last three years with no full career publication count, so cumulative productivity cannot be measured. This axis would track output volume over time across venues and formats.
Impact—
Insufficient data: h-index and i10-index both read 0 in the OpenAlex record and no citation counts are available, so research impact is not yet measurable. This axis would capture citations, uptake, and influence of the papers.
Focus—
Too sparse to score: OpenAlex topic tags exist, but with only 3 recent works the record cannot quantify research concentration. This axis would measure how tightly output clusters around core topics over time.
Collaboration72/100
Scores 72/100. OpenAlex records 36 distinct coauthors across just 3 recent works — an unusually wide network relative to the record's size — and the recurring co-author pool holds twelve names, among them Jiaheng Liu, Ge Zhang, Wenhao Huang, Shen Yan, Yuhao Wu, and Yuxuan Zhang. All three catalog papers are multi-authored arXiv collaborations, indicating team-embedded work rather than solo research.
Activity40/100
Scores 40/100. OpenAlex counts 3 works in 2026 and 3 works over the last three years, with no earlier years on record — all catalog output is 2026 arXiv preprints. The moderate score reflects a short observable window rather than inactivity; a longer horizon is needed before publication cadence can be judged.
Reproducibility—
No aggregate score yet: Evolve rates reproducibility readiness per reviewed paper (one medium, two low in this catalog), but no record-level metric exists. This axis would measure how available code, data, and experimental setups are for replicating results.
Research summary
Jiajun Shi is an early-stage researcher working on self-evolution and self-improvement in large language models, with an emerging footprint in benchmark design and pre-training data methods. The Evolve catalog holds three 2026 arXiv papers: S3Gym, first-authored, asking whether LLMs can turn self-testing and self-judging into self-improvement; Aspire, a middle-author collaboration testing whether models can self-evolve from vague goals against a hidden expert evaluation set; and REER-PT, a middle-author method adding reverse-engineered reasoning annotations as perplexity-guided pre-training data augmentation. A consistent thread is honest evaluation: Evolve's reviews credit sound protocols and candid negative results while flagging thin statistical power — rare gains, three-episode evaluations, and single-run results without significance tests. OpenAlex attributes 3 works in the last three years, all in 2026, alongside 36 distinct coauthors and h-index and i10-index of 0, consistent with a profile at its very beginning. Assigned topics span topic modeling, AI-based problem solving and planning, NLP techniques, reinforcement learning in robotics, and AI in games. The profile is name-resolved from public bibliographic data; no institution, country, or ORCID is on record.
Collaboration pattern
The record shows 36 distinct coauthors across only 3 recent works — a wide network for its size — with a recurring pool of twelve names, including Jiaheng Liu, Ge Zhang, Wenhao Huang, Shen Yan, Yuhao Wu, and Yuxuan Zhang. All three catalog papers are multi-author arXiv collaborations where Shi is first author once and middle author twice, consistent with working inside a stable, collaboration-heavy LLM research group.
Research keywords
Highlights
First-authored S3Gym: can self-testing and self-judging become self-improvement?
Aspire: benchmarking self-evolution from vague goals
REER-PT: perplexity-guided pre-training data augmentation
Timeline
- 2026Co-authored Aspire, an arXiv benchmark on whether models can self-evolve from vague goals under hidden expert evaluation
- 2026First-authored S3Gym, an arXiv benchmark on turning LLM self-testing and self-judging into self-improvement
- 2026Co-authored REER-PT, an arXiv method for perplexity-guided pre-training data augmentation with reverse-engineered reasoning annotations
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.
- Bibliometrics come from a single OpenAlex record; no institution, country, or ORCID is available, so the profile is name-resolved from public bibliographic data.
- h-index and i10-index read 0 and no citation counts are provided; impact is unmeasured in this data, not necessarily absent.
- All cataloged output is three 2026 arXiv preprints; any earlier or non-arXiv work falls outside the observed window.
- Reproducibility readiness is rated per reviewed paper only; no record-level reproducibility score exists.