Nanjing University researcher building benchmarks for LLM self-evolution, self-judging agents, and pre-training data
3
papers here
0
reviewed by Evolve
0
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—
No output score is computed: the OpenAlex record shows only 4 works (all in 2026) with h-index 0 and i10-index 0, too thin to measure publication volume, venue consistency, and productivity trends over time.
Impact—
Impact is not scored: OpenAlex reports h-index 0 and i10-index 0 with no citation time series, so this axis—which would measure citation accumulation, uptake by other researchers, and venue visibility—lacks usable evidence.
Focus—
No focus score is available: the five OpenAlex topics (Topic Modeling 3; NLP Techniques 2; AI Problem Solving and Planning 2; two single-count topics) are too sparse to measure how concentrated or dispersed the research agenda is.
Collaboration72/100
The OpenAlex record lists 36 distinct coauthors, and Evolve's catalog shows a stable recurring circle of 12 frequent collaborators (including Jiajun Shi, Ge Zhang, and Wenhao Huang). All three reviewed 2026 papers are multi-author arXiv preprints with middle-position authorship, consistent with a team-based, contribution-sharing working style. This breadth of co-authorship underpins a collaboration score of 72, reflecting broad but not yet self-led networks.
Activity47/100
OpenAlex records 4 works in the recent three-year window, with the counts-by-year series containing only 2026 (4 works). Combined with the three 2026 arXiv preprints reviewed by Evolve, this shows a concentrated, recently started publication rhythm rather than a sustained multi-year record. The resulting activity score of 47 sits near the middle of the scale and should be revisited as the 2026 output accumulates.
Reproducibility—
Reproducibility is not scored: no code releases, artifact availability, or replication records appear in the data. Evolve's per-paper repro-readiness tags for the three reviewed papers (low, medium, low) are qualitative notes, not a computed axis.
Research summary
Jiaheng Liu is a researcher affiliated with Nanjing University whose reviewed work centers on evaluating and improving large language models, with three 2026 arXiv preprints in the Evolve catalog. Two of these probe whether LLMs can genuinely improve themselves: Aspire asks whether models self-evolve from vague goals under a hidden expert evaluation set, and S3Gym tests whether self-testing and self-judging translate into self-improvement, finding gains that are task-dependent. A third, REER-PT, moves upstream to pre-training, using reverse-engineered reasoning annotations to guide perplexity-based data augmentation, with reported perplexity drops and gains for a 680M-parameter model. Across these papers Liu appears as a middle-position co-author on collaborative team efforts rather than corresponding author. OpenAlex records 36 distinct coauthors and a recurring circle of 12 frequent collaborators, supporting a collaboration score of 72, while output activity is early: 4 works in the recent three-year window, all landing in 2026, giving an activity score of 47. Citation-based metrics (h-index 0, i10-index 0) and a focus score are not yet computable from the available record. The profile should be read as that of an early-stage contributor to LLM self-evolution and evaluation research whose direction and independent leadership are still taking shape.
Collaboration pattern
Liu works in sizeable teams: the three reviewed 2026 papers are multi-author arXiv efforts with middle-position, non-corresponding roles. OpenAlex counts 36 distinct coauthors overall, and 12 recur across the record — Jiajun Shi, Ge Zhang, Wenhao Huang, Shen Yan, Yuhao Wu, Yuxuan Zhang, Xinping Lei, Zexuan Wang, Jinkai Liu, Haoran Que, Ting Huang, and Renming Pang — forming a stable cluster consistent with the collaboration score of 72.
Research keywords
Highlights
Aspire: testing whether LLMs self-evolve from vague goals
S3Gym: when self-testing and self-judging become self-improvement
REER-PT: perplexity-guided pre-training data from reverse-engineered reasoning
Timeline
- 2026Co-authored Aspire, a benchmark testing LLM self-evolution from vague goals under a hidden expert evaluation set (arXiv preprint, middle author).
- 2026Co-authored S3Gym, a benchmark asking whether LLM self-testing and self-judging translate into self-improvement.
- 2026Co-authored REER-PT, applying reverse-engineered reasoning annotations to perplexity-guided pre-training data augmentation.
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.
- Citation and output metrics come from OpenAlex only (h-index 0, i10-index 0, 4 works in 2026); the record is likely incomplete and does not establish absence of impact.
- Scope is limited to three 2026 arXiv preprints reviewed by Evolve; no peer-reviewed venue, grant, or award data is included.
- "Jiaheng Liu" is a common Chinese name; this profile is name-resolved from public bibliographic data (OpenAlex identity, Nanjing University), so record fragmentation or conflation is possible.
- No ORCID is provided, and all reviewed papers list the researcher as a middle-position, non-corresponding author, so individual contribution boundaries cannot be verified from this data.