Early-career researcher applying RL-trained LLMs to tool learning and text reranking, with two 2026 arXiv preprints
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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—
The output axis measures publication volume, venue quality, and authorship position over time. Evolve cannot compute it here: only two 2026 arXiv preprints are cataloged, with no peer-reviewed record yet, so this axis stays unscored.
Impact—
The impact axis aggregates citations, usage, and downstream uptake of a researcher's work. No citation metrics exist for these 2026 preprints in the supplied data, so impact is left unscored rather than estimated.
Focus0/100
Scored 0 because the available record is evenly split: 50% machine learning (LLM tool learning / reinforcement learning for agents) and 50% information retrieval / natural language processing, one paper each. The two lines do share a methodological core in GRPO-style reinforcement learning for LLMs, but across only two cataloged 2026 preprints Evolve cannot yet tell deliberate dual expertise from an unsettled early direction, so focus reads as very dispersed.
Collaboration48/100
Scored 48 from bibliographic co-authorship data: the record lists 10 distinct coauthors across the two reviewed 2026 preprints, with Lijun Liu, Yuanyuan Zhao, and Fei Huang recurring most often, alongside Shihao Liu, Hao Yin, and Wenyan Li. Chen appears as a middle-position, non-corresponding author on both papers, consistent with contributing within a small-to-mid-size team rather than leading an independent group; the moderate score reflects a still-thin, two-paper co-authorship base.
Activity—
The activity axis tracks publishing cadence and recency across years. With both cataloged papers dated 2026 and no multi-year metadata available, Evolve cannot compute a meaningful activity trend for this profile yet.
Reproducibility38/100
Scored 38, the mean repro-readiness across the two reviewed papers. MATCH rates medium: Evolve credits strong ablations but notes single-run results lacking variance tests. MERIT-Rank rates low: despite strong BRIGHT/TREC/BEIR numbers, its efficiency claims and reproducibility need scrutiny. Neither paper lists Chen as corresponding author, so artifact and code-release practice cannot be attributed personally; averaged over only two papers, this score should be read as preliminary.
Research summary
Zhengzong Chen is an early-stage researcher whose cataloged work sits at the intersection of reinforcement learning for large language models and information retrieval. Evolve has reviewed two 2026 arXiv preprints on which Chen is a middle-listed, non-corresponding author. The first, MATCH, couples curriculum scheduling with hierarchically gated rewards for model-aware tool learning; Evolve credits strong ablations but notes that single-run results lack variance tests. The second, MERIT-Rank ('Think Thrice Before Reranking'), integrates multi-perspective evidence and staged reinforcement learning to improve LLM text reranking on BRIGHT, TREC, and BEIR; the review judges the results strong but flags efficiency claims and reproducibility for scrutiny. Available bibliometric signals come from this two-paper record: 10 distinct coauthors and an even 50/50 split between LLM tool learning / RL for agents and IR / NLP, yielding a collaboration score of 48, a focus score of 0 reflecting the evenly divided two-field profile, and a mean reproducibility score of 38 (medium for MATCH, low for MERIT-Rank). Output, impact, and activity dimensions cannot be computed yet. Given the 2026-only, preprint-only record, this profile should be read as an emerging one in RL-based LLM training methods, pending peer-reviewed venues and citation evidence.
Collaboration pattern
Chen's co-authorship pattern, drawn from 10 distinct collaborators on two 2026 preprints, centers on a recurring group: Lijun Liu, Yuanyuan Zhao, and Fei Huang each appear twice in the co-occurrence list, with Shihao Liu, Hao Yin, and Wenyan Li also present. Chen holds middle, non-corresponding positions on both papers, suggesting a contributor role within a stable team spanning LLM tool learning and IR/NLP work rather than an independent research line.
Research keywords
Highlights
MATCH: curriculum scheduling with hierarchically gated rewards for RL tool learning
MERIT-Rank: multi-perspective reasoning for LLM text reranking
Timeline
- 2026First cataloged arXiv preprint, MATCH, introduces model-aware tool learning with curriculum scheduling and hierarchically gated rewards (middle author, repro-readiness medium).
- 2026Second cataloged preprint, MERIT-Rank ('Think Thrice Before Reranking'), applies multi-perspective evidence and staged reinforcement learning to LLM reranking on BRIGHT/TREC/BEIR (middle author, repro-readiness low).
- 2026Bibliometric snapshot of the record: 10 distinct coauthors, an even 50/50 split between LLM tool learning and IR/NLP, and a mean reproducibility score of 38 across the two papers.
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.
- Bibliometric indicators derive from OpenAlex-based public bibliographic data; no citation counts, h-index, or venue-level metrics were available for these 2026 arXiv preprints.
- Institution, country, and ORCID are unresolved; the profile is name-resolved from public bibliographic data and common-name ambiguity cannot be excluded.
- Output, impact, and activity dimensions are unscored because the record is limited to two 2026 preprints.
- The reproducibility score (38) averages only two reviewed papers and should be read as preliminary.