NTU researcher in LLM reinforcement learning and multi-agent systems — test-time RL for code, model-pool design
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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 scored. This axis would measure publication volume, venue tier, and authorship roles over time; the snapshot holds only two 2026 arXiv preprints, too small a base to score output reliably.
Impact—
Not yet scored. This axis would capture citation-based influence — citation counts, h-index, and uptake of the reviewed papers. No citation or h-index data was available in this snapshot, so no impact score is reported or inferred.
Focus0/100
Scores 0 on the 0-100 scale, and the raw field distribution shows why: the two reviewed papers split exactly 50/50 between LLM reinforcement learning (test-time RL for code generation) and LLM multi-agent systems (model-pool selection). Both sit under the same LLM-methods umbrella, but an even two-way split reads as breadth rather than a single concentrated research line on this metric.
Collaboration44/100
Scores 44/100 with 8 distinct coauthors recorded across the reviewed work. The frequent-coauthor list spans Feng Chen, Xiuneng Xu, Bo An, Sara Vera Marjanović, Aleksandr Laptev, Grigor Nalbandyan, Erik Arakelyan, and Evelina Bakhaturina, and authorship positions range from first author on the test-time RL paper to middle author on the multi-agent study.
Activity—
Not yet scored. This axis would measure publishing cadence and continuity — papers per year and sustained research activity. With only two 2026 preprints in scope and no multi-year record, activity cannot be estimated yet.
Reproducibility55/100
Averages 55/100 across the two papers with reproducibility assessments (reproPapers = 2), each rated medium readiness. Evolve's verdicts pinpoint the gaps: the 23-model multi-agent study's headline curves lack variance estimates, and the TTRL code-generation results carry no significance tests — systematic, well-ablated work whose statistical reporting could be stronger.
Research summary
Jiacheng Xu is a researcher at Nanyang Technological University working on methods for large language models, with a current profile that splits evenly between two subfields: LLM reinforcement learning and LLM multi-agent systems. The Evolve catalog contains two 2026 arXiv papers. The first, with Xu as first author, proposes entropy-regularized rank-masked policy optimization for test-time reinforcement learning in code generation, using a probe-consensus reward to counter reward hacking; Evolve's review calls the LiveCodeBench gains strong and well-ablated while noting the absence of significance tests. The second, a middle-author contribution, asks how to select model pools when designing multi-agent systems: across a systematic 23-model study it finds that growing heterogeneous pools can push performance below the best single model — judged credible overall, though headline curves lack variance estimates. Bibliometric signals are early-stage: 8 distinct coauthors, a collaboration score of 44/100, a focus score of 0 driven by the exact 50/50 field split, and a reproducibility score of 55/100 averaged over two medium-readiness papers. Output, impact, and activity remain unscored because the underlying data is insufficient. "Jiacheng Xu" is a common name, so this profile is name-resolved from public bibliographic data rather than a verified personal identity.
Collaboration pattern
Radar records 8 distinct coauthors and a collaboration score of 44/100. The frequent-coauthor list spans Feng Chen, Xiuneng Xu, Bo An, Sara Vera Marjanović, Aleksandr Laptev, Grigor Nalbandyan, Erik Arakelyan, and Evelina Bakhaturina. Xu's authorship positions range from first author on the test-time RL paper to middle author on the 23-model multi-agent study, indicating a mix of leading projects and contributing to larger collaborative efforts.
Research keywords
Highlights
First-author test-time RL method for code generation
23-model study on model-pool selection for multi-agent systems
Timeline
- 2026First-author arXiv paper on entropy-regularized rank-masked policy optimization for test-time reinforcement learning in code generation appears.
- 2026Co-authors (middle position) a systematic 23-model study on how to select model pools when designing multi-agent systems.
- 2026Both papers enter the Evolve catalog with medium reproducibility readiness; the radar records 8 distinct coauthors and an even 50/50 split across two LLM subfields.
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 signals come from a public OpenAlex-derived snapshot only; no citation counts, h-index, or venue rankings were available.
- "Jiacheng Xu" is a common name and no ORCID was provided; this profile is name-resolved from public bibliographic data, so paper attribution should be independently verified.
- The reproducibility score (55) averages only two assessed papers — a very small base.
- Both catalogued papers are 2026 arXiv preprints; peer-review status is not captured in this snapshot.