RL for LLM reasoning researcher at CAS: first-author InfoPPO (RLVR) and multi-agent debate co-author (R2-MAD)
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 measures total publications, authorship positions, and yearly output volume. The current snapshot lacks sufficient output data; only two Evolve-reviewed 2026 papers — one first-author — serve as verified signals so far.
Impact—
Not yet scored: this axis tracks citation counts, h-index-style aggregates, and influence on follow-up work. No citation or impact data is available in the current snapshot, so no estimate is made for Yongcheng Zeng.
Focus0/100
The focus score is 0, reflecting an even 50/50 split in the field distribution between reinforcement learning for large language models and natural language processing (LLM multi-agent systems), computed from two reviewed papers. The spread looks wide on paper, but the two lines are closely related — both target LLM reasoning — so the low score reads more as breadth within one problem area than scattered work.
Collaboration58/100
Collaboration scores 58, a mid-range value derived from 17 distinct coauthors across the two reviewed papers. The recurring coauthor list centers on Haifeng Zhang, Jun Wang, Yan Song, and Kaike Zhang; Yongcheng Zeng held a first-author slot on InfoPPO and a middle position on R2-MAD, suggesting both leading and supporting roles inside team-based projects.
Activity—
Not yet scored: this axis measures recent publication cadence and ongoing research momentum over time. The snapshot provides paper years for only two reviewed works, which is too little to compute an activity trend.
Reproducibility55/100
Reproducibility scores 55, averaged over 2 reviewed papers, and both carry a medium repro-readiness rating. Evolve's verdicts flag concrete gaps: InfoPPO relies on single-seed evaluations and shows theory-practice gaps, while R2-MAD reports consistent small gains across 4 benchmarks without significance testing — reproducible setups, but with verification caveats.
Research summary
Yongcheng Zeng is a researcher at the Chinese Academy of Sciences working on reinforcement learning and multi-agent methods for large language models. The current profile is built from two Evolve-reviewed 2026 arXiv papers, split evenly (50/50) between reinforcement learning for LLMs and NLP LLM multi-agent systems. As first author of "Information-Time Proximal Policy Optimization" (InfoPPO), Zeng re-times PPO updates by information density, drawing on predictive entropy, discounting, and credit assignment for RLVR-style training; Evolve finds the theory coherent and Qwen3 math gains consistent, though single-seed evaluations and theory-practice gaps temper the claims. On "Remember and Reweight" (R2-MAD), Zeng contributed as a middle author to a multi-agent debate framework adding experience memory, confidence estimation, and retrieval-augmented reasoning, yielding consistent small gains across 4 benchmarks without significance testing. Both papers carry a medium repro-readiness rating, and the reproducibility score is 55. Collaboration data shows 17 distinct coauthors (score 58), with recurring partners including Haifeng Zhang, Jun Wang, Yan Song, and Kaike Zhang. Output, impact, and activity axes are not yet scored because the underlying bibliometric data is insufficient. Name resolution is based on public bibliographic data; no ORCID is on file, so cross-database identity is not independently verified.
Collaboration pattern
Collaboration scores 58 with 17 distinct coauthors across two reviewed 2026 papers. Yongcheng Zeng works in team-based settings: first author on InfoPPO, then a middle-position co-author on R2-MAD. The recurring coauthor list centers on Haifeng Zhang, Jun Wang, Yan Song, Kaike Zhang, Xinyu Cui, and Guoqing Liu; "Jun Wang" appears twice in the frequent-coauthor data, which may reflect two different authors sharing a common name.
Research keywords
Highlights
First-author InfoPPO re-times PPO by information density
R2-MAD adds memory and confidence weighting to multi-agent debate
Dual-track work on LLM reasoning: RL training and agent collaboration
The two reviewed papers split the field distribution 50/50 between reinforcement learning for LLMs and LLM multi-agent systems, tracing one research thread that attacks LLM reasoning from both the training-algorithm side (first-author InfoPPO) and the multi-agent inference side (R2-MAD).
Timeline
- 2026First-author arXiv paper "Information-Time Proximal Policy Optimization" (InfoPPO): re-times PPO by information density for RLVR training, with consistent Qwen3 math gains.
- 2026Middle-position co-author on "Remember and Reweight" (R2-MAD): experience memory and confidence estimation for multi-agent debate, with consistent gains across 4 benchmarks.
- 2026Reviewed output spans two adjacent fields — reinforcement learning for LLMs and LLM multi-agent systems — at 50% each in the field distribution, with 17 distinct coauthors.
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 are computed from OpenAlex-based public data; the output, impact, and activity axes lack sufficient underlying data and are not scored.
- Only 2 papers from the Evolve catalog (both 2026 arXiv preprints) inform this profile; no citation-based impact is measured yet.
- No ORCID or country is on file; "Yongcheng Zeng" is resolved from public bibliographic data, so identity across databases is not independently verified.
- The reproducibility score (55) is averaged over only 2 papers, making it a coarse signal.