RL researcher in the GRPO lineage: group policy optimization for LLM post-training and multi-turn agentic tasks
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
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 scored: the catalog holds only the two Evolve-reviewed 2026 arXiv papers, so publication volume, venue spread, and yearly counts are insufficient to compute this axis. It would measure total and per-year paper output.
Impact—
Not yet scored: no citation counts, h-index, or usage signals are available in the provided data for these two 2026 preprints. This axis would measure citation-based influence and uptake of the researcher's work.
Focus0/100
Scored 0 on topical focus. The field distribution splits evenly at 0.5/0.5 between reinforcement learning for LLM agents and LLM post-training/reinforcement learning, so no single sub-field dominates this two-paper catalog. Both branches still share one technical core — GRPO-family group policy optimization — spanning training-time objectives (GVPO++) and multi-turn agent credit assignment (PGPO).
Collaboration50/100
Scored 50 on collaboration strength. Across the two reviewed 2026 papers, Yang Song worked with 11 distinct coauthors, including Junwei Bao, Yuyao Zheng, Haipeng Sun, Lemao Liu, Hongfei Jiang (the only repeated name), Dejing Dou, Kaichen Zhang, Yuzhong Hong, Dingqian Hong, and Hui Xiong. Both papers list a middle, non-corresponding author position, consistent with contributing inside a larger RL-for-LLM team rather than solo leadership.
Activity—
Not yet scored: with only two 2026 papers in the catalog and no full publication history, recent-year cadence and longitudinal productivity cannot be computed. This axis would measure publishing continuity and research momentum over time.
Reproducibility78/100
Scored 78 — the mean repro-readiness across 2 reviewed papers. GVPO++ is rated high: its zero-sum weights remove importance sampling under exact-optimum theory, though Evolve notes the 7B gains lack error bars. PGPO is rated medium: gains are largely credible but small-margin, three-seeded, and shown on only two benchmarks.
Research summary
Yang Song is a machine learning researcher working at the intersection of reinforcement learning and large language models, with a catalog profile built from two 2026 arXiv papers reviewed by Evolve. The work splits evenly — 0.5/0.5 in the field distribution — between reinforcement learning for LLM agents and LLM post-training/reinforcement learning, and both papers converge on one technical thread: group-based policy optimization in the GRPO family. GVPO++ extends group variance policy optimization to post-training and on-policy distillation, replacing importance sampling with zero-sum weights under an exact-optimum theory for KL-constrained reward maximization; Evolve rates its reproducibility readiness high while noting the 7B-scale gains lack error bars. PGPO adds potential-guided step-level credit assignment to group RL for multi-turn agentic tasks; its gains are judged largely credible but small-margin, three-seeded, and shown on only two benchmarks, with medium reproducibility readiness. Across the two papers Yang Song appears in a middle, non-corresponding author position and has worked with 11 distinct coauthors, including Junwei Bao, Yuyao Zheng, Haipeng Sun, Lemao Liu, Hongfei Jiang, Dejing Dou, Kaichen Zhang, Yuzhong Hong, Dingqian Hong, and Hui Xiong. Radar dimensions currently cover collaboration (50), focus (0), and reproducibility (78); output, impact, and activity are not yet scored because the underlying bibliometric data is insufficient. Because "Yang Song" is a common name, this profile is name-resolved from public bibliographic data.
Collaboration pattern
Yang Song's reviewed work is team-based: the two 2026 papers jointly involve 11 distinct coauthors, and both list a middle, non-corresponding author position. The recurring list includes Junwei Bao, Yuyao Zheng, Haipeng Sun, Lemao Liu, Hongfei Jiang, Dejing Dou, Kaichen Zhang, Yuzhong Hong, Dingqian Hong, and Hui Xiong; Hongfei Jiang is the only name appearing twice, suggesting overlap across the two projects. No first-author or sole-author paper is present in the catalog.
Research keywords
Highlights
GVPO++: variance-structured objectives for post-training and on-policy distillation
PGPO: potential-guided credit assignment for multi-turn LLM agents
Timeline
- 2026Publishes GVPO++ on arXiv, extending group variance policy optimization to LLM post-training and on-policy distillation.
- 2026Publishes PGPO on arXiv, introducing potential-guided step-level credit assignment for multi-turn agentic reinforcement learning.
- 2026Accumulates a co-author network of 11 distinct collaborators across the two papers, with Hongfei Jiang the only repeated name.
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.
- Name note: "Yang Song" is a common name; this profile is name-resolved from public bibliographic data, with no ORCID or institutional verification provided.
- Scope: all evidence comes from two 2026 arXiv preprints reviewed by Evolve; peer-review status and venue acceptance are not recorded.
- Radar scores (collaboration 50, focus 0, reproducibility 78) are computed from this small two-paper catalog; output, impact, and activity axes lack underlying data.
- Citation counts, h-index, and full publication history are not available and are therefore not quoted anywhere in this profile.