Jingxiao Yang

Zhejiang University · China

Unclaimed · AI-structured profileOpenAlex

Early-career Zhejiang University researcher unifying distillation and GRPO credit assignment for LLM post-training

papers here

reviewed by Evolve

first-author

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—

Output cannot be scored yet: the OpenAlex snapshot shows an h-index of 0 and no aggregate publication volume beyond the 2 catalog works, so there is no reliable record of total papers, venues, and yearly production.

Impact—

Impact cannot be scored yet: citation counts, i10-index, and h-index are absent or zero (h-index 0) in the OpenAlex snapshot, so the axis that would measure citation influence and uptake of the 2026 papers has no usable signal.

Focus100/100

Focus scores a perfect 100: the OpenAlex topic profile clusters on reinforcement learning in robotics (2 mentions) alongside single tags for topic modeling, AI-based problem solving and planning, machine learning and algorithms, and AI applications, and the field distribution lists one field, machine learning (LLM post-training, reinforcement learning). Both 2026 catalog papers attack the same problem — credit assignment in GRPO-style training — giving an unusually narrow, coherent agenda.

Collaboration46/100

Collaboration scores 46, a modest mid-range value. The OpenAlex window records 9 distinct coauthors, matching the frequent-collaborator list (Wangjie Gan, Yingxuan Zhuang, Wenqi Zhang, Jintao Chen, and others). Authorship spans first author on TASPO and middle author on UECR-GRPO, indicating participation within a small, stable group rather than broad independent networking.

Activity32/100

Activity scores 32, reflecting low recent volume: countsByYear shows only 2 works in 2026 and recent3yWorks is 2, both arXiv preprints in the Evolve catalog. The record is current but thin, an early-stage cadence rather than high-throughput output, so the score should be read alongside the missing output axis.

Reproducibility55/100

Reproducibility scores 55, computed from 1 assessed paper (reproMean 55). Both catalog papers carry medium reproducibility readiness: TASPO reports +10.6% average over 9 agentic RL settings with strong ablations but draft artifacts lower trust, while UECR-GRPO posts small gains backed by rigorous audits. Evidence of honest self-assessment exists, though artifact quality still limits independent replication.

Research summary

Jingxiao Yang is a researcher at Zhejiang University working on machine learning, with a sharply concentrated focus on reinforcement learning for large language model post-training and agents. The Evolve catalog contains two of the researcher's 2026 arXiv papers, both attacking credit assignment in GRPO-style training. The first-authored TASPO paper reweights GRPO credit using privileged action-level supervision, reporting a +10.6% average improvement across 9 agentic RL settings with strong ablations, although draft artifacts temper trust in the results. The second, UECR-GRPO, unifies on-policy distillation and GRPO through pre-normalized teacher utility and entropy-gated, budget-preserving credit redistribution, delivering small but rigorously audited gains on mathematical reasoning. Bibliometric data from OpenAlex reflects an early-stage record: an h-index of 0, 2 works in the recent three-year window, 9 distinct coauthors, and topic tags dominated by reinforcement learning in robotics. Radar scores show perfect focus (100), moderate reproducibility (55, from one assessed paper), modest collaboration (46), and low recent activity (32), while output and impact cannot yet be scored for lack of citation data. The overall picture is a promising early-career contributor whose published work is thematically coherent and methodologically self-critical.

Collaboration pattern

Jingxiao Yang works within a compact, stable team: the OpenAlex window lists 9 distinct coauthors, and the frequent-collaborator roster includes Wangjie Gan, Yingxuan Zhuang, Wenqi Zhang, Jintao Chen, Xuhong Zhang, Jie Zhang, Zhehao Huang, Yuhang Liu, and Xiaolin Huang. Authorship alternates between first (TASPO) and middle (UECR-GRPO) positions, consistent with contributing within a single Zhejiang University-based research group rather than an extended external network; the collaboration radar score is 46.

Research keywords

GRPOcredit assignmenton-policy distillationreinforcement learning with verifiable rewardsLLM agentsknowledge distillationmathematical reasoningon-policy self-distillationprivileged informationprocess supervisionagentic policy optimizationentropy calibration

Highlights

First-authored TASPO: privileged action-level supervision for GRPO credit in agentic policy optimization
Reconciles process supervision with outcome-based credit via on-policy self-distillation, reporting a +10.6% average improvement across 9 agent-RL settings with strong ablations, though draft artifacts lower trust (reproducibility readiness: medium).
UECR-GRPO unifies on-policy distillation and GRPO via entropy-calibrated credit assignment
Pre-normalized teacher utility decides when and where to trust the teacher, with entropy-gated, budget-preserving credit redistribution yielding small but rigorously audited gains on mathematical reasoning (reproducibility readiness: medium).
A coherent early agenda centered on credit assignment in GRPO-family training
Both 2026 papers and the OpenAlex topic profile converge on one field — machine learning (LLM post-training, reinforcement learning) — with credit assignment as the recurring thread, reflected in a perfect focus score of 100.

Timeline

2026First-authored TASPO paper posted to arXiv: privileged action-level supervision reconciled with outcome-based credit in agentic policy optimization.
2026Co-authored UECR-GRPO paper posted to arXiv, unifying on-policy distillation and GRPO through entropy-calibrated credit assignment.
2026OpenAlex snapshot records 2 works in 2026 with 9 distinct coauthors; an early-stage record with h-index 0.

Papers in the Evolve catalog

Frequent co-authors

Wangjie Gan×1 Yingxuan Zhuang×1 Wenqi Zhang×1 Jintao Chen×1 Xuhong Zhang×1 Jie Zhang×1 Zhehao Huang×1 Yuhang Liu×1 Xiaolin Huang×1

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 drawn from OpenAlex only; identity is name-resolved from public bibliographic data, and no ORCID was provided.
Output and impact axes are unscored because citation counts, h-index, and i10-index are absent or zero in the snapshot.
Both catalog papers are 2026 arXiv preprints and may not yet reflect peer review.
The reproducibility score derives from a single assessed paper (reproMean 55, reproPapers 1), a small sample.