Shihang Wang

Alibaba Group (China)

Unclaimed · AI-structured profileOpenAlex

Alibaba researcher on open-ended LLM agent RL — credit assignment, tournament ranking, rubric-grounded rewards

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—

No output data is available in this snapshot, so this axis is unscored. It would measure publication volume — total and recent paper counts, venue distribution, and authorship roles — none of which can be estimated from the current record.

Impact—

Impact is unscored because no citation or influence data exists in this snapshot. The axis would measure citations to the reviewed papers, downstream adoption of methods such as hierarchical credit propagation, and comparative influence within the agent-RL community.

Focus100/100

The focus dimension scores 100/100, the maximum, because the bibliometric field distribution contains exactly one field — machine learning (LLM agents and reinforcement learning) — and both Evolve-reviewed papers (ArenaFlow, ARISE-RL) sit inside it. Within that field the agenda is coherent: credit assignment, tournament ranking, and rubric-based reward design for open-ended agent RL. A single-field snapshot means the score measures concentration, not breadth; a longer publication history could change it.

Collaboration62/100

Collaboration scores 62/100, drawn from 22 distinct coauthors across the reviewed records. The pattern is a moderately wide industry-research network: recurring coauthors include Ruixue Ding, Boli Chen, Pengjun Xie, Zheng-Jun Zha, Xi Chen, Qiang Zhang, Fanrui Zhang, Yinfeng Huang, Yi Zheng, Kaipeng Zhang, and Jiawei Liu. On both catalog papers Wang is a middle-position, non-corresponding author, consistent with contributing inside larger Alibaba-led teams rather than leading small groups.

Activity—

Activity is unscored: this snapshot lacks the publication-timestamp data the axis needs. It would measure recent output tempo — papers per year, submission cadence, and continuity of work — which cannot be inferred from two preprints dated only to 2026.

Reproducibility55/100

Reproducibility scores 55/100, computed as reproMean 55 over just one paper with a recorded score (reproPapers: 1). Both reviewed papers carry a medium reproReadiness rating. Evolve's reviews flag concrete obstacles: ArenaFlow was criticized for judge circularity and missing variance tests, and ARISE-RL relies on self-built benchmarks rather than standard suites. Method descriptions appear detailed, but independent replication is not yet demonstrated.

Research summary

Shihang Wang is a researcher at Alibaba Group (China) working on machine learning for LLM agents and reinforcement learning. Evolve's catalog contains two 2026 arXiv papers by this author, both in a middle, non-corresponding author position. ArenaFlow extends tournament-based trajectory ranking for open-ended agent RL with hierarchical, step- and skill-level credit propagation; Evolve's review credits it with +19.5 over ArenaRL on Open-Travel while flagging judge circularity and missing variance tests. ARISE-RL claims state-of-the-art open-ended agent RL through rubric-grounded generator-solver co-evolution and reward-gated self-distillation, with strong reported gains on benchmarks that reviewers note were self-built. The bibliometric picture is unusually concentrated: the field distribution holds a single field, giving a focus score of 100, and records show 22 distinct coauthors (collaboration score 62) including Ruixue Ding, Boli Chen, Pengjun Xie, Zheng-Jun Zha, Xi Chen, Qiang Zhang, Fanrui Zhang, Yinfeng Huang, Yi Zheng, Kaipeng Zhang, and Jiawei Liu. Reproducibility scores 55, computed from one paper, with both works rated medium reproReadiness. Output, impact, and activity axes lack underlying data and remain unscored.

Collaboration pattern

Records show 22 distinct coauthors across the reviewed work (collaboration score 62), a moderately wide network typical of industrial ML teams. Recurring names include Ruixue Ding, Boli Chen, Pengjun Xie, Zheng-Jun Zha, Xi Chen, Qiang Zhang, Fanrui Zhang, Yinfeng Huang, Yi Zheng, Kaipeng Zhang, and Jiawei Liu. On both catalog papers Wang appears as a middle-position, non-corresponding author, suggesting contribution within larger Alibaba-led teams rather than leading small-group projects.

Research keywords

LLM agent RLagentic reinforcement learningcredit assignmenttournament-based rankingLLM-as-a-Judgeskill memoryprocess supervisionself-evolving agentsrubric-based rewardsself-distillationgenerator-solver co-evolutionopen-ended tasks

Highlights

Hierarchical credit propagation layered on tournament-based agent RL
Per Evolve's review, ArenaFlow extends tournament-based trajectory ranking with step- and skill-level credit propagation for open-ended agent RL, reporting +19.5 over ArenaRL on Open-Travel; reviewers flag judge circularity and the absence of variance tests.
Rubric-grounded generator-solver co-evolution for agent RL
ARISE-RL claims state-of-the-art open-ended agent RL through rubric-grounded generator-solver co-evolution and reward-gated self-distillation; Evolve judges the reported gains strong but notes that the benchmarks were self-built.
A consistent thread: making reward signals trustworthy in open-ended agent RL
Across both 2026 preprints, the shared move is replacing hand-designed scalar rewards with structured signals — tournament rankings, step/skill credit, rubrics — a direction Evolve reviews find promising while flagging judge circularity and self-built evaluation as open risks.

Timeline

2026ArenaFlow posted to arXiv, extending tournament-based agent RL with hierarchical step- and skill-level credit propagation (+19.5 over ArenaRL on Open-Travel per Evolve's review).
2026ARISE-RL posted to arXiv, presenting rubric-grounded generator-solver co-evolution with reward-gated self-distillation and claiming state-of-the-art open-ended agent RL.
2026Both catalog papers land in the same field — machine learning (LLM agents and reinforcement learning) — producing a concentrated single-field profile with a focus score of 100.

Papers in the Evolve catalog

Frequent co-authors

Ruixue Ding×2 Boli Chen×2 Pengjun Xie×2 Zheng-Jun Zha×2 Xi Chen×2 Qiang Zhang×1 Fanrui Zhang×1 Xi Chen×1 Yinfeng Huang×1 Yi Zheng×1 Kaipeng Zhang×1 Jiawei Liu×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.

Bibliometric dimension scores are computed from a public OpenAlex-based bibliometric snapshot only; no ORCID is attached and Shihang Wang is a common name, so identity resolution relies on public bibliographic data.
Only two Evolve-reviewed papers (both 2026 arXiv preprints) inform this profile; no peer-reviewed venue information is available.
Output, impact, and activity dimensions are unscored because citation counts, publication volume, and timestamped output data are absent from the snapshot.
The reproducibility score (55) is a mean over a single scored paper (reproPapers: 1) and should be read as an early, low-sample signal.