Ziheng Yi

—

Unclaimed · AI-structured profileOpenAlex

LLM-based multi-agent systems researcher: RL-learned agent topologies (MAGIC) and memory co-evolution (MACE)

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-15) 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 is not yet scored: this axis measures publication volume and sustained production rate, but the record currently contains only two 2026 arXiv preprints, too little data to compute a reliable value.

Impact—

Impact is not yet scored: this axis tracks citations and citation-based scholarly influence, and with h-index and i10-index both recorded at 0 there is no citation evidence available to compute a meaningful value.

Focus100/100

Focus scores a perfect 100: the field distribution contains exactly one field — machine learning (LLM-based multi-agent systems) — and both 2026 papers fall in it. Topic clustering reinforces this: multi-agent systems and negotiation (2 papers) and AI-based problem solving and planning (2) dominate, with secondary threads in reinforcement learning in robotics and evolutionary algorithms (1 each). This is an unusually tight, single-track specialization for such a short record.

Collaboration51/100

Collaboration scores 51, reflecting a moderate pattern. The record lists 12 distinct coauthors, with Kairui Yang, Xunkai Li, Rong-Hua Li, and Zekai Chen recurring across both papers, plus single appearances by Minghao An, Zhanke Liu, and Guangyuan He. On both 2026 papers Yi sits in a middle author position and is not the corresponding author, consistent with a contributing member of a larger research group rather than an independent lead.

Activity32/100

Activity scores 32, an early-stage signal. The by-year counts show 2 works, both in 2026, and the last-three-year window also contains just 2 works. No earlier publication years appear in the record, so the score reflects a short observable history rather than a decline; both outputs are arXiv preprints in the same field, suggesting an intensive but very recent burst of output.

Reproducibility55/100

Reproducibility scores 55, the mean across 2 scored papers, both rated medium readiness. Evolve's reviews flag a consistent pattern: MAGIC ranks first on 8 benchmarks but its single-run results need caution, while MACE's 81.11% versus SAGE's 78.97% rests on single-run evaluations with sub-0.1 percentage-point margins and uncited baselines. Multi-seed evaluation and clearer baselines would materially lift this score.

Research summary

Ziheng Yi is an early-stage researcher working at the center of LLM-based multi-agent systems, a specialization the bibliometric record confirms in a single-field distribution (machine learning, LLM-based multi-agent systems) and a perfect focus score of 100. Two 2026 arXiv preprints anchor this profile. MAGIC introduces mixed-granularity agent graphs built incrementally with dense-reward reinforcement learning, learning per-role atomic-or-group topologies that rank first on 8 benchmarks. MACE proposes memory-agent co-evolution with adaptive memory graphs, reporting 81.11% against SAGE's 78.97% across the same 8 benchmarks. Evolve's reviews flag a shared weakness: both papers rest on single-run evaluations, and MACE's sub-0.1 percentage-point margins plus uncited baselines leave its state-of-the-art claim unproven, giving a medium reproducibility readiness (mean 55/100 across 2 scored papers). Collaboratively, Yi works alongside 12 distinct coauthors, with Kairui Yang, Xunkai Li, Rong-Hua Li, and Zekai Chen recurring across records; both papers list Yi as a middle author rather than lead or corresponding, consistent with a contributing member of a larger group. Topic analysis places multi-agent systems and negotiation (2) and AI-based problem solving and planning (2) at the core, with secondary threads in reinforcement learning in robotics and evolutionary algorithms (1 each). With h-index and i10-index at 0 and only 2 works in the last three years, activity (32) and impact remain early-stage; this is a developing profile worth tracking.

Collaboration pattern

Ziheng Yi collaborates within a compact team: 12 distinct coauthors appear in the record, and four names recur across the 2026 papers — Kairui Yang, Xunkai Li, Rong-Hua Li, and Zekai Chen — with Minghao An, Zhanke Liu, and Guangyuan He appearing additionally. On both papers Yi holds a middle author position without corresponding-author status, suggesting a contributing member of a multi-person LLM multi-agent systems group rather than a solo or leading investigator.

Research keywords

LLM multi-agent systemsagent topology generationgraph construction policymixed granularityreinforcement learningreward shapingagent memorymemory graphsgraph-of-graphsexperience reusefeedback-driven adaptationmulti-agent negotiation

Highlights

MAGIC: RL-shaped mixed-granularity agent graphs rank first on 8 benchmarks
Per-role atomic-or-group topologies are learned by incremental construction with dense-reward reinforcement learning. Evolve judges the idea sound but cautions that single-run results warrant replication before accepting the ranking.
MACE: memory-agent co-evolution tops the same 8 benchmarks
Adaptive memory graphs co-evolve with agents to reuse experience and adapt through feedback, reporting 81.11% versus SAGE's 78.97%. Evolve notes sub-0.1 percentage-point margins and uncited baselines leave SOTA unproven.
A shared single-run evaluation pattern caps reproducibility at 55/100
Both papers are rated medium readiness, with a mean of 55 across 2 scored papers; Evolve flags single-run evaluations across both, plus MACE's sub-0.1pp margins and uncited baselines, as the main barrier to confirming state-of-the-art claims.

Timeline

2026First catalog paper appears: MAGIC on arXiv, learning per-role atomic-or-group agent topologies by incremental construction with dense-reward RL.
2026MACE posts to arXiv, extending the multi-agent line into memory: adaptive memory graphs co-evolve with agents, reporting 81.11% versus SAGE's 78.97%.
2026Evolve reviews both papers, rating each medium reproducibility readiness (mean 55/100) and cautioning that single-run evaluations underpin both benchmark claims.
2026Snapshot of the year: countsByYear records 2 works, both in 2026, alongside 12 distinct coauthors and an h-index of 0 — a concentrated first burst of output.

Papers in the Evolve catalog

Frequent co-authors

Kairui Yang×2 Xunkai Li×2 Minghao An×2 Zekai Chen×2 Rong-Hua Li×2 Zhanke Liu×1 Kairui Yang×1 Xunkai Li×1 Ziheng Yi×1 Zekai Chen×1 Guangyuan He×1 Rong-Hua Li×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 name-resolved from OpenAlex only; no ORCID or institutional affiliation is linked, so this profile should be read as resolved from public bibliographic data.
The record covers only two 2026 arXiv preprints; no peer-reviewed venue, citation record, or earlier career history is available, and h-index and i10-index are both 0.
Radar scores are computed from this small snapshot; the output and impact axes lack sufficient data and are not scored.
The frequent-coauthor list derives from catalog co-authorship data and includes a self-entry artifact; it may not reflect the researcher's full collaboration history.