Xipeng Cao

Alibaba Group (China)

Unclaimed · AI-structured profileORCID OpenAlex

Alibaba Group researcher building evaluation methods and efficiency techniques for self-evolving LLM agents.

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—

Not scored: the underlying data is missing — no total publication counts, venue types, or per-year volume are available. This axis would measure how much the researcher publishes and where; only two reviewed arXiv preprints are visible here.

Impact—

Not scored: citation counts, h-index, and downstream influence are not available in the data behind this profile. This axis would measure how widely the two reviewed papers are cited and taken up; no such metrics exist here yet.

Focus100/100

Focus scores 100: the field distribution behind this score is a single field — artificial intelligence (LLM agent evaluation) — with both reviewed 2026 papers landing in the same self-evolving-agent cluster: one on process-level benchmarking (EvoPathBench), the other on execution-aware skill compression (SkillZip Pro). With only two reviewed papers, this perfect concentration reflects a deliberately narrow, coherent agenda rather than a long-track record, and the value may shift as more papers enter the catalog.

Collaboration46/100

Collaboration scores 46: the reviewed papers record 9 distinct coauthors, with a recurring core spanning Xiaofan Bai, Chao Liu (listed twice among frequent collaborators), Hongqiang Lin, Xuan Jin, Yuhong Li, Di Wu, Mingli Song, and Nenggan Zheng. Cao is not corresponding author on either reviewed paper — last author on the benchmark paper and a middle author on the compression paper — a moderate, team-embedded collaboration footprint at Alibaba Group.

Activity—

Not scored: no multi-year publication cadence or recent-output timestamps are available, so ongoing activity cannot be measured. This axis would track publishing pace and recency over time; the record here shows only two 2026 arXiv preprints.

Reproducibility100/100

Reproducibility scores 100, computed as a mean of 100 across the single paper scored so far (reproPapers: 1) — the EvoPathBench benchmark paper, which carries high reproducibility readiness. Evolve's review highlights its rigor and its honesty about negative results, both favorable signals for reproducibility. The perfect score therefore rests on a very small base of one paper, while the second paper (SkillZip Pro) is rated medium because its evidence rests on small, private evaluations.

Research summary

Xipeng Cao is a researcher at Alibaba Group (China) whose reviewed work centers on self-evolving large language model (LLM) agents, spanning both how such agents should be evaluated and how they can be made cheaper to run. The Evolve catalog includes two 2026 arXiv papers. The first, Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents, introduces EvoPathBench, which evaluates self-evolving agents at intermediate checkpoints rather than only at final endpoints; Evolve's review calls it rigorous and honest about negative results, while noting single-domain scope and absent error bars. The second, SkillZip Pro, proposes execution-aware dynamic compression of progressively loaded agent skills, claiming evaluation-free skill-bundle compression that cuts bundles 38% without quality loss; the review praises the engineering but observes that evidence rests on small, private evaluations. On the measured dimensions, focus scores 100 — both catalog papers sit in the artificial intelligence (LLM agent evaluation) field cluster — and reproducibility scores 100, a mean taken across one reproducibility-scored paper, with the benchmark paper rated high for reproducibility readiness. Collaboration scores 46, with 9 distinct coauthors recorded, including Xiaofan Bai, Chao Liu, Hongqiang Lin, Xuan Jin, Yuhong Li, Di Wu, Mingli Song, and Nenggan Zheng. Output, impact, and activity dimensions are not scored because the underlying bibliometric data is insufficient. This profile is name-resolved from public bibliographic data (ORCID 0000-0001-8740-6369).

Collaboration pattern

Cao's reviewed work is team-embedded: 9 distinct coauthors appear across two 2026 papers, with a recurring circle including Xiaofan Bai, Chao Liu (listed twice among frequent collaborators), Hongqiang Lin, Xuan Jin, Yuhong Li, Di Wu, Mingli Song, and Nenggan Zheng. He is not corresponding author on either paper — last author on the benchmark, middle author on the compression work — consistent with a senior-but-supporting role within an Alibaba Group agent-systems team. The collaboration dimension scores 46.

Research keywords

self-evolving agentsLLM agentsbenchmarkprocess evaluationcontinual learningcatastrophic forgettingagent skillscontext compressionprogressive disclosuretoken cost optimization

Highlights

EvoPathBench: evaluating self-evolving agents at checkpoints, not just endpoints
As last author, Cao contributed to a process-level benchmark that evaluates self-evolving LLM agents at intermediate checkpoints across continual learning. Evolve's review finds it rigorous and honest about negative results, while flagging single-domain scope and missing error bars; reproducibility readiness is rated high.
SkillZip Pro: execution-aware compression of progressively loaded agent skills
The paper claims evaluation-free dynamic compression of progressively loaded skill bundles, cutting bundle size 38% without quality loss. Evolve credits strong engineering but notes the evidence rests on small, private evaluations; reproducibility readiness is medium. Cao is a middle-position author.
A coherent two-pronged agenda: measuring and economizing self-evolving agents
Together, the two 2026 papers pair a process-level evaluation benchmark (rated high for reproducibility) with a token-cost compression method for the same class of agents, indicating a focused push on both rigor of measurement and runtime efficiency within one research line.

Timeline

2026Co-authors SkillZip Pro, proposing execution-aware dynamic compression of progressively loaded skills for self-evolving agents (arXiv).
2026Appears as last author on Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents, introducing the EvoPathBench benchmark (arXiv).
2026Both reviewed outputs land in the artificial intelligence (LLM agent) field, with one paper rated high for reproducibility readiness and the profile's focus dimension scoring 100.

Papers in the Evolve catalog

Frequent co-authors

Xiaofan Bai×2 Chao Liu×2 Hongqiang Lin×2 Xuan Jin×2 Yuhong Li×2 Di Wu×1 Mingli Song×1 Nenggan Zheng×1 Chao 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 signals derive from OpenAlex-indexed public records; only the two Evolve-reviewed papers (both 2026 arXiv preprints) inform this profile.
Output, impact, and activity dimensions are unscored because citation counts, publication totals, and multi-year cadence data were unavailable.
The reproducibility score (100) is a mean over a single scored paper (reproPapers: 1), so it has a very small base.
Name resolved from public bibliographic data via ORCID 0000-0001-8740-6369; affiliation recorded as Alibaba Group (China), country field not provided.