Xuan Jin

Alibaba Group (China)

Unclaimed · AI-structured profileOpenAlex

Alibaba Group researcher on evaluation and efficiency 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 snapshot carries no publication counts or productivity metrics, so output — the volume and consistency of papers over time — cannot yet be measured; it becomes computable once a fuller bibliographic record is linked.

Impact—

Not scored: no citation counts, h-index, or downstream usage data exist in the provided material. Impact would measure how strongly the community cites, reuses, and builds on this researcher's evaluation and compression work for self-evolving agents.

Focus100/100

Focus scores 100/100, the maximum. The snapshot's field distribution lists a single field — artificial intelligence (LLM agent evaluation) — and both 2026 papers sit inside one coherent space, self-evolving LLM agents: one on process-level evaluation and benchmarks, the other on skill compression, progressive disclosure, and token cost optimization. Continual learning and catastrophic forgetting themes recur across both, so the profile reads as a specialist rather than a generalist.

Collaboration46/100

The snapshot records 9 distinct coauthors, producing a moderate collaboration score of 46/100. Recurring partners include Xiaofan Bai, Chao Liu, Hongqiang Lin, Xipeng Cao, Yuhong Li, Di Wu, Mingli Song, and Nenggan Zheng. On both reviewed 2026 papers Xuan Jin is a middle author without corresponding status, consistent with contributing inside a stable Alibaba team rather than leading as sole or corresponding author; the score leaves room for a broader network.

Activity—

Not scored: the data holds no year-by-year publication record, so activity — publishing cadence, recency, and continuity across a career — cannot be assessed. Only two 2026 preprints are on file; judging rhythm requires more longitudinal history.

Reproducibility100/100

Reproducibility scores 100/100 (reproMean 100), but the score rests on a single scored paper (reproPapers: 1): the EvoPathBench evaluation paper, whose checkpoint-based benchmark design and open reporting of negative results earned a 'high' reproducibility rating. The companion paper, SkillZip Pro, is rated 'medium' because its evidence rests on small, private evaluations. The perfect score is therefore one strong data point, not yet a verified pattern across the output.

Research summary

Xuan Jin is a researcher at Alibaba Group working on large language model (LLM) agents; this profile is name-resolved from public bibliographic data. The evidence base is an early-stage snapshot of two 2026 arXiv preprints, both on self-evolving agents — systems that learn and adapt continually. The first, EvoPathBench (Beyond Endpoint Performance), shifts evaluation from final task scores to process-level checks at checkpoints during an agent's evolution; it is reviewed as rigorous and honest about negative results, though single-domain scope and missing error bars limit generality. The second, SkillZip Pro, addresses efficiency: execution-aware dynamic compression of progressively loaded agent skills, claiming evaluation-free compression that cuts bundles by 38% without quality loss, though its evidence rests on small, private evaluations. Together they sketch one coherent agenda: making self-evolving agents both measurable and affordable. The snapshot records 9 distinct coauthors, a perfect 100/100 focus score, and a 100/100 reproducibility score based on a single scored paper. Xuan Jin is a middle author on both papers, consistent with a strong contributor role inside a team. Output, impact, and activity metrics are not yet available.

Collaboration pattern

The snapshot lists 9 distinct coauthors and a collaboration score of 46/100. Recurring partners include Xiaofan Bai, Chao Liu, Hongqiang Lin, Xipeng Cao, Yuhong Li, Di Wu, Mingli Song, and Nenggan Zheng (Chao Liu appears twice in the source list). On both 2026 papers Xuan Jin is a middle author without corresponding status, consistent with a dependable contributor inside a stable Alibaba group rather than a solo or correspondence-leading author.

Research keywords

self-evolving agentsLLM agentsprocess-level evaluationbenchmarkingcontinual learningcatastrophic forgettingagent skillscontext compressionprogressive disclosuretoken cost optimization

Highlights

EvoPathBench: process-level evaluation for self-evolving agents
Moves beyond end-task scores by evaluating self-evolving agents at checkpoints during evolution; reviewed as rigorous and honest about negative results, with single-domain scope and missing error bars limiting generality.
SkillZip Pro: 38% skill-bundle compression without quality loss
Proposes execution-aware dynamic compression of progressively loaded agent skills, claiming evaluation-free compression that cuts bundles by 38% with no quality loss; strong engineering, though evidence rests on small, private evaluations.
A paired agenda: measure and economize self-evolving agents
Read together, the two 2026 preprints cover complementary halves of one problem — checkpoint-level measurement of agent evolution and cheap, execution-aware skill loading — giving the profile a coherent specialist direction with a 100/100 focus score.

Timeline

2026Co-authored the EvoPathBench paper 'Beyond Endpoint Performance', introducing checkpoint-based process-level evaluation of self-evolving agents (arXiv, middle author).
2026Co-authored 'SkillZip Pro', proposing execution-aware dynamic compression of progressively loaded skills with a reported 38% bundle reduction (arXiv, middle author).
2026Both preprints land in the self-evolving LLM agent space within the same year, pairing measurement (evaluation) with efficiency (compression) in a single research agenda.

Papers in the Evolve catalog

Frequent co-authors

Xiaofan Bai×2 Chao Liu×2 Hongqiang Lin×2 Xipeng Cao×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.

Common-name risk: this profile is name-resolved from public bibliographic data and may conflate people sharing the name Xuan Jin.
Bibliometrics (coauthor counts, field distribution, radar scores) derive only from the provided OpenAlex-based snapshot and were not independently verified.
Output, impact, and activity axes are unscored because the underlying bibliometric data is insufficient.
Coverage is limited to two reviewed 2026 arXiv preprints; earlier or unpublished work is not reflected.