Alibaba researcher on self-evolving LLM agents: process-level benchmarking and execution-aware skill compression
2
papers here
1
reviewed by Evolve
1
first-author
0
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 is not scored: the snapshot provides no verified publication counts, venue history, or yearly production totals for this author. This axis would normally measure the volume, consistency, and venues of research output over time.
Impact—
Impact is not scored: citation counts, h-index, and downstream adoption are absent from the input data. This axis would measure how far the two agent papers shape follow-up work on self-evolving agent evaluation and skill compression.
Focus100/100
Focus scores 100, the maximum in this snapshot. The recorded field distribution maps to a single area — artificial intelligence (LLM agent evaluation) — and both reviewed 2026 papers target the same problem, self-evolving LLM agents, from complementary angles: one builds process-level evaluation at training checkpoints, the other compresses progressively loaded agent skills. Within the visible record, the research agenda is tightly concentrated rather than dispersed.
Collaboration46/100
Collaboration scores 46, computed from 9 distinct coauthors in the snapshot. The recurring partner list includes Chao Liu, Hongqiang Lin, Xuan Jin, Xipeng Cao, Yuhong Li, Di Wu, Mingli Song, and Nenggan Zheng, with Chao Liu appearing twice in the frequency list, suggesting the closest observed tie. Authorship positions vary — first author on SkillZip Pro and a middle position on the EvoPathBench paper — indicating a mix of leading and contributing roles in team-based work.
Activity—
Activity is not scored: only two 2026 arXiv preprints are visible, with no multi-year publication timeline. This axis would measure publication cadence and sustained research pace across recent years, which the current data cannot support.
Reproducibility100/100
Reproducibility scores 100, with a mean of 100 computed over 1 assessed paper. The EvoPathBench evaluation study carries high reproducibility readiness: it evaluates self-evolving agents at explicit checkpoints and is candid about negative results. The SkillZip Pro paper is rated medium — its evaluation-free compression claim of a 38% bundle reduction rests on small, private evaluations, which tempers how far the score should be generalized.
Research summary
Xiaofan Bai is a researcher at Alibaba Group (China) working on self-evolving LLM agents, with a visible catalog that spans both evaluation methodology and systems engineering for that problem area. The Evolve-reviewed record shows two 2026 arXiv preprints. The first, 'Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents' (2026, middle author position), introduces EvoPathBench, which evaluates self-evolving agents at training checkpoints rather than only at final endpoints; Evolve judges it rigorous and candid about negative results, though single-domain scope and missing error bars limit generality. The second, 'SkillZip Pro' (2026, first author), proposes execution-aware dynamic compression of progressively loaded skill bundles, reporting a 38% bundle reduction without quality loss; Evolve credits strong engineering but notes the evidence rests on small, private evaluations. Bibliometric signals from the available snapshot are narrow but internally consistent: the recorded field distribution maps to a single area, artificial intelligence (LLM agent evaluation), the focus dimension scores 100, and reproducibility scores 100 based on one assessed paper. Collaboration data covers 9 distinct coauthors with a collaboration score of 46, including recurring co-authorship with Chao Liu, Hongqiang Lin, Xuan Jin, Xipeng Cao, Yuhong Li, Di Wu, Mingli Song, and Nenggan Zheng. Output, impact, and activity dimensions lack sufficient underlying data and are not scored. This profile is name-resolved from public bibliographic data; no ORCID is linked.
Collaboration pattern
The snapshot records 9 distinct coauthors, with a collaboration score of 46. The recurring list includes Chao Liu (listed twice, the closest observed tie), Hongqiang Lin, Xuan Jin, Xipeng Cao, Yuhong Li, Di Wu, Mingli Song, and Nenggan Zheng. Bai takes first authorship on SkillZip Pro and a middle position on the EvoPathBench paper, suggesting a mix of leading and supporting roles. All visible work carries the Alibaba Group affiliation.
Research keywords
Highlights
EvoPathBench: process-level evaluation of self-evolving agents
SkillZip Pro: evaluation-free compression of agent skill bundles
One agenda, two angles: measure and compress self-evolving agents
Across both reviewed 2026 papers, Bai works the same problem from complementary sides — process-level checkpoints to evaluate how agents evolve, and execution-aware compression to keep progressively loaded skills cheap to run.
Timeline
- 2026First visible preprint: 'SkillZip Pro' (arXiv, first author) — execution-aware dynamic compression of progressively loaded skills, reporting a 38% bundle reduction without quality loss.
- 2026Second visible preprint: 'Beyond Endpoint Performance' (arXiv, middle author) — introduces EvoPathBench, checkpoint-based process-level evaluation for self-evolving agents.
- 2026Both papers enter the Evolve catalog under the artificial intelligence (LLM agent evaluation / systems) field, with reproducibility readiness rated high and medium respectively.
Papers in the Evolve catalog
Frequent co-authors
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 come from a narrow snapshot: only two Evolve-reviewed 2026 arXiv preprints and one mapped field; citation counts, h-index, and venue history are not available.
- Output, impact, and activity dimensions are unscored because the underlying bibliometric data is insufficient (dimMissing).
- The reproducibility score of 100 is computed from a single assessed paper (reproPapers: 1) and should not be generalized.
- No ORCID or country is linked; the profile is name-resolved from public bibliographic data, and 'Xiaofan Bai' may match multiple authors.