Alibaba Group researcher on self-evolving LLM agents: process-level evaluation benchmarks and skill compression.
2
papers here
1
reviewed by Evolve
0
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—
No publication-volume data is available yet, so this axis is unscored. It would measure total papers, venue spread, and first/corresponding-author share; only two 2026 arXiv papers are catalogued, too few to compute.
Impact—
Citation-based influence cannot be assessed: the snapshot provides no citation counts or h-index. This axis would capture how widely the two 2026 agent papers are cited and adopted, so it is left unscored rather than estimated.
Focus100/100
The focus score is 100, the maximum, reflecting a field distribution concentrated in a single labeled area, artificial intelligence (LLM agent evaluation). Both catalogued 2026 papers sit on self-evolving agents — one on process-level evaluation (EvoPathBench), the other on execution-aware skill-bundle compression — confirming tight, deliberate specialization rather than scattered topics.
Collaboration46/100
The collaboration score is 46, a mid-range value, with 9 distinct coauthors recorded. Authorship roles vary: a middle position on the process-level evaluation paper and the last position on SkillZip Pro (not corresponding on either). Frequent co-authors include Xiaofan Bai, Chao Liu, Hongqiang Lin, Xuan Jin, Xipeng Cao, Di Wu, Mingli Song, and Nenggan Zheng, suggesting a stable team.
Activity—
Publication cadence cannot be measured here: the snapshot holds only two catalogued papers, both 2026, with no yearly output series. This axis would track output frequency and continuity across years, so it stays unscored.
Reproducibility100/100
The reproducibility score is 100 (reproMean), computed from a single assessed paper. That paper, EvoPathBench, carries a 'high' reproducibility rating and was praised as rigorous and honest about negative results. The second paper, SkillZip Pro, is rated only 'medium', its evidence described as small, private evaluations — so the perfect score rests on one data point.
Research summary
Yuhong Li is a researcher at Alibaba Group (China) working on self-evolving large language model (LLM) agents, with a catalogued record that concentrates on two complementary problems: how to evaluate such agents rigorously and how to run them efficiently. The first catalogued paper, 'Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents' (arXiv, 2026), introduces EvoPathBench, which scores agents at intermediate checkpoints rather than only on final answers; Evolve's review calls it rigorous and honest about negative results while noting single-domain scope and missing error bars. The second, 'SkillZip Pro' (arXiv, 2026), proposes execution-aware dynamic compression of progressively loaded skill bundles, reporting a 38% reduction without quality loss; the review praises the engineering but finds the evidence rests on small, private evaluations. Bibliometric signals echo this specialization: the focus score is 100 with the field distribution concentrated in AI (LLM agent evaluation), and reproducibility scores 100 on the single paper assessed so far. Collaboration sits mid-range at 46 across 9 distinct coauthors. Output, impact, and activity lack sufficient data and remain unscored, making this an early-stage but coherent profile of an agent-systems builder.
Collaboration pattern
Yuhong Li publishes in multi-author teams: both 2026 papers are collaborative, with a middle position on the evaluation paper and the last position on SkillZip Pro, neither as corresponding author. Nine distinct coauthors are recorded; recurring names include Xiaofan Bai, Chao Liu, Hongqiang Lin, Xuan Jin, Xipeng Cao, Di Wu, Mingli Song, and Nenggan Zheng — a stable pool consistent with an industrial lab at Alibaba Group. The collaboration score is 46, mid-range.
Research keywords
Highlights
EvoPathBench: process-level evaluation of self-evolving agents
SkillZip Pro: execution-aware compression of progressively loaded agent skills
Timeline
- 2026SkillZip Pro is posted to arXiv (ID 2608-30785): execution-aware dynamic compression of progressively loaded skill bundles for self-evolving agents, authored in the last position.
- 2026The EvoPathBench paper is posted to arXiv (ID 2609-24663): process-level evaluation of self-evolving agents at intermediate checkpoints, authored in a middle position.
- 2026Both papers enter the Evolve catalog under AI (LLM agent evaluation / agent systems), consolidating a focus on self-evolving agents, with reproducibility rated 'high' on the evaluation paper.
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 public bibliometric snapshot (OpenAlex-style data); output, impact, and activity dimensions lack underlying data and are unscored.
- The reproducibility score of 100 is computed from a single assessed paper (reproPapers = 1); the second catalogued paper is rated 'medium'.
- 'Yuhong Li' is a common name and no ORCID is provided; this profile is name-resolved from public bibliographic data and may merge or split identities.
- The evidence base is two 2026 arXiv preprints reviewed by Evolve; no peer-reviewed venue information is available yet.