Trustworthy infrastructure for LLM agents — verifiable agent environments and benchmark-integrity instrumentation.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
7
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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 output score yet: the platform lacks sufficient publication-volume data on this axis. It would normally measure annual paper counts, venue-normalized productivity, and total scholarly output; until the data is complete, the 2017-2026 yearly counts remain unrated.
Impact—
Impact is unscored: citation-side evidence in this snapshot is insufficient for the platform's impact axis. It would normally aggregate citation counts and h-index trajectory (the record lists h-index 7) into a measure of influence on downstream research.
Focus0/100
The radar scores focus at 0, and the underlying field distribution explains why: the record splits evenly, 50/50, between AI safety/systems security for LLM agents and machine learning/AI for science. Topic counts reinforce the spread — privacy-preserving technologies (8), adversarial robustness (7), ML and data classification (5), AI ethics (4), and AI planning (4). The low score reflects breadth across two tracks rather than an absence of direction.
Collaboration84/100
Collaboration scores 84/100. The OpenAlex record lists 65 distinct coauthors, twelve of whom recur frequently — including Shenghan Zheng, Yimin Liu, Kyoung Whan Choe, Jiankai Sun, and Yifeng He. Both cataloged papers are multi-author works with Di in a middle position and no corresponding-author role, consistent with a heavily networked, team-based working pattern across both of the record's fields.
Activity87/100
Activity scores 87/100, driven by a sharp recent ramp in the yearly counts: 1 to 5 works per year from 2017 through 2024, a dip to 1 in 2025, then 13 works recorded in 2026 — the record's highest annual total. The 19 works in the trailing three years confirm that current output sits at its peak.
Reproducibility38/100
Reproducibility scores 38/100, the mean over the 2 papers Evolve has reviewed. ScienceIDE is rated medium reproducibility readiness, while BenchShield is rated low — no code release is stated and its ground truth is self-adjudicated. Public code, released environments, and third-party-verifiable evaluation setups are the clearest levers for lifting this score.
Research summary
Zonglin Di (University of Zenica; ORCID 0000-0002-3875-2676) is an early-stage researcher whose OpenAlex-resolved record spans two evenly weighted areas — AI safety and systems security for LLM agents, and machine learning/AI for science centered on agent environments and LLM training infrastructure. The snapshot lists an h-index of 7 and i10-index of 7, 65 distinct coauthors, and 19 works in the trailing three years. Publication activity ramps sharply: yearly counts of 1-5 works from 2017 through 2024 give way to 13 works recorded in 2026, the record's highest annual count. Topic frequencies place privacy-preserving technologies (8) and adversarial robustness (7) at the top, followed by ML and data classification (5), AI ethics and social impacts (4), and AI-based problem solving and planning (4). Two 2026 arXiv papers reviewed by Evolve anchor the profile. ScienceIDE turns scientific codebases into agent-learnable environments, with evaluation, SFT, and reinforcement-learning gains Evolve calls credible though selectively reported; BenchShield adds TLA+ model checking and runtime instrumentation against reward hacking in agent benchmarks, far outperforming BenchJack but with self-adjudicated ground truth and no stated code release. Radar scores run high for activity (87) and collaboration (84), low for reproducibility (38); output and impact are unscored.
Collaboration pattern
This is a highly collaborative record: OpenAlex lists 65 distinct coauthors, twelve of whom recur — including Shenghan Zheng, Yimin Liu, Kyoung Whan Choe, Jiankai Sun, Heguang Lin, and Yifeng He. Both Evolve-reviewed 2026 papers place Di in a middle author position without corresponding-author status, indicating a contributor within larger teams rather than a solo line. The collaborator pool appears distributed across the record's two fields, agent safety and AI-for-science infrastructure.
Research keywords
Highlights
ScienceIDE turns scientific codebases into agent-learnable environments
BenchShield: formal, model-backed defense against reward hacking in agent benchmarks
Timeline
- 2017First work appears in the OpenAlex record (1 work), opening a low-volume early period that reaches only 2 works in 2018.
- 20194 works recorded — the highest annual count of the record's first several years.
- 20245 works recorded as output climbs again after quieter years (2021: 1; 2022: 2; 2023: 3).
- 202613 works recorded, the record's highest annual count; both Evolve-reviewed papers — ScienceIDE and BenchShield — appear on arXiv this year.
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 counts, topics, and coauthor lists come from OpenAlex; citation-side metrics are incomplete, leaving the output and impact axes unscored.
- Only 2 papers in this record carry Evolve reviews (both 2026 arXiv preprints, middle authorship); quality and reproducibility signals rest on that small sample.
- The 13 works counted for 2026 reflect a partial year of OpenAlex indexing and may shift as records are added.
- Affiliation and country metadata are as recorded in the input snapshot and were not independently verified; this profile is name-resolved from public bibliographic data.