Ansel Kaplan Erol
Georgia Institute of Technology · United States
Georgia Tech researcher focused on LLM agent safety benchmarks: compliance under pressure and agent-memory security
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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 output score yet: this axis would measure total indexed publications and per-year production volume, but the record shows an h-index of 0 and an i10-index of 0, and only two papers have been reviewed, leaving insufficient data to score it reliably.
Impact—
No impact score yet: this axis would measure citation-derived influence such as citation counts, h-index trajectory, and field-weighted impact, but h-index 0 and i10-index 0 mean no citation signal exists in the snapshot, so it cannot be scored or estimated.
Focus100/100
Scores 100 for focus. The field distribution assigns the reviewed work to a single field — natural language processing / AI safety (LLM agent evaluation) — and both catalog papers target the same problem class: benchmarking LLM agents under compliance pressure and memory-based authorization corruption. Topic counts reinforce this: explainable AI leads with 6 mentions, alongside ethics and social impacts of AI (3), artificial intelligence in law (3), and scientific computing and data management (3), consistent with a tightly held trustworthy-AI agenda.
Collaboration22/100
Scores 22 for collaboration, based on only 2 distinct coauthors in the record. Both reviewed papers place Erol in the last (senior) author position, working with Mika Okamoto and Tommaso Cerruti as the recurring partners. The small, stable two-person circle suggests a compact working group rather than a broad cross-institutional network; the low score reflects the limited number of unique collaborators visible in the current snapshot, not necessarily weak teamwork.
Activity75/100
Scores 75 for activity. The publication record is young but steep: countsByYear shows 1 work in 2025 jumping to 11 in 2026, with recent3yWorks totaling 12. That single-year acceleration from 1 to 11 is the main driver of the score. Both reviewed 2026 papers are arXiv preprints, indicating a fast, repository-first output rhythm typical of an early-career researcher shipping evaluation artifacts quickly.
Reproducibility100/100
Scores 100 for reproducibility, computed as a mean across 1 scored paper. Both Evolve-reviewed papers carry high reproducibility readiness: the agent-memory benchmark is credited with strong causal isolation and open artifacts, while PACT's benchmark construction is described as rigorous, though its items and labels rest entirely on LLM generation and judging. Open, re-runnable evaluation artifacts appear to be a deliberate design goal across the catalog.
Research summary
Ansel Kaplan Erol is an early-stage researcher at Georgia Institute of Technology working on the safety and evaluation of LLM-based agents. The bibliometric snapshot (OpenAlex identity) records 12 works in the trailing three years — 1 in 2025 rising to 11 in 2026 — with an h-index and i10-index both at 0, consistent with a first- or second-year publishing career. Two Evolve-reviewed papers, both 2026 arXiv preprints with Erol in the last author position, define a focused agenda: PACT benchmarks whether enterprise AI assistants hold their compliance rules under user pressure, and a second paper treats agent persistent memory as a surface for "endogenous authorization laundering." Both carry high reproducibility readiness, and the reproducibility dimension scores 100 (mean across 1 scored paper), while focus also scores 100, with work assigned to a single field: natural language processing / AI safety (LLM agent evaluation). Topic counts widen the lens slightly: explainable AI (6), ethics and social impacts of AI (3), scientific computing and data management (3), artificial intelligence in law (3), and IoT and edge/fog computing (2). Collaboration is the narrow axis — score 22, with 2 distinct coauthors, Mika Okamoto and Tommaso Cerruti. Output and impact remain unscored pending more citation data.
Collaboration pattern
The record shows a tightly scoped collaboration pattern: 2 distinct coauthors, with Mika Okamoto and Tommaso Cerruti as the recurring partners. On both Evolve-reviewed papers Erol sits in the last (senior) author position without corresponding-author designation, consistent with a senior-contributor role in a small team. The collaboration score of 22 reflects this small visible network; it is based on a limited snapshot and may change as more papers are indexed.
Research keywords
Highlights
PACT: benchmarking enterprise AI assistants under user pressure
Agent memory as a surface for authorization laundering
Timeline
- 2025First work appears in the indexed record: countsByYear shows 1 publication in 2025, marking the start of a visible research footprint.
- 2026Output accelerates sharply to 11 indexed works in a single year, bringing the three-year total to 12 and driving the activity score to 75.
- 2026PACT preprint released on arXiv with Erol as last author, benchmarking enterprise-rule compliance of LLM assistants under user pressure.
- 2026Second arXiv preprint frames agent persistent memory as a surface for endogenous authorization laundering, rated high for reproducibility readiness.
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.
- Bibliometrics and identity resolution come from OpenAlex only; no ORCID is linked, so this profile is name-resolved from public bibliographic data.
- Only 2 papers have been Evolve-reviewed; the reproducibility score of 100 is a mean over 1 scored paper.
- Output and impact dimensions are unscored: h-index and i10-index are both 0 and no citation data is available yet.
- All quantified facts come from the provided snapshot; no external career, award, or biographical information was used.