Ion Stoica

University of California, Berkeley

Unclaimed · AI-structured profileORCID OpenAlex

UC Berkeley researcher in AI-agent software engineering and agent security evaluation, from probes to position papers

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-15) 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 yet scored: this profile's underlying data lacks the publication-volume record needed to measure output. Once assembled, this axis would count papers per year and total production across venues to gauge sustained research productivity.

Impact—

Not yet scored: citation-based evidence is unavailable in the current record. This axis would measure citation counts, h-index-style influence, and uptake of the two catalog papers once a fuller bibliometric trail is linked.

Focus0/100

Scores 0 on focus because this catalog's records split evenly across two adjacent fields — software engineering (AI agents) at 0.5 and AI agent security evaluation at 0.5 — with no single dominant area. The two reviewed papers sit on either side of that divide: an Android vulnerability-discovery benchmark and a conceptual position paper on agentic software engineering. With only two catalog entries, this even split says as much about sample size as about research direction.

Collaboration68/100

Scores 68 on collaboration, drawn from 30 distinct coauthors in the available record, with 12 recurring partners — including Alexander Krentsel, Shubham Agarwal, Mert Cemri, Ziming Mao, and Matei Zaharia — appearing across recent work. Both catalog papers are multi-author efforts in which Stoica holds neither first authorship nor the corresponding role (middle author on one, last on the other), a pattern consistent with a senior contributor joining large, active team projects rather than leading small groups.

Activity—

Not yet scored: the data needed to track publishing cadence over recent years is missing. This axis would measure recent-year output rhythm and continuity, which two same-year 2026 entries cannot establish.

Reproducibility38/100

Scores 38 on reproducibility, a mean computed over exactly 2 papers with reproducibility assessments. The spread is visible in the catalog: MobileCybench, an executable-probe benchmark, rates medium readiness since its probes are runnable but rest on 13 apps without significance testing, while 'Reality Is the Final Verifier' is a purely conceptual position paper rated low, as it offers no empirical artifacts to re-run. More assessed papers would be needed to stabilize this score.

Research summary

Ion Stoica is a researcher at the University of California, Berkeley whose recent recorded work centers on two fast-moving corners of the AI-agent stack: agentic software engineering and AI agent security evaluation. Evolve's catalog holds two 2026 arXiv papers. MobileCybench introduces executable probes for scoring how well coding agents discover vulnerabilities on Android; Evolve's review calls it well-designed and honest while flagging a 13-app scope, absent significance tests, and refusal confounds. The second paper, 'Reality Is the Final Verifier', is a position paper — Stoica is last author — that reframes agentic software-engineering failures through requirement gaps and model gaps; it is judged sharp and well-grounded but purely conceptual. Bibliometric radar scores reflect this early catalog footprint: collaboration 68, drawn from 30 distinct coauthors with 12 recurring partners including Matei Zaharia and Alexander Krentsel; reproducibility 38 averaged over two assessed papers; and focus 0, an artifact of the record splitting evenly (0.5/0.5) between the two fields. Output, impact, and activity are not scored because the underlying data is insufficient. This is a full but narrow profile: it is name-resolved from public bibliographic data (ORCID 0000-0002-5373-0088), and the two preprints may represent only a slice of the researcher's actual output.

Collaboration pattern

The record shows 30 distinct coauthors and 12 frequent partners — Alexander Krentsel, Shubham Agarwal, Mert Cemri, Shu Liu, Sidharth Sankhe, Ziming Mao, Matei Zaharia, Andy K. Zhang, Ava Huang, Joey Ji, Wai Han, and Thomas Qin. In both catalog papers Stoica is neither first nor corresponding author (middle on MobileCybench, last on the position paper), a pattern consistent with a senior collaborator contributing to large, team-driven agent projects rather than leading small first-author efforts.

Research keywords

AI agent vulnerability discoveryAndroid securityexecutable security probesbenchmarkexploit validationcoding agentsagentic software engineeringreward hackinghallucinationrequirement gapmodel gapevaluation gap

Highlights

MobileCybench: executable probes for agent-found Android vulnerabilities
The 2026 benchmark scores how well AI agents discover real vulnerabilities on Android using runnable, executable probes. Evolve's review calls the design well-designed and honest, while noting the 13-app scope, missing significance tests, and refusal-behavior confounds; Stoica is a middle author and reproducibility readiness is rated medium.
'Reality Is the Final Verifier': two gaps in agentic software engineering
In this 2026 arXiv position paper Stoica is the last author. It reframes agentic software-engineering failures through the requirement gap and the model gap, arguing reality is the final verifier. Evolve judges it sharp and well-grounded but purely conceptual, with no empirical validation — hence a low reproducibility readiness rating.
A wide collaboration footprint: 30 distinct coauthors
The bibliometric record lists 30 distinct coauthors and 12 frequent partners — including Matei Zaharia, Alexander Krentsel, Shubham Agarwal, Mert Cemri, and Ziming Mao — supporting a collaboration score of 68. In both catalog papers Stoica takes neither the first nor corresponding slot, consistent with a senior collaborator on team-driven agent projects.

Timeline

2026MobileCybench appears on arXiv, introducing executable probes to evaluate AI agents' vulnerability discovery on Android; Stoica authors in the middle position.
2026'Reality Is the Final Verifier' appears on arXiv, a position paper reframing agentic software-engineering failures as requirement and model gaps; Stoica is the last author.
2026Both catalog papers land in the same year, splitting the record evenly between AI-agent software engineering and AI agent security evaluation.

Papers in the Evolve catalog

Frequent co-authors

Alexander Krentsel×1 Shubham Agarwal×1 Mert Cemri×1 Shu Liu×1 Sidharth Sankhe×1 Ziming Mao×1 Matei Zaharia×1 Andy K. Zhang×1 Ava Huang×1 Joey Ji×1 Wai Han×1 Thomas Qin×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.

Radar scores are computed from a bibliometric snapshot; output, impact, and activity lack sufficient underlying data and are intentionally not scored.
Focus (0) and reproducibility (38) rest on very small samples: a 0.5/0.5 field split across two papers, and only two reproducibility-assessed papers.
Both catalog papers are arXiv preprints; peer-reviewed venue status is not captured in the provided data.
This profile is name-resolved from public bibliographic data (ORCID 0000-0002-5373-0088); only the affiliation and these two papers are claimed, with no career, award, or full-publication facts.