Omer Tafveez

—

Unclaimed · AI-structured profilename-matched identity

First author of a 2026 SAFE-benchmark study on whether frontier LLMs seek safety evidence before acting

papers here

reviewed by Evolve

first-author

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—

Unscored: the snapshot contains no publication-count or productivity data beyond one cataloged paper, so this axis cannot be computed. It would measure total scholarly output over time, including venues, arXiv preprints, and publication cadence.

Impact—

Unscored: no citation counts or influence signals are available in the current data, so this axis is left blank. It would measure citation uptake, downstream use of the SAFE benchmark, and the paper's reach within AI safety research.

Focus100/100

A perfect focus score of 100 reflects a field distribution in which the single cataloged paper falls entirely within AI safety / LLM behavioral evaluation — one field, one paper, zero dispersion. The profile is therefore maximally concentrated, though this purity comes from a one-paper evidence base rather than demonstrated sustained specialization, and each new publication will re-weight this score.

Collaboration0/100

The collaboration score of 0 comes from a raw count of zero distinct co-authors in the bibliometric snapshot, despite the researcher appearing as first author on the 2026 paper. No frequent co-authors are on file, so the collaborative footprint — team size, recurring partners, and network reach — is not yet visible in Evolve's data.

Activity—

Unscored: the snapshot lacks the multi-year publication timestamps needed to compute research activity, since only one paper year (2026) is on record. This axis would measure publication cadence and sustained research engagement across recent years.

Reproducibility55/100

A reproducibility score of 55 is the mean across one reviewed paper (reproPapers = 1), whose reproducibility readiness is rated medium. The underlying study is a behavioral benchmark with explicit coding rules, and Evolve notes that its strongest claims depend partly on its own payoff structure and coding choices — conditions that make independent replication non-trivial.

Research summary

Omer Tafveez appears in Evolve's catalog as an early-stage, name-resolved profile anchored by a single first-authored 2026 arXiv paper, "Do Frontier Models Seek Safety Evidence Before Acting?". The paper introduces the SAFE benchmark, which finds that four frontier models differ sharply in whether they voluntarily inspect safety evidence before acting, with reported rates spanning 48-96%. Its behavioral analysis attributes the variation to severity and cost far more than to probability, and Evolve's review credits the work as careful and unusually self-critical while cautioning that its strongest claims are partly artifacts of the study's own payoff structure and coding rules. The bibliometric snapshot is correspondingly narrow: the field distribution places 100% of cataloged output in AI safety / LLM behavioral evaluation, yielding a perfect focus score of 100, while the reproducibility score of 55 is the mean over one paper rated medium for reproducibility readiness. Output, impact, and activity lack sufficient underlying data and are left unscored, and zero distinct co-authors are recorded in the current snapshot despite the first-author position. As further publications accrue, this profile should be read as a directional snapshot rather than a settled record of the researcher's trajectory.

Collaboration pattern

The snapshot records zero distinct co-authors and no frequent collaborators, so Evolve cannot yet characterize this researcher's collaboration pattern. The one cataloged paper is first-authored (non-corresponding), consistent with a hands-on lead role, but the co-author network behind that arXiv paper is not represented in the current bibliometric data. Collaboration metrics should be revisited as more papers enter the catalog.

Research keywords

AI safety evaluationEvidence acquisitionInformation avoidanceFrontier language modelsBenchmark designRationale faithfulnessSAFE benchmarkVoluntary evidence seekingSeverity and cost sensitivityBehavioral evaluation

Highlights

Cross-model gap in voluntary safety-evidence inspection
The SAFE benchmark shows four frontier models differ sharply in whether they voluntarily inspect safety evidence before acting, with reported rates spanning 48-96%.
Severity and cost outweigh probability
Behavioral analysis finds that evidence-seeking is driven by severity and cost far more than by probability, a comparative result across the study's manipulated decision factors.
A deliberately self-critical study design
Evolve's review calls the study careful and unusually self-critical, while cautioning that its strongest claims are partly artifacts of its own payoff structure and coding rules.

Timeline

2026First-authors the arXiv paper "Do Frontier Models Seek Safety Evidence Before Acting?".
2026Introduces the SAFE benchmark, reporting 48-96% voluntary safety-evidence inspection across four frontier models.
2026Evolve reviews the paper with a medium reproducibility-readiness rating, feeding a reproducibility score of 55 and a focus score of 100.

Papers in the Evolve catalog

2026-09-17first author
Do Frontier Models Seek Safety Evidence Before Acting?

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 scores are computed from a public bibliometric snapshot only; no institution, country, or ORCID was available to verify identity.
The evidence base is a single reviewed paper (2026, arXiv), so the focus, reproducibility, and collaboration scores rest on n = 1.
Output, impact, and activity axes are unscored because the underlying bibliometric data is not yet available.
"Omer Tafveez" is name-resolved from public bibliographic data; this profile may not cover all authors sharing the name.