UT Dallas researcher on LLM safety and privacy: reward-channel honesty tests and membership-inference evidence
2
papers here
2
reviewed by Evolve
2
first-author
1
corresponding
—
works · OpenAlex
—
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 metric is available yet — the radar marks this axis missing. It would measure publication volume, venue quality, and first-author counts over time; only two cataloged arXiv papers exist so far, so no score is estimated.
Impact—
Impact data is insufficient: no citation counts, h-index, or usage signals are recorded in this snapshot. This axis would measure citations and downstream influence of the two 2026 arXiv preprints once citation coverage exists.
Focus0/100
The focus score is 0, grounded in a raw field distribution split exactly 50/50 between natural language processing (LLM privacy / membership inference) and artificial intelligence (LLM evaluation / AI safety). With only two papers, this reads less as scattered work than as two halves of one agenda: stress-testing what language models will and will not reliably report, whether about a reward channel or about their own pretraining data.
Collaboration0/100
The collaboration score is 0, and the underlying raw count of distinct co-authors is 0: both cataloged papers list Arman Nik Khah as first author (one also corresponding), and the frequent-co-author list is empty. Within this snapshot the cataloged work is entirely solo, so collaboration habits beyond these two 2026 arXiv preprints cannot yet be characterized.
Activity—
Activity is not yet measurable: the radar marks this axis missing because no per-year publication cadence or recency indicator is available. It would measure publishing rhythm and recent output; the only dated evidence in the snapshot is two arXiv preprints, both from 2026.
Reproducibility100/100
Reproducibility is the standout axis: the radar scores it 100, with a reproducibility mean of 100 computed across 2 papers, both rated high reproducibility readiness. Evolve's verdicts call one study "a rigorous, highly transparent empirical study" and the other "rigorous," consistent with a working style built on verified records, counted duplications, and explicit failure modes.
Research summary
Arman Nik Khah is a researcher at The University of Texas at Dallas whose cataloged work concentrates on two adjacent fronts of large language model research: AI safety evaluation and LLM privacy. Both papers in the Evolve catalog are 2026 arXiv preprints with him in the first-author position — one also corresponding — and both carry a high reproducibility rating. The first study asks whether language models can diagnose a corrupted reward channel when given a verified record. Evolve's review found that models catch a lying reward reporter near-perfectly yet often convict honest ones over irrelevant prompt details, calling the work "a rigorous, highly transparent empirical study." The second examines what verified duplication counts say about membership evidence, claiming loss-based membership signals are faint below roughly 1,000 verified copies and confounded above — findings the review deems rigorous while noting they rest on tiny banks and six novels. Bibliometrically, the snapshot is early-stage: zero distinct co-authors recorded, an even 50/50 field distribution between natural language processing (LLM privacy / membership inference) and artificial intelligence (LLM evaluation / AI safety), and a reproducibility mean of 100 across two papers. Output, impact, and activity metrics are not yet available, so this profile should be read as an early record of a solo, rigor-oriented LLM evaluation researcher.
Collaboration pattern
Within the current snapshot, Arman Nik Khah's cataloged work records no co-authors at all: the distinct co-author count is 0, the frequent-co-author list is empty, and both papers carry him as first author (one also corresponding). The recorded pattern is therefore fully independent authorship on 2026 arXiv preprints. Whether this reflects a deliberate solo style or simply the narrow scope of the two cataloged papers cannot be determined from the available data.
Research keywords
Highlights
Near-perfect at catching a lying reward reporter
But honest agents are hard to clear
Membership evidence is faint below ~1,000 verified copies
Rigorous, but built on tiny banks and six novels
Timeline
- 2026Debuts in the Evolve catalog with a first-authored arXiv study on language models diagnosing a corrupted reward channel from a verified record.
- 2026Publishes a second first-authored arXiv paper — this time as corresponding author — on verified duplication counts and membership evidence in language models.
- 2026Both papers receive high reproducibility readiness ratings, yielding a reproducibility mean of 100 across 2 papers in the radar.
Papers in the Evolve catalog
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.
- All radar scores and counts come only from the supplied bibliometric snapshot of 2 cataloged arXiv papers (both 2026), not a full publication record.
- Output, impact, and activity axes are marked missing — no citation, volume, or cadence data was provided, and none is estimated here.
- The 100 reproducibility score is computed over only 2 papers, so it should be read as high readiness on a very small sample.
- Identity is name-resolved from public bibliographic data (ORCID 0009-0009-4296-5250, The University of Texas at Dallas); no country or career-history data was supplied.