AI safety researcher red-teaming morally trained LLM agents; first-author arXiv study on persona-attack robustness
1
papers here
1
reviewed by Evolve
1
first-author
1
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-21) 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—
Insufficient data: the catalog holds only one paper for this name, so publication volume, venue mix, and yearly output trends cannot be computed. This axis would measure total scholarly production and its growth over time.
Impact—
No citation metrics are available yet: the single 2026 arXiv paper is too recent for citation capture. This axis would measure citation counts, h-index, and uptake of the work by other researchers.
Focus100/100
The focus score of 100 reflects perfect topical concentration: all captured bibliographic output — currently one paper — sits in a single field, machine learning / AI safety (LLM alignment and adversarial robustness). The first-author arXiv study combines reinforcement learning, persona attacks, and activation steering within one research question. As the catalog holds just one record, this concentration should be read as an early snapshot rather than an established specialization.
Collaboration0/100
The collaboration score of 0 mirrors the raw record: zero distinct co-authors captured and an empty frequent-coauthor list. The reviewed paper lists this researcher as first and corresponding author, consistent with a solo or minimally attributed early-stage work. No co-authorship network can be drawn yet; institutional affiliation and team composition remain unknown from the available bibliometric data.
Activity—
With one captured paper, research activity cadence — publications per year, preprint pacing, and ongoing output — cannot be assessed. This axis would measure publishing rhythm and sustained research engagement over recent years.
Reproducibility100/100
Reproducibility is the strongest signal in this profile: a reproducibility mean of 100 across one assessed paper, matching the Evolve review's high reproducibility-readiness rating for the 2026 red-teaming study. The review still flags one material limitation — the evaluation pipeline depends on an LLM judge that has not yet been validated — so future readers should treat judge-based metrics as the main replication risk.
Research summary
Arth Singh is an early-stage researcher whose Evolve profile currently rests on a single, highly focused output: a 2026 arXiv first-author and corresponding-author study, "Does Moral Reasoning Training Help or Hurt? Red-Teaming RL-Trained Ethical Agents with Persona Attacks." The paper sits in machine learning / AI safety, targeting LLM alignment and adversarial robustness — a concentration reflected in a focus score of 100, with all captured bibliographic activity in one field. Evolve's review finds the work credible: moral-reward reinforcement learning yields 3.9-5.8x persona-attack robustness under strong experimental controls, although fiction-based role-play still defeats the trained agents, and the evaluation hinges on an as-yet unvalidated LLM judge. Reproducibility is the clearest strength, with a reproducibility mean of 100 across the one assessed paper and a high reproducibility-readiness rating. Collaboration data is empty — zero distinct co-authors captured and no frequent collaborators — so team-science patterns cannot yet be characterized. Output, impact, and activity axes lack sufficient data to score. This profile should be read as a provisional, early-stage snapshot, name-resolved from public bibliographic data.
Collaboration pattern
The available record shows no co-authors: the distinct co-author count is zero and the frequent-coauthor list is empty. The cataloged paper is first-authored and corresponding-authored, consistent with independent early-career work. Because bibliometric linkage for this name is limited, real-world collaborations may exist but are not captured; institutional affiliation is also unknown.
Research keywords
Highlights
Moral-reward RL lifts persona-attack robustness 3.9-5.8x
Fiction role-play still defeats morally trained agents
High reproducibility, hinged on an unvalidated LLM judge
Timeline
- 2026First-authored and corresponding-authored arXiv preprint released: red-teaming RL-trained ethical agents with persona attacks, in LLM alignment and adversarial robustness.
- 2026Evolve review finds moral-reward RL yields 3.9-5.8x persona-attack robustness under strong controls, while fiction role-play still wins.
- 2026The paper receives a high reproducibility-readiness rating, with a reproducibility mean of 100 in Evolve's assessment.
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.
- Bibliometric scores are computed from public bibliographic data only; institution, country, and ORCID are not provided for this name.
- Only one paper (2026, arXiv) is in the Evolve catalog, so this is an early-stage profile that will change as more outputs are captured.
- Scores of 0 and 100 on collaboration, focus, and reproducibility reflect a single-record basis, not mature minima or maxima.
- Name resolution is based on public bibliographic data; 'Arth Singh' may correspond to multiple researchers, and identity confidence is limited without an ORCID.