Tsinghua NLP researcher evaluating how AI assistants respond to repeated verbal abuse in multi-turn dialogue.
1
papers here
1
reviewed by Evolve
1
first-author
0
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—
Not yet scoreable: the snapshot contains only one catalogued 2026 paper, so publication volume cannot be measured. This axis would track output quantity, venue spread, and productivity trends over time.
Impact—
Not yet scoreable: no citation counts or influence metrics appear in the provided data. This axis would measure citations, uptake, and the broader scholarly influence of the researcher's publications.
Focus100/100
Focus scores a perfect 100: the field distribution contains a single bucket, 'NLP / AI safety evaluation', covering the one reviewed paper. Read cautiously — with only one catalogued output, this indicates early, sharp specialization in conversational AI safety evaluation rather than a demonstrated long-term research concentration.
Collaboration44/100
The collaboration score is 44 out of 100, computed from 8 distinct coauthors in the bibliometric snapshot. The frequent-coauthor list is led by Wei Zhang and Yi Wang, each appearing twice, alongside Pierrick Bougault, Agoston Bodo, Vitor D. de Moura, and José O. Gomes — a compact, recurring circle centered on the first-authored 2026 arXiv paper rather than a wide network.
Activity—
Not yet scoreable: only a single 2026 arXiv paper is visible, so recent publishing cadence cannot be computed. This axis would measure publication frequency and sustained research engagement across recent years.
Reproducibility100/100
Reproducibility scores 100 out of 100, the mean readiness across exactly 1 reviewed paper. Evolve rates 'How AI Assistants Respond to Repeated Abuse' as high reproducibility readiness, matching the review verdict of careful, transparent measurement; the perfect score should still be read on a single-paper basis.
Research summary
William Guey is a researcher at Tsinghua University working in natural language processing, with a catalogued focus on AI safety evaluation of conversational systems. Evolve's records center on the 2026 arXiv paper 'How AI Assistants Respond to Repeated Abuse', on which Guey is first author, though not the corresponding author. The study's central claim — echoed in Evolve's review verdict — is that refusal is not a single axis: different AI assistant configurations diverge sharply when subjected to repeated verbal abuse. The measurement approach is described as careful and transparent, combining multi-turn dialogue evaluation, an LLM-as-a-judge protocol, and a bilingual benchmark, while remaining exploratory and deployment-bound. Methodological rigor carries into the scores: reproducibility reaches a perfect 100, the mean readiness across one reviewed paper rated 'high'. Focus likewise scores 100 because the reviewed output sits entirely within NLP / AI safety evaluation — an early signal of specialization rather than an established pattern. Collaboration scores a moderate 44, built on 8 distinct coauthors, with Wei Zhang and Yi Wang recurring most often alongside Pierrick Bougault, Agoston Bodo, Vitor D. de Moura, and José O. Gomes. Three of six radar axes — output, impact, and activity — cannot yet be scored for lack of data, making this an early-stage profile whose quantitative picture rests on a single reviewed paper.
Collaboration pattern
Collaboration is compact and recurring: 8 distinct coauthors appear in the snapshot, giving a collaboration score of 44. Wei Zhang and Yi Wang recur most often (each listed twice among frequent coauthors), alongside Pierrick Bougault, Agoston Bodo, Vitor D. de Moura, and José O. Gomes — a small, stable circle around the first-authored 2026 paper. No seniority or role information is available.
Research keywords
Highlights
Refusal is not one axis under repeated abuse
Careful, transparent measurement via LLM-as-a-judge
High reproducibility readiness
Timeline
- 2026First-authored 'How AI Assistants Respond to Repeated Abuse' released on arXiv (non-corresponding).
- 2026The paper enters the Evolve review catalog in the NLP / AI safety evaluation field.
- 2026Evolve's review calls the measurement careful and transparent and rates reproducibility readiness 'high'.
Papers in the Evolve catalog
- first author
How AI Assistants Respond to Repeated Abuse
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.
- Name-resolved from public bibliographic data: no ORCID or country field was available, so identity matches rest on the provided snapshot.
- Bibliometric dimensions (collaboration 44, focus 100, reproducibility 100) are computed from real bibliometric data but rest on a single reviewed 2026 paper and one field bucket.
- Output, impact, and activity are unscored: no citation counts, publication counts, or venue history were provided, and none are inferred.
- Catalog coverage is limited to papers Evolve has reviewed; the full publication record may be broader.