Early-stage, UK-based researcher on AI governance benchmarks and sandbagging in model organisms
3
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
Not yet scored: Evolve lacks a stable total-output record for this author. The axis would measure cumulative works, annual output tempo, and authorship share; countsByYear currently shows 1 work (2025) and 10 (2026).
Impact—
Not yet scored. Citation coverage is essentially empty in the snapshot — h-index 0 and i10-index 0 — so no reliable impact estimate exists. This axis would measure citations, citation-weighted influence, and uptake of the 2026 benchmark and sandbagging work.
Focus100/100
Scored 100 — maximal topical concentration. Topic counts form a single tight cluster: adversarial robustness in machine learning (10 works), ethics and social impacts of AI (6), explainable AI (5), with topic modeling (4) and hate speech detection (3) smaller, and exactly one classified field (AI governance / computational policy analysis). Every catalog paper sits in AI safety or governance, showing a deliberately narrow agenda.
Collaboration36/100
Scored 36, in the below-mid range. The recent three-year window records only 5 distinct coauthors, and the recurring set (Linh Le, Hong Kiat Tan, Melanie Bui, My Chiffon Nguyen, Zachary Schlosser) reads as one compact team. All three Evolve-catalog papers place him as last author, none corresponding, consistent with a senior-listed but non-corresponding role in a small group.
Activity72/100
Scored 72. OpenAlex records 11 works in the last three years, and countsByYear shows output accelerating sharply: 1 work in 2025 versus 10 in 2026. All three catalog papers appeared on arXiv in 2026, so the score reflects a steep single-year ramp rather than a long publication history; sustainability beyond 2026 is not yet observable.
Reproducibility55/100
Scored 55 as the mean over just 1 scored paper. The scored item is GPS-Bench (medium reproducibility readiness), whose Evolve review praises exemplary leakage discipline while noting headline result tables were absent from the excerpt and the passage is largely prospective. The two sandbagging papers carry no reproducibility assessment yet, so this score rests on a single data point.
Research summary
David Williams-King (listed affiliation: Oldham Council, United Kingdom; name-resolved from OpenAlex bibliographic records, no ORCID linked) is an early-stage researcher whose indexed work concentrates sharply on AI safety and governance. OpenAlex records 11 works in the last three years — 1 in 2025 rising to 10 in 2026 — with h-index and i10-index both 0, indicating citation coverage has not yet accumulated. Topic counts form a tight cluster: 10 works on adversarial robustness in machine learning, 6 on ethics and social impacts of AI, 5 on explainable AI, plus topic modeling (4) and hate speech and cyberbullying detection (3). In the Evolve catalog, three 2026 arXiv papers list him as last author, none corresponding: GPS-Bench, a governance policy benchmark for automating policy analysis, judged to show exemplary leakage discipline though headline tables were absent from the excerpt and the passage is largely prospective; and two sandbagging studies — a causal model for locating and unlocking sandbagging in model organisms, and evidence that reference-grafting matches fine-tuning at eliciting sandbagged capabilities. Radar scores place focus at 100 (maximal concentration), activity at 72, reproducibility at 55 (mean of one scored paper), and collaboration at 36 across five distinct coauthors. Output and impact axes remain unscored for lack of data.
Collaboration pattern
Collaboration is compact and stable: the recent three-year window records just 5 distinct coauthors, and the recurring set — Linh Le, Hong Kiat Tan, Melanie Bui, My Chiffon Nguyen, Zachary Schlosser — reads as one core team. Across all three Evolve-catalog papers he holds the last author position while serving as corresponding author on none, a senior-listed but non-corresponding role consistent with the below-mid collaboration score of 36.
Research keywords
Highlights
GPS-Bench: an evidence-grounded governance policy benchmark for automating policy analysis
A causal model for locating and unlocking sandbagging in model organisms
2026 arXiv paper with last-author position. Evolve has not reviewed it: no verdict, keywords, or reproducibility score are on record, so only the title-level contribution — a causal account of where sandbagging resides in model organisms and how it can be unlocked — can be stated here.
Reference-grafting matches fine-tuning at eliciting sandbagged capabilities
The second 2026 sandbagging paper, again last-authored and unreviewed by Evolve. Its title-level claim — that grafting reference signals elicits hidden capabilities on par with fine-tuning — fits an author topic profile dominated by adversarial robustness in machine learning (10 counted works).
Timeline
- 2025First indexed output appears: OpenAlex countsByYear records a single work in 2025, with h-index and i10-index still at 0.
- 2026Output accelerates to 10 works in a single year, lifting the recent three-year total to 11 and driving the activity score to 72.
- 2026GPS-Bench, a governance policy benchmark for automating policy analysis, is posted to arXiv with him as last author; Evolve rates its reproducibility readiness medium.
- 2026Two sandbagging studies follow on arXiv — a causal model for locating and unlocking sandbagging, and reference-grafting versus fine-tuning — both last-authored and both awaiting Evolve review.
Papers in the Evolve catalog
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.
- Bibliometrics come from OpenAlex only; no ORCID is linked and identity is name-resolved from public bibliographic records — the listed affiliation (Oldham Council, UK local government) is atypical for this output, so entity resolution warrants caution.
- Output and impact radar axes are unscored; h-index and i10-index are both 0, meaning citation coverage has not accumulated in this snapshot.
- The reproducibility score (55) is a mean over one scored paper (GPS-Bench); the two sandbagging papers lack verdicts, keywords, and slugs in the catalog.
- Work counts reflect OpenAlex indexing and may under- or over-count actual output.