AI safety & governance researcher: sandbagging in model organisms and evidence-grounded LLM policy benchmarks
3
papers here
1
reviewed by Evolve
2
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: the snapshot provides no aggregate output measure — h-index and i10-index both record 0 and no total works count is given — so this axis, which would measure publication volume and its growth over time, cannot be assessed yet.
Impact—
Not yet scored: citation data is absent from the snapshot, so this axis — measuring citation uptake, citing works, and influence on follow-up research — cannot be assessed; the h-index and i10-index both record 0 at present.
Focus100/100
A focus score of 100 reflects a strongly concentrated profile. OpenAlex topics cluster tightly: 'Ethics and Social Impacts of AI' (4 papers), 'Adversarial Robustness in Machine Learning' (3) and 'Explainable Artificial Intelligence' (2) dominate, with single counts for multi-agent systems and computational text analysis, and the field distribution lists only 'AI governance / computational policy analysis' (1). Sandbagging alignment work plus governance benchmarking reads as one coherent agenda.
Collaboration36/100
A collaboration score of 36 reflects a still-narrow network: the snapshot records only 5 distinct co-authors across recent work. Repeat collaborators include David Williams-King, Hong Kiat Tan, Melanie Bui, My Chiffon Nguyen and Zachary Schlosser, suggesting small-team authorship on the sandbagging and benchmark papers rather than large consortium-style collaborations.
Activity47/100
An activity score of 47 matches a compact but current record: countsByYear lists 4 works in 2026 and recent3yWorks is also 4, meaning all indexed output falls inside the latest window. No works are recorded before 2026 in this snapshot, so the publishing cadence beyond that single year cannot yet be characterized.
Reproducibility55/100
Reproducibility scores 55, computed as a mean of 55 over 1 assessed paper (reproPapers: 1). The single reviewed paper, GPS-Bench, carries a 'medium' repro-readiness rating; Evolve's review called its leakage discipline exemplary while noting the headline tables were not in the excerpt — a good methods-hygiene signal from very limited coverage.
Research summary
Linh Le is an early-career researcher working at the intersection of AI safety and AI governance, with a profile name-resolved from public OpenAlex bibliographic data; no institutional affiliation, country, or ORCID is recorded in the current snapshot. The indexed record is concentrated in 2026, with four works in the last three years and both the h-index and i10-index recorded as 0, consistent with a very recent publication record. Two threads stand out. The first is alignment research on sandbagging — models deliberately withholding capabilities: Le is first author of 'Reference-Grafting Matches Fine-Tuning at Eliciting Sandbagged Capabilities' and a middle author of 'A Causal Model for Locating and Unlocking Sandbagging in Model Organisms'. The second is governance tooling: as first author of 'GPS-Bench: A Governance Policy Benchmark for Automating Policy Analysis', Le proposes an evidence-grounded benchmark for LLM-agent policy simulation; Evolve's review singled out exemplary leakage discipline while noting the headline tables were absent from the excerpt. OpenAlex topics corroborate this mix — 'Ethics and Social Impacts of AI' (4), 'Adversarial Robustness in Machine Learning' (3), 'Explainable Artificial Intelligence' (2) — with single counts for multi-agent systems and computational text analysis. Collaboration remains narrow (5 distinct co-authors), anchored by repeat co-authors David Williams-King, Hong Kiat Tan, Melanie Bui, My Chiffon Nguyen, and Zachary Schlosser.
Collaboration pattern
The recorded pattern is small-team research: 5 distinct co-authors appear in the recent snapshot, and the same group recurs — David Williams-King, Hong Kiat Tan, Melanie Bui, My Chiffon Nguyen and Zachary Schlosser. Le takes first-author roles on both GPS-Bench and the reference-grafting sandbagging study, and a middle authorship on the causal sandbagging-model paper, consistent with a compact core group spanning alignment experiments and governance tooling rather than large multi-institution consortia.
Research keywords
Highlights
First-authored governance policy analysis benchmark
First-author study on eliciting sandbagged capabilities
'Reference-Grafting Matches Fine-Tuning at Eliciting Sandbagged Capabilities' (2026, arXiv) is a first-authored contribution to sandbagging research; as the title states, it reports that a reference-grafting approach matches fine-tuning for eliciting capabilities a model withholds.
Co-authorship on a causal model of sandbagging
'A Causal Model for Locating and Unlocking Sandbagging in Model Organisms' (2026, arXiv) places Le in a middle author position, extending the sandbagging agenda toward causal localization within model organisms alongside the same co-author group.
Timeline
- 2026Releases 'A Causal Model for Locating and Unlocking Sandbagging in Model Organisms' on arXiv as a middle author, on the sandbagging alignment thread.
- 2026Publishes first-authored 'Reference-Grafting Matches Fine-Tuning at Eliciting Sandbagged Capabilities' on arXiv.
- 2026First-authors GPS-Bench, a governance policy benchmark for automating policy analysis, on arXiv; OpenAlex records 4 works in total for this year.
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.
- Bibliometric scores are computed from OpenAlex only; the h-index and i10-index both record 0 and no citation counts are available in the snapshot.
- No institutional affiliation, country, or ORCID is recorded; 'Linh Le' is a name with many homonyms, so this profile is name-resolved from public bibliographic data and may merge or split identities.
- Only one paper has been reviewed by Evolve (reproPapers: 1), so the reproducibility score rests on a single assessment.
- Two of the three cataloged papers lack Evolve review data, keywords, and repro-readiness ratings, and countsByYear covers only 2026.