Stanford researcher in LLM interpretability, AI safety and evaluation auditing, plus applied learning analytics.
3
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
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—
Output is not scored: the snapshot lacks publication counts and per-year productivity, so total output volume cannot be measured. This axis would aggregate the number of papers over time, including first- and corresponding-author output.
Impact—
Impact is not scored: citation counts, h-index and venue-level influence are absent from the snapshot. This axis would measure citations to the reviewed work and its uptake across the LLM evaluation and interpretability communities.
Focus0/100
Focus scores 0, the lowest reading on this radar, and the field distribution explains why: the reviewed record splits exactly 50/50 between machine learning (LLM interpretability & AI safety) and natural language processing (LLM evaluation and auditing), with a further educational-data-mining paper outside both. The papers are thematically linked around LLM behavior, but formally span multiple field labels, and that breadth dilutes the focus score.
Collaboration51/100
Collaboration scores 51, a middling band, on 12 distinct coauthors in the snapshot. The frequent-coauthor list is anchored by repeat collaborator Nathan Hu, listed twice, alongside Christopher Potts, Daniel E. Ho, Benjamin W. Domingue, Nick Haber, Sang T. Truong, Jennifer Wang, Fagun Patel, Duc Q. Nguyen, Kazunori Fukuhara and Joachim Baumann. Authorship positions — middle on two papers, last on one — indicate both contributing and senior roles.
Activity—
Activity is not scored: the snapshot has no per-year publication series, so research tempo and recent output cadence cannot be computed. This axis would track publication frequency and recency across the recorded years.
Reproducibility78/100
Reproducibility is the strongest scored axis at 78, the mean across 2 scored papers. Per-paper readiness is consistent with that reading: high for the subliminal-learning verbalization work, and medium for both the API-versus-chatbot audit and the CodeInsight dataset paper. The audit reports test-retest reliability checks and the 3.4pp accuracy gap, though unverified checkpoint identity is flagged as a confound, and the dataset is gated.
Research summary
Sanmi Koyejo (Stanford University) is profiled here from public bibliographic data and Evolve's reviewed catalog rather than a verified biography. The reviewed record centers on large language models: Evolve has reviewed three arXiv papers, all dated 2026, split evenly between machine learning (LLM interpretability & AI safety) and natural language processing (LLM evaluation and auditing), plus one educational-data-mining study. The interpretability paper, where Koyejo is a middle author, introduces SALVE, which verbalizes teacher traits hidden in subliminal data via soft prompts plus beam search. As last author on the evaluation audit, the record argues that API benchmark scores do not reliably transfer to chatbot interfaces, reporting a 3.4pp accuracy gap. A third paper, again middle-author, releases CodeInsight, a dataset of 3M+ code submissions with a shared benchmark for modeling iterative problem-solving. Bibliometric scores put reproducibility at 78 (mean across 2 scored papers), collaboration at 51 (12 distinct coauthors), and focus at 0, reflecting the evenly split field distribution. Output, impact and activity lack sufficient underlying data and are left unscored.
Collaboration pattern
The snapshot records 12 distinct coauthors and a collaboration score of 51. Nathan Hu appears twice in the frequent-coauthor list, marking the closest repeat tie; other recurring names — Christopher Potts, Daniel E. Ho, Benjamin W. Domingue, Nick Haber, Sang T. Truong, Jennifer Wang, Fagun Patel, Duc Q. Nguyen, Kazunori Fukuhara and Joachim Baumann — point to project-based teams mixing NLP, evaluation-audit and learning-analytics collaborators rather than one fixed group.
Research keywords
Highlights
SALVE verbalizes traits hidden in subliminal data
API benchmark scores diverge from chatbot interfaces
CodeInsight: 3M+ submissions for modeling iterative problem-solving
Timeline
- 2026Middle author on the SALVE paper verbalizing subliminal learning effects via text optimization (soft prompts plus beam search).
- 2026Last author on the audit showing API benchmark scores do not reliably transfer to chatbot interfaces, with a 3.4pp accuracy gap.
- 2026Middle author on the CodeInsight release: a 3M+ submission dataset and shared benchmark for modeling iterative problem-solving.
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.
- Radar scores are computed from a public bibliometric snapshot only; output, impact and activity lack the underlying counts and are intentionally left unscored.
- Only the three arXiv papers in the Evolve catalog are covered here; other work by this researcher may be absent from this profile.
- Country and affiliation details beyond the listed institution are not provided; the profile is name-resolved from public bibliographic data (ORCID 0000-0002-4023-419X).
- Paper-level facts (Evolve verdicts, reproducibility readiness, authorship position) come from Evolve's reviews, not independent verification.