Stanford researcher in mechanistic interpretability and LLM safety: attribution methods and subliminal learning.
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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 metrics are not yet available for this profile, so this axis is unscored. It would measure total publications, venue types, and output volume over time once a fuller bibliographic record is linked.
Impact—
Impact data such as citation counts and h-index is absent from the input, so this axis is unscored. It would measure citations and uptake of the two 2026 papers once citation records are attached.
Focus0/100
The focus score is 0, derived from a field distribution split exactly 50/50 between machine learning (LLM interpretability & AI safety) and machine learning (mechanistic interpretability). The two cataloged papers sit one in each bucket, so the metric reads maximal dispersion. In substance the labels are adjacent subfields of the same interpretability agenda, so the low score reflects label-level splitting rather than genuinely scattered research interests.
Collaboration46/100
Collaboration scores 46, computed from 9 distinct coauthors in the bibliographic snapshot. Named frequent collaborators include Sanmi Koyejo, Christopher Potts, Aryaman Arora, Kirill Acharya, Yanzhe Zhang, Noah Goodman, and Dan Jurafsky, a cluster consistent with a Stanford-centered NLP/ML network. The mid-range score reflects a small collaborator pool so far; authorship spanning first author on one paper and middle author on the other indicates active team participation.
Activity—
Activity data is insufficient to score this axis. It would measure publication cadence and continuity over recent years; both cataloged papers are dated 2026, too narrow a window to establish an activity trajectory.
Reproducibility78/100
Reproducibility scores 78 (reproMean 78), assessed across 2 papers. The first-author SALVE paper is rated high repro-readiness, while Matryoshka Attribution is rated medium, with Evolve noting the work is rigorous but partly measured by its own metrics. Within a two-paper base this is a solid signal of releasable setups, though the sample is small.
Research summary
Nathan Hu is a Stanford University researcher whose cataloged work sits squarely in machine learning interpretability, split evenly between machine learning (LLM interpretability & AI safety) and machine learning (mechanistic interpretability) — a 50/50 field distribution that yields a focus score of 0. The Evolve catalog holds two 2026 arXiv papers. As first author, Hu contributed SALVE (Verbalizing Subliminal Learning Effects Using Text Optimization), which verbalizes teacher traits hidden in subliminal data via soft prompts plus beam search; Evolve rates the results strong, with reliability dropping on steered and preference data, and its reproducibility readiness as high. As a middle author, Hu contributed Matryoshka Attribution, which learns one nested attribution ordering via sigmoid top-k, tops MIB (5.6 vs 1.95) and ablates refusal behavior with 1% of weights; Evolve calls it rigorous but partly self-metric'd, with medium repro-readiness. Together the two papers carry a reproducibility score of 78 (reproMean 78 across 2 assessed papers). Collaboration scores 46, grounded in 9 distinct coauthors including Sanmi Koyejo, Christopher Potts, Aryaman Arora, Kirill Acharya, Yanzhe Zhang, Noah Goodman, and Dan Jurafsky — a Stanford-centered NLP/ML network. Output, impact, and activity axes remain unscored because citation and volume data are missing from the current snapshot.
Collaboration pattern
Nathan Hu's collaboration pattern is small and Stanford-centered: 9 distinct coauthors appear in the snapshot, with a collaboration score of 46. Frequent collaborators — Sanmi Koyejo, Christopher Potts, Aryaman Arora, Kirill Acharya, Yanzhe Zhang, Noah Goodman, and Dan Jurafsky — span Stanford NLP and machine learning groups. Authorship positions span both ends: first author on the SALVE paper and middle author on Matryoshka Attribution, indicating both led and supporting roles within team projects.
Research keywords
Highlights
SALVE surfaces hidden teacher traits from subliminal data
Matryoshka Attribution beats MIB with nested attribution
Timeline
- 2026Posts first-author arXiv paper SALVE (Verbalizing Subliminal Learning Effects Using Text Optimization) in LLM interpretability and AI safety, rated high repro-readiness.
- 2026Contributes as a middle author to the arXiv paper Matryoshka Attribution in mechanistic interpretability, rated medium repro-readiness.
- 2026Both papers enter the Evolve catalog; the profile snapshot records a 50/50 field split, 9 distinct coauthors, and a reproducibility score of 78 across the two papers.
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 come from a bibliographic snapshot; publication counts, citations, and activity data are missing, so output, impact, and activity axes are unscored.
- The profile rests on only two Evolve-reviewed papers, both 2026 arXiv preprints; peer-review status is not stated.
- No ORCID or country is provided and Nathan Hu is a moderately common name; identity is name-resolved from public bibliographic data, and the duplicate Nathan Hu entry in the coauthor list was treated as the researcher himself.
- The reproducibility score of 78 is computed from just 2 assessed papers and should not be read as a career-wide benchmark.