First-author researcher at UIUC in mechanistic interpretability and language-model evaluation
2
papers here
1
reviewed by Evolve
2
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 underlying publication-count, venue, and output-volume data is insufficient in this snapshot, which lists only two 2026 arXiv preprints. This axis would measure total publications, venue quality distribution, and yearly volume.
Impact—
Impact is not scored: no citation counts, citation-per-year trajectories, or downstream-adoption signals are available for these two preprints. This axis would measure citations, influential citations, and uptake of the proposed Definition 4 by later interpretability work.
Focus100/100
Focus scores 100, but the base is tiny: the field distribution contains exactly one labeled field — machine learning interpretability (mechanistic interpretability) — from one tagged paper. Within that field, the survey's keyword set (linear probing, sparse autoencoders, activation steering, causal abstraction) shows a coherent conceptual agenda, while the unlabeled MineCEraft preprint points to language-model evaluation as a second direction. Treat the perfect score as concentration-by-default, not demonstrated depth across venues.
Collaboration36/100
Collaboration scores 36 from a record of 5 distinct coauthors: Julia Hockenmaier, Marc E. Canby, Ikhyun Cho, Risham Sidhu, and Yoonhwa Jung. Both catalog papers are first-author and non-corresponding, so the visible footprint is that of a lead author writing preprints within a small circle around the listed frequent coauthors. No senior-collaborator counts, cross-institution ties, or team-size trends can be derived from this snapshot; the network picture will firm up as more indexed work appears.
Activity—
Activity is not scored: the snapshot lacks multi-year publication timing data, so cadence and continuity cannot be computed. This axis would measure recent publication frequency, preprint-to-venue conversion, and sustained research pace beyond the two 2026 entries.
Reproducibility20/100
Reproducibility scores 20, computed as the mean over 1 rated paper. The rated paper is the linear representation hypothesis survey, which Evolve classifies as low reproducibility readiness: the review credits a rigorous taxonomy and a falsifiable Definition 4 but notes the paper adds no new experiments. With only one rated item, this score describes that survey rather than the researcher's broader practice; the MineCEraft preprint is not yet rated.
Research summary
Sewoong Lee is a researcher at the University of Illinois Urbana-Champaign whose visible record in this snapshot consists of two first-author arXiv preprints, both dated 2026. The anchor output, 'A Survey on the Linear Representation Hypothesis', is a conceptual contribution: it argues that the linear representation hypothesis is ill-defined unless tied to a fixed model-layer-feature-dataset tuple, and proposes a falsifiable Definition 4. Evolve's review credits its taxonomy as rigorous while noting that it adds no new experiments, and rates its reproducibility readiness as low. The second preprint, 'MineCEraft', moves toward evaluation, testing language models as construction engineers inside Minecraft and suggesting a second, applied direction. Mechanistic interpretability is the only labeled field (one paper), so the focus score of 100 reflects concentration on a very small base rather than demonstrated breadth or depth. Collaboration is modest: 5 distinct coauthors, including Julia Hockenmaier, give a collaboration score of 36. Output, impact, and activity metrics are unavailable and are not scored. This is an early-stage profile built from a narrow bibliographic snapshot.
Collaboration pattern
The record shows 5 distinct coauthors — Julia Hockenmaier, Marc E. Canby, Ikhyun Cho, Risham Sidhu, and Yoonhwa Jung — yielding a collaboration score of 36. Both catalog papers are first-author and non-corresponding, consistent with lead-authored preprints written within a small circle around the listed frequent coauthors. No senior-advisor roles, cross-institution ties, or team-size trends can be inferred from this snapshot, so the collaboration picture remains provisional.
Research keywords
Highlights
A falsifiable definition for the linear representation hypothesis
Conceptual rigor with an explicit empirical gap
MineCEraft: language models as construction engineers
A second 2026 first-author arXiv preprint moves from conceptual analysis to evaluation, benchmarking language models as construction engineers inside Minecraft — an early signal of a second direction in interactive, environment-based language-model evaluation.
Timeline
- 2026First-author survey 'A Survey on the Linear Representation Hypothesis' posted to arXiv, arguing the hypothesis is ill-defined without a fixed model-layer-feature-dataset tuple and proposing falsifiable Definition 4.
- 2026Second first-author arXiv preprint 'MineCEraft' released, evaluating language models as construction engineers in the world of Minecraft.
- 2026Both catalog entries appear in one year as first-author, non-corresponding arXiv work with 5 distinct coauthors, marking the visible start of this profile's publication record.
Papers in the Evolve catalog
- first author
A Survey on the Linear Representation Hypothesis
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 dimensions are computed from OpenAlex-derived bibliometric data only; no ORCID or country is attached in this snapshot.
- 'Sewoong Lee' is a relatively common name; this profile is name-resolved from public bibliographic data and may be incomplete or conflated with same-name authors.
- Output, impact, and activity axes are not scored because only two 2026 arXiv preprints are in scope; citation and timing data are absent.
- Focus (100) and reproducibility (20) each rest on a single labeled or rated paper, so both scores are low-confidence.