NLP researcher at TTIC (Chicago) working across LLM mechanistic interpretability and computational sociolinguistics.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
13
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—
Not yet scored: the snapshot lacks validated per-work authorship and venue data needed to measure total production and seniority of roles. Year-level counts exist, but the pipeline could not compute an output score from them.
Impact—
Not yet scored: no citation-level evidence (citations per paper, citing works, normalized field impact) is available in this snapshot, so the impact axis, which measures scholarly uptake and influence of the work, is left unscored rather than estimated.
Focus0/100
Scores 0/100, the floor of the focus axis: fieldDist assigns exactly 0.5 to LLM mechanistic interpretability and 0.5 to computational sociolinguistics, an even two-way split. The two 2026 catalog papers mirror this - one a mechanistic study of domain jargon in Llama-3.1, the other a Reddit slang-diffusion analysis. The split reads as a deliberate dual track rather than drift, but topical concentration is currently nil.
Collaboration28/100
Scores 28/100, near the low end: only 3 distinct coauthors appear in the snapshot, and the recurring set is just Darin Keng, Xiaoning Wang, and Ted Underwood. Both 2026 catalog papers place Sun in the last (senior) author position, though neither flags a corresponding-author role. The pattern suggests a small, stable core team producing supervised work rather than a broad external collaboration network.
Activity90/100
Scores 90/100, the strongest axis. Counts by year run 5 (2019), 6 (2020), 8 (2021), 5 (2022), 6 (2023), 4 (2024), 8 (2025), and 9 (2026) - a dip in 2024 followed by the two busiest years on record, and the last three years total 21 works (recent3yWorks), the highest three-year stretch in the 2019-2026 window.
Reproducibility78/100
Scores 78/100, a reproMean averaged over 2 papers reviewed by Evolve. The jargon-benchmark paper ('Domain-Specific Jargon in Large Language Models') is rated high readiness, while the Reddit slang-diffusion study ('From Utterances to Networks') sits at medium. Both are arXiv-hosted, and Evolve's verdicts flag analytical caveats - a single model pair in the first, survivorship effects in the second - rather than replication blockers.
Research summary
Zhewei Sun is a researcher at Toyota Technological Institute at Chicago whose profile is name-resolved from OpenAlex bibliographic data. The snapshot splits attributed output evenly (0.5/0.5) between two NLP subfields: LLM mechanistic interpretability and computational sociolinguistics. The 2026 Evolve catalog reflects both tracks, with Sun as last author on each paper. In 'Domain-Specific Jargon in Large Language Models', the group built two jargon benchmarks where a fine-tuned Llama-3.1 underperformed its base model; Evolve rates the mechanistic study solid while cautioning that a single model pair overstates the 'closes the gap' claim. In 'From Utterances to Networks', LLM-based slang annotation combined with fixed-effects and survival models ties Reddit slang spread to users' network position and context, judged sound with survivorship caveats. Activity is the strongest axis (90/100): annual works rise from 5 in 2019 to 9 in 2026, with the last three years totaling 21 works. Collaboration is thin (28/100) — 3 distinct coauthors, with Darin Keng, Xiaoning Wang, and Ted Underwood recurring — and focus scores 0/100 given the exact two-way field split. Reproducibility averages 78/100 across the 2 reviewed papers (one high, one medium readiness). The snapshot reports an h-index of 13 (i10: 18), though citation-level data was insufficient to score the output and impact axes.
Collaboration pattern
Collaboration is narrow but stable: the snapshot lists only 3 distinct coauthors, with Darin Keng, Xiaoning Wang, and Ted Underwood as recurring collaborators. On both 2026 Evolve-reviewed papers Sun sits in the last (senior) author position without a corresponding-author flag, consistent with a supervisory role over a small core group. The low collaboration score (28/100) reflects this concentrated team rather than absent partnerships; no cross-institution or industry collaboration data is available in the snapshot.
Research keywords
Highlights
Two jargon benchmarks show fine-tuned Llama-3.1 lagging its base model
LLM annotation ties Reddit slang spread to network position
Timeline
- 2019Earliest year in the snapshot's counts-by-year series, with 5 works recorded.
- 2021Output reaches 8 works, the highest annual mark of the window's first half.
- 2024A dip to 4 works, the lowest annual count in the 2019-2026 window.
- 2025Rebound to 8 works, tying the 2021 high and starting the strongest stretch on record.
- 2026Busiest year on record with 9 works; both Evolve-reviewed papers appear - the Llama-3.1 jargon-benchmark study and the Reddit slang-diffusion study - each with Sun as last author.
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 and radar scores come from an OpenAlex name-resolved record (identity: openalex); 'Zhewei Sun' is not a rare name, so identity confidence depends on OpenAlex disambiguation.
- The OpenAlex topic labels on this record (antibiotic-resistance and bacterial-genetics themes) contradict the paper-level NLP fields and are treated as name-resolution noise; they were excluded from the narratives above.
- Output and impact axes are unscored because citation-level data is insufficient; the reported h-index (13) and i10-index (18) are snapshot values not reflected in the radar.
- Reproducibility (78/100) averages only the 2 Evolve-reviewed papers; the remaining works in the counts-by-year series have not been assessed.