LLM interpretability researcher probing persona representations and linear structure in dialogue-conditioned LLMs
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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 scored: this snapshot contains no publication counts or per-year output totals for the researcher beyond the two Evolve-cataloged 2026 papers. This axis would measure total research output and how it is distributed over time and across venue types.
Impact—
Not scored: no citation counts, h-index, or other influence metrics are present in the input data. This axis would measure how widely the two cataloged studies are cited and adopted within the LLM interpretability community.
Focus0/100
Focus scores 0/100 in this snapshot. The recorded field distribution splits evenly (0.5/0.5) between two nearly identical labels — 'natural language processing (LLM interpretability)' and 'NLP — LLM interpretability / representation analysis' — so the low score reflects fragmentation of field labels in the underlying data rather than demonstrated topical breadth. Substantively, both catalog papers target one niche: representation-level analysis of persona in dialogue-conditioned and decoder-only LLMs.
Collaboration41/100
Collaboration scores 41/100, based on 7 distinct coauthors recorded across the two cataloged papers. The frequent-coauthor list is dominated by a small recurring core: Yu Sun, Cong Feng, and Guangming Lu appear on both entries, with Mengyin Lu also listed. Han takes the last author position on both papers, with no corresponding-author flag recorded in either, consistent with a senior role inside a compact, stable team rather than a large distributed network.
Activity—
Not scored: no per-year publication cadence or recency signal is available beyond two arXiv entries dated 2026. This axis would measure ongoing output and research engagement in recent years, which the current snapshot cannot support.
Reproducibility55/100
Reproducibility scores 55/100, computed as a mean (reproMean = 55) over 2 papers (reproPapers = 2), each rated 'medium' for repro-readiness in Evolve review. The cosine-similarity study is described as careful and well-controlled, including a PCA control; the K/V-cache intervention study candidly states limits but rests on a single model/pair (Llama-3.1-8B) and a proxy behavioral measure. With only two papers behind it, this score is preliminary.
Research summary
Huimin Han is profiled here as an early-stage record in natural language processing, focused on LLM interpretability and representation analysis. The profile is name-resolved from public bibliographic data: no institution, country, or ORCID is attached, and the name is common, so identity matching is probabilistic. The Evolve catalog holds two papers, both 2026 arXiv preprints, with Han in the last author position on each — a senior-author signature — though neither carries a corresponding-author flag. The first argues that cosine similarity underestimates linearly decodable persona structure in dialogue-conditioned LLMs, using linear probing, supervised subspace projection, and a PCA control; Evolve's review calls it careful and well-controlled, with modest novelty. The second applies mid-layer K/V-cache interventions to Llama-3.1-8B to dissociate representation alignment from persona expression, candidly stating limits while relying on a proxy behavioral measure. Radar scores: collaboration 41/100 (7 distinct coauthors; Yu Sun, Cong Feng, and Guangming Lu recur on both papers), reproducibility 55/100 (mean over 2 papers, both rated medium), and focus 0/100, which tracks a field distribution split across two near-identical interpretability labels rather than demonstrated breadth. Output, impact, and activity are unscored — the snapshot lacks publication counts, citations, and per-year cadence. With a two-paper base, all figures should be read as provisional.
Collaboration pattern
Han sits in the last author position on both cataloged papers, with no corresponding-author flag recorded — a senior-author pattern in a small team. Seven distinct coauthors appear across the two entries; Yu Sun, Cong Feng, and Guangming Lu recur on both, and Mengyin Lu on one. A recurring trio plus a compact roster (collaboration score 41/100) points to a stable, close-knit working group rather than a broad collaborative network.
Research keywords
Highlights
Where cosine similarity fails persona structure
K/V-cache interventions split representation from persona
Timeline
- 2026First Evolve-cataloged paper appears: an arXiv study with Han as last author showing when cosine similarity fails to reflect linearly accessible persona structure in dialogue models.
- 2026Second cataloged paper appears: K/V-cache interventions dissociate representation alignment from persona expression in decoder-only language models, again with Han as last author.
- 2026Evolve profile coverage opens on this basis: two 2026 arXiv entries in LLM interpretability, 7 distinct coauthors recorded, and a reproducibility mean of 55 across 2 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.
- Profile is name-resolved from public bibliographic data; no institution, country, or ORCID is attached, and 'Huimin Han' is a common name, so identity matching is probabilistic.
- All radar scores rest on only two Evolve-cataloged papers (reproPapers = 2); collaboration 41 and reproducibility 55 are small-sample figures.
- No citation counts, publication totals, or per-year output data are available; the output, impact, and activity axes are unscored, not zero.
- Both papers are 2026 arXiv preprints; peer-review status and venue acceptance are not recorded in the snapshot.