LLM evaluation and mechanistic interpretability researcher; last author on two 2026 arXiv studies in NLP
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-28) 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—
No publication counts are available in this snapshot, so output is unscored. This axis would measure total research production — paper volume over time and across venues — and it needs a fuller bibliographic record than the two cataloged 2026 arXiv papers can provide.
Impact—
No citation counts are available, so impact is unscored. This axis would track how often the researcher's work is cited — total citations and citation-based aggregates — and it cannot be estimated from the Evolve catalog alone.
Focus100/100
The focus score of 100 comes from a field distribution with a single entry — natural language processing / mechanistic interpretability. The two 2026 studies approach one theme from complementary angles: one measures model behaviour through likelihood ranking versus prompting across 95 LLMs, the other inspects part-of-speech structure inside LLaMA-3 SAE latents. Such one-bucket concentration signals a deliberately narrow research agenda.
Collaboration32/100
Collaboration scores 32, and the raw count explains why: across both cataloged papers there are only 4 distinct coauthors, with Alessandro Bondielli, Lucia Passaro, Serena Auriemma and Davide Bacciu recurring as frequent collaborators. This is a compact, stable team profile rather than a wide international network. Lenci holds the last author position on both papers without corresponding-author flags, consistent with a senior advisory role.
Activity—
Year-by-year publication activity is not in this snapshot, so activity is unscored. It would measure publishing cadence and recency across the career; the only dated evidence here is the pair of 2026 preprints.
Reproducibility100/100
reproMean is 100, but it is computed over a single paper (reproPapers=1), so the perfect score carries limited statistical weight. The SAE interpretability study carries Evolve's 'high' repro-readiness rating, while the 95-model evaluation study is rated 'medium'. Both verdicts praise methodological care and honest hedging — family clustering and probe dependence are flagged openly — which supports confidence in the workflow despite the thin base.
Research summary
Alessandro Lenci appears in the Evolve catalog as a natural language processing researcher publishing on arXiv in 2026, in the last author position on both reviewed papers — a placement typically associated with senior or supervisory roles, though no corresponding-author designation is recorded. The two studies bracket one question from two sides: how we measure large language models, and what is inside them. 'Likelihood Ranking doesn't Scale Like Prompting in LLMs' (2026) examines 95 LLMs and finds statement-likelihood accuracy stays flat with scale while prompted accuracy soars; Evolve calls it a careful, large study, while noting that family clustering tempers certainty. 'Parts-of-Speech as Emergent Categories in SAE Latent Space' (2026) argues that part-of-speech structure in LLaMA-3 SAE latents is distributed and category-dependent, with Evolve flagging single-model/layer and probe-dependence caveats while praising its honest hedging. The radar mirrors this narrowness and compactness: focus scores 100 on a single-field distribution, reproducibility averages 100 over one paper, and collaboration scores 32 with 4 distinct coauthors — Alessandro Bondielli, Lucia Passaro, Serena Auriemma and Davide Bacciu. Output, impact and activity metrics are absent from this snapshot, so the profile stays deliberately thin on volume and citation standing.
Collaboration pattern
The cataloged record shows a small, stable circle: 4 distinct coauthors across two 2026 papers, with Alessandro Bondielli, Lucia Passaro, Serena Auriemma and Davide Bacciu recurring as frequent collaborators. Lenci sits in the last author position on both papers without corresponding-author flags, consistent with a senior advisory role within a compact team rather than a broad international network (radar collaboration score: 32).
Research keywords
Highlights
Across 95 LLMs, likelihood ranking doesn't keep pace with prompting
Part-of-speech structure in LLaMA-3 SAE latents is distributed, not localized
A consistent last-author, evidence-first pattern
On both cataloged papers Lenci appears as last author (non-corresponding), and both earn verdicts praising care and honest hedging — one rated medium and one high for reproducibility readiness — suggesting a senior role in a small, method-oriented team.
Timeline
- 2026Last-author arXiv preprint on likelihood ranking versus prompting across 95 LLMs enters the Evolve catalog, rated medium for reproducibility readiness.
- 2026Last-author arXiv preprint on part-of-speech structure in LLaMA-3 SAE latents is cataloged, earning Evolve's 'high' repro-readiness rating.
- 2026The year's two cataloged preprints yield a focus score of 100 from a single-field distribution and a collaboration score of 32 from 4 distinct coauthors.
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.
- Name resolution: institution, country and ORCID are absent from the input data, so this profile is name-resolved from public bibliographic data and 'Alessandro Lenci' may be ambiguous.
- Radar scores are computed from a very small snapshot — two 2026 arXiv papers — so the focus and reproducibility maxima rest on thin evidence.
- Output, impact and activity are unscored because publication counts and citation metrics are missing from this extract.
- The reproducibility mean of 100 is based on a single paper (reproPapers=1); the second cataloged paper is rated medium for reproducibility readiness.