Sapienza (Rome) researcher on LLM reliability: model merging, quantization-robust retrieval, homomorphic inference.
3
papers here
2
reviewed by Evolve
0
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—
This axis would measure publication volume — total papers, per-year counts, and venue quality. The snapshot carries no such data for this researcher, so it cannot be scored; only the three Evolve-reviewed 2026 arXiv papers are currently on record.
Impact—
This axis would measure citation-based influence — citation counts, h-index, and uptake of the cataloged work by other papers. No such metrics exist in the snapshot, so the axis is left unscored rather than estimated from the three 2026 papers.
Focus0/100
Scores 0/100. The raw field distribution splits output evenly — 0.5 to machine learning (LLM interpretability and safety) and 0.5 to machine learning (model compression and information retrieval) — and a third paper extends into applied cryptography / privacy-preserving ML. Read as deliberate breadth across adjacent efficiency-and-safety topics rather than a single lane; the zero reflects measured dispersion, not inactivity.
Collaboration41/100
Scores 41/100. The snapshot records 7 distinct coauthors across the three 2026 cataloged papers, matching the frequent-coauthor list exactly (Luca Zhou, Alessandro Zirilli, Davide Marincione, Evgenios M. Kornaropoulos, Giuseppe Ateniese, Daniele Solombrino, Roberto Dessì), so repeat collaboration is high. The pattern points to compact, stable core teams rather than wide external networking, consistent with a modest breadth score.
Activity—
This axis would measure publishing cadence — outputs per year and the gaps between them. The snapshot lists only the 2026 paper years with no multi-year history, so recent activity levels cannot be scored.
Reproducibility55/100
Scores 55/100, a mean over 2 scored papers (reproPapers: 2). All three cataloged papers carry a medium reproReadiness rating from Evolve review: the quantization paper rests on large sweeps; the merging paper is robust within its LoRA testbeds; HEAT is well-evaluated but single-model. The mid-range score fits — solid empirical support, scoped testbeds, and some restated theory limit independent reuse.
Research summary
Emanuele Rodolà is a researcher at Sapienza University of Rome whose reviewed work centers on the reliability and efficiency of large language models. Three 2026 arXiv papers in the Evolve catalog, on all of which he holds the last author position, divide evenly between two declared subfields — machine learning for LLM interpretability and safety, and machine learning for model compression and information retrieval — while a third line reaches into applied cryptography and privacy-preserving machine learning. 'On Emergent Capabilities and Model Merging' maps three regimes in which merging preserves, cannot create, or preferentially dilutes emergent capabilities. 'The Undetected Damage of Quantization on Retrieval' shows quantization can preserve headline accuracy while flipping 14-46% of retrieval top-1 results. 'HEAT' reports 3.1x fewer iterations and 1.4x faster encrypted GPT-2 decoding through learnable per-site FHE approximation depth. Evolve's reproducibility rating averages 55/100 over two scored papers, with all three rated medium readiness. The collaboration footprint is compact — 7 distinct coauthors, exactly matching the frequent-coauthor list — pointing to small, stable teams. Output, impact, and activity metrics are unavailable and left unscored.
Collaboration pattern
The snapshot records 7 distinct coauthors across the three 2026 papers — exactly matching the frequent-coauthor list (Luca Zhou, Alessandro Zirilli, Davide Marincione, Evgenios M. Kornaropoulos, Giuseppe Ateniese, Daniele Solombrino, Roberto Dessì). Everyone repeats, pointing to a small, stable core team rather than an extended network; the collaboration axis scores 41/100. All three papers place him in the last author position, the conventional senior slot, though none flags him as corresponding author.
Research keywords
Highlights
Three regimes of emergent capabilities under model merging
Quantization flips 14-46% of retrieval top-1 while accuracy looks intact
HEAT: 1.4x faster encrypted GPT-2 decoding via co-adapted FHE approximations
Timeline
- 2026Last-author arXiv paper 'On Emergent Capabilities and Model Merging': maps three regimes of how merging preserves, cannot create, or preferentially dilutes emergent capabilities.
- 2026Last-author arXiv paper 'The Undetected Damage of Quantization on Retrieval': reports 14-46% top-1 flips under quantization, with a mixed-precision fix.
- 2026Last-author arXiv paper 'HEAT': reports 3.1x fewer iterations and 1.4x faster encrypted GPT-2 decoding via learnable per-site FHE approximation depth.
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: built solely from the Evolve-reviewed catalog (3 arXiv papers, all 2026) and the accompanying radar snapshot; no ORCID or country field was available for disambiguation.
- Output, impact, and activity axes have no underlying data in the snapshot (no publication counts, citation metrics, or multi-year history) and are deliberately left unscored.
- The reproducibility mean (55/100) covers only 2 scored papers; the third carries a medium readiness rating but no numeric score, so the average may shift.
- Last-author placement with no corresponding-author flags is read as a conventional senior role; actual supervision or funding roles are not confirmed by the data.