Sapienza researcher bridging compression and cryptography: quantization-aware retrieval, faster homomorphic inference
2
papers here
1
reviewed by Evolve
1
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 measures publication volume and authorship roles over time, but output counts are not yet available in the bibliometric snapshot; only the two reviewed 2026 arXiv papers are known, so no score is estimated.
Impact—
Impact would aggregate citation counts and uptake by follow-up work, yet no citation metrics exist in this snapshot; both papers are recent 2026 arXiv preprints with citations not yet accumulated, so no score is estimated.
Focus100/100
Focus is scored 100, the maximum concentration: the bibliometric field distribution assigns the entire detected record to a single field, machine learning (model compression and information retrieval). Within that narrow lane the two reviewed papers still cover compression's two application faces — quantization damage to retrieval accuracy and encrypted inference speed — reading as deliberate specialization in efficient, trustworthy inference rather than scattered topics.
Collaboration41/100
The collaboration score of 41 reflects a compact network: 7 distinct coauthors appear in the record, anchored by Emanuele Rodolà, Giuseppe Ateniese, Evgenios M. Kornaropoulos, Davide Marincione, Luca Zhou, Daniele Solombrino and Roberto Dessì. Both reviewed papers are multi-author efforts, with Zirilli first author on HEAT and middle author on the quantization study — a small, repeat team rather than a broad collaboration footprint.
Activity—
This axis measures publishing cadence across years, but yearly output counts beyond the two 2026 papers are unavailable, so the timeline cannot yet be characterized and no activity score is estimated.
Reproducibility55/100
Reproducibility scores 55, computed as a mean over just one scored paper — an early, single-sample signal rather than a track record. Both reviewed papers carry medium reproducibility readiness: the quantization study's large sweeps support replication, while HEAT is well-evaluated but rests on GPT-2 as a single model, so conclusions may not transfer across architectures.
Research summary
Alessandro Zirilli is a researcher at Sapienza University of Rome whose reviewed work sits at the intersection of model compression and applied cryptography, unified by one question: when models are quantized or encrypted to make inference cheaper or more private, what silently breaks — and can it be repaired without retraining? In 'The Undetected Damage of Quantization on Retrieval and How to Fix It' (2026) he is a middle author on work showing that quantization can preserve headline accuracy while flipping 14–46% of top-1 retrieval results, addressed through mixed-precision allocation and label-free calibration; Evolve's review judged the large sweeps well-supported while noting the core theory restates a known condition. In 'HEAT: Faster Fully Homomorphic Inference via Approximations-Weights Co-Adaptation' (2026) he is first author, proposing learnable per-site approximation depth for CKKS-based fully homomorphic encryption, reported to cut iterations 3.1x and speed encrypted GPT-2 decoding 1.4x, with solid but single-model evaluation. The radar profile shows maximal field concentration (focus 100), a compact 7-coauthor network (collaboration 41), and a moderate reproducibility signal (55, from one scored paper).
Collaboration pattern
Zirilli's recorded network is compact: 7 distinct coauthors, with the frequent-coauthor list — Emanuele Rodolà, Davide Marincione, Evgenios M. Kornaropoulos, Giuseppe Ateniese, Luca Zhou, Daniele Solombrino, Roberto Dessì — centered on Sapienza University of Rome. His authorship roles span first author (HEAT, 2026) and middle author (the quantization-retrieval study, 2026). The collaboration score of 41 reflects this small, repeat team rather than a broad, externally distributed collaboration footprint.
Research keywords
Highlights
Quantization silently flips retrieval rankings
First author on faster encrypted inference
One thesis across two fields: repair compression damage cheaply
Across both 2026 papers, the recurring pattern is diagnosing a hidden cost of compression or encryption — flipped retrieval rankings, bootstrapping overhead — and repairing it with lightweight calibration-style adaptation rather than full retraining.
Timeline
- 2026Published (middle author) 'The Undetected Damage of Quantization on Retrieval and How to Fix It' on arXiv, reporting 14–46% top-1 flips under quantization.
- 2026First-authored 'HEAT: Faster Fully Homomorphic Inference via Approximations-Weights Co-Adaptation' on arXiv, reporting 3.1x fewer iterations and 1.4x faster encrypted GPT-2 decoding.
- 2026Both papers entered the Evolve catalog with medium reproducibility readiness, forming the basis of this early-stage profile.
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.
- Bibliometric scores (focus, collaboration, reproducibility) derive from OpenAlex-based snapshots; the field distribution reflects a single matched record.
- The reproducibility mean (55) averages over only 1 scored paper — a single-sample signal, not a track record.
- Output, impact and activity axes are unscored: no publication counts, citation metrics, or multi-year output data were available.
- No ORCID or country field was provided; the profile is name-resolved from public bibliographic data plus the two reviewed arXiv papers.